A self-adaptive trimming and testing system and method for welding seams of chemical pressure vessels

Abstract

The present application relates to the technical field of weld finishing and nondestructive testing, and discloses a chemical pressure vessel weld self-adaptive finishing and testing system and method, which comprises the following steps: obtaining a three-dimensional model of a chemical pressure vessel and weld distribution, establishing a weld topology graph, and generating a full-traversal path of the weld; a magnetic adsorption wall-climbing robot collects weld images, and improved YOLOv11n weld visual detection model forms weld operation feature data; a finishing boundary position is determined, continuous weld sections are divided, and segmented finishing tasks are formed; adaptive finishing parameters are generated and finishing is implemented, and weld finishing track results are formed; a detection time interval is determined according to an installation distance and the finishing track, ultrasonic signals are collected to form segmented testing results; and the segmented testing results are matched with the weld finishing track results to form weld finishing and testing results. The present application realizes weld finishing and testing integration through visual identification and segmented finishing.

Description

Technical Field

[0001] This invention relates to the field of weld repair and non-destructive testing technology, and in particular to an adaptive repair and inspection system and method for welds of chemical pressure vessels. Background Technology

[0002] Current methods for repairing weld seams on chemical pressure vessels typically involve manually observing or visually identifying the weld seam location, followed by surface treatment using grinding equipment at set speeds, travel speeds, and contact pressures. After repair, ultrasonic testing is used to obtain information about internal weld defects. Some wall-climbing robots can move along the vessel surface and perform weld seam identification, grinding, and inspection tasks separately.

[0003] In existing technologies, the extent of weld corrosion coverage, weld reinforcement, and surface undulation vary along the weld length, making it difficult to correlate fixed trimming parameters with different weld sections. Robotic deviations and changes in contact force can cause trimming trajectory offsets or inconsistent material removal. Furthermore, the trimming process data and ultrasonic testing results lack segment-time correspondence, making it difficult to distinguish between coupling anomalies, trimming deviations, and weld defects.

[0004] Therefore, how to provide an adaptive repair and inspection system and method for weld seams of chemical pressure vessels is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an adaptive repair and inspection system and method for welds in chemical pressure vessels. This invention fully utilizes weld visual recognition, segmented repair, constant force control, and diffraction time-of-flight ultrasonic testing technology to adaptively determine repair parameters according to the weld condition and correlate the weld repair trajectory with the segmented inspection results. It has the advantages of strong repair adaptability, high operational stability, accurate inspection results, and process traceability.

[0006] An adaptive repair and inspection method for weld seams of chemical pressure vessels according to an embodiment of the present invention includes the following steps:

[0007] Obtain the 3D model and weld distribution information of the chemical pressure vessel, establish the weld topology map, pair the odd-degree topological vertices in the weld topology map, and generate the full traversal path of the weld.

[0008] Control the magnetic adsorption wall-climbing robot to move along the full traverse path of the weld and collect weld images. Input the weld images into the improved YOLOv11n weld visual inspection model to form weld operation feature data.

[0009] Arrange the weld operation characteristic data along the weld centerline, determine the repair boundary position, divide the continuous weld section according to the repair boundary position, and form a segmented repair task.

[0010] Based on the segmented trimming task, the grinding head speed, trimming times and robot walking speed are controlled, and each continuous weld seam segment is associated with the corresponding parameter combination to form adaptive trimming parameters.

[0011] The weld seam sections are trimmed according to the adaptive trimming parameters to form the weld seam trimming trajectory result.

[0012] Obtain the installation distance between the grinding head and the ultrasonic testing probe group using the diffraction time-of-flight method, determine the testing time interval based on the weld seam trimming trajectory results, and establish a corresponding relationship between trimming and inspection.

[0013] According to the corresponding relationship between repair and inspection, coupling agent is supplied to each continuous weld section, and ultrasonic signals are collected in the corresponding inspection time interval to form segmented inspection results;

[0014] The segmented inspection results are matched with the weld repair trajectory results according to the continuous weld sections to form weld repair and inspection results.

[0015] Optionally, the step of acquiring the 3D model of the chemical pressure vessel and weld distribution information, establishing a weld topology map, pairing the odd-degree topological vertices in the weld topology map, and generating a full traversal path for the weld specifically includes:

[0016] The weld distribution information is mapped onto the 3D model of the chemical pressure vessel, and a weld topology map is established based on the weld intersection positions and adjacent weld segments.

[0017] For each topological vertex in the weld topology diagram, determine the odd-degree topological vertices and form a set of odd-degree topological vertices;

[0018] Pair up the odd-degree topological vertices in the set of odd-degree topological vertices without repetition to determine the pairing results of the odd-degree topological vertices.

[0019] The weld segment to be repeated is determined based on the odd-degree topology vertex pairing results, and a copy of the topology edge corresponding to the weld segment to be repeated is added to the weld topology graph to form an augmented weld topology graph.

[0020] Generate a full traversal path for the weld seam based on the augmented weld seam topology.

[0021] Optionally, the controlled magnetic adsorption wall-climbing robot moves along the entire traversal path of the weld seam and acquires weld seam images. The weld seam images are then input into the improved YOLOv11n weld seam visual inspection model to form weld seam operation feature data, specifically including:

[0022] Acquire and process weld images of the weld to be processed to form weld input images;

[0023] The weld seam input image is input into the feature extraction module of the improved YOLOv11n weld seam visual inspection model. The coordinate attention mechanism is used to enhance the spatial location features of the weld seam in the weld seam input image, forming coordinate-enhanced features.

[0024] Multi-branch heterogeneous convolutional structures are used to extract multi-level weld seam features from coordinate enhancement features, forming multi-branch weld seam features.

[0025] The low-frequency semantic components and high-frequency edge components in the multi-branch weld features are processed separately using the frequency domain feature processing module to form frequency domain fusion features;

[0026] A progressive feature pyramid network is used to perform cross-scale feature fusion of frequency domain fusion features to form cross-scale weld features;

[0027] Based on the characteristics of cross-scale welds, the weld contour is identified and the weld centerline is fitted. The rust coverage, weld reinforcement status and surface undulation status are extracted to form weld operation characteristic data.

[0028] Optionally, the step of arranging weld operation characteristic data along the weld centerline, determining the trimming boundary position, and dividing continuous weld sections according to the trimming boundary position to form a segmented trimming task specifically includes:

[0029] Arrange the weld operation feature data along the weld centerline to form a weld feature position sequence;

[0030] The stable positions are determined from the weld feature position sequence, and the fluctuation range between adjacent stable positions is calculated to form the stable position analysis results;

[0031] Based on the results of stable position analysis, positions with continuously changing states are selected to form candidate trimming boundary positions;

[0032] The candidate trimming boundary positions are merged to form the trimming boundary positions;

[0033] The continuous weld seam sections are divided according to the repair boundary positions, forming a segmented repair task.

[0034] Optionally, the step of controlling the grinding head speed, the number of dressing operations, and the robot walking speed based on the segmented dressing task, and associating each continuous weld segment with the corresponding parameter combination to form adaptive dressing parameters specifically includes:

[0035] Collect statistical data on the weld operation characteristics of each continuous weld section to form the section characteristic statistical results;

[0036] Based on the statistical results of the section characteristics, the degree of corrosion deviation, the degree of reinforcement deviation, and the degree of fluctuation deviation of each continuous weld section are determined to form the section state deviation results;

[0037] Establish the calibration relationship between weld condition and calibration parameters, and generate calibration results for calibration parameters.

[0038] Based on the deviation results of the section state and the calibration results of the dressing parameters, the grinding head speed, robot walking speed and dressing times are determined to form the material removal parameters;

[0039] The target normal pressure and the normal pressure adjustment range are determined based on the degree of fluctuation and deviation, thus forming the normal pressure parameters;

[0040] By associating material removal parameters and normal pressure parameters with the corresponding continuous weld sections, adaptive trimming parameters are formed.

[0041] Optionally, the step of adjusting each continuous weld section according to adaptive adjustment parameters to form a weld adjustment trajectory result specifically includes:

[0042] The adaptive trimming parameters corresponding to each continuous weld section are retrieved and the walking deviation of the magnetic adsorption wall-climbing robot is determined to form walking deviation data.

[0043] Adjust the running speed of the left and right drive wheels based on the walking deviation data to form drive wheel control parameters;

[0044] The actual contact force between the grinding head and the weld to be treated is collected and the feed position of the grinding head is adjusted to form feed control parameters;

[0045] The current continuous weld section is trimmed according to the drive wheel control parameters and feed control parameters to form the section trimming execution result;

[0046] The data of the trimming process of the current continuous weld section are recorded synchronously to form a section trimming record;

[0047] Summarize the section repair records corresponding to each continuous weld section to form the weld repair trajectory result.

[0048] Optionally, the step of obtaining the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group, determining the testing time interval based on the weld seam trimming trajectory results, and forming a trimming-inspection correspondence specifically includes:

[0049] Determine the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group to generate installation position data for the operating mechanism;

[0050] Extract the position and time when each continuous weld section completes the repair from the weld repair trajectory results to form section repair boundary data;

[0051] The starting detection time for the ultrasonic testing probe group to reach the detection position of the starting position of each continuous weld section is determined based on the installation distance.

[0052] The termination time of the ultrasonic testing probe group for diffraction time difference method is determined based on the installation distance to the termination position of each continuous weld section.

[0053] The inspection time intervals for each continuous weld section are determined based on the start and end inspection times, thus forming section inspection time data.

[0054] Each continuous weld section is associated with its corresponding inspection time interval to form a repair and inspection correspondence.

[0055] Optionally, the step of supplying coupling agent to each continuous weld section according to the repair and inspection correspondence, and collecting ultrasonic signals in the corresponding inspection time interval to form segmented inspection results specifically includes:

[0056] According to the corresponding relationship between repair and inspection, coupling agent is supplied to each continuous weld section to form coupling execution data;

[0057] Ultrasonic signals are collected within the detection time interval corresponding to each continuous weld section to form ultrasonic acquisition data;

[0058] The ultrasonic signals in the acquired ultrasonic data are filtered and their gain is adjusted to form processed ultrasonic signals.

[0059] Identify the through waves in the ultrasonic signal and determine the validity of the ultrasonic signal to form an ultrasonic signal validity result;

[0060] Defect diffraction waves are extracted from effective ultrasonic signals to determine the defect location, defect burial depth, and defect length, thus forming defect characterization data.

[0061] The ultrasonic signal validity results and defect characterization data are correlated with the corresponding continuous weld sections to form segmented inspection results.

[0062] Optionally, the step of correlating the segmented inspection results with the weld repair trajectory results according to the continuous weld sections to form weld repair and inspection results specifically includes:

[0063] The segmented inspection results are correlated with the weld repair trajectory results according to the continuous weld section to form joint judgment data for the section;

[0064] From the joint judgment data of the sections, select the continuous weld sections with abnormal locations of direct wave or whose ultrasonic signals are determined to be invalid, resupply the coupling agent and re-inspect, and form the coupling re-inspection data;

[0065] Based on the joint judgment data of the sections, the weld trimming trajectory results of each continuous weld section are compared with the corresponding adaptive trimming parameters to form the trimming deviation result;

[0066] Based on the correction deviation results, the adaptive correction parameters for the corresponding continuous weld section are redefined, and supplementary correction and re-inspection are completed to form supplementary correction and re-inspection data.

[0067] Based on the joint judgment data of the sections, the coupled re-inspection data, and the supplementary repair re-inspection data, each continuous weld section is classified to form a section classification result.

[0068] The results of the segment classification are summarized to form the weld repair and inspection results.

