A method for welding h-shaped high-frequency steel with intelligent positioning

Abstract

The present application relates to the technical field of welding, in particular to a kind of intelligent positioning H type high-frequency steel welding method.The method is deployed line laser profile sensor and high-frequency power voltage current acquisition module on the upper part of feed path, extracts the initial assembly angle of wing plate and web as feedforward feature, real-time obtains the high-frequency current phase difference and impedance modulus of welding area as feedback feature.Feedforward feature and feedback feature are input into the mapping relationship model established, and the output V type angle dynamic deviation value is output.According to the deviation value, the displacement compensation amount of the servo motor of extrusion roller is calculated, and the real-time adaptive adjustment of the position of extrusion roller is completed before welding point.The present application deeply couples surface geometry data and internal electromagnetic physical characteristics, solves the problem of V type angle convergence point deviation caused by single dependence on surface geometry detection unable to reflect internal electromagnetic distribution change.

Description

Technical Field

[0001] This invention relates to the field of welding technology, specifically to a method for welding H-shaped high-frequency steel with intelligent positioning. Background Technology

[0002] In existing high-frequency welding technologies, the assembly and positioning of the flanges and webs of H-beams typically employs fixed mechanical stops or pneumatic clamps. During the pre-welding preparation stage, these rigid structures fix the relative positions of the flanges and webs. Once the continuous feeding high-frequency welding process begins, the positioning mechanism no longer interferes with the relative posture of the flanges and webs. Some improved solutions introduce vision sensors for weld seam tracking. By acquiring two-dimensional or three-dimensional image data of the workpiece surface, the assembly angle and misalignment are extracted, and the position of the extrusion rollers is controlled based on surface geometric deviations.

[0003] The core problem with the aforementioned existing technologies is that neither static mechanical rigid positioning nor visual tracking relying on surface geometric images can cope with the dynamic deformation of long H-beams caused by differences in surface flatness during continuous feeding. This leads to dynamic convergence misalignment of the V-angle between the flange and web during high-frequency welding. Visual sensors can only acquire the geometric contour information of the workpiece's exterior and cannot penetrate the metal surface to perceive the distribution changes of the high-frequency electromagnetic field inside the V-angle. When minute gap changes or uneven edge heating occur inside the V-angle, surface geometric detection cannot provide corresponding feedback on the internal physical state. This causes the adjustment action of the extrusion rollers to become disconnected from the actual required electromagnetic convergence state, resulting in convergence point offset and high-frequency current shunting along the plate surface. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent positioning method for welding H-shaped high-frequency steel, which can solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for welding H-beam high-frequency steel with intelligent positioning, comprising: A line laser profile sensor and a high-frequency power supply voltage and current acquisition module are deployed on the feed path of H-beam high-frequency welding. The initial assembly angle and initial misalignment of the wing plate and web plate are extracted using the line laser profile sensor. The high-frequency power supply voltage and current acquisition module acquires the high-frequency current phase difference and impedance modulus of the welding zone in real time. Establish a mapping relationship model between the initial assembly angle, the high-frequency current phase difference, the impedance magnitude, and the actual convergence state of the V-angle; The initial assembly angle is used as a feedforward feature input to the mapping relationship model, and the high-frequency current phase difference and the impedance modulus are used as feedback features input to the mapping relationship model. The V-angle dynamic deviation value is output through the mapping relationship model. The displacement compensation amount of the servo motor of the extrusion roller is calculated based on the dynamic deviation value of the V-angle, and the position of the extrusion roller is adjusted in real time before the welding point based on the displacement compensation amount of the servo motor.

[0006] Preferably, the initial assembly angle and initial misalignment of the wingplate and web plate are extracted using the line laser profile sensor, including: The line laser profile sensor is controlled to project structured light strips onto the splicing area of ​​the wing and the web to acquire raw point cloud data containing the surface topography of the wing and the web; The original point cloud data is subjected to denoising processing based on statistical filtering to remove outlier points caused by oxide scale on the surfaces of the wing plate and the web plate, thereby obtaining denoised point cloud data. The first plane equation of the wing plate and the second plane equation of the web plate are extracted using the random sampling consensus algorithm on the denoised point cloud data. The spatial angle between the first plane equation and the second plane equation is calculated as the initial assembly angle, and the distance difference between the first plane equation and the second plane equation in the direction perpendicular to the feed direction is calculated as the initial value of the misalignment amount.

[0007] Preferably, the high-frequency current phase difference and impedance modulus of the welding zone are acquired in real time through the high-frequency power supply voltage and current acquisition module, including: A high-frequency voltage transformer is connected in parallel at the output terminal of the high-frequency power supply, and a high-frequency current transformer is connected in series in the circuit from which the high-frequency power supply flows to the welding electrode. The instantaneous voltage signal and instantaneous current signal of the welding area are acquired synchronously by the high-frequency voltage transformer and the high-frequency current transformer. The instantaneous voltage signal and the instantaneous current signal are subjected to quadrature phase-locked amplification processing to eliminate the coupling interference of the background electromagnetic noise in the welding area on the amplitude and phase of the instantaneous voltage signal and the instantaneous current signal; Based on the instantaneous voltage signal and the instantaneous current signal after orthogonal lock-in amplification, the high-frequency current phase difference and the impedance magnitude of the welding area are calculated.

[0008] Preferably, a mapping model is established between the initial assembly angle, the high-frequency current phase difference, the impedance magnitude, and the actual convergence state of the V-angle, including: The initial assembly angle is converted into a theoretical arrival timestamp based on the fixed physical distance between the line laser profile sensor and the welding point. The high-frequency current phase difference and the impedance magnitude are marked as the measured arrival timestamps according to their respective data acquisition timestamps. Using the theoretical arrival timestamp and the measured arrival timestamp as indexes, interpolation and alignment operations are performed on the time dimension for the initial assembly angle, the high-frequency current phase difference, and the impedance magnitude. The initial assembly angle, the high-frequency current phase difference, and the impedance modulus after time-dimensional interpolation and alignment are input into a long short-term memory network for feature-level fusion training to generate the mapping relationship model.

[0009] Preferably, the output of the V-angle dynamic deviation value through the mapping relationship model includes: The mapping relationship model contains multiple sets of parallel convolution kernels, and each set of convolution kernels corresponds to a specific V-shaped angle convergence failure mode. The mapping relationship model inputs the initial assembly angle, the high-frequency current phase difference, and the impedance modulus after time-dimension interpolation and alignment operations into the multiple sets of parallel convolution kernels; The multiple sets of parallel convolutional kernels respectively output local deviation feature vectors for the V-shaped angle convergence failure mode; The local deviation feature vector is weighted and spliced ​​and nonlinearly activated to filter out the nonlinear error components caused by mechanical transmission clearance in the local deviation feature vector, and the V-angle dynamic deviation value is output as the final result.

[0010] Preferably, the servo motor displacement compensation amount of the extrusion roller is calculated based on the dynamic deviation value of the V-angle, and the position of the extrusion roller is adjusted in real time according to the servo motor displacement compensation amount before the welding point, including: Obtain the current real-time feed speed of the H-beam, and divide the dynamic deviation value of the V-angle by the real-time feed speed to obtain the theoretical adjustment time difference of the extrusion roller; The theoretical adjustment time difference is advanced to the physical position of the weld point and converted into a feedforward displacement; The servo motor displacement compensation amount is obtained by superimposing the feedforward displacement amount with the preset extrusion roller reference position. A sliding mode control strategy with dead zone compensation is adopted. A drive signal is generated based on the displacement compensation of the servo motor and sent to the servo motor rigidly connected to the extrusion roller to eliminate the static friction dead zone of the extrusion roller during the reversing process.

