Submerged arc welding seam dynamic calibration method

Abstract

The application relates to the technical field of welding automation, and discloses a submerged-arc welding seam dynamic calibration method, which comprises the following steps: collecting an initial welding seam image and performing pretreatment, determining initial welding seam feature points according to an initial reference image, and establishing a reference coordinate according to the initial welding seam feature points; collecting a current welding seam image and performing denoising processing, determining mechanical coordinates of current welding seam feature points based on the denoised current welding seam image; judging whether to update the reference image based on the similarity of the current welding seam image and the reference image; calculating the deviation between the mechanical coordinates and the reference coordinates, processing the deviation by using a dimension control algorithm, generating and outputting calibration instructions for driving the welding gun to move; acquiring actual pose feedback data of the welding gun, calculating an error correction value based on the actual pose feedback data and the expected pose of the calibration instructions, adjusting parameters according to the error correction value; and repeating the above steps until welding is completed. The application significantly improves the tracking precision and welding quality under complex working conditions.

Description

Technical Field

[0001] This application relates to the field of welding automation technology, and in particular to a dynamic calibration method for submerged arc welding seams. Background Technology

[0002] Submerged arc welding is widely used in heavy manufacturing fields such as shipbuilding, bridges, and pressure vessels due to its high efficiency and high quality. In these critical structures, the quality of the weld directly determines the strength and lifespan of the product, requiring extremely high welding precision.

[0003] However, in actual welding processes, factors such as workpiece assembly errors, welding thermal deformation, surface oxide scale interference, and welding torch movement deviation often cause the actual position of the weld to deviate from the preset path. If not corrected in time, defects such as weld misalignment, undercut, and incomplete penetration will occur, affecting structural safety. Therefore, achieving real-time accurate tracking and calibration of the weld is crucial to ensuring welding quality. Currently, existing weld tracking technologies mainly suffer from the following shortcomings: First, in terms of image processing, most rely on single filtering algorithms and fixed threshold segmentation, which are insufficient to handle high-frequency noise, pulse interference, and complex conditions such as workpiece surface reflection and oxide scale during the welding process, leading to inaccurate weld area identification. Second, in terms of feature localization, most rely on template matching based on fixed reference images, and the reference update uses a timed or global refresh mode, which cannot adapt to the dynamic changes in weld morphology during the welding process, easily resulting in feature point positioning drift, and lacking verification of the reliability of the localization results. Third, regarding control strategies, a single PID controller is commonly used to uniformly handle deviations in multiple dimensions such as level and height. The control parameters are fixed or switch roughly, making it difficult to adapt to the differences in working conditions under different materials, plate thicknesses, and welding speeds. This results in slow response, insufficient correction accuracy, and even overshoot or lag. Fourth, regarding data fusion, even when multiple sensors are introduced, it is mostly a simple weighted average, lacking dynamic evaluation of data reliability. The rate of outlier data removal is low, affecting the reliability of the correction values.

[0004] In summary, existing technologies have significant shortcomings in terms of robustness of image processing, reliability of feature localization, accuracy of multi-dimensional control, and effectiveness of data fusion, making it difficult to meet the high-precision welding requirements of high-end manufacturing. Summary of the Invention

[0005] The purpose of this invention is to address at least one deficiency in the existing technology and to provide a dynamic calibration method for submerged arc welding seams. This invention achieves adaptive optimization of submerged arc welding seams in terms of image processing, feature localization, and dimensional control, significantly improving tracking accuracy and welding quality under complex working conditions.

[0006] To achieve the above objectives, the present invention provides a dynamic calibration method for submerged arc welding seams, the method comprising the following steps:

[0007] S1: Acquire an initial weld image and preprocess it to obtain an initial reference image; determine the initial weld feature points based on the initial reference image, and establish reference coordinates based on the initial weld feature points; S2: Acquire the current weld seam image and perform denoising processing; determine the mechanical coordinates of the current weld seam feature points based on the denoised current weld seam image. S3: Calculate the similarity between the current weld image and the reference image. If the similarity is less than a preset threshold, update the reference image and use the updated reference image as the reference image for similarity comparison in the next frame image processing cycle. S4: Calculate the deviation between the mechanical coordinates and the reference coordinates, and use a multi-dimensional control algorithm to process the deviation, generate and output a calibration command to drive the welding torch movement, wherein the deviation includes horizontal deviation, height deviation and slope deviation; S5: Obtain the actual pose feedback data of the welding torch, and calculate the error correction value based on the actual pose feedback data and the expected pose of the calibration command. According to the magnitude of the error correction value, dynamically adjust the control parameters of the multi-dimensional control algorithm and / or the image sampling frequency of the industrial camera. S6: Repeat steps S2-S5 until welding is complete.

[0008] Furthermore, step S1 specifically includes: S100: The initial weld image is processed using an adaptive threshold segmentation algorithm to separate the weld area from the background area, resulting in a segmented binary image; S101: Perform morphological opening operation on the segmented binary image to obtain a denoised weld area image; S102: Apply an edge detection operator to the denoised weld area image to extract the weld edge contour and obtain a contour image; S103: Based on the geometric features of the contour image, select the effective weld area from the contour image, and use the image containing the effective area as the reference image; S104: Locate initial weld feature points in the reference image. The initial weld feature points include at least one of the weld centerline endpoint, inflection point, and weld edge contour turning point. S105: Using the initial weld feature points as a reference, based on the preset mapping relationship between the image pixel coordinate system and the actual welding coordinate system, the pixel coordinates of the initial weld feature points are converted into mechanical coordinates under the actual welding coordinate system to establish the reference coordinates.