[0069] An adaptive repair and inspection system for weld seams in chemical pressure vessels includes the following modules:

[0070] The path planning module is used to build a weld topology graph, pair up the odd-degree topological vertices in the weld topology graph, and generate a full traversal path for the weld.

[0071] The visual recognition module is used to acquire weld images and input the weld images into the improved YOLOv11n weld visual inspection model to form weld operation feature data.

[0072] The task segmentation module is used to arrange weld operation feature data along the weld centerline, determine the trimming boundary position, divide the continuous weld section according to the trimming boundary position, and form a segmented trimming task.

[0073] The parameter generation module is used to generate adaptive trimming parameters based on the segmented trimming task;

[0074] The trimming control module is used to trim each continuous weld section according to adaptive trimming parameters, and to generate weld trimming trajectory results.

[0075] The inspection mapping module is used to determine the inspection time interval based on the weld repair trajectory results and to form a correspondence between repair and inspection.

[0076] The ultrasonic testing module is used to supply coupling agent to each continuous weld section according to the corresponding relationship of repair and inspection, and to collect ultrasonic signals in the corresponding detection time interval to form segmented inspection results.

[0077] The result determination module is used to match the segmented inspection results with the weld repair trajectory results according to the continuous weld sections, and form weld repair and inspection results.

[0078] The beneficial effects of this invention are:

[0079] This invention establishes a weld topology map based on the three-dimensional model of a chemical pressure vessel and weld distribution information. By pairing odd-degree topological vertices, it determines the weld segments to be repeated and forms a full traversal path for the weld. This enables the magnetic adsorption wall-climbing robot to cover the weld to be processed along a continuous path, reducing invalid movement and missed segments, and providing a unified position reference for subsequent weld identification, repair and inspection.

[0080] Compared to existing YOLOv11n models, this invention integrates a coordinate attention mechanism and a multi-branch heterogeneous convolutional structure into the feature extraction module. It enhances the spatial location features of the weld seam through lateral and longitudinal coordinate attention weights, and utilizes convolutional branches with different receptive ranges to extract local edges, textures, main regions, and extended morphologies of the weld seam. Simultaneously, it enhances low-frequency semantic components and high-frequency edge components using a frequency domain feature processing module, and progressively fuses weld seam features at different scales through a progressive feature pyramid network. These improvements enhance the distinguishability between the weld seam region and the adjacent base material regions, improve the feature integrity of rust textures, weld boundaries, and local undulations, and provide a stable data foundation for weld contour recognition, weld centerline fitting, and weld operation feature data extraction.

[0081] This invention arranges weld operation characteristic data along the weld centerline, determines the trimming boundary position based on the continuous changes in rust coverage, weld reinforcement status, and surface undulation, and divides continuous weld sections. Combining the degree of deviation of each continuous weld section's state with the trimming parameter calibration results, it determines the grinding head rotation speed, robot travel speed, trimming times, target normal pressure, and normal pressure adjustment range, ensuring that the trimming parameters correspond to the actual state of different weld sections. By continuously adjusting the left and right drive wheel speeds and grinding head feed position through lateral deviation, angular deviation, and actual contact force feedback, the impact of robot travel deviation and contact force fluctuations on the trimming trajectory and material removal process can be reduced.

[0082] This invention, based on the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group, converts the travel displacement, travel direction, and processing time in the weld seam finishing trajectory results into detection time intervals corresponding to each continuous weld seam segment, thus establishing a segmental correspondence between the finishing process data and the ultrasonic inspection results. The effectiveness of the ultrasonic signal is distinguished by the continuous state of the direct-through wave, and combined with whether the finishing process deviates from the adaptive finishing parameters, a re-inspection is performed on segments with coupling anomalies, and supplementary finishing and re-inspection are performed on segments with finishing deviations. This results in a weld seam finishing and inspection result that includes defective weld seam segments, qualified inspection segments, and re-inspection results, improving the traceability of the finishing process and inspection results. Attached Figure Description

[0083] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0084] Figure 1 This is a flowchart of an adaptive repair and inspection method for weld seams in chemical pressure vessels proposed in this invention.

[0085] Figure 2 This is a schematic diagram of the functional module structure of an adaptive repair and inspection system for weld seams of chemical pressure vessels proposed in this invention.

[0086] Figure 3 This is a schematic diagram of the network structure of the improved YOLOv11n weld visual inspection model proposed in this invention;

[0087] Figure 4 This is a schematic diagram of the weld full traversal path planning results proposed in this invention. Detailed Implementation

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0089] refer to Figures 1-3 An adaptive repair and inspection method for weld seams in chemical pressure vessels includes the following steps:

[0090] Obtain the 3D model and weld distribution information of the chemical pressure vessel, establish the weld topology map, pair the odd-degree topological vertices in the weld topology map, and generate the full traversal path of the weld.

[0091] Control the magnetic adsorption wall-climbing robot to move along the full traverse path of the weld and collect weld images. Input the weld images into the improved YOLOv11n weld visual inspection model to form weld operation feature data.

[0092] Arrange the weld operation characteristic data along the weld centerline, determine the repair boundary position, divide the continuous weld section according to the repair boundary position, and form a segmented repair task.

[0093] Based on the segmented trimming task, the grinding head speed, trimming times and robot walking speed are controlled, and each continuous weld seam segment is associated with the corresponding parameter combination to form adaptive trimming parameters.

[0094] The weld seam sections are trimmed according to the adaptive trimming parameters to form the weld seam trimming trajectory result.

[0095] Obtain the installation distance between the grinding head and the ultrasonic testing probe group using the diffraction time-of-flight method, determine the testing time interval based on the weld seam trimming trajectory results, and establish a corresponding relationship between trimming and inspection.

[0096] According to the corresponding relationship between repair and inspection, coupling agent is supplied to each continuous weld section, and ultrasonic signals are collected in the corresponding inspection time interval to form segmented inspection results;

[0097] The segmented inspection results are matched with the weld repair trajectory results according to the continuous weld sections to form weld repair and inspection results.

[0098] In this embodiment, acquiring the 3D model of the chemical pressure vessel and weld distribution information, establishing a weld topology map, and pairing the odd-degree topological vertices in the weld topology map to generate a full traversal path for the weld specifically includes:

[0099] Establish a weld topology map; specifically, map the weld distribution information to the 3D model of the chemical pressure vessel and extract the weld space centerline of each weld to be processed.

[0100] The intersection points of the intersecting weld space centerlines and the end points that do not intersect with other weld space centerlines are defined as topological vertices.

[0101] The weld space centerline is divided according to the adjacent topological vertices. The resulting weld segments are determined as topological edges, and the weld lengths corresponding to each topological edge are recorded to form a weld topology graph.

[0102] Determine the odd-degree topological vertices in the weld topology graph; specifically, for each topological vertex in the weld topology graph, count the number of topological edges connected to the corresponding topological vertex.

[0103] Topological vertices with an odd number of connected topological edges are defined as odd-degree topological vertices, and arranged according to their spatial positions in the 3D model of the chemical pressure vessel to form a set of odd-degree topological vertices.

[0104] Determine the shortest connection path between all odd-degree topological vertices;

[0105] Specifically, each topological vertex in the weld topology graph is used as a path node, and the weld length corresponding to each topological edge is used as the path weight. For any odd-degree topological vertex in the set of odd-degree topological vertices, the Dijkstra shortest path algorithm is used to calculate the shortest connection path from the odd-degree topological vertex to the other odd-degree topological vertices and the corresponding weld length. The same processing is performed on each odd-degree topological vertex to form the odd-degree topological vertex connection result.

[0106] Pair up the odd-degree topological vertices; specifically, based on the odd-degree topological vertex connection results, read the shortest connection path and the corresponding weld length between any two odd-degree topological vertices, and use the weld length as the connection cost of the corresponding pairing relationship.

[0107] The pairing relationship of odd-degree topological vertices is determined step by step by using the branch and bound method. Each time, an odd-degree topological vertex is selected from the odd-degree topological vertices that have not yet been paired, and a candidate pairing branch is formed with the remaining unpaired odd-degree topological vertices respectively.

[0108] Accumulate the connection costs of the established pairings, and determine the lower bound of the remaining pairing length based on the minimum connection cost between the remaining odd-degree topological vertices;

[0109] When the sum of the connection cost and the lower bound of the remaining pairing length is not less than the cumulative weld length corresponding to the current complete pairing result, stop expanding the corresponding candidate pairing branch; otherwise, continue to determine the pairing relationship of the remaining odd-degree topology vertices.

[0110] Compare the cumulative weld lengths of each complete pairing result, and determine the smallest one as the odd-degree topology vertex pairing result.

[0111] The segment to be repeated is determined based on the odd-degree topology vertex pairing results; specifically, the shortest connection path between each pair of odd-degree topology vertices is retrieved.

[0112] The topological edges contained in each shortest connection path are identified as weld segments to be repeated, and copies of the topological edges corresponding to the weld segments to be repeated are added to the weld topology graph so that the number of topological edges connected to each topological vertex is even, thus forming an augmented weld topology graph.

[0113] Generate a full traversal path for the weld seam based on the augmented weld seam topology graph. Specifically, select the topology vertex corresponding to the starting position of the magnetic adsorption wall-climbing robot in the augmented weld seam topology graph, move along the topology edge or copy of the topology edge connected to the current topology vertex and not yet traversed to the adjacent topology vertex, until return to the topology vertex corresponding to the starting position of the magnetic adsorption wall-climbing robot, thus forming a closed traversal path.

[0114] When a closed traversal path contains a topological vertex that connects an untraversed topological edge or a copy of a topological edge, a new closed traversal path is formed from that topological vertex, and the new closed traversal path is embedded into the already formed closed traversal path, until all topological edges and copies of topological edges have been traversed. The final traversal order is then converted into a sequence of weld space coordinates in the 3D model of the chemical pressure vessel, generating a full traversal path for the weld.

[0115] In this embodiment, the magnetic adsorption wall-climbing robot is controlled to move along the full traverse path of the weld seam and collect weld seam images. The weld seam images are then input into the improved YOLOv11n weld seam visual inspection model to form weld seam operation feature data, specifically including:

[0116] Training was performed on the improved YOLOv11n weld visual inspection model;

[0117] Specifically, weld images were collected under different weld types, rust coverage conditions, weld reinforcement conditions, surface undulation conditions, and imaging brightness conditions. The weld area bounding boxes and weld categories were manually annotated and verified to form a weld model dataset. The weld model dataset was divided into training dataset, validation dataset, and test dataset in an 8:1:1 ratio, and continuously collected images corresponding to the same continuous weld segment were included in the same dataset.

[0118] Data augmentation processing was performed on the weld seam images in the training dataset, including brightness adjustment, contrast adjustment, scale adjustment, and small-angle rotation. The processed training data was then input into the improved YOLOv11n weld seam visual detection model. The total training loss of the model was calculated based on the model prediction results and the manually labeled results.

[0119] ;

[0120] In the formula, This represents the total training loss of the model. The regression loss of the weld area bounding box is represented by the intersection-union regression loss, which is used to determine the positional difference between the predicted bounding box and the manually labeled bounding box. This represents the weld category loss, used to determine the difference between the predicted category and the manually labeled category; This represents the bounding box distribution loss, used to constrain the discrete distribution of the boundary positions of the weld region; , and These represent the loss weights of the corresponding loss terms;

[0121] Backpropagation is performed based on the total training loss of the model, and the network parameters in the improved YOLOv11n weld visual detection model are updated. During model training, the size of the input weld images is uniformly adjusted to the range of 512×512 pixels to 768×768 pixels, the training batch size is set to 8 to 32, the initial learning rate is set to 0.0005 to 0.005, the weight decay coefficient is set to 0.0001 to 0.001, and the maximum number of training rounds is set to 150 to 300 rounds.