[0011] Preferably, before performing statistical filtering-based denoising processing on the original point cloud data, the method further includes performing multi-frame joint processing on the original point cloud data: During the feeding process of the H-beam, multiple frames of the original point cloud data are continuously acquired at a fixed sampling frequency; The continuous multiple frames of the original point cloud data are projected into the same three-dimensional spatial coordinate system according to the feed speed of the H-beam to form a fused point cloud set; Calculate the overlapping region of two adjacent frames of the original point cloud data in the fused point cloud set, and perform point cloud registration operation in the overlapping region to eliminate the spatial coordinate drift error of a single frame of the original point cloud data caused by the vibration of the line laser profile sensor itself during installation. The fused point cloud set after point cloud registration is subjected to voxel downsampling to obtain the original point cloud data with uniform spatial distribution density.

[0012] Preferably, after performing quadrature phase-locked amplification on the instantaneous voltage signal and the instantaneous current signal, the method further includes an outlier removal mechanism for the high-frequency current phase difference and the impedance magnitude: A sliding time window is set, and the high-frequency current phase difference and the impedance magnitude are continuously extracted from multiple time sampling points within the sliding time window; Calculate the first-order difference between the phase difference of the high-frequency current and the first-order difference between the impedance magnitude within the sliding time window; When the first-order difference of the high-frequency current phase difference or the first-order difference of the impedance magnitude exceeds the preset physical change extreme value, the high-frequency current phase difference or the impedance magnitude at the corresponding time sampling point is determined to be abnormal data caused by the transient breakdown of the internal switching transistor of the high-frequency power supply. The abnormal data is removed from the sliding time window and supplemented by linear interpolation using data from adjacent normal time sampling points.

[0013] Preferably, after generating the mapping relationship model, the method further includes performing an online self-learning update mechanism on the mapping relationship model: An infrared thermal imager is deployed behind the weld joint to collect data on the actual temperature field distribution of the weld at the weld joint. The offset of the highest temperature point in the actual temperature field distribution data of the weld is used as the online correction error index of the mapping relationship model; Using the online correction error index as a supervision signal, calculate the gradient descent value of each hidden layer node of the Long Short-Term Memory network in the mapping relationship model; The connection weight matrix inside the long short-term memory network is iteratively updated based on the gradient descent value to compensate for the prediction deviation of the mapping relationship model caused by the V-angle convergence state drift due to batch changes in the H-beam material.

[0014] Preferably, after sending the drive signal to the servo motor rigidly connected to the extrusion roller, the system further includes coordinated control of feedforward compensation and extrusion roller torque monitoring: A torque sensor is installed at the output shaft end of the servo motor, and the actual extrusion torque applied by the extrusion roller to the wing plate and the web plate is collected in real time through the torque sensor; The actual extrusion torque is compared with the preset reference extrusion torque to obtain the torque deviation value; When the torque deviation value is greater than the allowable fluctuation threshold, it is determined that relative sliding has occurred between the extrusion roller and the wing plate or the web plate, triggering the torque limiting protection. When the torque limiting protection is triggered, the displacement compensation of the servo motor at the current moment is frozen, and a power reduction command is sent to the high-frequency power supply voltage and current acquisition module to forcibly reduce the actual welding power corresponding to the high-frequency current phase difference in the welding area.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses the initial assembly angle extracted by a line laser contour sensor as a feedforward feature, and the high-frequency current phase difference and impedance modulus obtained by a high-frequency power supply voltage and current acquisition module as feedback features. After interpolating and aligning these features in the time dimension based on theoretical and measured arrival timestamps, the data is input into a mapping model, achieving deep coupling between surface geometric data and internal electromagnetic physical characteristics. This closed-loop control logic, which integrates feedforward and feedback, enables the mapping model to output a dynamic V-angle deviation value reflecting the true physical convergence state. This, in turn, generates a drive signal to adjust the extrusion roller position through a sliding mode control strategy with dead-zone compensation, solving the problem of V-angle convergence point offset caused by the inability to reflect changes in internal electromagnetic distribution when relying solely on surface geometric detection.

[0016] 2. At the signal acquisition level, this invention eliminates background electromagnetic noise interference through orthogonal phase-locked loop amplification and uses first-order difference values ​​within the sliding time window to eliminate abnormal data caused by transient breakdown of the switching transistor, ensuring the purity of the input feature data. At the model calculation level, multiple sets of parallel convolutional kernels are used to extract local deviation feature vectors. Nonlinear activation calculations are used to filter out nonlinear error components caused by mechanical transmission gaps. Simultaneously, temperature field distribution data behind the weld point is used as a monitoring signal to iteratively update the model weight matrix, compensating for prediction deviations caused by batch variations in material. At the execution level, a torque sensor monitors the extrusion torque in real time. When relative slippage is detected, the displacement compensation is frozen, and a power reduction command is sent to the high-frequency power supply, preventing mechanical overload and abnormal current. Attached Figure Description

[0017] Figure 1 A flowchart illustrating an intelligent positioning method for welding H-shaped high-frequency steel according to an embodiment of the present invention; Figure 2 A flowchart of feedforward feature extraction and multi-frame joint processing provided in an embodiment of the present invention; Figure 3 A flowchart illustrating the feedback feature acquisition and abnormal data removal mechanism provided in this embodiment of the invention; Figure 4 A flowchart illustrating the mapping relationship model construction and online self-learning update mechanism provided in this embodiment of the invention; Figure 5 A flowchart for calculating the dynamic deviation value of the V-shaped angle and filtering out nonlinear errors provided in this embodiment of the invention; Figure 6 The flowchart of servo motor displacement compensation and torque coordinated protection control provided in the embodiment of the present invention is shown. Detailed Implementation

[0018] This embodiment describes in detail the intelligent positioning process for welding H-shaped high-frequency steel. The described embodiments are feasible technical solutions. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1This embodiment provides an intelligent positioning method for welding H-beam high-frequency steel. The intelligent positioning process for welding H-beam high-frequency steel relies on a detection and execution mechanism deployed on the high-frequency welding feed path of the H-beam. The feed path is a linear path that continuously moves towards the high-frequency welding point after the H-beam flange and web are pre-assembled. The extension direction of the feed path is consistent with the length direction of the H-beam. Along the feed path, a line laser profile sensor, a high-frequency welding electrode, and an extrusion roller group are sequentially deployed along the feed direction of the H-beam. The detection window of the line laser profile sensor covers the pre-assembled area of ​​the flange and web. The working end of the high-frequency welding electrode is attached to both sides of the V-shaped opening formed by the flange and web. The roller surface of the extrusion roller group is in rigid contact with the outer surface of the flange and the end face of the web, respectively. The adjustment end of the extrusion roller group is rigidly connected to the output shaft of the servo motor. The displacement output direction of the servo motor is perpendicular to the feed direction of the H-beam. The signal acquisition end of the high-frequency power supply voltage and current acquisition module is electrically connected to the output circuit of the high-frequency power supply and the power supply circuit of the welding electrode. The output end of the high-frequency power supply is electrically connected to the power supply end of the welding electrode to provide high-frequency welding current to the welding area.