[0009] Furthermore, the denoising in step S2 includes sequentially performing Gaussian filtering and median filtering, and determining the filtering parameters using an iterative optimization method, specifically including: S200: Set the initial filtering parameters, and perform Gaussian filtering and median filtering on the current weld image in sequence to obtain the denoised image; S201: Calculate the denoising effect evaluation value F of the denoised image relative to the image before filtering, wherein the expression for F is as follows:

[0010] in, The signal-to-noise ratio of the image before filtering; The signal-to-noise ratio of the filtered image; The equivalent noise level of the image before filtering; The equivalent noise level of the filtered image; S202: Compare the calculated F value with the preset denoising effect threshold. If the F value is less than the denoising effect threshold, adjust the filtering parameters and re-execute Gaussian filtering and median filtering on the current weld image until the calculated F value is greater than or equal to the threshold. Output the final image as the denoised current weld image.

[0011] Furthermore, updating the reference image in step S3 specifically includes: (1). In each frame image processing cycle, calculate the similarity between the current weld seam image after composite filtering and denoising and the currently stored reference image; (2). If the similarity is lower than the image similarity threshold, the LAE network is triggered to locally update the reference image: extract the core features of the weld in the current weld image, locate the local area where the weld shape has changed; perform adaptive correction on the local area, and smoothly merge the corrected area with the unchanged area in the reference image to obtain the updated reference image; use the updated reference image as the reference image for similarity comparison in the next frame image processing cycle.

[0012] Furthermore, step S4 specifically includes: S400: Calculate the deviation between the mechanical coordinates and the reference coordinates. For the horizontal deviation, use a PID controller to correct the deviation and dynamically adjust the proportional coefficient, integral time, and derivative time of the PID controller according to the absolute value of the horizontal deviation and its rate of change. S401: For the height deviation, a fuzzy PID controller is used for correction, and the fuzzy rule base of the fuzzy PID controller is dynamically adjusted according to the weld thickness and welding current process parameters. S402: For the slope deviation, a sliding mode controller is used for correction, and the system chattering in the sliding mode control is weakened by designing an exponential approach law; S403: At the end of each control cycle, calculate the overall response speed V of the current control cycle, the expression of which is:

[0013] in, This refers to the time for correcting horizontal deviations. To control the weights horizontally; The time for correcting height deviation; For high-level control of weights; The slope deviation correction time; As the slope control weight, the comprehensive response speed V is used to guide the online optimization of the control parameters of the multidimensional control algorithm.

[0014] Further, step S5 involves calculating the error correction value, specifically including: S501: Uses a laser rangefinder and a photoelectric encoder to acquire actual distance data and displacement error data, respectively; S502: Synchronize and align the actual distance data and the displacement error data according to the timestamp; S503: The two sets of aligned data are input into the Kalman filter algorithm for fusion to obtain the actual pose feedback data; S504: Calculate the error correction value based on the actual pose feedback data and the expected pose of the calibration command.

[0015] Furthermore, before step S504, the method further includes: calculating the standard deviation of the two sets of data before fusion, and calculating the fusion reliability R based on the preset weighting coefficients of the laser ranging data and the photoelectric encoder data, the expression of which is as follows:

[0016] in, These are the weighting coefficients for laser ranging data; These are the weighting coefficients for the photoelectric encoder data; The standard deviation of the first two sets of data is used to merge the data; if R is greater than or equal to the preset confidence threshold, the correction value is calculated.

[0017] Furthermore, in step S5, parameter adjustments are made based on the error correction value. Specifically, the rule is: if the error correction value... If the difference is 0.05mm, then maintain the current control parameters and sampling frequency; if the difference is 0.05mm... Error correction value When the error is 0.1mm, the sampling frequency should be increased by 20%-50%; if the error correction value is... When the value is 0.1mm, the sampling frequency is increased and the control parameters of the dimensionality control algorithm are adjusted.

[0018] Furthermore, after step S2, the following steps are also included: (1). Based on the motion trajectory of weld feature points in the previous N frames of images, the predicted position of weld feature points in the current frame image is predicted by a long short-term memory network, and a reduced search area is defined with the predicted position as the center. (2). Perform feature point localization within the narrowed search area to obtain the actual location of the weld feature points in the current frame; and calculate the location confidence of the feature points in the current frame. (3) Calculate the first deviation between the actual location and the predicted location, compare the first deviation with a preset deviation threshold, and compare the confidence level with a preset confidence threshold; if the first deviation exceeds the deviation threshold or the confidence level is lower than the confidence threshold, the current frame feature point positioning result is determined to be unreliable, trigger the image re-acquisition and feature point re-positioning process, and repeat steps (1)-(2) until a reliable positioning result is obtained; if the first deviation exceeds the deviation threshold and the confidence level is not lower than the confidence threshold, the actual location is used as the pixel coordinate of the weld feature point in the current frame, and converted into real-time mechanical coordinate output through the camera calibration parameters.