[0122] In the first 3 to 10 rounds after the start of training, a learning rate warm-up method is used to gradually increase the learning rate from a starting value lower than the initial learning rate to the initial learning rate; after the learning rate warm-up is completed, the learning rate is gradually reduced according to the number of training rounds, so that the model parameters are updated at a learning rate lower than the initial learning rate in the later stage of model training.

[0123] After each round of model training, the validation loss corresponding to the current model is calculated using the validation dataset, and the recognition result of the weld area in the validation dataset is obtained; the accuracy of weld area recognition is determined based on the degree of overlap between the manually labeled weld area and the weld area recognized by the model, and the positional deviation of the weld area is determined based on the image distance between the center position of the weld area recognized by the model and the center position of the manually labeled weld area.

[0124] When adjusting the model training parameters, first keep the training batch size, loss weights and maximum number of training rounds unchanged, change the initial learning rate within the range of 0.0005 to 0.005, complete the model training and compare the corresponding validation loss, and take the initial learning rate corresponding to the lowest validation loss and the stable decrease process as the current initial learning rate.

[0125] With the initial learning rate unchanged, the loss weights corresponding to the bounding box regression loss, weld category loss and bounding box distribution loss are adjusted respectively. The accuracy of weld region recognition and the positional deviation of weld region corresponding to different combinations of loss weights are compared. The loss weight combination with higher accuracy of weld region recognition and smaller positional deviation of weld region is used as the model training loss weight.

[0126] When the validation loss continues to decrease relative to the previous lowest validation loss, save the current model parameters and update the lowest validation loss; when the validation loss does not fall below the previous lowest validation loss for 20 consecutive rounds, stop model training and determine the model parameters corresponding to the lowest validation loss as the parameters of the improved YOLOv11n weld visual inspection model after training is completed.

[0127] The improved YOLOv11n weld visual inspection model was independently tested using a test dataset. The number of weld images that correctly identified the weld area in the test data was counted, and the positional deviation between the weld area identified by the model and the manually labeled weld area was calculated. When the test results met the requirements for weld visual inspection, the improved YOLOv11n weld visual inspection model was deployed to the control system of the magnetic adsorption wall-climbing robot to process the weld images continuously acquired by the magnetic adsorption wall-climbing robot during actual operation.

[0128] Acquire and process weld images of the weld to be processed to form weld input images;

[0129] Specifically, the movement of the magnetic adsorption wall-climbing robot is controlled according to the spatial coordinate sequence of the weld in the full traversal path of the weld, so that the acquisition area of ​​the binocular vision camera covers the weld to be processed and the base material area on both sides. The binocular vision camera has a left imaging unit and a right imaging unit with fixed positions. Before the weld image is acquired, the internal parameters, relative position relationship and imaging reference of the two imaging units are calibrated, and the left and right weld images are stereo-corrected according to the calibration results.

[0130] The acquisition area is continuously illuminated by a supplementary lighting component. Weld seam images are acquired according to the continuous acquisition time. The weld seam images are then subjected to size uniformity, brightness equalization, contrast enhancement, and noise suppression. The processed weld seam images are then associated with the corresponding acquisition time, robot walking displacement, and weld seam spatial coordinates to form a weld seam input image.

[0131] The coordinate attention mechanism is used to enhance the spatial location features of the weld in the weld input image, forming coordinate-enhanced features;

[0132] Specifically, the weld seam input image is input into the feature extraction module of the improved YOLOv11n weld seam visual inspection model, and the feature values ​​in each feature channel are averaged and aggregated along the horizontal and vertical directions of the image to form horizontal position description data and vertical position description data.

[0133] The horizontal position description data is concatenated with the vertical position description data, and the feature channels of the concatenation result are compressed by a 1-row 1-column convolution kernel. After non-linear activation, the data is reseparated into horizontal position branches and vertical position branches.

[0134] Convolution processing is performed on the horizontal and vertical position branches respectively, and the Sigmoid function is used to perform independent nonlinear mapping on each convolution output value in the horizontal and vertical position branches, transforming each convolution output value to the range of 0 to 1, forming horizontal coordinate attention weights and vertical coordinate attention weights respectively.

[0135] The Sigmoid function performs mapping independently for the convolution output values ​​corresponding to each spatial location, does not perform normalization between different feature channels, and does not perform summation and normalization between the attention weights corresponding to different feature channels;

[0136] The horizontal coordinate attention weights are extended along the image width direction, and the vertical coordinate attention weights are extended along the image height direction, so that both the horizontal and vertical coordinate attention weights have spatial dimensions corresponding to the original features, while maintaining the correspondence between each attention weight and the original feature channels;

[0137] The feature values ​​of each spatial location in the original feature are multiplied sequentially by the corresponding horizontal coordinate attention weights and vertical coordinate attention weights to form coordinate-enhanced features;

[0138] Multi-branch heterogeneous convolutional structures are used to extract multi-level weld seam features from coordinate enhancement features, forming multi-branch weld seam features.

[0139] Specifically, the coordinate enhancement features are input into four parallel convolutional branches, and the number of feature channels is adjusted while preserving the weld position relationship using a 1-row, 1-column convolutional kernel.

[0140] A 3x3 convolution kernel is used to extract the local edges and fine-grained texture of the weld, and a 5x5 convolution kernel is used to extract the regional variations between the main body of the weld and the base material regions on both sides.

[0141] A 7-row, 7-column convolution kernel was used to extract the weld extension morphology, large-area rust distribution, and continuous contour changes.

[0142] Edge supplementation ensures that the output features of the four parallel convolutional branches maintain the same feature size. The output features are then connected according to the feature channel direction. A 1-row, 1-column convolutional kernel is used to compress the connected feature channels. The compression result is then added to the coordinate enhancement feature position by position to form a multi-branch weld feature.

[0143] The low-frequency semantic components and high-frequency edge components in the multi-branch weld features are processed separately using the frequency domain feature processing module to form frequency domain fusion features;

[0144] Specifically, for each feature channel in the multi-branch weld feature, a 3x3 mean filter kernel is used for local smoothing, with the filter step size set to 1. Feature values ​​with the same characteristics as the adjacent edge feature values ​​are added to the feature map edge to keep the feature size consistent before and after local smoothing. For the current filtering position, the average of the 9 feature values ​​in the 3x3 range is used as the smoothing feature value corresponding to the current filtering position.

[0145] The features after 3 rows and 3 columns of mean filtering are identified as low-frequency semantic components that reflect the overall area of ​​the weld, the direction of weld extension, and the main outline of the weld. The multi-branch weld features are subtracted from the corresponding low-frequency semantic components position by position to obtain high-frequency edge components that reflect the weld boundary, rust texture, and local abrupt change locations.

[0146] Continuous pooling and convolution are performed on low-frequency semantic components to enhance the spatial continuity of the weld area, and edge enhancement convolution is performed on high-frequency edge components to enhance the response intensity of weld boundaries and rust textures. The processed low-frequency semantic components and high-frequency edge components are adjusted to have the same feature size and number of feature channels, connected according to the feature channel direction, and the connection results are fused using convolution to form frequency domain fusion features.

[0147] A progressive feature pyramid network is used to perform cross-scale feature fusion of frequency domain fusion features to form cross-scale weld features;

[0148] Specifically, the frequency domain fusion features corresponding to the shallow, middle, and deep feature layers in the improved YOLOv11n weld visual detection model are obtained respectively. The number of feature channels in adjacent feature layers is unified by convolution processing, and the feature size of adjacent feature layers is unified by upsampling or downsampling. The overall semantic information of the weld in the deep feature layer is fused with the weld region information in the middle feature layer for the first time, and then the result of the first fusion is fused with the weld edge information in the shallow feature layer for the second time.

[0149] For each spatial location, shallow feature layer, middle feature layer and deep feature layer are generated respectively. The spatial fusion weights are adjusted to a weight combination with a sum of 1. Each weight is multiplied with the corresponding feature layer position by position, and the multiplication result is added position by position to form a cross-scale weld feature.

[0150] Based on the characteristics of cross-scale welds, the weld contour is identified and the weld centerline is fitted. The rust coverage, weld reinforcement status and surface undulation status are extracted to form weld operation characteristic data.

[0151] Specifically, the cross-scale weld features are input into the detection output layer of the improved YOLOv11n weld visual detection model, which outputs the weld region location and weld region confidence. Weld regions that overlap at the same image location are sorted from high to low according to their weld region confidence. The weld region with the highest confidence is retained. The corresponding weld region is extracted from the weld input image according to the weld region location. The boundary positions between the weld region and the base material regions on both sides with opposite gray-scale change directions and continuous distribution are detected. The boundary positions are connected to form the weld contour.

[0152] Perform Hough line transformation on the boundary points in the weld contour, determine the weld extension direction according to the extension direction with the most boundary point votes, set continuous transverse sections along the weld extension direction, connect the middle positions between the two boundary positions of each transverse section and smooth them to form the weld centerline.

[0153] Within each transverse section, the continuous area where the color depth and texture roughness are higher than the corresponding state of the base material area on both sides is defined as the rust area. The rust coverage range is determined according to the correspondence between the width occupied by the rust area along the transverse section and the weld outline width.

[0154] For the same transverse section, the weld contour and the corresponding image positions in the base material regions on both sides are extracted from the left and right weld images corresponding to the binocular vision camera. According to the positional relationship between the two imaging units pre-calibrated by the binocular vision camera, the corresponding image positions in the left and right weld images are matched, and the three-dimensional spatial coordinates of the corresponding positions are recovered by binocular triangulation to form the three-dimensional contour of the transverse section.

[0155] Select the base material three-dimensional contour points located on both sides of the weld contour from the three-dimensional contour of the transverse section, and fit the base material reference plane of the current transverse section according to the base material three-dimensional contour points; determine the normal distance from each three-dimensional contour point within the weld contour range to the base material reference plane, and determine the largest normal distance as the weld reinforcement height of the current transverse section, and arrange the weld reinforcement heights of each transverse section according to the extension direction of the weld centerline to form the weld reinforcement height state.

[0156] The surface undulation state is determined based on the weld contour width, weld centerline position, weld reinforcement change direction, and weld boundary change direction of adjacent transverse sections. Each transverse section is then associated with the corresponding rust coverage area, weld reinforcement state, surface undulation state, robot walking displacement, and weld spatial coordinates to form weld operation characteristic data.

[0157] In this embodiment, the weld operation characteristic data is arranged along the weld centerline to determine the trimming boundary position. Based on the trimming boundary position, continuous weld sections are divided to form a segmented trimming task, which specifically includes:

[0158] Arrange the weld operation feature data along the weld centerline to form a weld feature position sequence;

[0159] Specifically, for each weld to be processed, the transverse sections are sorted according to the extension direction of the weld centerline;

[0160] When the weld to be processed is a closed weld, the position where the magnetic adsorption wall-climbing robot first arrives at the weld to be processed is taken as the starting point for arrangement. The rust coverage range, weld height status, surface undulation status, robot walking displacement and weld space coordinates corresponding to each transverse section are written into the corresponding position in sequence, and the position is corrected according to the actual distance between adjacent transverse sections on the weld center line.

