An accurate and rapid repositioning method for ABB welding robot tool coordinates

Abstract

The application relates to the field of industrial robot control and discloses an ABB welding robot tool coordinate accurate and rapid repositioning method, which comprises the following steps: acquiring an end tool equivalent transverse flexibility matrix and an elastic deformation limit; synchronously collecting visual depth features and force control sequences, aligning and extracting an optimal regularized generation value by using a dynamic time warping algorithm, and calculating a stiffness and an approaching speed to issue a control instruction in real time; when stress reaches a trigger threshold, extracting a visual surface gradient and a force control derivative peak value to generate an abnormal contact confidence coefficient; combining instantaneous shear force to calculate a microscopic side slip prediction compensation vector, superimposing the vector to a Kalman filtering model to solve an initial deviation, and calling a nonlinear optimization algorithm to approach a residual vector to complete repositioning updating. By aligning the phases of heterogeneous data and smoothing the stiffness, splash impurity interference is filtered and contact side slip error is compensated, so that the calibration stability and absolute accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of industrial robot control technology, specifically to a method for precise and rapid repositioning of tool coordinates on an ABB welding robot. Background Technology

[0002] When industrial robots perform automated welding tasks, the accuracy of the tool's center point coordinates directly determines the tracking accuracy of the weld seam trajectory. During long-term production operations, factors such as accidental collisions, mechanical wear, or welding torch replacements can cause spatial shifts in the robot's tool coordinate system, necessitating periodic repositioning. Existing automatic calibration or repositioning methods typically rely on a single vision or force sensor. However, with increasing demands for calibration accuracy and automation, multimodal fusion methods combining vision and force sensors are increasingly being applied.

[0003] However, existing relocation methods still have some limitations in practical industrial calibration scenarios. On the one hand, there are significant differences in the sampling frequency and underlying communication link between the 3D point cloud acquired by the vision sensor and the 1D force data acquired by the force sensor. This natural asynchrony in the time phase of this heterogeneous data makes it difficult for the robot control system to achieve smooth damping and stiffness adjustment during the high-speed approach of the end effector to the calibration plane. This can easily lead to a rigid collision between the end effector and the calibration object at the moment of contact, reducing the safety and stability of the calibration process.

[0004] On the other hand, real welding worktables or calibration reference surfaces inevitably have tiny impurities such as welding spatter and slag attached to them. When the end tool touches these impurities, existing methods often rely solely on set single-dimensional force control thresholds or visual depth thresholds for judgment, which can easily misjudge local impurity protrusions as real calibration reference surfaces, leading to incorrect contact point extraction and direct calibration failure.

[0005] Furthermore, at the instant the end-effector establishes physical contact with the calibration surface and applies a normal probing force, the center point of the end-effector will experience microscopic lateral deformation and slippage due to the structural flexibility of the welding torch and the surface friction state. Existing coordinate iterative solution models typically treat the robot end-effector as an absolutely rigid body, directly using the joint position at the moment of contact triggering to calculate spatial coordinates, completely ignoring the physical errors caused by this microscopic lateral slippage and elastic deformation. These uncompensated local slippage amounts are directly coupled into the calculation of coordinate residuals, resulting in a systematic deviation in the final fitted tool coordinate system that is difficult to eliminate, failing to meet the absolute accuracy requirements for repositioning in high-precision welding processes. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a precise and rapid repositioning method for ABB welding robot tool coordinates. This method solves the problems of unstable contact control caused by asynchronous data from heterogeneous sensors, false contact judgments caused by surface impurities in complex welding environments, and decreased calibration accuracy due to microscopic lateral slippage of the tool end at the moment of force contact during the repositioning process of robot tool coordinates.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method for precise and rapid repositioning of tool coordinates of an ABB welding robot, applicable to a hardware environment including an industrial computer, a robot controller, a vision sensor, and a six-dimensional force sensor.

[0008] The method includes: acquiring the equivalent lateral flexibility matrix, maximum allowable elastic deformation limit, and estimated contact point of the robot's current end-effector; determining the system's overall worst-case response time; simultaneously acquiring visual depth feature sequences and force-controlled normal force sequences within a set sliding time window; using a dynamic time warping algorithm to align the visual depth feature sequences and force-controlled normal force sequences in time phase and extract the optimal warping generation value; calculating the equivalent lateral stiffness parameters and dynamic target approach velocity in real time based on the optimal warping generation value; and issuing control commands to the robot controller to execute actions. When the normal force fed back by the six-dimensional force sensor reaches the preset normal force trigger threshold, the peak values ​​of the surface gradient in visual space and the normal derivative of force control are extracted to generate an abnormal contact confidence coefficient. An adaptive Kalman filter model is constructed, and the micro-sideslip prediction compensation vector is calculated by combining the instantaneous shear force component output by the six-dimensional force sensor, the equivalent lateral compliance matrix, and the abnormal contact confidence coefficient. The micro-sideslip prediction compensation vector is superimposed on the prior state estimation equation of the adaptive Kalman filter model to solve for the initial coordinate deviation. A nonlinear optimization algorithm is called to perform residual approximation to obtain the coordinate residual vector. When the coordinate residual vector converges, the repositioning update of the robot tool coordinate system is completed.