[0019] Furthermore, it also includes: Acquire a calibrated weld cross-sectional image and extract weld quality indicators from the cross-sectional image; the weld quality indicators include at least one of weld reinforcement height, weld width and undercut depth. The extracted weld quality indicators are compared with preset standard values ​​to determine whether the quality indicators exceed the allowable deviation range. If the quality index exceeds the allowable deviation range, the cause of the quality deviation is analyzed; based on the analysis results, at least one of the following parameters is selectively adjusted: the relevant parameters for denoising, the baseline image update threshold, or the control parameters of the dimensionality control algorithm. The adjusted parameters are then applied to the corresponding steps in the subsequent welding process.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively suppresses welding noise interference and adapts to weld morphology changes through composite filtering denoising and dynamic reference image updates, ensuring the reliability of feature point localization. Furthermore, it employs a multi-dimensional control strategy to specifically optimize horizontal, height, and slope deviations, enabling quantitative evaluation of control performance. This invention significantly improves the anti-interference capability, positioning accuracy, and control adaptability of weld tracking under complex working conditions, effectively guaranteeing welding quality in high-end manufacturing fields. Attached Figure Description

[0021] Figure 1 This is a flowchart of a submerged arc welding weld dynamic calibration method according to Embodiment 1 of the present invention; Figure 2 This is a flowchart illustrating a dynamic calibration method for submerged arc welding seams according to Embodiment 1 of this aspect. Figure 3 This is a schematic block diagram of the composite filtering and LAE network dynamic update logic in Embodiment 1 of the present invention; Figure 4 This is a schematic block diagram of the dual feedback data fusion and error correction logic in Embodiment 1 of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0023] Example 1 Please see Figure 1 and 2 A preferred embodiment of the present invention provides a dynamic calibration method for submerged arc welding seams, the method comprising the following steps: S1: Acquire an initial weld image and preprocess it to obtain an initial reference image; determine the initial weld feature points based on the initial reference image, and establish reference coordinates based on the initial weld feature points; S2: Acquire the current weld seam image and perform denoising processing; determine the mechanical coordinates of the current weld seam feature points based on the denoised current weld seam image. S3: Calculate the similarity between the current weld image and the reference image. If the similarity is less than a preset threshold, update the reference image and use the updated reference image as the reference image for similarity comparison in the next frame image processing cycle. S4: Calculate the deviation between the mechanical coordinates and the reference coordinates, and use a multi-dimensional control algorithm to process the deviation, generate and output a calibration command to drive the welding torch movement, wherein the deviation includes horizontal deviation, height deviation and slope deviation; S5: Obtain the actual pose feedback data of the welding torch, and calculate the error correction value based on the actual pose feedback data and the expected pose of the calibration command. According to the magnitude of the error correction value, dynamically adjust the control parameters of the multi-dimensional control algorithm and / or the image sampling frequency of the industrial camera. S6: Repeat steps S2-S5 until welding is complete.

[0024] This embodiment effectively suppresses welding noise interference and adapts to weld morphology changes through composite filtering denoising and dynamic reference image updates, ensuring the reliability of feature point positioning. Furthermore, it employs a multi-dimensional control strategy to specifically optimize horizontal, height, and slope deviations, enabling quantitative evaluation of control performance. This invention significantly improves the anti-interference capability, positioning accuracy, and control adaptability of weld tracking under complex working conditions, effectively guaranteeing welding quality in high-end manufacturing.

[0025] In an optional embodiment, step S1 specifically includes: S100: Use an industrial camera to acquire an initial weld seam image, and use an adaptive threshold segmentation algorithm to process the initial weld seam image to separate the weld seam area and the background area, and obtain a segmented binary image. S101: Perform morphological opening operation on the segmented binary image to obtain a denoised weld area image; S102: Apply an edge detection operator to the denoised weld area image to extract the weld edge contour and obtain a contour image; S103: Based on the geometric features of the contour image, select the effective weld area from the contour image, and use the image containing the effective area as the reference image; S104: Locate initial weld feature points in the reference image. The initial weld feature points include at least one of the weld centerline endpoint, inflection point, and weld edge contour turning point. S105: Using the initial weld feature points as a reference, based on the preset mapping relationship between the image pixel coordinate system and the actual welding coordinate system, the pixel coordinates of the initial weld feature points are converted into mechanical coordinates under the actual welding coordinate system to establish the reference coordinates.

[0026] This embodiment employs adaptive threshold segmentation and morphological opening operations to remove background interference and minor noise in sequence. Then, it accurately extracts the effective weld area through edge detection and geometric feature filtering, fundamentally eliminating the influence of interference factors such as spatter and oxide scale on subsequent positioning. On this basis, by locating key feature points such as the endpoints and inflection points of the weld centerline in the reference image, and using a preset mapping relationship to convert their pixel coordinates into mechanical coordinates in the actual welding coordinate system, a precise and reliable initial reference image and reference coordinates are established for the entire tracking system.

[0027] In an optional embodiment, in step S2, the denoising includes sequentially performing Gaussian filtering and median filtering, and determining the filtering parameters using an iterative optimization approach, specifically including: S200: Set the initial filtering parameters, and perform Gaussian filtering and median filtering on the current weld image in sequence to obtain the denoised image; S201: Calculate the denoising effect evaluation value F of the denoised image relative to the image before filtering, wherein the expression for F is as follows:

[0028] in, The signal-to-noise ratio of the image before filtering; The signal-to-noise ratio of the filtered image; The equivalent noise level of the image before filtering; The equivalent noise level of the filtered image; S202: The calculated F value is compared with a preset denoising effect threshold. If the F value is less than the denoising effect threshold, the filtering parameters are adjusted, and Gaussian filtering and median filtering are re-applied to the current weld image until the calculated F value is greater than or equal to the threshold. The final image is then output as the denoised current weld image. Finally, the mechanical coordinates of the current weld feature points are determined based on the denoised current weld image: the denoised current weld image is compared with the currently stored reference image to locate the pixel coordinates of the weld feature points in the current frame, and then converted into real-time mechanical coordinates using the camera calibration parameters.