[0161] The stable positions are determined from the weld feature position sequence, and the fluctuation range between adjacent stable positions is calculated to form the stable position analysis results;

[0162] Specifically, five consecutive transverse sections are used as local analysis units. The rust coverage area, weld reinforcement status and surface undulation status are read respectively. The five status values ​​corresponding to each status are arranged in ascending order. The maximum and minimum status values ​​are removed. The difference between the maximum and minimum status values ​​among the remaining three status values ​​is determined as the local fluctuation amount of the corresponding status.

[0163] The absolute values ​​of the differences between the corresponding state values ​​of adjacent transverse sections are compared in turn. The position where the absolute values ​​of the three consecutive differences do not exceed the corresponding local fluctuation amount is determined as the stable position. The stable positions that are consecutive to each other are merged into a stable position group. The corresponding fluctuation range is determined by the difference between the maximum state value and the minimum state value of each state in each stable position group.

[0164] Based on the results of stable position analysis, positions with continuously changing states are selected to form candidate trimming boundary positions;

[0165] Specifically, along the characteristic position sequence of the weld, the state difference between the rust coverage range, weld reinforcement state and surface undulation state of adjacent transverse sections is calculated sequentially. A state difference greater than 0 is determined as the increasing direction, a state difference less than 0 is determined as the decreasing direction, and a state difference equal to 0 is determined as the unchanged direction.

[0166] When any state changes in three or more consecutive adjacent positions in either an increasing or decreasing direction, the state change amount between the start and end positions of the continuous change is accumulated. The state change amount is then compared with the fluctuation range of the nearest stable position group before and after the start and end positions of the continuous change. The position where the state change amount exceeds the corresponding fluctuation range on both sides is determined as the candidate trimming boundary position.

[0167] The candidate trimming boundary positions are merged to form the trimming boundary positions;

[0168] Specifically, candidate trimming boundary positions located within three consecutive transverse cross-sections are grouped into the same candidate position group. The state change corresponding to each candidate trimming boundary position is divided by the fluctuation range of the adjacent stable position of the corresponding state to obtain the state change degree of the corresponding candidate trimming boundary position. When the fluctuation range of the adjacent stable position is 0, the corresponding state change is taken as the state change degree.

[0169] In the same candidate position group, retain the candidate trimming boundary position with the greatest degree of state change. When multiple candidate trimming boundary positions correspond to the same degree of state change, retain the candidate trimming boundary position that appears first along the moving direction of the magnetic adsorption wall-climbing robot and determine the retention result as the trimming boundary position.

[0170] The continuous weld seam sections are divided according to the repair boundary positions, forming a segmented repair task;

[0171] Specifically, the trimming boundary positions are arranged according to the extension direction of the weld centerline. The starting position of the weld to be treated, the first trimming boundary position, the adjacent trimming boundary positions, and the ending position of the weld to be treated are used as the section boundaries. The welds to be treated between the adjacent section boundaries are determined as continuous weld sections.

[0172] The weld operation feature data located within each continuous weld section are associated with the starting weld space coordinates, ending weld space coordinates, starting robot movement displacement, and ending robot movement displacement of the corresponding continuous weld section. The continuous weld sections are then arranged according to the traversal order of the full weld traversal path to form a segmented trimming task.

[0173] In this embodiment, the grinding head speed, number of dressing operations, and robot walking speed are controlled based on the segmented dressing task. Each continuous weld seam segment is associated with a corresponding parameter combination to form adaptive dressing parameters, which specifically include:

[0174] Collect statistical data on the weld operation characteristics of each continuous weld section to form the section characteristic statistical results;

[0175] Specifically, according to the order of the segmented repair tasks, the rust coverage range, weld reinforcement status and surface undulation status of the transverse cross section in each continuous weld section are read.

[0176] Arrange the state values ​​of each continuous weld section in ascending order, select the state value in the middle position of the arrangement as the section state value of the corresponding continuous weld section, and record the maximum state value, minimum state value and corresponding weld space coordinates among the state values.

[0177] Determine the degree of deviation in the state of each continuous weld section to form the section state deviation result;

[0178] Specifically, the most recent stable position groups located before and after each continuous weld section are retrieved from the stable position analysis results. The average value of the state values ​​corresponding to the two most recent stable position groups is calculated, and the obtained average value is used as the state reference value of the corresponding continuous weld section.

[0179] Subtract the corresponding state reference values ​​from the rust coverage area, weld reinforcement height state, and surface undulation state in the section feature statistics results to obtain the rust state difference, reinforcement height state difference, and undulation state difference. Then, divide the absolute value of each state difference by the fluctuation range of the corresponding most recent stable position group to obtain the rust deviation degree, reinforcement height deviation degree, and undulation deviation degree.

[0180] When the corresponding fluctuation range is 0, the absolute value of the state difference is directly used as the corresponding degree of deviation.

[0181] Establish the calibration relationship between weld condition and calibration parameters, and generate calibration results for calibration parameters.

[0182] Specifically, a calibration weld section with the same material and weld structure as the chemical pressure vessel was selected, and the grinding head speed, robot walking speed and target normal pressure were calibrated using three factors and three levels.

[0183] Three grinding head rotation speed levels of 8000r / min, 9000r / min and 10000r / min are set according to the stable working range of the grinding head; three robot walking speed levels of 3mm / s, 4mm / s and 5mm / s are set according to the stable walking range of the magnetic adsorption wall-climbing robot; and three target normal pressure levels of 50N, 60N and 70N are set according to the stable contact range between the grinding head and the weld.

[0184] Nine sets of trimming calibration parameters were formed by combining the grinding head speed, robot walking speed and target normal pressure according to the three-factor three-level orthogonal test table. For each set of trimming calibration parameters, three independent calibration weld positions were selected, and trimming was performed once according to the same trimming path length.

[0185] For each repair, record the rust coverage area before repair, the weld reinforcement status before repair, the surface undulation status before repair, the rust coverage area after repair, the weld reinforcement status after repair, and the surface undulation status after repair.

[0186] The difference between the rust coverage area before and after the repair corresponding to the same repair calibration parameter combination is determined as the rust removal amount per operation. The difference between the weld reinforcement height before and after the repair is determined as the reinforcement height removal amount per operation. For the three repair results corresponding to the same repair calibration parameter combination, the three rust removal amounts per operation are arranged in ascending order and the median value is taken to form the rust calibration removal amount for the corresponding repair calibration parameter combination. The three reinforcement height removal amounts per operation are arranged in ascending order and the median value is taken to form the reinforcement height calibration removal amount for the corresponding repair calibration parameter combination. This is to reduce the impact of a single abnormal measurement result on the calibration result.

[0187] For a continuous weld section for which adaptive trimming parameters need to be determined, the corrosion fluctuation range and reinforcement height fluctuation range corresponding to the continuous weld section are read. Based on the rust removal amount, reinforcement height removal amount, and corresponding fluctuation range, the single trimming action amount of the current trimming calibration parameter combination is determined:

[0188] ;

[0189] In the formula, , representing the single-time adjustment effect of the j-th group of adjustment calibration parameters on the current continuous weld section. , representing the amount of rust removal corresponding to the j-th combination of calibration parameters. This indicates the corresponding residual high calibration removal amount. This indicates the corrosion fluctuation range corresponding to the most recent stable position group in the current continuous weld section. This indicates the range of residual height fluctuations corresponding to the most recent stable position group; when any fluctuation range is 0, the absolute value of the difference between the states before and after the adjustment is directly used to determine the single adjustment action amount for the corresponding state.

[0190] Record the single dressing action, the surface undulation state after dressing, the grinding head speed, the robot walking speed and the target normal pressure corresponding to each dressing calibration parameter combination, and form the dressing parameter calibration results.

[0191] Based on the deviation results of the section state and the calibration results of the dressing parameters, the grinding head speed, robot walking speed and dressing times are determined to form the material removal parameters;

[0192] Specifically, the degree of rust deviation and the degree of residual height deviation are compared, and the larger deviation is determined as the degree of material removal in the corresponding continuous weld section; from the calibration results of the dressing parameters, the combination of dressing calibration parameters that does not exceed the degree of material removal in a single dressing action and has the smallest difference from the degree of material removal is selected, and the grinding head speed and robot walking speed in the selected dressing calibration parameter combination are determined as the grinding head speed and robot walking speed in the corresponding continuous weld section;

[0193] The single trimming action amount corresponding to the selected trimming calibration parameter combination is deducted from the material removal degree one by one. When the remaining material removal degree is not greater than 1, it means that the remaining state deviation has entered the fluctuation range of the corresponding most recent stable position group. The number of times required to complete the deduction is determined as the trimming number.

[0194] The target normal pressure and the normal pressure adjustment range are determined based on the degree of fluctuation and deviation, thus forming the normal pressure parameters;

[0195] Specifically, from the calibration results of the calibration parameters, the combination of calibration parameters that the surface undulation state after calibration falls within the fluctuation range of the adjacent stable position is selected, and the target normal pressure in the calibration parameter combination that is closest to the state of the adjacent stable position after calibration is determined as the target normal pressure of the corresponding continuous weld section.

[0196] The actual contact force of the magnetic adsorption wall-climbing robot before entering the corresponding continuous weld section is obtained. The pressure difference between the target normal pressure and the actual contact force is calculated. The pressure difference is compared with the maximum pressure change that the feed actuator can stably complete in a single control cycle. The smaller one is determined as the normal pressure adjustment range.

[0197] The material removal parameters and normal pressure parameters are correlated with the corresponding continuous weld sections to form adaptive trimming parameters;

[0198] Specifically, according to the arrangement order of each continuous weld section in the segmented repair task, the corresponding starting weld space coordinates, ending weld space coordinates, grinding head speed, robot walking speed, repair times, target normal pressure and normal pressure adjustment range are recorded, and the parameter switching position between adjacent continuous weld sections is set as the corresponding repair boundary position.

[0199] In this embodiment, the welding seam trimming results, which are generated by trimming each continuous weld seam segment according to adaptive trimming parameters, specifically include:

[0200] The adaptive trimming parameters corresponding to each continuous weld section are retrieved and the walking deviation of the magnetic adsorption wall-climbing robot is determined to form walking deviation data.

[0201] Specifically, according to the order of the segmented repair tasks, the starting weld space coordinates, ending weld space coordinates, grinding head speed, robot walking speed, repair times, target normal pressure and normal pressure adjustment range of the current continuous weld section are read, and the magnetic adsorption wall climbing robot is controlled to move to the starting weld space coordinates.

[0202] The center line of the weld in the current continuous weld section is continuously acquired by a binocular vision camera. The projection position of the geometric center of the magnetic adsorption wall-climbing robot on the weld surface is taken as the current position of the robot, and the forward direction of the magnetic adsorption wall-climbing robot is taken as the walking direction of the robot.

[0203] The shortest lateral distance from the robot's current position to the weld centerline is defined as the lateral deviation, and the angle between the robot's walking direction and the tangent direction of the weld centerline corresponding to the robot's current position is defined as the angular deviation. The offset directions of the lateral deviation and the angular deviation are recorded.

[0204] Adjust the running speed of the left and right drive wheels based on the walking deviation data to form drive wheel control parameters;

[0205] Specifically, the robot's walking speed in the adaptive adjustment parameters is used as the base running speed of the left and right drive wheels. The lateral deviation and angular deviation of the current control cycle, the cumulative result of the corresponding deviation in the continuous control cycle, and the deviation change between the current control cycle and the previous control cycle are calculated respectively. The lateral deviation, angular deviation, cumulative result and deviation change are multiplied by the corresponding adjustment coefficient determined in the magnetic adsorption wall climbing robot walking calibration process, and then summed to obtain the wheel speed adjustment amount.