[0020] Furthermore, during the continuous feeding of the H-beam along the feed path, the line laser profile sensor continuously detects the pre-assembly area of ​​the flange and web, extracting the initial assembly angle and initial misalignment values ​​between the flange and web. The detection reference coordinate system of the line laser profile sensor is pre-calibrated with the global coordinate system of the feed path, and the detected geometric feature data can be directly mapped to the global coordinate system, eliminating systematic errors caused by installation posture deviations.

[0021] Specifically, the high-frequency power supply voltage and current acquisition module synchronously samples the output circuit of the high-frequency power supply and the power supply circuit of the welding electrode during the welding process, acquiring the high-frequency current phase difference and impedance magnitude in the welding area in real time. The sampling timing of the high-frequency power supply voltage and current acquisition module is synchronized with the switching timing of the high-frequency power supply, ensuring that the acquired electrical parameter data has a corresponding timing relationship with the actual electromagnetic state of the welding area.

[0022] Furthermore, a mapping model is established between the initial assembly angle, high-frequency current phase difference, impedance modulus, and the actual convergence state of the V-angle. The V-angle is the opening structure formed by the flange and web in front of the welding point. The actual convergence state of the V-angle is the deviation between the actual contact gap and contact position of the flange and web at the welding point and the preset reference state. The convergence state of the V-angle directly determines the weld formation quality and welding strength of high-frequency welding. The mapping model is used to establish a quantitative correspondence between the input multi-dimensional features and the actual convergence state of the V-angle, realizing accurate characterization of the convergence state inside the V-angle that cannot be directly observed based on detectable features.

[0023] Specifically, the initial assembly angle is used as the input feature of the mapping model, while the high-frequency current phase difference and impedance modulus are used as the input features of the feedback feature. The feedforward feature is the geometric feature of the pre-assembly area in front of the welding point, which can characterize the initial pose deviation of the H-beam before it enters the welding zone, providing advance for the adjustment of the extrusion roller. The feedback feature is the electromagnetic physical feature inside the welding zone, which can characterize the current convergence state and contact condition inside the V-angle in real time, reflecting the actual convergence state at the current moment. The mapping model fuses the input feedforward and feedback features to output the dynamic deviation value of the V-angle. The dynamic deviation value of the V-angle is the quantified deviation between the actual convergence state of the V-angle and the preset benchmark convergence state, which can be directly used to drive the position adjustment of the extrusion roller.

[0024] Furthermore, the servo motor displacement compensation amount of the extrusion roller is calculated based on the dynamic deviation value of the V-angle. The servo motor displacement compensation amount is the linear displacement value that the servo motor needs to output, which is used to adjust the spatial position of the extrusion roller, thereby changing the relative posture of the flange and web at the V-shaped opening and correcting the convergence state of the V-angle. Before the welding point, the position of the extrusion roller is adjusted in real time based on the servo motor displacement compensation amount. The timing of the adjustment action is earlier than the time when the corresponding H-beam segment reaches the welding point, ensuring that the corrected posture of the flange and web is in the preset reference convergence state when it reaches the welding point.

[0025] In this embodiment, to clarify the spatial deployment relationship and timing matching constraints of each detection and execution unit, the following table is provided: Table 1. Spatial Deployment Relationship of Detection and Execution Units on the Feed Path for High-Frequency Welding of H-Beams ; Regarding the table above, the specific explanations are as follows: The relative spatial positional relationships defined in the table clarify the upstream and downstream sequence of each unit along the feed direction, ensuring that the acquisition time of the feedforward feature is earlier than the time when the corresponding workpiece segment enters the welding zone, providing sufficient adjustment time for the feedforward control; the timing matching constraint clarifies the timing correspondence between data acquisition and action execution of each unit, eliminating the feature timing misalignment caused by feed speed fluctuations, and ensuring the time dimension matching of the feedforward feature and the feedback feature; the core function column clarifies the role of each unit in the intelligent positioning process, forming a complete closed-loop control link of feature extraction - state representation - deviation calculation - position adjustment.

[0026] In this embodiment, by using the geometric features of the pre-assembly area as feedforward input and the electromagnetic physical features of the welding area as feedback input, deep coupling between the external geometric pose of the workpiece and its internal electromagnetic convergence state is achieved, forming a complete closed-loop control link. Through the fusion processing of multi-dimensional features using a mapping model, accurate characterization of the convergence state inside the V-angle, which cannot be directly observed, is achieved. The output dynamic deviation value of the V-angle can be directly used to drive the position adjustment of the extrusion roller, ensuring that the flange and web are in a preset reference convergence state at the welding point, thus avoiding the problems of V-angle convergence point offset and high-frequency current shunting along the plate surface.

[0027] refer to Figure 2 In a preferred embodiment, the extraction process for the initial assembly angle and initial misalignment of the flange and web is refined. During the continuous feeding of the H-beam along the feed path, a control line laser profile sensor projects structured light strips onto the assembly area of ​​the flange and web. These structured light strips create diffuse reflection on the surfaces of the flange and web. The photosensitive unit of the line laser profile sensor receives the reflected light signal and converts it into raw point cloud data containing the surface topography of the flange and web. The raw point cloud data consists of multiple three-dimensional coordinate points, each corresponding to a surface sampling point within the assembly area. The coordinate values ​​are defined based on the detection reference coordinate system of the line laser profile sensor.

[0028] Furthermore, before performing statistical filtering-based denoising on the original point cloud data, multi-frame joint processing is performed on the original point cloud data. During the feeding process of the H-beam, multiple frames of original point cloud data are continuously acquired at a fixed sampling frequency. Each frame of original point cloud data corresponds to the surface morphology of a continuous cross-section within the splicing area. The continuous multiple frames of original point cloud data are projected onto the same three-dimensional spatial coordinate system according to the feeding speed of the H-beam, forming a fused point cloud set. The three-dimensional spatial coordinate system is the global coordinate system of the feeding path. The origin of the coordinate system is located at the projection point of the welding point on the feeding path. The positive direction of the X-axis is opposite to the feeding direction of the H-beam, the Y-axis is the horizontal direction perpendicular to the feeding direction, and the Z-axis is the vertical direction. The projection position of each frame of original point cloud data is determined according to the feeding displacement of the H-beam corresponding to the acquisition time of that frame of data. The feeding displacement is the distance the H-beam moves relative to the reference position at the acquisition time, which is obtained in real time by the encoder of the feeding mechanism.

[0029] Specifically, the overlapping region of two adjacent frames of original point cloud data in the fused point cloud set is calculated. The overlapping region is the intersection of the spatial regions covered by the two adjacent frames of point cloud data in the global coordinate system, corresponding to the same physical region on the surface of the H-beam. Point cloud registration is performed within the overlapping region. The iterative nearest-point algorithm is used to register the two adjacent frames of point cloud data, solving for the spatial transformation matrix between the two frames. The coordinates of the subsequent frame of point cloud data are corrected based on the spatial transformation matrix to eliminate spatial coordinate drift errors caused by the vibration of the line laser profile sensor itself in the single frame of original point cloud data. Voxel downsampling is then performed on the fused point cloud set after point cloud registration. The three-dimensional space in the global coordinate system is divided into a fixed-size voxel grid. Only the sampling point closest to the voxel center is retained in each voxel grid, resulting in original point cloud data with a uniform spatial distribution density. The size of the voxel grid is determined based on the surface roughness of the H-beam and the geometric dimensions of the spliced ​​area. While preserving key surface geometric features, this reduces the redundancy of the point cloud data and improves the computational efficiency of subsequent processing.