[0029] This embodiment uses a composite filter of Gaussian filtering and median filtering. First, Gaussian filtering is used to weaken high-frequency noise, and then median filtering is used to remove impulse noise. The filtering parameters are also dynamically adjusted by the noise reduction effect evaluation value F to achieve precise suppression of different noises and more accurate separation of weld seam areas.

[0030] Optionally, after step S2, the method further includes: (1). Based on the motion trajectory of weld feature points in the previous N frames of images, the predicted position of weld feature points in the current frame image is predicted by a long short-term memory network, and a reduced search area is defined with the predicted position as the center. (2) Perform feature point localization within the narrowed search area to obtain the actual location of the weld feature points in the current frame; and calculate the location confidence C of the feature points in the current frame, the calculation expression of which is as follows:

[0031] in, The effective weld area; The total area of ​​the candidate region; This is the regional effectiveness weight, with a value ranging from 0.3 to 0.4; This represents the number of feature points matched with the contour. This represents the total number of feature points extracted. The matching weight is 0.4-0.5; D is the deviation between the feature point and the predicted position. The maximum allowable deviation is 0.1mm by default; This is the deviation weight, with a value ranging from 0.1 to 0.3.

[0032] (3) Calculate the first deviation between the actual location and the predicted location, compare the first deviation with a preset deviation threshold, and compare the confidence level C with a preset confidence threshold; if the first deviation exceeds the deviation threshold or the confidence level C is lower than the confidence threshold, then the current frame feature point positioning result is determined to be unreliable, trigger the image re-acquisition and feature point re-positioning process, and repeat steps (1)-(2) until a reliable positioning result is obtained; if the first deviation exceeds the deviation threshold and the confidence level C is not lower than the confidence threshold, then the actual location is used as the pixel coordinates of the weld feature points in the current frame, and converted into real-time mechanical coordinates output through the camera calibration parameters.

[0033] This embodiment uses an LSTM network to predict the location of feature points in the next frame, narrowing the search range. It also calculates the feature point positioning confidence C to verify the positioning reliability. When the deviation exceeds the threshold, it triggers reacquisition and repositioning, avoiding positioning failures caused by welding spatter and workpiece vibration, effectively improving anti-interference capability and positioning speed.

[0034] In an optional embodiment, in step S3, please refer to Figure 3 First, set the corresponding image similarity threshold according to the welding conditions; specifically, set it to 0.75 for normal conditions and 0.65 for complex conditions. Then, the baseline image update specifically includes: (1). In each frame image processing cycle, the similarity between the current weld seam image after composite filtering and denoising and the currently stored reference image is calculated. The similarity can be pixel comparison, structural similarity SSIM, etc., and the present invention does not limit it. (2) If the similarity is lower than the image similarity threshold, the LAE network is triggered to locally update the reference image: extract the core features of the weld in the current weld image, locate the local area where the weld morphology has changed; adaptively correct the local area, and smoothly merge the corrected area with the unchanged area in the reference image to obtain the updated reference image; use the updated reference image as the reference image for similarity comparison in the next frame image processing cycle. The updated reference image retains the core features of the weld while adapting to the dynamic changes in the weld morphology.

[0035] The core function of the LAE network in this embodiment is to keep the reference image and the real-time weld image locally aligned, adapt to dynamic changes in weld morphology (such as the scenario where the weld is easily deformed in the welding of thin-walled parts), and at the same time enhance the recognizability of the core features of the weld to avoid positioning drift.

[0036] In an optional embodiment, step S4 specifically includes: S400: Calculate the deviation between the mechanical coordinates and the reference coordinates. For the horizontal deviation, a PID controller is used for correction, and the proportional coefficient, integral time, and derivative time of the PID controller are dynamically adjusted according to the absolute value and rate of change of the horizontal deviation. Specifically: First, the initial horizontal PID parameters (proportional coefficient, integral time, derivative time) adapted to the current welding material, plate thickness, and welding speed are retrieved from the process parameter database. Then, the weld feature points are located using an industrial camera, and the real-time horizontal deviation is calculated in conjunction with the reference coordinates. The parameters are dynamically adjusted according to the magnitude of the deviation, the rate of change, and the dual feedback error correction value. Finally, the correction effect is evaluated by the comprehensive response speed V. If the target is not met, the adjustment is repeated until the requirements are met.

[0037] S401: For the aforementioned height deviation, a fuzzy PID controller is used for correction, and the fuzzy rule base of the fuzzy PID controller is dynamically adjusted according to the weld thickness and welding current process parameters. Specifically: first, an initial fuzzy PID rule base adapted to the current weld thickness and welding current is retrieved from the process parameter database; then, the height deviation and error correction value are collected through dual feedback; the membership function and inference rules of the fuzzy rule base are dynamically updated according to changes in operating conditions; finally, closed-loop fine-tuning is performed by combining the comprehensive response speed V and the weld height index to ensure the adaptability of height control.