[0206] When the robot's current position is biased to the left of the weld centerline or the robot's walking direction is biased to the left relative to the tangent direction of the weld centerline, the wheel speed adjustment of the left drive wheel is increased and the wheel speed adjustment of the right drive wheel is decreased on the basic running speed. When the corresponding offset direction is to the right, the wheel speed adjustment of the left drive wheel is decreased and the wheel speed adjustment of the right drive wheel is increased.

[0207] The actual contact force between the grinding head and the weld to be treated is collected and the feed position of the grinding head is adjusted to form feed control parameters;

[0208] Specifically, the grinding head is controlled to run according to the grinding head speed in the adaptive dressing parameters, and the normal force of the grinding head acting on the current continuous weld section is collected by the force sensor according to the continuous control cycle, and the collected normal force is determined as the actual contact force.

[0209] The pressure deviation is obtained by subtracting the actual contact force from the target normal pressure. The pressure deviation of the current control cycle, the cumulative result of the pressure deviation in the continuous control cycle, and the change in pressure deviation between the current control cycle and the previous control cycle are calculated respectively. These three are multiplied by the corresponding adjustment coefficient determined in the calibration process of the feed actuator and then summed to obtain the feed position adjustment amount.

[0210] When the actual contact force is less than the target normal pressure, the grinding head is controlled to move toward the weld surface; when the actual contact force is greater than the target normal pressure, the grinding head is controlled to move away from the weld surface, and the feed position adjustment within a single control cycle is limited to the feed displacement range corresponding to the normal pressure adjustment amplitude.

[0211] The current continuous weld section is trimmed according to the drive wheel control parameters and feed control parameters to form the section trimming execution result;

[0212] Specifically, the left and right drive wheels are controlled to move the magnetic adsorption wall-climbing robot along the weld centerline from the starting weld space coordinates to the ending weld space coordinates at the corresponding running speed. At the same time, the grinding head is controlled to trim the current continuous weld section according to the corresponding grinding head speed and feed position. During the movement, the walking deviation data, drive wheel control parameters and feed control parameters are continuously updated.

[0213] When the magnetic adsorption wall-climbing robot reaches the spatial coordinates of the termination weld, it completes one trimming. When the number of trimmings completed is less than the number of trimmings in the adaptive trimming parameters, the grinding head is controlled to detach from the weld surface, so that the magnetic adsorption wall-climbing robot returns to the spatial coordinates of the starting weld along the walking trajectory that has been trimmed, and performs trimming of the current continuous weld section again until the corresponding number of trimmings is reached.

[0214] The data of the trimming process of the current continuous weld section are recorded synchronously to form a section trimming record;

[0215] Specifically, within each control cycle, the segment identifier, weld space coordinates, actual contact force, grinding head speed, running speed of left and right drive wheels, robot walking speed, walking displacement, walking direction, lateral deviation, angular deviation, grinding head feed position, number of finishing operations completed, and finishing time of the current continuous weld segment are recorded. Data with the same acquisition time are written into the same finishing data unit, and the finishing data units are arranged in order of finishing time.

[0216] Summarize the section trimming records corresponding to each continuous weld section to form the weld trimming trajectory result;

[0217] Specifically, the moment when the actual contact force first enters the pressure range centered on the target normal pressure and bounded by the normal pressure adjustment range is determined from the section trimming record, and the moment when the actual contact force leaves the pressure range due to the grinding head leaving the weld surface is determined as the trimming data unit between the two moments.

[0218] According to the arrangement order of each continuous weld section in the segmented repair task, the effective repair data is associated with the starting weld space coordinates, ending weld space coordinates and adaptive repair parameters of the corresponding continuous weld section, and arranged according to the repair time and robot walking displacement to form the weld repair trajectory result.

[0219] In this embodiment, obtaining the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group, determining the testing time interval based on the weld seam trimming trajectory results, and forming a trimming-inspection correspondence specifically includes:

[0220] Determine the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group to generate installation position data for the operating mechanism;

[0221] Specifically, the center position of the contact area between the grinding head and the weld to be treated is determined as the trimming reference position, the projection position of the midpoint of the line connecting the two probes in the diffraction time-of-flight ultrasonic testing probe group on the surface of the weld to be treated is determined as the inspection reference position, the trajectory length between the trimming reference position and the inspection reference position is measured along the walking trajectory of the magnetic adsorption wall-climbing robot, the trajectory length is determined as the installation distance, and the front-back direction of the inspection reference position relative to the trimming reference position is recorded.

[0222] Extract the position and time when each continuous weld section completes the repair from the weld repair trajectory results to form section repair boundary data;

[0223] Specifically, according to the number of repairs corresponding to each continuous weld section, select the effective repair data corresponding to the last repair from the weld repair trajectory results, read the starting repair time and starting travel displacement corresponding to the starting weld spatial coordinates of the continuous weld section, as well as the ending repair time and ending travel displacement corresponding to the ending weld spatial coordinates of the continuous weld section, and record the corresponding travel direction.

[0224] The starting detection time for the ultrasonic testing probe group to reach the detection position of the starting position of each continuous weld section is determined based on the installation distance.

[0225] Specifically, starting from the trimming data unit corresponding to the initial trimming time, subsequent trimming data units are read sequentially along the same walking direction. The walking displacement of each subsequent trimming data unit is compared with the initial walking displacement. The cumulative displacement difference along the walking direction is calculated, and the trimming data unit where the cumulative displacement difference first reaches the installation distance is selected.

[0226] When the cumulative displacement difference between two adjacent trimming data units is less than and greater than the installation distance, the remaining distance between the installation distance and the smaller cumulative displacement difference is divided by the walking displacement increment between the two adjacent trimming data units to obtain the time allocation ratio. The time allocation ratio is multiplied by the time interval between the two adjacent trimming data units and added to the trimming time corresponding to the smaller cumulative displacement difference to determine the starting detection time.

[0227] The termination time of the ultrasonic testing probe group for diffraction time difference method is determined based on the installation distance to the termination position of each continuous weld section.

[0228] Specifically, starting from the trimming data unit corresponding to the termination trimming time, subsequent trimming data units are read sequentially along the same walking direction. The walking displacement of each subsequent trimming data unit is compared with the termination walking displacement. The cumulative displacement difference along the walking direction is calculated. According to the method for determining the starting detection time, the trimming time corresponding to when the cumulative displacement difference reaches the installation distance is determined. The obtained trimming time is determined as the termination detection time.

[0229] The inspection time intervals for each continuous weld section are determined based on the start and end inspection times, thus forming section inspection time data.

[0230] Specifically, the start time of the inspection is determined as the start time of the inspection of the corresponding continuous weld section, the end time of the inspection is determined as the end time of the inspection of the corresponding continuous weld section, and the continuous time range between the two times is determined as the inspection time interval.

[0231] When the inspection time intervals of adjacent continuous weld sections are connected, the time when the corresponding trimming boundary position reaches the inspection reference position is used as the common boundary of the two inspection time intervals. When the walking direction of the magnetic adsorption wall-climbing robot changes, the cumulative displacement data before the direction change is stopped, and the walking displacement is re-accumulated from the first trimming data unit after the direction change.

[0232] Each continuous weld section is associated with its corresponding inspection time interval to form a repair and inspection correspondence;

[0233] Specifically, according to the arrangement order of each continuous weld section in the segmented repair task, the spatial coordinates of the starting weld, the spatial coordinates of the ending weld, the walking direction corresponding to the last repair, the starting repair time, the ending repair time, the starting inspection time, the ending inspection time, and the installation distance of the corresponding continuous weld section are recorded. Data with the same continuous weld section identifier are written into the same repair and inspection corresponding unit, and each repair and inspection corresponding unit is arranged according to the order of the inspection time interval to form a repair and inspection correspondence.

[0234] In this embodiment, coupling agent is supplied to each continuous weld section according to the repair and inspection correspondence, and ultrasonic signals are collected during the corresponding inspection time interval to form segmented inspection results, specifically including:

[0235] According to the corresponding relationship between repair and inspection, coupling agent is supplied to each continuous weld section to form coupling execution data;

[0236] Specifically, according to the order of the inspection time intervals in each repair and inspection unit, the start inspection time, end inspection time, start weld space coordinates, end weld space coordinates, and the walking direction corresponding to the last repair are read for the current continuous weld section.

[0237] Before the start detection time arrives, the coupling agent supply component is activated to deliver the coupling agent between the diffraction time-of-flight ultrasonic testing probe group and the weld surface, and continuously covers the contact area corresponding to the two probes along the walking direction. After the end detection time arrives, the coupling agent supply for the current continuous weld section is stopped, and the start time of coupling agent supply, the end time of coupling agent supply, and the corresponding continuous weld section are recorded.

[0238] Ultrasonic signals are collected within the detection time interval corresponding to each continuous weld section to form ultrasonic acquisition data;

[0239] Specifically, within the detection time interval, the transmitting probe in the diffraction time-difference ultrasonic testing probe group is controlled to emit ultrasonic pulses according to a continuous detection cycle, and the receiving probe is used to collect the ultrasonic signals propagating through the weld to be processed, and the ultrasonic pulse emission time, ultrasonic signal reception time and signal amplitude corresponding to each ultrasonic signal are recorded.

[0240] Read the robot's walking displacement and direction at the same acquisition time from the weld seam repair trajectory results, and associate each ultrasonic signal with the corresponding robot walking displacement, walking direction, weld seam spatial coordinates and continuous weld seam sections;

[0241] The ultrasonic signals in the acquired ultrasonic data are filtered and their gain is adjusted to form processed ultrasonic signals.

[0242] Specifically, for each ultrasound detection cycle, the signal amplitude acquired before the ultrasound pulse is emitted is determined as the background signal, the amplitude variation range of the background signal is calculated, and the signal components within the amplitude variation range are removed from the ultrasound signal of the current detection cycle.

[0243] The retention frequency range is determined according to the working frequency of the ultrasonic detection probe group based on the diffraction time difference method. Signal components below and above the retention frequency range are removed. The signal gain is gradually increased from early to late according to the ultrasonic signal reception time, so that the signal amplitude corresponding to different propagation distances is adjusted to the same amplitude detection range. The reception time, signal amplitude and robot walking displacement corresponding to each signal peak are retained.

[0244] Identify the through waves in the ultrasonic signal and determine the validity of the ultrasonic signal to form an ultrasonic signal validity result;

[0245] Specifically, in each ultrasonic testing cycle, from the ultrasonic processing signal after the ultrasonic pulse emission time, the continuous signal peak with the earliest reception time and signal amplitude exceeding the range of background signal amplitude variation is extracted, and the continuous signal peak is identified as the through wave.

[0246] Arrange the direct waves corresponding to adjacent ultrasonic testing cycles according to the robot's walking displacement, and record the reception time and signal amplitude of each direct wave in sequence.

[0247] When determining the direct wave reception time for the current ultrasound detection cycle, read the direct wave reception time corresponding to the 20 consecutive ultrasound detection cycles with effective direct waves before the current position, calculate the average and standard deviation of the 20 direct wave reception times, use the average as the local reference reception time corresponding to the current detection position, and use the standard deviation as the local reception time fluctuation.

[0248] When the absolute difference between the direct wave reception time and the local reference reception time in the current ultrasound detection cycle is no more than 3 times the local reception time fluctuation, the current direct wave reception time is determined as the continuous reception time; when the absolute difference is greater than 3 times the local reception time fluctuation, the current position is determined as the abnormal position of the direct wave reception time.

[0249] Locations where the direct wave is missing, where there is an abnormal direct wave reception time, or where the direct wave signal amplitude drops to the range of the background signal amplitude variation are defined as abnormal direct wave locations. Locations where the direct wave reception time and signal amplitude are continuous and the above-mentioned abnormal states do not occur are defined as continuous direct wave locations.