[0030] Furthermore, statistical filtering-based denoising is performed on the original point cloud data after multi-frame joint processing to remove outliers caused by oxide scale on the surfaces of the wingplate and web, resulting in denoised point cloud data. The statistical filtering process is as follows: for each sampling point in the original point cloud data, the average distance from that sampling point to a preset number of its nearest neighbor sampling points is calculated. The average distance of all sampling points follows a Gaussian distribution. The mean and standard deviation of this Gaussian distribution are calculated, and a distance threshold is determined based on a preset standard deviation coefficient. Outliers whose average distance exceeds the distance threshold are removed, and the remaining sampling points form the denoised point cloud data. Outliers are caused by defects such as oxide scale, rust, and scratches on the surfaces of the wingplate and web. Their spatial coordinates deviate significantly from the actual surface morphology of the workpiece, interfering with the subsequent plane fitting process. Statistical filtering effectively removes these outliers, ensuring the authenticity of the point cloud data.

[0031] The core calculation process of statistical filtering is quantitatively described by the following formula:

[0032]

[0033] in, Let N be the average distance from the i-th sampling point to its k nearest neighbor sampling points, and N be the total number of sampling points in the point cloud data. The mean of the average distances between all sampling points. This represents the standard deviation of the average distance across all sampling points. The distance threshold is set to... , The standard deviation coefficient is a preset value. Sampling points whose average distance exceeds this threshold are identified as outliers and removed.

[0034] Specifically, the random sample consensus algorithm is used to extract the first plane equation of the airfoil and the second plane equation of the web from the denoised point cloud data. The processing procedure of the random sample consensus algorithm is as follows: In the denoised point cloud data, based on the pre-assembled pose of the airfoil and web, the point cloud data is divided into a first point cloud subset corresponding to the airfoil and a second point cloud subset corresponding to the web; three non-collinear sampling points are randomly selected from the first point cloud subset, the initial plane equation is solved, and the number of interior points in the first point cloud subset that satisfy the initial plane equation is counted. Interior points are sampling points whose distance to the initial plane is less than a preset distance threshold; through multiple iterations, the plane equation with the most interior points is selected as the first plane equation corresponding to the airfoil; the same processing procedure is used to solve for the second plane equation corresponding to the web in the second point cloud subset.

[0035] The equations for the first and second planes use the standard expressions for three-dimensional planes, as follows:

[0036] Where A, B, and C are the three components of the plane normal vector, D is the constant term of the plane, and the unit normal vector of the plane. Satisfying the constraints .

[0037] Furthermore, the spatial angle between the equations of the first and second planes is calculated as the initial assembly angle. The spatial angle between the two planes is the angle between the normal vectors of the two planes or their supplementary angle. An acute angle or a right angle is taken as the final initial assembly angle. The initial assembly angle characterizes the relative inclination of the flange and web during the pre-assembly stage and directly determines the initial opening size of the V-angle. The calculation process of the initial assembly angle is quantitatively described by the following formula:

[0038] in, The initial assembly angle, Let be the unit normal vector component of the equation of the first plane. Let be the unit normal vector component of the equation of the second plane.

[0039] The distance difference between the equations of the first and second planes perpendicular to the feed direction is calculated as the initial value of the misalignment. This distance difference perpendicular to the feed direction represents the relative positional deviation of the flange and web in the Y-axis direction. The initial misalignment value characterizes the horizontal alignment deviation of the flange and web during the pre-assembly stage. The calculation process of the initial misalignment value is quantified by the following formula:

[0040] in, This is the initial value for the offset amount. The coordinates of the reference point within the splicing area are given, and the reference point is located on the center line of the end face of the web.

[0041] In this embodiment, to clarify the algorithm parameter definitions and constraints for each stage of point cloud data processing, the following table is provided: Table 2. Key Algorithm Parameter Configuration Table for Point Cloud Data Processing ; Regarding the table above, the specific explanations are as follows: Each algorithm parameter defined in the table is a core adjustable parameter in the point cloud data processing process. The constraints clearly define the value range of each parameter, avoiding a decrease in processing accuracy or computational efficiency due to improper parameter values. The adaptation logic clarifies the adjustment basis for each parameter, allowing dynamic adaptation based on different specifications and surface conditions of H-beam welding scenarios, ensuring the robustness and accuracy of the point cloud data processing process. Through the reasonable configuration of the above parameters, interference from factors such as sensor vibration, surface oxide scale, and point cloud redundancy on geometric feature extraction can be effectively eliminated, ensuring the extraction accuracy of the initial assembly angle and misalignment amount.

[0042] In this embodiment, the spatial coordinate drift error caused by the installation vibration of the line laser contour sensor is eliminated through multi-frame joint processing and point cloud registration. Voxel downsampling is used to achieve uniform processing of point cloud data and reduce data redundancy. Statistical filtering is used to remove outliers caused by surface oxide scale, ensuring the authenticity of point cloud data. The random sampling consensus algorithm is used to achieve accurate fitting of the corresponding planes of the wing plate and the web plate, thereby achieving high-precision extraction of the initial assembly angle and the initial value of the misalignment, providing reliable input features for subsequent feedforward control and avoiding the deviation in the V-angle convergence state representation caused by the pre-assembly geometric feature extraction error.

[0043] refer to Figure 3 In another preferred embodiment, the acquisition, processing, mapping model establishment, and online updating of high-frequency electrical parameters in the welding zone are refined. A high-frequency voltage transformer is connected in parallel at the output of the high-frequency power supply, and a high-frequency current transformer is connected in series in the circuit from the high-frequency power supply to the welding electrodes. The high-frequency voltage transformer converts the instantaneous high-voltage signal in the welding zone into a low-voltage signal suitable for the range of the signal acquisition unit, and the high-frequency current transformer converts the instantaneous large current signal in the welding circuit into a small current signal suitable for the range of the signal acquisition unit. The operating frequency bands of the high-frequency voltage transformer and the high-frequency current transformer are consistent with the output frequency band of the high-frequency power supply to avoid distortion of the amplitude and phase of the acquired signal.

[0044] Furthermore, instantaneous voltage and current signals of the welding area are synchronously acquired using high-frequency voltage and current transformers. The sampling trigger signal of the signal acquisition unit is synchronized with the switching drive signal of the high-frequency power supply, ensuring that each sampling moment corresponds to a fixed phase point within the switching cycle of the high-frequency power supply, thus eliminating phase measurement errors caused by asynchronous sampling timing. The acquired instantaneous voltage and current signals are discrete digital signals, with each sampling point corresponding to a fixed acquisition timestamp. The timing reference of the acquisition timestamp is consistent with that of the line laser profile sensor, ensuring that feature data from different sources have a unified timing reference.

[0045] The time-domain expressions for the acquired instantaneous voltage and instantaneous current signals are as follows:

[0046]

[0047] in, The amplitude of the voltage signal. The amplitude of the current signal. The angular frequency of the high-frequency power supply. This represents the initial phase of the voltage signal. This represents the initial phase of the current signal. The background electromagnetic noise component of the voltage signal. This represents the background electromagnetic noise component of the current signal.