[0038] S402: For the slope deviation, a sliding mode controller is used for correction, and an exponential reaching law is designed to reduce system chattering in the sliding mode control. Initial sliding mode control parameters adapted to the current working conditions are retrieved from the process parameter database. The slope deviation is calculated by locating weld feature points. A reaching law is designed to reduce system chattering. Control parameters are dynamically optimized using dual feedback error correction values ​​to ensure that the weld direction matches the welding torch posture. The effect is evaluated using the comprehensive response speed V.

[0039] S403: At the end of each control cycle, calculate the overall response speed V of the current control cycle, the expression of which is:

[0040] in, This refers to the time for correcting horizontal deviations. To control the weights horizontally; The time for correcting height deviation; For high-level control of weights; The slope deviation correction time; As the slope control weight, the comprehensive response speed V is used to guide the online optimization of the control parameters of the multidimensional control algorithm.

[0041] This embodiment employs a multi-dimensional control algorithm, dynamically adjusting PID parameters for correction in the horizontal direction, dynamically adjusting fuzzy PID control based on process parameters in the vertical direction using a fuzzy rule base, and using sliding mode control to reduce system chattering in the slope direction. It also evaluates the algorithm's response capability through a comprehensive response speed V, achieving accurate and rapid correction of multi-dimensional deviations and adapting to the differences in working conditions for different welding materials, plate thicknesses, and welding speeds.

[0042] In an optional embodiment, in step S5, please refer to Figure 4 Specifically, it includes: S501: A laser rangefinder and a photoelectric encoder are used to acquire actual distance data and displacement error data, respectively. Specifically, the laser rangefinder is first configured such that the angle between its emitted laser beam and the welding torch axis is 30 to 45 degrees to ensure that the laser spot is projected onto the center line of the weld seam, avoiding welding spatter from obstructing the laser path. The laser rangefinder then collects the actual distance data between the welding torch and the weld seam surface, while the photoelectric encoder simultaneously collects the displacement error data of the welding torch drive motor. S502: Synchronize and align the actual distance data and the displacement error data according to the timestamp; S503: The two sets of aligned data are input into the Kalman filter algorithm for fusion to obtain the actual pose feedback data; during the Kalman filter process, the process noise covariance matrix Q and the observation noise covariance matrix R are dynamically adjusted; wherein, the process noise covariance matrix Q is dynamically updated according to the current welding speed, and the observation noise covariance matrix R is adaptively adjusted based on the historical error statistics of the sensor, so that the rejection rate of abnormal data reaches more than 95%; S504: Based on the actual pose feedback data and the expected pose of the calibration command, calculate the error correction value. Output the error correction value for use in subsequent steps to dynamically adjust the control parameters of the multi-dimensional control algorithm and / or the image sampling frequency of the industrial camera. The specific adjustment rule is: if the error correction value... If the difference is 0.05mm, then maintain the current control parameters and sampling frequency; if the difference is 0.05mm... Error correction value When the error is 0.1mm, the sampling frequency should be increased by 20%-50%; if the error correction value... When the value is 0.1mm, the sampling frequency is increased and the control parameters of the dimensional control algorithm are adjusted.

[0043] Optionally, before step S504, the method further includes: calculating the standard deviation of the two sets of data before fusion, and calculating the fusion confidence level R based on the preset weighting coefficients of the laser ranging data and the photoelectric encoder data, the expression of which is as follows:

[0044] in, This is the weighting coefficient for laser ranging data, with a value of 0.6-0.8 (the upper limit is used when the weld thickness is >10mm). The weighting coefficient for photoelectric encoder data ranges from 0.2 to 0.4. The weighting coefficient is adjusted in real time based on the surface roughness feedback of the weld (the weighting of laser ranging data is increased when the roughness Ra > 3.2 μm). To merge the standard deviations of the first two sets of data; if R is greater than or equal to a preset confidence threshold, then calculate the correction value. If R is less than the preset confidence threshold, then the current error correction value is temporarily distrusted, and the system can take one of the following measures: (1) Keep the current control parameters unchanged and wait for the data in the next cycle; (2) Trigger sensor self-test or recalibration; (3) Temporarily increase the weight difference between the two types of sensors (e.g., give more trust to the one with more stable historical performance); (4) Issue a sensor malfunction warning to the operator.

[0045] Optionally, a welding process parameter database can be established to associate and store welding materials, plate thickness, welding speed parameters with optimal control parameters. During dynamic optimization, the initial control parameters are retrieved from the database based on the current welding process parameters, and secondary optimization is performed by combining the error correction value obtained from the dual feedback verification, thereby shortening the response time of parameter adjustment and improving calibration stability.

[0046] Optionally, it also includes: Acquire a calibrated weld cross-sectional image and extract weld quality indicators from the cross-sectional image; the weld quality indicators include at least one of weld reinforcement height, weld width and undercut depth. The extracted weld quality indicators are compared with preset standard values ​​to determine whether the quality indicators exceed the allowable deviation range. If the quality index exceeds the allowable deviation range, the cause of the quality deviation is analyzed; based on the analysis results, at least one of the following parameters is selectively adjusted: the relevant parameters for denoising, the baseline image update threshold, or the control parameters of the dimensionality control algorithm. The adjusted parameters are then applied to the corresponding steps in the subsequent welding process.