[0250] When the through wave in the current continuous weld section is maintained continuously from the start time to the end time, the corresponding ultrasonic signal is determined to be a valid ultrasonic signal; otherwise, the corresponding ultrasonic signal is determined to be an invalid ultrasonic signal.

[0251] Defect diffraction waves are extracted from effective ultrasonic signals to determine the defect location, defect burial depth, and defect length, thus forming defect characterization data.

[0252] Specifically, after the direct wave of each effective ultrasonic signal, local signal peaks whose signal amplitude exceeds the range of background signal amplitude variation are extracted. Local signal peaks in adjacent ultrasonic detection cycles are arranged according to the robot's walking displacement. Local signal peaks whose reception time changes continuously with the robot's walking displacement and appear continuously in adjacent ultrasonic detection cycles are connected to form candidate diffraction wave trajectories. Candidate diffraction wave trajectories located in the same continuous weld section and corresponding to the same weld space range are determined as defect diffraction waves.

[0253] Before using defect diffraction waves to determine the defect burial depth, the ultrasonic propagation velocity in the chemical pressure vessel material is calibrated.

[0254] Specifically, an ultrasonic calibration test block with the same material and material state as the parent material of the chemical pressure vessel to be tested and a known propagation path size is selected. The diffraction time-of-flight ultrasonic testing probe group is set on the surface of the ultrasonic calibration test block according to the probe spacing and installation method during actual testing. The system delays corresponding to the transmitting probe, receiving probe and probe wedge are first corrected.

[0255] An ultrasonic pulse is emitted to the ultrasonic calibration test block, the received signal after the ultrasonic wave propagates along a known propagation path is obtained, the actual propagation time after deducting the system delay is recorded, multiple measurements are performed continuously, and the actual propagation times are arranged in ascending order and the median value is taken as the calibration propagation time.

[0256] The ultrasonic propagation velocity in chemical pressure vessel materials is determined based on the ratio of the known propagation path length to the calibration propagation time in the ultrasonic calibration test block.

[0257] After completing the ultrasonic propagation velocity calibration, the ultrasonic propagation velocity is used as the defect location parameter for the corresponding chemical pressure vessel, and is used to determine the subsequent defect diffraction wave propagation path length and defect burial depth.

[0258] The weld space between the first and last appearance locations of the defect diffraction wave is defined as the defect location. The time difference between the defect diffraction wave reception time and the ultrasonic pulse emission time is multiplied by the ultrasonic propagation speed in the chemical pressure vessel material to obtain the propagation path length from the transmitting probe to the receiving probe via the defect endpoint. Based on the propagation path length, the installation positions of the two probes, and the distance between the two probes, the vertical distance from the defect endpoint to the weld surface is determined, and the vertical distance is defined as the defect burial depth.

[0259] The difference between the robot's walking displacement corresponding to the first and last occurrence positions of the defect diffraction wave is determined as the defect length.

[0260] The ultrasonic signal validity results and defect characterization data are correlated with the corresponding continuous weld sections to form segmented inspection results;

[0261] Specifically, according to the order of each continuous weld section in the correspondence of the finishing inspection, record the starting weld spatial coordinates, ending weld spatial coordinates, inspection time interval, continuous position of the straight wave, abnormal position of the straight wave, and the validity of the ultrasonic signal for the corresponding continuous weld section.

[0262] For continuous weld sections with valid ultrasonic signals, the corresponding defect diffraction wave, defect location, defect burial depth, and defect length are recorded. For continuous weld sections where no defect diffraction wave is extracted, the corresponding defect diffraction wave is recorded as empty. Data with the same continuous weld section identifier are written into the same segment inspection unit to form segment inspection results.

[0263] In this embodiment, the segmented inspection results are correlated with the weld repair trajectory results according to the continuous weld sections to form weld repair and inspection results, specifically including:

[0264] Based on the segment inspection results and weld repair trajectory results of the continuous weld section, joint judgment data of the section is formed.

[0265] Specifically, the process involves reading the continuous weld section identifier, starting weld spatial coordinates, ending weld spatial coordinates, detection time interval, continuous position of the through wave, abnormal position of the through wave, ultrasonic signal validity, defect diffraction wave, defect location, defect burial depth, and defect length corresponding to each segment inspection unit in the segment inspection results.

[0266] Read the actual contact force, grinding head speed, robot walking speed, walking displacement, walking direction, repair time and adaptive repair parameters corresponding to the same continuous weld section in the weld repair trajectory results, write the data with the same starting weld spatial coordinates and ending weld spatial coordinates into the same section joint judgment unit, and arrange the joint judgment units of each section according to the arrangement order of the segment repair tasks.

[0267] For continuous weld sections with abnormal direct wave locations, the coupling agent is resupplied and re-inspected to generate coupling re-inspection data;

[0268] Specifically, the continuous weld seam sections with abnormal direct wave locations or whose ultrasonic signals are determined to be invalid ultrasonic signals are selected from the joint judgment data of the sections. The magnetic adsorption wall-climbing robot is controlled to move the diffraction time difference ultrasonic detection probe group to the starting weld seam spatial coordinates of the corresponding continuous weld seam section, remove the residual coupling agent between the two probes and the weld seam surface, and resupply the coupling agent.

[0269] The detection time interval is redefined according to the spatial coordinates of the starting weld, the spatial coordinates of the ending weld, and the direction of travel of the corresponding continuous weld section. Ultrasonic signals are collected again within the redefined detection time interval. The continuous position of the straight wave, the abnormal position of the straight wave, the validity of the ultrasonic signal, the defect position, the defect burial depth, and the defect length are redefined according to the processing order of filtering, gain adjustment, straight wave identification, and defect diffraction wave extraction.

[0270] The weld trimming trajectory results of each continuous weld section are compared with the adaptive trimming parameters to form the trimming deviation results;

[0271] Specifically, the effective repair data corresponding to each continuous weld section is arranged according to the repair time, and the actual contact force is compared with the target normal pressure.

[0272] The position where the actual contact force is continuously outside the pressure range centered on the target normal pressure and bounded by the normal pressure adjustment range is defined as the pressure deviation position;

[0273] The actual recorded grinding head speed and robot walking speed are compared with the grinding head speed and robot walking speed in the adaptive dressing parameters, respectively.

[0274] The positions where the difference between the two continuously exceeds the stable operating fluctuation range of the corresponding actuator are respectively defined as the speed deviation position and the velocity deviation position;

[0275] The number of completed dressings is compared with the number of dressings in the adaptive dressing parameters. Continuous weld sections with fewer completed dressings than the number of dressings are identified as number deviation sections. The pressure deviation position, rotation speed deviation position, speed deviation position, and number deviation section are associated with the corresponding continuous weld sections.

[0276] The adaptive trimming parameters are redefined based on the trimming deviation results, and supplementary trimming and re-verification are completed to generate supplementary trimming and re-verification data.

[0277] Specifically, the single adjustment action amount corresponding to each parameter combination is retrieved from the adjustment parameter calibration results, targeting continuous weld sections with pressure deviation, rotation speed deviation, or speed deviation.

[0278] The actual contact force, actual grinding head speed and actual robot walking speed are compared with the various parameter combinations, and the parameter combination with the smallest sum of parameter differences is selected.

[0279] Multiply the single trimming action corresponding to the selected parameter combination by the actual number of trimmings completed to obtain the trimming action completed.

[0280] Subtract the amount of dressing action already completed from the amount of material removed to obtain the remaining material removal degree. From the dressing parameter calibration results, reselect the parameter combination that does not exceed the remaining material removal degree in a single dressing action and has the smallest difference from the remaining material removal degree, and redetermine the grinding head speed, robot walking speed, dressing number and target normal pressure.

[0281] The normal pressure adjustment range is redefined based on the pressure difference between the redefined target normal pressure and the current actual contact force, thus forming the redefined adaptive trimming parameters.

[0282] The corresponding continuous weld section is supplemented and repaired according to the newly determined adaptive repair parameters, the weld repair trajectory results are updated, and the corresponding detection time interval is redefined. The coupling agent is supplied, the ultrasonic signal is collected, and the continuous position of the direct wave, the abnormal position of the direct wave, the effectiveness of the ultrasonic signal, the defect position, the defect burial depth, and the defect length are determined.

[0283] Based on the segmented inspection results, coupled re-inspection data, and supplementary repair re-inspection data, each continuous weld segment is classified to form a segment classification result.

[0284] Specifically, for continuous weld sections that have not been re-inspected, the corresponding through-wave continuity state and defect diffraction wave are read from the segmented inspection results; for continuous weld sections that have been re-inspected, the through-wave continuity state and defect diffraction wave formed by the last inspection in the corresponding coupled re-inspection data or supplemented and modified re-inspection data are read.

[0285] A continuous weld section in which the direct wave is maintained continuously from the start of the detection time to the end of the detection time and a defective diffraction wave is present is defined as a weld defect section. A continuous weld section in which the direct wave is maintained continuously from the start of the detection time to the end of the detection time and no defective diffraction wave is present is defined as an inspection qualified section. A continuous weld section in which the direct wave still has an abnormal position after the coupling agent is supplied again or after the supplementary repair is completed is defined as a re-inspection abnormal section.

[0286] Summarize the defective sections, qualified sections, and re-inspection results of the weld to form the weld repair and inspection results;

[0287] Specifically, the data from the re-inspection, the supplementary repair re-inspection, and the data corresponding to the abnormal sections in the re-inspection are summarized into the re-inspection results. According to the order of each continuous weld segment in the segmented repair task, the spatial coordinates of the starting weld, the spatial coordinates of the ending weld, the defect location, the defect burial depth, and the defect length corresponding to the defective weld segment are recorded. The spatial coordinates of the starting weld, the spatial coordinates of the ending weld, and the last inspection time corresponding to the qualified weld segment are recorded. The abnormal position of the direct wave, the last coupling agent supply time, the last supplementary repair time, and the last inspection time corresponding to the abnormal section in the re-inspection are recorded. Each continuous weld segment is associated with the corresponding adaptive repair parameters, weld repair trajectory results, and segmented inspection results to form the weld repair and inspection results.

[0288] An adaptive repair and inspection system for weld seams in chemical pressure vessels includes the following modules:

[0289] The path planning module is used to build a weld topology graph, pair up the odd-degree topological vertices in the weld topology graph, and generate a full traversal path for the weld.

[0290] The visual recognition module is used to acquire weld images and input the weld images into the improved YOLOv11n weld visual inspection model to form weld operation feature data.

[0291] The task segmentation module is used to arrange weld operation feature data along the weld centerline, determine the trimming boundary position, divide the continuous weld section according to the trimming boundary position, and form a segmented trimming task.

[0292] The parameter generation module is used to generate adaptive trimming parameters based on the segmented trimming task;

[0293] The trimming control module is used to trim each continuous weld section according to adaptive trimming parameters, and to generate weld trimming trajectory results.

[0294] The inspection mapping module is used to determine the inspection time interval based on the weld repair trajectory results and to form a correspondence between repair and inspection.

[0295] The ultrasonic testing module is used to supply coupling agent to each continuous weld section according to the corresponding relationship of repair and inspection, and to collect ultrasonic signals in the corresponding detection time interval to form segmented inspection results.

[0296] The result determination module is used to match the segmented inspection results with the weld repair trajectory results according to the continuous weld sections, and form weld repair and inspection results.