[0048] Specifically, the instantaneous voltage and current signals are subjected to quadrature phase-locked amplification to eliminate the coupling interference of background electromagnetic noise in the welding area on the amplitude and phase of the instantaneous voltage and current signals. Background electromagnetic noise in the welding area is caused by factors such as the switching action of the high-frequency power supply, electromagnetic radiation from surrounding electrical equipment, and electromagnetic coupling in the welding circuit. This noise covers a wide frequency range and interferes with the amplitude and phase of the acquired voltage and current signals, leading to deviations in the subsequently calculated phase difference and impedance magnitude. Quadrature phase-locked amplification multiplies the acquired signal with a quadrature reference signal of the same frequency, then performs low-pass filtering to extract the useful signal component with the same frequency as the reference signal, while filtering out noise components of different frequencies, thus achieving accurate extraction of weak useful signals against a strong noise background.

[0049] The core calculation process of quadrature lock-in amplification is as follows: An orthogonal reference signal with the same frequency as the high-frequency power supply output signal is set, in the form of:

[0050]

[0051] Multiplying the instantaneous voltage signal by the two orthogonal reference signals respectively, we get:

[0052]

[0053] The multiplied signals are then subjected to low-pass filtering to remove the second harmonic component and noise components, yielding the in-phase component of the voltage signal. Orthogonal components :

[0054]

[0055] Using the same processing procedure, quadrature lock-in amplification is performed on the instantaneous current signal to obtain the in-phase component of the current signal. Orthogonal components :

[0056]

[0057] Furthermore, based on the instantaneous voltage and current signals after orthogonal lock-in amplification, the high-frequency current phase difference and impedance magnitude of the welding zone are calculated. The high-frequency current phase difference, the phase difference between the instantaneous voltage and current signals in the welding zone, reflects the reactive characteristics of the load in the welding circuit. Changes in the contact gap and current convergence state within the V-angle directly cause changes in the phase difference. The impedance magnitude is the amplitude of the circuit impedance in the welding zone, reflecting the resistive characteristics of the load in the welding circuit. Changes in the contact area and heating temperature within the V-angle directly cause changes in the impedance magnitude. The high-frequency current phase difference and impedance magnitude directly characterize the electromagnetic convergence state and metal contact condition within the V-angle, and are core physical characteristics reflecting the actual convergence state of the V-angle.

[0058] The calculation process of high-frequency current phase difference is quantitatively described by the following formula:

[0059] in, This represents the phase difference of the high-frequency current.

[0060] The calculation process of impedance magnitude is quantitatively described by the following formula:

[0061] in, This represents the impedance modulus of the circuit in the welding zone.

[0062] Specifically, an outlier removal mechanism is implemented for the high-frequency current phase difference and impedance magnitude obtained after quadrature lock-in amplification. A sliding time window is set, which is a time-series data interval of fixed length. Multiple time sampling points within the sliding time window continuously extract the high-frequency current phase difference and impedance magnitude. The first-order difference of the high-frequency current phase difference and the first-order difference of the impedance magnitude are calculated within the sliding time window. The first-order difference is the ratio of the parameter difference between two adjacent sampling points to the sampling time interval, characterizing the instantaneous rate of change of the parameters. During normal welding, the electromagnetic state of the welding zone changes continuously, and the rate of change of the high-frequency current phase difference and impedance magnitude has a physical upper limit. When the first-order difference of the high-frequency current phase difference or the first-order difference of the impedance magnitude exceeds the preset physical extreme value, the high-frequency current phase difference or impedance magnitude at the corresponding time sampling point is determined to be abnormal data caused by transient breakdown of the internal switching transistor of the high-frequency power supply. The abnormal data is removed from the sliding time window, and linear interpolation is performed using data from adjacent normal time sampling points to ensure the continuity and integrity of the time-series data.

[0063] Furthermore, a mapping model is established between the initial assembly angle, high-frequency current phase difference, impedance modulus, and the actual convergence state of the V-angle. The initial assembly angle is converted into a theoretical arrival timestamp based on the fixed physical distance between the line laser profile sensor and the welding point. The theoretical arrival timestamp is the moment when the H-beam segment corresponding to the initial assembly angle arrives at the welding point, calculated by dividing the fixed physical distance between the line laser profile sensor and the welding point by the real-time feed speed of the H-beam, and adding the acquisition timestamp of the initial assembly angle. The high-frequency current phase difference and impedance modulus are marked as measured arrival timestamps based on their respective data acquisition timestamps. The measured arrival timestamps are the acquisition times of the corresponding electrical parameter data and the times when the H-beam segment is at the welding point.

[0064] refer to Figure 4Specifically, using the theoretical and measured arrival timestamps as indices, interpolation alignment operations are performed on the initial assembly angle, high-frequency current phase difference, and impedance modulus in the time dimension. Since there is a fixed physical distance between the acquisition position of the initial assembly angle and the welding point, and a time difference exists between the acquisition time of the initial assembly angle and the arrival time of the corresponding H-beam segment at the welding point, while the acquisition times of the high-frequency current phase difference and impedance modulus correspond in real-time to the actual state of the welding point, interpolation alignment is necessary to align the feature data acquired at different times to the time axis corresponding to the same H-beam segment. This ensures that the multi-dimensional feature data of the input mapping model corresponds to different states of the same H-beam segment. The interpolation alignment operation uses a linear interpolation method. For each theoretical arrival timestamp, two adjacent sampling times are found in the sequence of measured arrival timestamps. Based on the feature values ​​and time intervals corresponding to the two sampling times, the interpolated feature value corresponding to the theoretical arrival timestamp is calculated, achieving time dimension alignment between feedforward and feedback features.

[0065] The calculation process of linear interpolation is quantitatively described by the following formula:

[0066] in, These are the interpolated feature values ​​corresponding to the theoretical arrival timestamp. For the theoretical arrival timestamp, , For two adjacent sampling times in the measured arrival timestamp sequence, and satisfying the following conditions: , for The measured feature value corresponding to time . for The measured feature value corresponding to the time.

[0067] Furthermore, the initial assembly angle, high-frequency current phase difference, and impedance magnitude, after time-dimensional interpolation and alignment, are input into the Long Short-Term Memory (LSTM) network for feature-level fusion training to generate a mapping model. The LSM network, a type of temporal recurrent neural network, effectively processes long-sequence temporal data, captures temporal dependencies between feature data, and avoids the gradient vanishing problem of traditional recurrent neural networks. The input layer of the LSM network is an interpolated and aligned three-dimensional feature vector containing feature data in three dimensions: initial assembly angle, high-frequency current phase difference, and impedance magnitude. The hidden layer consists of multiple cascaded LSM units, each containing a forget gate, input gate, output gate, and cell state, used to extract and memorize the input temporal features. The output layer is a fully connected layer, outputting a quantized representation of the actual convergence state of the V-shaped angle. The training process adopts a supervised learning approach. The training dataset consists of feature data collected under different welding conditions and corresponding V-angle actual convergence state label data. The label data is obtained through post-weld detection and online monitoring. The weight matrix and bias terms inside the network are optimized through iterative training to minimize the error between the predicted value and the label value, and the final mapping relationship model is generated.

[0068] The core gating calculation process of the Long Short-Term Memory (LSTM) unit is quantitatively described by the following formula:

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] in, Let be the input feature vector at time t. Let be the hidden layer output vector at time t-1. Let be the cell state vector at time t-1. For the output of the forget gate, The output of the input gate, Candidate cell state, This represents the updated cell state. For the output of the output gate, Let be the hidden layer output vector at time t; , , , Here are the weight matrices for each gate. , , , For each gate, there is a corresponding bias term; It is the sigmoid activation function. It is the hyperbolic tangent activation function.