[0047] This embodiment associates and stores welding materials, plate thickness, welding speed and optimal control parameters. During dynamic optimization, it first calls the initial parameters and then combines them with error correction values ​​for secondary optimization, which shortens the parameter adjustment response time. By comparing quality indicators such as weld reinforcement height, weld width and undercut depth with preset standards, it adjusts the filter parameters, LAE network threshold or control parameters in a targeted manner to improve calibration stability.

[0048] In an optional embodiment, in step S6, steps S2-S5 are repeated until the welding is completed.

[0049] Example 2 To verify the effectiveness of the method proposed in Embodiment 1 of this invention, two sets of experiments were conducted in this embodiment: I. Calibration of Submerged Arc Welding of Medium and Thick Carbon Steel Plates (Routine Working Conditions): This experiment is applied to submerged arc welding scenarios of medium and thick carbon steel plates in the shipbuilding industry.

[0050] Welding parameter settings: Welding material is Q345 carbon steel, plate thickness is 12mm, welding speed is 6mm / s, welding current is 600A.

[0051] 1. Initial Reference Establishment: Initial weld seam images are acquired using an industrial camera. An adaptive threshold segmentation algorithm is used to separate the weld seam region from the background region, resulting in a segmented binary image. Minor noise in the image is eliminated by morphological opening operations of a 3×3 structuring element. The Canny edge detection operator is applied to extract the weld seam edge contour. Based on the geometric features of the contour, effective weld seam regions are selected, and the image containing the effective regions is used as the reference image. Initial weld seam feature points are located in the reference image, including the endpoints of the weld seam centerline and the inflection points of the edge contour. Based on the preset mapping relationship between the image pixel coordinate system and the actual welding coordinate system, the pixel coordinates of the initial weld seam feature points are converted into mechanical coordinates to establish the reference coordinates.

[0052] 2. Real-time Sensing: The industrial camera continuously acquires real-time weld seam images at a frame rate of 50fps. Each frame undergoes composite filtering for denoising: first, a Gaussian filter with a standard deviation of 1.2 is used to smooth the image's grayscale distribution; then, a 3×3 window median filter is used to remove impulse noise. The denoising effect evaluation value F is calculated and compared with the target value of 1.8. If F is lower than 1.8, the filtering parameters are dynamically adjusted until the target is met. Based on the motion trajectory of weld seam feature points from the previous N frames, the position of the feature points in the current frame is predicted using an LSTM network, narrowing the localization search range. Feature points are located within the narrowed area, and the localization confidence C is calculated. When the C value is lower than a threshold or the deviation between the localization result and the predicted position exceeds 0.5mm, a re-acquisition process is triggered.

[0053] 3. Baseline Image Update: Calculate the similarity between the denoised current image and the current baseline image, with a similarity threshold set to 0.75. When the similarity is below 0.75, trigger the LAE network to locally update the baseline image: extract the core features of the weld seam in the current image, locate the local area where the weld seam morphology changes, adaptively correct only this area and smoothly merge it with the unchanged area in the baseline image, and use the updated image as the new baseline image for comparison in subsequent frames.

[0054] 4. Decision Control: Based on the real-time mechanical coordinates of the current frame feature points and the reference coordinates established by S1, the horizontal deviation, height deviation, and slope deviation are calculated. For the horizontal deviation, a PID controller is used for correction, with the initial value of the PID proportional coefficient set to 0.8 and dynamically adjusted according to the magnitude and rate of change of the deviation. For the height deviation, a fuzzy PID controller is used, with the fuzzy rule base selected as the medium-thick plate adaptation mode. For the slope deviation, a sliding mode controller is used, with the reaching law parameter set to 0.05 to reduce system chattering. At the end of each control cycle, the correction time for the three dimensions is recorded, and the comprehensive response speed V is calculated. In this embodiment, the V value is stable above 0.08, meeting the target requirement.

[0055] 5. Verification and Adaptation: The angle between the laser rangefinder and the welding torch axis is set to 35° to ensure that the laser spot is projected onto the center line of the weld. The laser rangefinder collects actual distance data, and the photoelectric encoder collects displacement error data. Both are synchronized according to timestamps and then input into the Kalman filter algorithm. Since the welding speed is ≥5mm / s, the process noise covariance matrix Q increases by 30%; the observation noise covariance matrix R is adaptively adjusted based on the historical error statistics of the sensor, maintaining an abnormal data rejection rate of over 98%. The weight w1 of the laser rangefinder data is set to 0.8, and the weight w2 of the photoelectric encoder data is set to 0.2. The fusion reliability R is calculated; in this embodiment, the R value is stable above 0.85. After fusion, the error correction value is obtained. In this embodiment, the error correction value is stable in the range of 0.03~0.04mm, which falls within the small error range of ≤0.05mm. Therefore, the current control parameters and sampling frequency are kept unchanged.

[0056] 6. Cyclic execution: Repeat steps S2 to S5 until welding is completed. The final weld inspection results show that the weld reinforcement height and weld width deviation are both ≤0.1mm, meeting the welding accuracy requirements for medium and thick plate structural components.

[0057] II. Calibration of Submerged Arc Welding of Thin-Walled Stainless Steel (Complex Working Conditions): This experiment is applied to submerged arc welding scenarios of thin-walled stainless steel in the field of medical devices.