[0297] Example 1: To verify the feasibility of this invention in practice, it was applied to the repair and inspection of the inner wall welds of a chemical pressure vessel. This chemical pressure vessel is made of ferromagnetic pressure vessel steel, with an inner diameter of 4200 mm and a wall thickness of 18 mm. The welds to be treated include circumferential welds, longitudinal welds, and cross-connecting welds, with a total weld length of 42.6 m. Initial inspection revealed that the rust coverage of the welds to be treated accounted for 8% to 64% of the weld outline width, the weld reinforcement height ranged from 1.2 mm to 4.6 mm, and the surface undulation along the weld centerline ranged from 0.4 mm to 2.8 mm. The rust coverage, weld reinforcement height, and surface undulation varied continuously at different locations. When repairing with a fixed grinding head speed, a fixed robot walking speed, and a fixed normal pressure, locations with larger rust coverage areas showed insufficient repair, locations with smaller weld reinforcement heights showed excessive material removal, and locations with varying surface undulations also caused fluctuations in the actual contact force.

[0298] The magnetic adsorption wall-climbing robot enters the interior of the chemical pressure vessel through the manhole and adheres to the inner wall of the vessel via a non-contact area array permanent magnet adsorption structure. The robot's total weight is approximately 100N, and its working load, including a grinding head, binocular vision camera, diffraction time-of-flight ultrasonic testing probe assembly, force sensor, feed actuator, and coupling agent supply assembly, is approximately 76N. The total adsorption force provided by the permanent magnet adsorption structure is approximately 450N. Movement tests were conducted on the straight wall of the vessel, at weld protrusions, and at locations where the vessel's curvature changes. The left and right drive wheels maintained continuous operation without detaching from the inner wall of the vessel.

[0299] The control system acquires the 3D model of the chemical pressure vessel and the weld distribution information, maps the weld distribution information to the 3D model of the chemical pressure vessel, and extracts the weld space centerline of each weld to be processed. The intersection positions of intersecting weld space centerlines and the end positions that do not intersect with other weld space centerlines are determined as topological vertices. The weld space centerlines are divided according to adjacent topological vertices, and the resulting weld segments are determined as topological edges to establish a weld topology graph.

[0300] like Figure 4As shown, the weld topology graph established in this embodiment includes 20 topological vertices, which are connected by circumferential and longitudinal weld segments. The control system counts the number of topological edges connected to each topological vertex and determines 8 odd-degree topological vertices. Using the weld length corresponding to each topological edge as the path weight, Dijkstra's shortest path algorithm is used to calculate the shortest connection path between each odd-degree topological vertex and the corresponding weld length, forming the odd-degree topological vertex connection result. Based on the odd-degree topological vertex connection result, the branch and bound method is used to determine the pairing relationship of odd-degree topological vertices step by step. The candidate pairing branches are pruned according to the connection cost of the determined pairing relationship and the lower bound of the remaining pairing length. The complete pairing result with the smallest cumulative weld length is determined as the odd-degree topological vertex pairing result, and the topological edges contained in the corresponding shortest connection path are determined as weld segments to be repeated.

[0301] The control system adds topological edge copies corresponding to the weld segments to be repeated to the weld topology graph, ensuring that the number of topological edges connected to each topological vertex is even, thus forming an augmented weld topology graph. Starting from the topological vertex corresponding to the initial parking position of the magnetically adsorbed wall-climbing robot, the robot continuously moves along untraversed topological edges or topological edge copies. After returning to the starting position, a closed traversal path still containing untraversed topological edges is embedded into an existing closed traversal path, until all topological edges and edge copies in the augmented weld topology graph have been traversed, generating... Figure 4 The weld seam full traversal path is shown.

[0302] In this embodiment, the actual total length of the weld to be processed is 42.6m, the total length of the determined weld segment to be repeated is 2.5m, and the total walking length of the full weld traversal path is 45.1m. Using the conventional path of traveling back and forth along each weld, the total walking length required to cover the same weld to be processed is 57.8m, of which the return walking distance without trimming or inspection tasks is 15.2m. After adopting the full weld traversal path, the return walking distance without trimming or inspection tasks is reduced to 2.5m, and the magnetic adsorption wall-climbing robot can sequentially complete visual recognition, weld trimming, and ultrasonic inspection along a continuous path.

[0303] A magnetically adsorbed wall-climbing robot moves along the entire traversal path of the weld seam, simultaneously acquiring left and right weld seam images of the weld seam to be processed and the base material regions on both sides using a binocular vision camera. The weld seam images are then processed by size unification, brightness equalization, contrast enhancement, and noise suppression before being input into an improved YOLOv11n weld seam visual detection model. The improved YOLOv11n weld seam visual detection model aggregates features along both the horizontal and vertical axes of the image, forming horizontal and vertical coordinate attention weights, which are then used to enhance the spatial location features of the weld seam.

[0304] The weld features enhanced with coordinate attention are input into multi-branch heterogeneous convolutional structures composed of 1x1, 3x3, 5x5, and 7x7 convolutional kernels to extract weld positional relationships, local edges, fine-grained textures, the main weld region, rust distribution, and weld extension morphology. The frequency domain feature processing module decomposes the obtained multi-branch weld features into low-frequency semantic components and high-frequency edge components, enhancing the overall continuity of the weld region, weld boundaries, rust textures, and local abrupt changes, respectively. A progressive feature pyramid network fuses features from shallow, medium, and deep feature layers hierarchically to identify the weld contour and fit the weld centerline.

[0305] In this embodiment, 680 weld seam images were continuously acquired. The improved YOLOv11n weld seam visual inspection model correctly identified the weld seam region in 628 images, achieving a weld seam recognition accuracy of 92.4%. The average deviation of the weld seam centerline from the manually checked centerline was 0.14 mm, and the maximum deviation was 0.41 mm. When processing the same weld seam images using a YOLOv11n model without a coordinate attention mechanism, a multi-branch heterogeneous convolutional structure, a frequency domain feature processing module, and a progressive feature pyramid network, the weld seam region was correctly identified in 571 images, achieving a weld seam recognition accuracy of 84.0%. The average deviation of the weld seam centerline was 0.47 mm, and the maximum deviation was 1.36 mm.

[0306] The control system sets continuous transverse sections along the weld centerline, extracts the rust coverage area, weld reinforcement status, and surface undulation status corresponding to each transverse section, and correlates them with the robot's walking displacement and weld spatial coordinates to form weld operation characteristic data. The control system arranges the weld operation characteristic data along the weld centerline and determines the trimming boundary positions based on the direction of change of various status values ​​within continuous positions and the fluctuation range of adjacent stable positions. In this embodiment, a total of 26 trimming boundary positions are determined, and the 42.6m weld to be treated is divided into 27 continuous weld sections, forming a segmented trimming task.

[0307] The control system reads the weld operation characteristic data corresponding to each continuous weld section, calculates the degree of rust deviation, the degree of weld reinforcement deviation, and the degree of undulation deviation, and determines the grinding head speed, robot walking speed, number of repairs, target normal pressure, and normal pressure adjustment range based on the calibration results of the repair parameters, forming adaptive repair parameters. The grinding head speed is adjusted within the range of 8000 r / min to 10000 r / min, the robot walking speed is adjusted within the range of 3 mm / s to 5 mm / s, and the target normal pressure is adjusted within the range of 50 N to 70 N. For continuous weld sections with increased rust coverage or weld reinforcement, a higher grinding head speed, a lower robot walking speed, and an increased number of repairs are used. For continuous weld sections with significant changes in surface undulation, the target normal pressure and normal pressure adjustment range are used corresponding to the degree of undulation deviation.

[0308] During the finishing process, the control system adjusts the running speed of the left and right drive wheels based on the lateral and angular deviations of the magnetically adsorbed wall-climbing robot relative to the weld centerline, keeping the grinding head within the weld contour range. Force sensors continuously collect the actual contact force between the grinding head and the weld surface, and the feed actuator adjusts the feed position of the grinding head based on the deviation between the actual contact force and the target normal pressure. The average deviation of the actual contact force for 27 consecutive weld sections is 0.32N, with a maximum deviation of 0.49N. When using a fixed feed position for finishing, the average deviation of the actual contact force is 3.7N, with a maximum deviation of 8.4N.

[0309] The control system synchronously records the actual contact force, grinding head speed, robot walking speed, walking displacement, walking direction, and finishing time for each continuous weld section, forming the weld finishing trajectory result. After finishing, the surface roughness of the weld decreased from Ra38.6μm to Ra76.4μm before finishing to Ra17.1μm to Ra21.8μm. Of the 27 continuous weld sections, 26 sections reached the target surface state after the first finishing, while one continuous weld section was identified as requiring additional finishing because the actual contact force continued to deviate from the target normal pressure.

[0310] The diffraction time-of-flight ultrasonic testing probe group is set behind the grinding head. The control system obtains the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group, and determines the detection time interval of each continuous weld section after the grinding head completes the repair and reaches the detection position of the diffraction time-of-flight ultrasonic testing probe group, based on the robot's walking displacement, walking direction and repair time in the weld repair trajectory results, thus forming a repair and inspection correspondence.

[0311] The coupling agent supply component supplies coupling agent to the corresponding continuous weld section before the detection time interval arrives, and the diffraction time-of-flight ultrasonic testing probe group acquires ultrasonic signals within the detection time interval. The control system filters, adjusts the gain, identifies the direct wave, and extracts the defect diffraction wave from the ultrasonic signal. The validity of the ultrasonic signal is determined based on the continuity of the direct wave, and the defect location, defect burial depth, and defect length are determined based on the defect diffraction wave in the valid ultrasonic signal, forming a segmented inspection result.

[0312] During the initial inspection, two continuous weld sections exhibited abnormal direct-through waves. After resupplying the coupling agent to these two sections and re-inspecting, the direct-through wave of one section returned to continuity, while the other section still exhibited abnormalities. At this point, 26 continuous weld sections, representing 96.3% of all continuous weld sections, had valid ultrasonic signals. The control system correlated the sections with abnormal direct-through waves with the weld trimming trajectory results, determining that the actual contact force of these sections had consistently deviated from the adaptive trimming parameters. The adaptive trimming parameters were then redefined and supplementary trimming was performed. After supplementary trimming and re-inspection, the direct-through waves of the continuous weld sections returned to continuity, and ultimately, all 27 continuous weld sections had valid ultrasonic signals.

[0313] Ultimately, two continuous weld sections exhibiting defective diffraction waves were detected. One defect had a depth of 4.6 mm and a length of 12.8 mm, compared to 4.4 mm and 12.5 mm in the comparison test block. The other defect had a depth of 7.1 mm and a length of 18.6 mm, compared to 7.3 mm and 18.2 mm in the comparison test block. The remaining 25 continuous weld sections showed continuous direct waves and no defective diffraction waves, and were therefore identified as acceptable inspection sections. The control system summarized the defective weld sections, acceptable inspection sections, and re-inspection results to form the weld repair and inspection results.

[0314] Table 1. Results of Weld Repair and Inspection of Chemical Pressure Vessels

[0315] As can be seen from the table above, under the condition that the total length of the welds to be processed is 42.6m, the method of the present invention has achieved relatively stable comparative results in terms of walking path planning, weld visual recognition, finishing process control, and ultrasonic inspection. In terms of path planning, the total walking length of the conventional processing method is 57.8m, and the invalid return distance is 15.2m; the method of the present invention generates a full traversal path of the welds by establishing a weld topology map, pairing odd-degree topology vertices, and adding weld segments to be repeated, thereby reducing the total walking length to 45.1m and the invalid return distance to 2.5m, both reduced by 12.7m. This indicates that the path can cover all welds to be processed while reducing the amount of time spent walking without working.