[0075] Specifically, after generating the mapping relationship model, an online self-learning update mechanism is implemented on the model. An infrared thermal imager is deployed behind the weld joint, with its detection window covering the weld area at the joint. The imager collects data on the actual temperature field distribution of the weld at the joint. The temperature field distribution of the weld is directly determined by the convergence state of the high-frequency current. When the convergence point of the V-angle shifts, the highest temperature point of the weld also shifts. Therefore, the shift of the highest temperature point in the actual temperature field distribution data of the weld can directly characterize the deviation of the actual convergence state of the V-angle and can serve as an online correction error index for the mapping relationship model.

[0076] Furthermore, the offset of the highest temperature point is used as the online correction error index for the mapping relationship model. Using this index as a monitoring signal, the gradient descent values ​​of each hidden layer node in the Long Short-Term Memory (LSTM) network within the mapping relationship model are calculated. An error backpropagation algorithm is employed to propagate the prediction error of the output layer back to each hidden layer, calculating the gradient of the loss function with respect to the weight matrices of each layer. Based on the gradient descent values, the connection weight matrices within the LSM network are iteratively updated. The learning rate for this iteration is set to a small, fixed value to avoid overfitting and to slowly compensate for the prediction bias caused by V-angle convergence state drift due to batch variations in H-beam material and environmental temperature changes, ensuring the model's prediction accuracy during long-term continuous welding processes.

[0077] The loss function and weight update formula for the online self-learning process are as follows:

[0078]

[0079] in, Let the mean squared error loss function be . This represents the number of samples used in the batch training. This represents the label value of the m-th sample, i.e., the offset of the highest temperature point in the weld. These are the model's predicted values; This is the weight matrix of the network. For learning rate, The loss function is applied to the weight matrix. The gradient value.

[0080] In this embodiment, the following table is provided to clarify the network structure configuration and training hyperparameter definition of the mapping relationship model: Table 3. Network Structure and Training Hyperparameter Configuration of Mapping Relationship Model ; Regarding the table above, the specific explanations are as follows: The network structure parameters defined in the table clarify the hierarchical structure and neuron configuration of the Long Short-Term Memory network. The constraints ensure a balance between the network's feature extraction capability and inference real-time performance, avoiding excessive inference latency due to an overly complex network structure or insufficient feature extraction capability due to an overly simple network structure. The configuration of training hyperparameters and online update hyperparameters ensures the convergence speed and accuracy of the model during offline training, while enabling slow weight iteration during online updates to avoid overfitting or prediction oscillations, thus ensuring the model's adaptability and long-term stability under different welding conditions.

[0081] In this embodiment, synchronous sampling and orthogonal phase-locked amplification eliminate the interference of background electromagnetic noise in the welding zone on voltage and current signals, achieving high-precision extraction of high-frequency current phase difference and impedance modulus. An outlier removal mechanism eliminates abnormal data caused by transient breakdown of high-frequency power switching transistors, ensuring the stability and continuity of feedback feature data. Interpolation alignment in the time dimension achieves temporal matching between feedforward and feedback features, ensuring that the feature data input to the model corresponds to different states of the same H-beam segment. A long short-term memory network enables the fusion and mapping of multi-dimensional temporal features, accurately capturing the temporal dependencies between feature data. An online self-learning update mechanism uses weld temperature field distribution data as a monitoring signal to iteratively update model weights, effectively compensating for model prediction biases caused by material batch variations and ensuring the model's prediction accuracy during long-term continuous welding.

[0082] refer to Figure 5 In another preferred embodiment, the extraction of the dynamic deviation value of the V-angle, the servo control of the extrusion roller position, and the collaborative protection process are refined. The mapping relationship model contains multiple sets of parallel convolutional kernels, each set of kernels applying a specific V-angle convergence failure mode. V-angle convergence failure modes include various typical failure modes such as V-angle opening asymmetry caused by unilateral tilting of the flange, V-angle vertex misalignment caused by web offset, current shunting caused by excessive contact gap, and convergence point offset caused by uneven edge heating. The structure and parameters of each set of convolutional kernels are optimized for the characteristics of the corresponding failure mode, which can effectively extract the local deviation characteristics of the corresponding failure mode.

[0083] Furthermore, the mapping model inputs the initial assembly angle, high-frequency current phase difference, and impedance magnitude after time-dimension interpolation and alignment into multiple sets of parallel convolutional kernels. The input multi-dimensional feature data is first transformed into feature maps adapted to the input dimensions of the convolutional kernels, and then input into each set of parallel convolutional kernels. The multiple sets of parallel convolutional kernels perform convolution operations on the input feature maps, extracting local deviation features corresponding to the V-angle convergence failure mode, and outputting local deviation feature vectors for the V-angle convergence failure mode. The local deviation feature vector output by each set of convolutional kernels characterizes the degree and type of deviation in the V-angle convergence state under the corresponding failure mode.

[0084] Specifically, the local deviation feature vectors output by multiple sets of convolutional kernels are weighted and concatenated, and nonlinear activation is calculated. The weighted concatenation process is as follows: a corresponding weight coefficient is assigned to the local deviation feature vector output by each set of convolutional kernels. The weight coefficient is determined based on the degree of influence of the corresponding failure mode on the weld quality. The sum of the weight coefficients corresponding to all convolutional kernels is 1. The weighted local deviation feature vectors are then concatenated to form a fused global deviation feature vector.

[0085] The calculation process for weighted splicing is quantified and described by the following formula:

[0086] in, This is the fused global bias feature vector. The number of parallel convolution kernels. This is the local bias feature vector output by the k-th convolutional kernel. The weight coefficients corresponding to the k-th convolutional kernel satisfy the constraints. .

[0087] Nonlinear activation calculation is performed on the global deviation feature vector. A nonlinear activation function is used to transform the global deviation feature vector, filtering out nonlinear error components caused by mechanical transmission backlash. Mechanical transmission backlash exists in the servo motor's transmission mechanism and the extrusion roller's connection mechanism, causing nonlinear errors in the extrusion roller's position adjustment. These errors form fixed nonlinear components in the feature data. Nonlinear activation calculation effectively filters out these error components, improving the accuracy of the deviation value output. The feature vector after nonlinear activation calculation is then linearly transformed by a fully connected layer, outputting the final result: the V-angle dynamic deviation value.

[0088] refer to Figure 6Furthermore, the servo motor displacement compensation amount of the extrusion roller is calculated based on the V-angle dynamic deviation value. Real-time adaptive adjustment of the extrusion roller position is then performed based on this servo motor displacement compensation amount before the welding point. The current real-time feed speed of the H-beam is acquired; this speed is collected in real-time by the encoder of the feed mechanism and represents the movement speed of the H-beam along the feed path. The theoretical adjustment time difference of the extrusion roller is obtained by dividing the V-angle dynamic deviation value by the real-time feed speed. This theoretical adjustment time difference is the time interval between the moment the extrusion roller position adjustment action is performed and the moment the corresponding H-beam segment reaches the welding point, ensuring that the adjustment action is completed before the corresponding H-beam segment reaches the welding point.