[0058] Welding parameter settings: Welding material is 304 stainless steel, plate thickness is 3mm, welding speed is 4mm / s, welding current is 350A (low spatter process).

[0059] 1. Initial benchmark establishment: Considering the reflective strength of stainless steel, the segmentation coefficient of the adaptive threshold segmentation algorithm is adjusted to 1.5 to improve the distinction between the weld area and the background area; the morphological opening operation adopts a 2×2 structuring element to avoid the loss of weld features due to excessive smoothing caused by the thin plate thickness; the remaining steps are the same as in Experiment 1.

[0060] 2. Real-time perception: The industrial camera continuously acquires real-time weld seam images at a frame rate of 60fps. Composite filtering employs a 2×2 window Gaussian filter and median filter to avoid excessive smoothing that could lead to the loss of thin-walled weld seam features; the target value for the denoising effect evaluation value F is set to 1.6. The LSTM prediction and confidence evaluation mechanism is the same as in Experiment 1.

[0061] 3. Baseline Image Update: Considering the morphology of thin-walled stainless steel welds is easily changed, the similarity threshold of the LAE network is reduced to 0.65, and the local update frequency is increased by 20% compared with Experiment 1, so as to adapt to the rapid changes of the weld more timely.

[0062] 4. Decision Control: For horizontal deviation, the PID proportional coefficient is adjusted to 0.6 to reduce overshoot, which is prone to occur in thin-wall welding; for height deviation, the fuzzy PID controller switches to thin-wall high-precision mode; for slope deviation, the control weight is maintained at 0.3. The target value of the overall response speed V is set to above 0.1. In this embodiment, the V value reaches 0.1, which meets the requirement of fast response for thin-wall welding.

[0063] 5. Verification and Adaptation: The angle between the laser rangefinder sensor and the welding torch axis is set to 30°. Since the welding speed is 4 mm / s, which is less than 5 mm / s, the process noise covariance matrix Q remains at its initial value; the observation noise covariance matrix R is optimized for the stainless steel surface roughness Ra = 2.8 μm. Considering the susceptibility of laser reflection signals to interference under thin-walled conditions, the laser rangefinder data weight w1 is set to 0.6, and the photoelectric encoder data weight w2 is set to 0.4; the target value for the fusion reliability R is set to above 0.8, and in this embodiment, the R value meets the requirement. After fusion, the error correction value is ≤0.03 mm, which falls within the small error range, and the current control parameters and sampling frequency remain unchanged.

[0064] 6. Cyclic execution: Repeat steps S2 to S5 until welding is completed. Final weld inspection results show that common problems in thin-walled welding, such as burn-through and undercut, are effectively avoided, and the weld cross-sectional quality indicators meet the standards for precision component welding.

[0065] The two experiments above demonstrate that the method of the present invention can adaptively adjust image processing parameters, benchmark update threshold, control strategy parameters, and data fusion weights according to different welding conditions (material, plate thickness, welding speed, etc.). It can achieve high-precision weld tracking and calibration under both conventional and complex conditions, and has good adaptability and technical effectiveness.

[0066] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for dynamic calibration of submerged arc weld seams, characterized in that, The method includes the following steps: S1: Acquire an initial weld image and preprocess it to obtain an initial reference image; determine the initial weld feature points based on the initial reference image, and establish reference coordinates based on the initial weld feature points; S2: Acquire the current weld seam image and perform denoising processing; determine the mechanical coordinates of the current weld seam feature points based on the denoised current weld seam image. S3: Calculate the similarity between the current weld image and the reference image. If the similarity is less than a preset threshold, update the reference image and use the updated reference image as the reference image for similarity comparison in the next frame image processing cycle. S4: Calculate the deviation between the mechanical coordinates and the reference coordinates, and use a multi-dimensional control algorithm to process the deviation, generate and output a calibration command to drive the welding torch movement, wherein the deviation includes horizontal deviation, height deviation and slope deviation; S5: Obtain the actual pose feedback data of the welding torch, and calculate the error correction value based on the actual pose feedback data and the expected pose of the calibration command. According to the magnitude of the error correction value, dynamically adjust the control parameters of the multi-dimensional control algorithm and / or the image sampling frequency of the industrial camera. S6: Repeat steps S2-S5 until welding is complete.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S100: The initial weld image is processed using an adaptive threshold segmentation algorithm to separate the weld area from the background area, resulting in a segmented binary image; S101: Perform morphological opening operation on the segmented binary image to obtain a denoised weld area image; S102: Apply an edge detection operator to the denoised weld area image to extract the weld edge contour and obtain a contour image; S103: Based on the geometric features of the contour image, select the effective weld area from the contour image, and use the image containing the effective weld area as the reference image; S104: Locate initial weld feature points in the reference image. The initial weld feature points include at least one of the weld centerline endpoint, inflection point, and weld edge contour turning point. S105: Using the initial weld feature points as a reference, based on the preset mapping relationship between the image pixel coordinate system and the actual welding coordinate system, the pixel coordinates of the initial weld feature points are converted into mechanical coordinates under the actual welding coordinate system to establish the reference coordinates.