[0316] In weld seam recognition and finishing control, the improved YOLOv11n weld seam visual inspection model increased the weld seam recognition accuracy from 84.0% to 92.4%, and reduced the average deviation of the weld seam centerline from 0.47 mm to 0.14 mm. These results demonstrate that the coordinate attention mechanism enhances the lateral and longitudinal positional features of the weld seam, the multi-branch heterogeneous convolutional structure can simultaneously extract local edges, main regions, and extended morphology of the weld seam, the frequency domain feature processing module can retain low-frequency semantic components and high-frequency edge components respectively, and the progressive feature pyramid network can improve the fusion degree between weld seam features at different scales. After implementing segmented trimming based on weld centerline and weld operation characteristic data, the average deviation of actual contact force decreased from 3.7N to 0.32N, the number of qualified segments in the first trimming increased from 21 to 26, and the surface roughness after trimming converged from Ra19.6μm to Ra29.8μm to Ra17.1μm to Ra21.8μm, indicating that adaptive trimming parameters and feed position feedback can reduce trimming fluctuations between different weld segments.

[0317] In ultrasonic testing, the method of this invention correlates the weld seam trimming trajectory results with the detection time interval. After coupling and re-inspection, the proportion of effective ultrasonic signal segments increases from 88.9% to 96.3%, the average error of defect burial depth decreases from 0.62 mm to 0.20 mm, and the average error of defect length decreases from 0.85 mm to 0.35 mm. By identifying coupling anomalies through the continuous state of the direct wave and combining adaptive trimming parameters to identify trimming deviations, re-inspection or supplementary trimming can be performed on the corresponding continuous weld seam segments. The total time for trimming and inspection is reduced from 286 min to 198 min, a reduction of 88 min, indicating that this invention can improve the continuity of the trimming and inspection process and the degree of data correspondence while maintaining complete weld seam coverage.

[0318] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for adaptive repair and inspection of weld seams in chemical pressure vessels, characterized in that, Includes the following steps: Obtain the 3D model and weld distribution information of the chemical pressure vessel, establish the weld topology map, pair the odd-degree topological vertices in the weld topology map, and generate the full traversal path of the weld. Control the magnetic adsorption wall-climbing robot to move along the full traverse path of the weld and collect weld images. Input the weld images into the improved YOLOv11n weld visual inspection model to form weld operation feature data. Arrange the weld operation characteristic data along the weld centerline, determine the repair boundary position, divide the continuous weld section according to the repair boundary position, and form a segmented repair task. Based on the segmented trimming task, the grinding head speed, trimming times and robot walking speed are controlled, and each continuous weld seam segment is associated with the corresponding parameter combination to form adaptive trimming parameters. The weld seam sections are trimmed according to the adaptive trimming parameters to form the weld seam trimming trajectory result. Obtain the installation distance between the grinding head and the ultrasonic testing probe group using the diffraction time-of-flight method, determine the testing time interval based on the weld seam trimming trajectory results, and establish a corresponding relationship between trimming and inspection. According to the corresponding relationship between repair and inspection, coupling agent is supplied to each continuous weld section, and ultrasonic signals are collected in the corresponding inspection time interval to form segmented inspection results; The segmented inspection results are matched with the weld repair trajectory results according to the continuous weld sections to form weld repair and inspection results.

2. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The process of acquiring the 3D model and weld distribution information of the chemical pressure vessel, establishing a weld topology map, pairing up the odd-degree topological vertices in the weld topology map, and generating a full traversal path for the weld specifically includes: The weld distribution information is mapped onto the 3D model of the chemical pressure vessel, and a weld topology map is established based on the weld intersection positions and adjacent weld segments. For each topological vertex in the weld topology diagram, determine the odd-degree topological vertices and form a set of odd-degree topological vertices; Pair up the odd-degree topological vertices in the set of odd-degree topological vertices without repetition to determine the pairing results of the odd-degree topological vertices. The weld segment to be repeated is determined based on the odd-degree topology vertex pairing results, and a copy of the topology edge corresponding to the weld segment to be repeated is added to the weld topology graph to form an augmented weld topology graph. Generate a full traversal path for the weld seam based on the augmented weld seam topology.

3. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The controlled magnetic adsorption wall-climbing robot moves along the entire traverse path of the weld seam and collects weld seam images. The weld seam images are input into the improved YOLOv11n weld seam visual inspection model to form weld seam operation feature data, specifically including: Acquire and process weld images of the weld to be processed to form weld input images; The weld seam input image is input into the feature extraction module of the improved YOLOv11n weld seam visual inspection model. The coordinate attention mechanism is used to enhance the spatial location features of the weld seam in the weld seam input image, forming coordinate-enhanced features. Multi-branch heterogeneous convolutional structures are used to extract multi-level weld seam features from coordinate enhancement features, forming multi-branch weld seam features. The low-frequency semantic components and high-frequency edge components in the multi-branch weld features are processed separately using the frequency domain feature processing module to form frequency domain fusion features; A progressive feature pyramid network is used to perform cross-scale feature fusion of frequency domain fusion features to form cross-scale weld features; Based on the characteristics of cross-scale welds, the weld contour is identified and the weld centerline is fitted. The rust coverage, weld reinforcement status and surface undulation status are extracted to form weld operation characteristic data.

4. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The process of arranging weld operation characteristic data along the weld centerline to determine the trimming boundary position, and dividing continuous weld sections according to the trimming boundary position to form segmented trimming tasks, specifically includes: Arrange the weld operation feature data along the weld centerline to form a weld feature position sequence; The stable positions are determined from the weld feature position sequence, and the fluctuation range between adjacent stable positions is calculated to form the stable position analysis results; Based on the results of stable position analysis, positions with continuously changing states are selected to form candidate trimming boundary positions; The candidate trimming boundary positions are merged to form the trimming boundary positions; The continuous weld seam sections are divided according to the repair boundary positions, forming a segmented repair task.

5. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The adaptive trimming parameters, which control the grinding head speed, trimming cycles, and robot walking speed based on segmented trimming tasks, associate each continuous weld seam segment with corresponding parameter combinations to form adaptive trimming parameters, specifically include: Collect statistical data on the weld operation characteristics of each continuous weld section to form the section characteristic statistical results; Based on the statistical results of the section characteristics, the degree of corrosion deviation, the degree of reinforcement deviation, and the degree of fluctuation deviation of each continuous weld section are determined to form the section state deviation results; Establish the calibration relationship between weld condition and calibration parameters, and generate calibration results for calibration parameters. Based on the deviation results of the section state and the calibration results of the dressing parameters, the grinding head speed, robot walking speed and dressing times are determined to form the material removal parameters; The target normal pressure and the normal pressure adjustment range are determined based on the degree of fluctuation and deviation, thus forming the normal pressure parameters; By associating material removal parameters and normal pressure parameters with the corresponding continuous weld sections, adaptive trimming parameters are formed.

6. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The specific steps of adjusting each continuous weld section according to adaptive adjustment parameters to form a weld adjustment trajectory result include: The adaptive trimming parameters corresponding to each continuous weld section are retrieved and the walking deviation of the magnetic adsorption wall-climbing robot is determined to form walking deviation data. Adjust the running speed of the left and right drive wheels based on the walking deviation data to form drive wheel control parameters; The actual contact force between the grinding head and the weld to be treated is collected and the feed position of the grinding head is adjusted to form feed control parameters; The current continuous weld section is trimmed according to the drive wheel control parameters and feed control parameters to form the section trimming execution result; The data of the trimming process of the current continuous weld section are recorded synchronously to form a section trimming record; Summarize the section repair records corresponding to each continuous weld section to form the weld repair trajectory result.

7. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The process of obtaining the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group, determining the testing time interval based on the weld seam trimming trajectory results, and establishing a trimming-inspection correspondence specifically includes: Determine the installation distance between the grinding head and the diffraction time-of-flight ultrasonic testing probe group to generate installation position data for the operating mechanism; Extract the position and time when each continuous weld section completes the repair from the weld repair trajectory results to form section repair boundary data; The starting detection time for the ultrasonic testing probe group to reach the detection position of the starting position of each continuous weld section is determined based on the installation distance. The termination time of the ultrasonic testing probe group for diffraction time difference method is determined based on the installation distance to the termination position of each continuous weld section. The inspection time intervals for each continuous weld section are determined based on the start and end inspection times, thus forming section inspection time data. Each continuous weld section is associated with its corresponding inspection time interval to form a repair and inspection correspondence.

8. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The process of supplying coupling agent to each continuous weld section according to the repair and inspection correspondence, and collecting ultrasonic signals during the corresponding inspection time interval to form segmented inspection results specifically includes: According to the corresponding relationship between repair and inspection, coupling agent is supplied to each continuous weld section to form coupling execution data; Ultrasonic signals are collected within the detection time interval corresponding to each continuous weld section to form ultrasonic acquisition data; The ultrasonic signals in the acquired ultrasonic data are filtered and their gain is adjusted to form processed ultrasonic signals. Identify the through waves in the ultrasonic signal and determine the validity of the ultrasonic signal to form an ultrasonic signal validity result; Defect diffraction waves are extracted from effective ultrasonic signals to determine the defect location, defect burial depth, and defect length, thus forming defect characterization data. The ultrasonic signal validity results and defect characterization data are correlated with the corresponding continuous weld sections to form segmented inspection results.

9. The adaptive repair and inspection method for weld seams of chemical pressure vessels according to claim 1, characterized in that, The process of correlating the segmented inspection results with the weld repair trajectory results according to the continuous weld sections to form weld repair and inspection results specifically includes: The segmented inspection results are correlated with the weld repair trajectory results according to the continuous weld section to form joint judgment data for the section; For continuous weld sections with abnormal direct wave locations, the coupling agent is resupplied and re-inspected to generate coupling re-inspection data; The weld trimming trajectory results of each continuous weld section are compared with the adaptive trimming parameters to form the trimming deviation results; The adaptive trimming parameters are redefined based on the trimming deviation results, and supplementary trimming and re-verification are completed to generate supplementary trimming and re-verification data. Based on the segmented inspection results, coupled re-inspection data, and supplementary repair re-inspection data, each continuous weld segment is classified to form a segment classification result. The results of weld repair and inspection are summarized, including defective sections, qualified sections, and re-inspection results.

10. An adaptive repair and inspection system for weld seams of chemical pressure vessels, comprising the adaptive repair and inspection method for weld seams of chemical pressure vessels as described in any one of claims 1 to 9, characterized in that, Includes the following modules: The path planning module is used to build a weld topology graph, pair up the odd-degree topological vertices in the weld topology graph, and generate a full traversal path for the weld. The visual recognition module is used to acquire weld images and input the weld images into the improved YOLOv11n weld visual inspection model to form weld operation feature data. The task segmentation module is used to arrange weld operation feature data along the weld centerline, determine the trimming boundary position, divide the continuous weld section according to the trimming boundary position, and form a segmented trimming task. The parameter generation module is used to generate adaptive trimming parameters based on the segmented trimming task; The trimming control module is used to trim each continuous weld section according to adaptive trimming parameters, and to generate weld trimming trajectory results. The inspection mapping module is used to determine the inspection time interval based on the weld repair trajectory results and to form a correspondence between repair and inspection. The ultrasonic testing module is used to supply coupling agent to each continuous weld section according to the corresponding relationship of repair and inspection, and to collect ultrasonic signals in the corresponding detection time interval to form segmented inspection results. The result determination module is used to match the segmented inspection results with the weld repair trajectory results according to the continuous weld sections, and form weld repair and inspection results.

Images