[0089] Specifically, the theoretical adjustment time difference is converted into a feedforward displacement value by adjusting the physical position of the welding point in advance. This feedforward displacement is a pre-adjustment displacement based on the dynamic deviation value of the V-angle, which can correct the V-angle convergence deviation caused by the pre-assembly posture deviation of the H-beam in advance, avoiding the lag in feedback control. The feedforward displacement is superimposed with the preset extrusion roller reference position to obtain the servo motor displacement compensation value. The extrusion roller reference position is the preset initial position corresponding to the H-beam specification, serving as the reference point for extrusion roller position adjustment. The servo motor displacement compensation value is the total linear displacement value that the servo motor needs to output, used to drive the extrusion roller from the reference position to the target position, correcting the convergence state of the V-angle.

[0090] Furthermore, a sliding mode control strategy with dead-zone compensation is adopted, generating drive signals based on the servo motor displacement compensation. The sliding mode control strategy is a nonlinear control strategy, exhibiting strong robustness to system parameter variations and external disturbances. It can effectively cope with external disturbances such as workpiece reaction force and mechanical resistance experienced by the extrusion rollers during welding, ensuring the position tracking accuracy of the servo motor. The sliding mode control strategy employs an exponential reaching law design, which effectively reduces chattering phenomena and improves the smoothness of the control process.

[0091] The exponential reaching law of sliding mode control is quantitatively described by the following formula:

[0092] in, For sliding surface functions, For the coefficients of the transition term of the reaching law, The coefficients of the linear term in the approach law, It is a symbolic function.

[0093] Dead zone compensation is used to eliminate the static friction dead zone of the servo motor during commutation. The static friction dead zone will cause the servo motor to experience position tracking lag during commutation. By superimposing the corresponding dead zone compensation amount into the drive signal, the influence of the static friction dead zone can be effectively offset, thereby improving the response speed and position accuracy of the servo motor during commutation.

[0094] The calculation process for dead zone compensation is quantified using the following formula:

[0095] in, This is the amount of compensation for the dead zone. This refers to the drive voltage threshold corresponding to the static friction dead zone of the servo motor. This refers to the rotational speed of the servo motor's output shaft.

[0096] The generated drive signal is sent to the servo motor that is rigidly connected to the extrusion roller. The servo motor outputs the corresponding linear displacement according to the drive signal, which drives the extrusion roller to complete the position adjustment.

[0097] Specifically, after the drive signal is sent to the servo motor rigidly connected to the extrusion roller, coordinated control of feedforward compensation and extrusion roller torque monitoring is performed. A torque sensor is installed at the output shaft end of the servo motor. The detection end of the torque sensor is rigidly connected to the output shaft of the servo motor and the adjustment end of the extrusion roller. The torque sensor collects the actual extrusion torque applied by the extrusion roller to the flange and web in real time. The actual extrusion torque directly reflects the contact state and clamping force between the extrusion roller and the flange and web. When relative sliding occurs between the extrusion roller and the flange or web, the actual extrusion torque will change abruptly, exceeding the normal fluctuation range.

[0098] Furthermore, the actual extrusion torque is compared with the preset reference extrusion torque to obtain the torque deviation value. The reference extrusion torque is the preset extrusion torque corresponding to the H-beam specification and welding conditions, ensuring a tight fit between the flange and web at the weld point while avoiding workpiece deformation caused by excessive extrusion pressure. When the torque deviation value exceeds the allowable fluctuation threshold, it is determined that relative slippage has occurred between the extrusion roller and the flange or web, triggering torque limiting protection. The allowable fluctuation threshold is the preset fluctuation range of the reference extrusion torque, used to distinguish between normal torque fluctuations and abnormal torque abrupt changes.

[0099] Specifically, when the torque limiting protection is triggered, the servo motor displacement compensation is frozen at the current moment, and the position adjustment of the extrusion roller is stopped. This prevents further position adjustments from exacerbating the relative slippage between the extrusion roller and the workpiece, thus preventing defects such as scratches and deformation on the workpiece surface. Simultaneously, a power reduction command is sent to the high-frequency power supply voltage and current acquisition module. This module forwards the power reduction command to the high-frequency power supply control unit, forcibly reducing the actual welding power corresponding to the high-frequency current phase difference in the welding zone. This prevents high-frequency current shunting along the plate surface caused by abnormal V-angle convergence, thus preventing workpiece breakdown due to current and high-frequency power supply overload. When the torque deviation value returns to within the allowable fluctuation threshold range, the torque limiting protection is released, and the adaptive position adjustment of the extrusion roller and the normal welding power of the high-frequency power supply are restored.

[0100] In this embodiment, the core algorithm parameter configurations for the servo control and collaborative protection of the extrusion roller are provided in the following table: Table 4. Key Algorithm Parameter Configuration Table for Squeeze Roll Servo Control and Collaborative Protection ; Regarding the table above, the specific explanations are as follows: The algorithm parameters for each control link defined in the table clarify the core control parameters of the servo control and collaborative protection process. The constraints ensure the stability and safety of the control process, avoiding system oscillation, response lag, or protection malfunctions caused by improper parameter values. The adaptation logic clarifies the adjustment basis for each parameter, allowing for dynamic adaptation based on different specifications of H-beams and different welding conditions, ensuring the position tracking accuracy of the servo system and the safety of the welding process. Through the reasonable configuration of the above parameters, rapid and accurate adjustment of the extrusion roller position can be achieved, while effectively responding to abnormal conditions during the welding process and avoiding mechanical overload and electrical system failures.

[0101] In this embodiment, multiple sets of parallel convolution kernels are used to accurately extract the local deviation characteristics of different V-angle convergence failure modes. Through weighted splicing and nonlinear activation calculation, nonlinear error components caused by mechanical transmission gaps are effectively filtered out, ensuring the output accuracy of the V-angle dynamic deviation value. Through a sliding mode control strategy with dead zone compensation, the influence of external disturbances and static friction dead zones during the welding process is effectively addressed, enabling rapid and accurate adjustment of the extrusion roller position and ensuring the real-time performance and stability of the adjustment action. Through torque monitoring and collaborative control mechanisms, the actual extrusion torque of the extrusion roller is monitored in real time. When relative slippage is detected, torque limiting protection is triggered, the displacement compensation amount is frozen, and the welding power is reduced. This effectively avoids workpiece defects caused by relative slippage between the extrusion roller and the workpiece, and also prevents electrical faults caused by abnormal V-angle convergence, ensuring the stability and safety of the welding process.

Claims

1. A method for welding H-shaped high-frequency steel with intelligent positioning, characterized in that, include: A line laser profile sensor and a high-frequency power supply voltage and current acquisition module are deployed on the feed path of H-beam high-frequency welding. The initial assembly angle and initial misalignment of the wing plate and web plate are extracted using the line laser profile sensor. The high-frequency power supply voltage and current acquisition module acquires the high-frequency current phase difference and impedance modulus of the welding zone in real time. Establish a mapping relationship model between the initial assembly angle, the high-frequency current phase difference, the impedance magnitude, and the actual convergence state of the V-angle; The initial assembly angle is used as a feedforward feature input to the mapping relationship model, and the high-frequency current phase difference and the impedance magnitude are used as feedback features input to the mapping relationship model. The V-angle dynamic deviation value is output through the mapping relationship model. The displacement compensation amount of the servo motor of the extrusion roller is calculated based on the dynamic deviation value of the V-angle, and the position of the extrusion roller is adjusted in real time before the welding point based on the displacement compensation amount of the servo motor.

2. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 1, characterized in that, The initial assembly angle and initial misalignment of the wingplate and web plate are extracted using the line laser profile sensor, including: The line laser profile sensor is controlled to project structured light strips onto the splicing area of ​​the wing and the web to acquire raw point cloud data containing the surface topography of the wing and the web; The original point cloud data is subjected to denoising processing based on statistical filtering to remove outlier points caused by oxide scale on the surfaces of the wing plate and the web plate, thereby obtaining denoised point cloud data. The first plane equation of the wing plate and the second plane equation of the web plate are extracted using the random sampling consensus algorithm on the denoised point cloud data. The spatial angle between the first plane equation and the second plane equation is calculated as the initial assembly angle, and the distance difference between the first plane equation and the second plane equation in the direction perpendicular to the feed direction is calculated as the initial value of the misalignment amount.

3. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 1, characterized in that, The high-frequency power supply voltage and current acquisition module acquires the high-frequency current phase difference and impedance modulus of the welding area in real time, including: A high-frequency voltage transformer is connected in parallel at the output terminal of the high-frequency power supply, and a high-frequency current transformer is connected in series in the circuit from which the high-frequency power supply flows to the welding electrode. The instantaneous voltage signal and instantaneous current signal of the welding area are acquired synchronously by the high-frequency voltage transformer and the high-frequency current transformer. The instantaneous voltage signal and the instantaneous current signal are subjected to quadrature phase-locked amplification processing to eliminate the coupling interference of the background electromagnetic noise in the welding area on the amplitude and phase of the instantaneous voltage signal and the instantaneous current signal; Based on the instantaneous voltage signal and the instantaneous current signal after orthogonal lock-in amplification, the high-frequency current phase difference and the impedance magnitude of the welding area are calculated.

4. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 1, characterized in that, Establish a mapping model between the initial assembly angle, the high-frequency current phase difference, the impedance magnitude, and the actual convergence state of the V-angle, including: The initial assembly angle is converted into a theoretical arrival timestamp based on the fixed physical distance between the line laser profile sensor and the welding point. The high-frequency current phase difference and the impedance magnitude are marked as the measured arrival timestamps according to their respective data acquisition timestamps. Using the theoretical arrival timestamp and the measured arrival timestamp as indexes, interpolation and alignment operations are performed on the time dimension for the initial assembly angle, the high-frequency current phase difference, and the impedance magnitude. The initial assembly angle, the high-frequency current phase difference, and the impedance modulus after time-dimensional interpolation and alignment are input into a long short-term memory network for feature-level fusion training to generate the mapping relationship model.

5. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 1, characterized in that, The V-angle dynamic deviation value is output through the mapping relationship model, including: The mapping relationship model contains multiple sets of parallel convolution kernels, and each set of convolution kernels corresponds to a specific V-shaped angle convergence failure mode. The mapping relationship model inputs the initial assembly angle, the high-frequency current phase difference, and the impedance modulus after time-dimension interpolation and alignment operations into the multiple sets of parallel convolution kernels; The multiple sets of parallel convolutional kernels respectively output local deviation feature vectors for the V-shaped angle convergence failure mode; The local deviation feature vector is weighted and nonlinearly activated to filter out the nonlinear error components caused by mechanical transmission clearance in the local deviation feature vector, and the V-angle dynamic deviation value is output as the final result.

6. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 1, characterized in that, The servo motor displacement compensation amount of the extrusion roller is calculated based on the dynamic deviation value of the V-angle. Real-time adaptive adjustment of the extrusion roller position is then performed before the welding point based on the servo motor displacement compensation amount, including: Obtain the current real-time feed speed of the H-beam, and divide the dynamic deviation value of the V-angle by the real-time feed speed to obtain the theoretical adjustment time difference of the extrusion roller; The theoretical adjustment time difference is advanced to the physical position of the weld point and converted into a feedforward displacement; The servo motor displacement compensation amount is obtained by superimposing the feedforward displacement amount with the preset extrusion roller reference position. A sliding mode control strategy with dead zone compensation is adopted. A drive signal is generated based on the displacement compensation of the servo motor, and the drive signal is sent to the servo motor that is rigidly connected to the extrusion roller.

7. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 2, characterized in that, Before performing statistical filtering-based denoising on the raw point cloud data, the method further includes performing multi-frame joint processing on the raw point cloud data: During the feeding process of the H-beam, multiple frames of the original point cloud data are continuously acquired at a fixed sampling frequency; The continuous multiple frames of the original point cloud data are projected into the same three-dimensional spatial coordinate system according to the feed speed of the H-beam to form a fused point cloud set; Calculate the overlapping region of two adjacent frames of the original point cloud data in the fused point cloud set, and perform point cloud registration operation in the overlapping region to eliminate the spatial coordinate drift error of a single frame of the original point cloud data caused by the vibration of the line laser profile sensor itself during installation. The fused point cloud set after point cloud registration is subjected to voxel downsampling to obtain the original point cloud data with uniform spatial distribution density.

8. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 3, characterized in that, After performing quadrature phase-locked amplification on the instantaneous voltage signal and the instantaneous current signal, the method further includes an outlier removal mechanism for the high-frequency current phase difference and the impedance magnitude: A sliding time window is set, and the high-frequency current phase difference and the impedance magnitude are continuously extracted from multiple time sampling points within the sliding time window; Calculate the first-order difference between the phase difference of the high-frequency current and the first-order difference between the impedance magnitude within the sliding time window; When the first-order difference of the high-frequency current phase difference or the first-order difference of the impedance magnitude exceeds the preset physical change extreme value, the high-frequency current phase difference or the impedance magnitude at the corresponding time sampling point is determined to be abnormal data caused by the transient breakdown of the internal switching transistor of the high-frequency power supply. The abnormal data is removed from the sliding time window and supplemented by linear interpolation using data from adjacent normal time sampling points.

9. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 4, characterized in that, After generating the mapping relationship model, the method further includes performing an online self-learning update mechanism on the mapping relationship model: An infrared thermal imager is deployed behind the weld joint to collect data on the actual temperature field distribution of the weld at the weld joint. The offset of the highest temperature point in the actual temperature field distribution data of the weld is used as the online correction error index of the mapping relationship model; Using the online correction error index as a supervision signal, calculate the gradient descent value of each hidden layer node of the Long Short-Term Memory network in the mapping relationship model; The connection weight matrix inside the Long Short-Term Memory network is iteratively updated based on the gradient descent value.

10. The intelligent positioning method for welding H-shaped high-frequency steel according to claim 6, characterized in that, After sending the drive signal to the servo motor rigidly connected to the extrusion roller, the system also includes coordinated control of feedforward compensation and extrusion roller torque monitoring. A torque sensor is installed at the output shaft end of the servo motor, and the actual extrusion torque applied by the extrusion roller to the wing plate and the web plate is collected in real time through the torque sensor; The actual extrusion torque is compared with the preset reference extrusion torque to obtain the torque deviation value; When the torque deviation value is greater than the allowable fluctuation threshold, it is determined that relative sliding has occurred between the extrusion roller and the wing plate or the web plate, triggering the torque limiting protection. When the torque limiting protection is triggered, the displacement compensation amount of the servo motor at the current moment is frozen, and a power reduction command is sent to the high-frequency power supply voltage and current acquisition module.

Images