3. The method according to claim 1, characterized in that, The denoising process in step S2 includes sequentially performing Gaussian filtering and median filtering, and determining the filtering parameters using an iterative optimization method, specifically including: S200: Set the initial filtering parameters, and perform Gaussian filtering and median filtering on the current weld image in sequence to obtain the denoised image; S201: Calculate the denoising effect evaluation value F of the denoised image relative to the image before filtering, wherein the expression for F is as follows: in, The signal-to-noise ratio of the image before filtering; The signal-to-noise ratio of the filtered image; The equivalent noise level of the image before filtering; The equivalent noise level of the filtered image; S202: Compare the calculated F value with the preset denoising effect threshold. If the F value is less than the denoising effect threshold, adjust the filtering parameters and re-execute Gaussian filtering and median filtering on the current weld image until the calculated F value is greater than or equal to the threshold. Output the final image as the denoised current weld image.

4. The method according to claim 1, characterized in that, Step S3 involves updating the reference image, specifically including: (1). In each frame image processing cycle, calculate the similarity between the current weld seam image after composite filtering and denoising and the currently stored reference image; (2). If the similarity is lower than the image similarity threshold, the LAE network is triggered to locally update the reference image: extract the core features of the weld in the current weld image, locate the local area where the weld shape has changed; perform adaptive correction on the local area, and smoothly merge the corrected area with the unchanged area in the reference image to obtain the updated reference image; use the updated reference image as the reference image for similarity comparison in the next frame image processing cycle.

5. The method according to claim 4, characterized in that, Step S4 specifically includes: S400: Calculate the deviation between the mechanical coordinates and the reference coordinates. For the horizontal deviation, use a PID controller to correct the deviation and dynamically adjust the proportional coefficient, integral time, and derivative time of the PID controller according to the absolute value of the horizontal deviation and its rate of change. S401: For the height deviation, a fuzzy PID controller is used for correction, and the fuzzy rule base of the fuzzy PID controller is dynamically adjusted according to the weld thickness and welding current process parameters. S402: For the slope deviation, a sliding mode controller is used for correction, and the system chattering in the sliding mode control is weakened by designing an exponential approach law; S403: At the end of each control cycle, calculate the overall response speed V of the current control cycle, the expression of which is: in, This refers to the time for correcting horizontal deviations. To control the weights horizontally; The time for correcting height deviation; For high-level control of weights; The slope deviation correction time; As the slope control weight, the comprehensive response speed V is used to guide the online optimization of the control parameters of the multidimensional control algorithm.

6. The method according to claim 5, characterized in that, Step S5 involves calculating the error correction value, specifically including: S501: Uses a laser rangefinder and a photoelectric encoder to acquire actual distance data and displacement error data, respectively; S502: Synchronize and align the actual distance data and the displacement error data according to the timestamp; S503: The two sets of aligned data are input into the Kalman filter algorithm for fusion to obtain the actual pose feedback data; S504: Calculate the error correction value based on the actual pose feedback data and the expected pose of the calibration command.

7. The method according to claim 6, characterized in that, Before step S504, the method further includes: calculating the standard deviation of the two sets of data before fusion, and calculating the fusion confidence level R based on the preset weighting coefficients of the laser ranging data and the photoelectric encoder data, the expression of which is as follows: in, These are the weighting coefficients for laser ranging data; These are the weighting coefficients for the photoelectric encoder data; The standard deviation of the first two sets of data is used to merge the data; if R is greater than or equal to the preset confidence threshold, the correction value is calculated.

8. The method according to claim 7, characterized in that, In step S5, parameter adjustments are made based on the error correction value. The specific rule is: if the error correction value... If the difference is 0.05mm, then maintain the current control parameters and sampling frequency; if the difference is 0.05mm... Error correction value When the error is 0.1mm, the sampling frequency should be increased by 20%-50%; if the error correction value is... When the value is 0.1mm, the sampling frequency is increased and the control parameters of the dimensionality control algorithm are adjusted.

9. The method according to claim 1, characterized in that, Following step S2, the following is also included: (1). Based on the motion trajectory of weld feature points in the previous N frames of images, the predicted position of weld feature points in the current frame image is predicted by a long short-term memory network, and a reduced search area is defined with the predicted position as the center. (2). Perform feature point localization within the narrowed search area to obtain the actual location of the weld feature points in the current frame; and calculate the location confidence of the feature points in the current frame. (3). Calculate the first deviation between the actual location and the predicted location, compare the first deviation with a preset deviation threshold, and compare the confidence level with a preset confidence threshold; if the first deviation exceeds the deviation threshold or the confidence level is lower than the confidence threshold, the current frame feature point positioning result is determined to be unreliable, trigger the image re-acquisition and feature point re-positioning process, and repeat steps (1)-(2) until a reliable positioning result is obtained; if the first deviation exceeds the deviation threshold and the confidence level is not lower than the confidence threshold, the actual location is used as the pixel coordinates of the weld feature points in the current frame, and converted into real-time mechanical coordinates output through camera calibration parameters.

10. The method according to any one of claims 1 to 9, characterized in that, Also includes: Acquire a calibrated weld cross-sectional image and extract weld quality indicators from the cross-sectional image; the weld quality indicators include at least one of weld reinforcement height, weld width and undercut depth. The extracted weld quality indicators are compared with preset standard values ​​to determine whether the quality indicators exceed the allowable deviation range. If the quality index exceeds the allowable deviation range, the cause of the quality deviation is analyzed; based on the analysis results, at least one of the following parameters is selectively adjusted: the relevant parameters for denoising, the baseline image update threshold, or the control parameters of the dimensionality control algorithm. The adjusted parameters are then applied to the corresponding steps in the subsequent welding process.

Images