[0009] The steps for obtaining the equivalent lateral compliance matrix and maximum permissible elastic deformation limit of the robot's current end-effector, and determining the system's overall worst-case response time, specifically include: applying a known gradient lateral static test force to the end-effector; simultaneously recording the lateral force components in the robot's flange coordinate system and the actual microscopic displacement vector generated by the end-effector; and calculating the equivalent lateral compliance matrix using linear fitting via the least squares method. The material yield strength parameters, section moment of inertia, and assembly extension length of the end-effector are obtained. The critical yield displacement value is calculated using a cantilever beam stress model, and the maximum permissible elastic deformation limit is calculated by multiplying the critical yield displacement value by a set safety margin coefficient. Based on the system's internal timestamp comparison mechanism, the processing time for the visual sensor to acquire 3D point cloud data and transmit it to the industrial computer, the delay time for the six-dimensional force sensor to feed back force data and perform noise reduction filtering, the maximum delay time due to network communication jitter, the single position control cycle time, and the electromechanical response time are calculated and recorded, and summed to obtain the system's overall worst-case response time.

[0010] The steps for extracting the optimal normalization value specifically include: calculating the rate of change of the visual depth feature sequence and the first derivative of the force control normal sequence; applying linear mapping normalization to obtain a normalized visual feature sequence and a normalized force control feature sequence with a unified value range; constructing a two-dimensional cumulative distance matrix, initializing the starting node of the two-dimensional cumulative distance matrix to zero, and initializing the remaining nodes in the first row and first column to infinity; traversing the two-dimensional cumulative distance matrix in ascending order of time index, and performing dynamic programming recursive calculation by combining the Euclidean distance of local differences and the minimum cumulative distance value among adjacent preceding nodes; and extracting the cumulative distance of the ending node of the two-dimensional cumulative distance matrix as the optimal normalization value.

[0011] The steps for real-time calculation of equivalent lateral stiffness parameters and dynamic target approximation velocity based on the optimal warping algebraic value include: introducing a basic stiffness adjustment coefficient and a cost weight penalty coefficient; adjusting the inverse matrix of the equivalent lateral flexibility matrix based on the optimal warping algebraic value; and outputting the equivalent lateral stiffness parameters by constraining the lower and upper stiffness limits through a saturation limiting function. A velocity decay rate constant is introduced, and the optimal warping algebraic value is substituted into an exponential decay function. Combining the ratio of the maximum allowable elastic deformation limit to the system's worst-case comprehensive response time, the smoothly decreasing dynamic target approximation velocity is calculated and output.

[0012] The steps for sending control commands to the robot controller to execute actions include: substituting the current force state fed back by the six-dimensional force sensor, the calculated equivalent lateral stiffness parameter, and the preset damping coefficient and inertial mass parameter into the damped mass spring model to calculate the target position correction vector for the next control cycle. The target position correction vector and the dynamic target approximation velocity are packaged and encapsulated into a fixed-length control command data frame containing control timestamps and spatial coordinate information. This fixed-length control command data frame is then sent to the underlying communication process within the robot controller via the User Datagram Protocol (UDP) in a local area network environment, and subsequently transmitted to the servo drives of each joint via the internal bus.

[0013] When the normal force fed back by the six-dimensional force sensor reaches a preset normal force trigger threshold, the steps for extracting the visual space surface gradient and the peak value of the force-controlled normal derivative include: Firstly, setting a preset normal force trigger threshold based on the amplitude of the base noise under no-load, static conditions. When the normal force reaches the preset threshold, a data freezing mechanism is triggered, capturing the 3D point cloud data and the force-controlled normal force sequence collected by the visual sensor within the currently set sliding time window and storing them in static buffer memory. From the frozen 3D point cloud data, a local neighborhood point cloud set centered on the estimated contact point is extracted, and the principal component analysis algorithm is used to calculate the normal vector variance as the visual space surface gradient. A first-order difference operation is performed on the frozen force-controlled normal force sequence, iterating through and finding the maximum absolute value in the difference sequence, and using the maximum absolute value as the peak value of the force-controlled normal derivative.

[0014] The steps for generating anomaly contact confidence coefficients specifically include: constructing cross-validation logic based on the physical consistency of heterogeneous data, introducing a sensitivity adjustment coefficient and a minimum positive number to prevent denominator calculation anomalies; multiplying the peak value of the force control normal derivative by the sensitivity adjustment coefficient as the numerator, and summing the visual spatial surface gradient with the minimum positive number as the denominator to obtain the ratio term; performing a negative exponentiation operation on the ratio term with the natural constant as the base, calculating the difference between 1 and the negative exponentiation operation result to generate the anomaly contact confidence coefficient.

[0015] The steps for calculating the micro-sideslip prediction compensation vector by combining the instantaneous shear force component output by the six-dimensional force sensor, the equivalent lateral compliance matrix, and the abnormal contact confidence coefficient are as follows: First, set the coordinate deviation of the end-effector center point in three-dimensional space as the system state vector. Second, multiply the basic observation noise matrix of the adaptive Kalman filter model by a dynamic scaling factor that includes the abnormal contact confidence coefficient and a preset scaling gain, amplifying the observation noise variance setting value for the current observation period. Third, obtain the instantaneous shear force component, which includes the lateral force vectors in the X and Y directions of the robot's flange coordinate system, output in real time from the six-dimensional force sensor. Fourth, calculate the difference between step 1 and the abnormal contact confidence coefficient, and multiply the obtained difference, the equivalent lateral compliance matrix, and the instantaneous shear force component to obtain the micro-sideslip prediction compensation vector.

[0016] The steps for obtaining the coordinate residual vector by calling a nonlinear optimization algorithm for residual approximation specifically include: mapping the initial coordinate deviation to the target compensation angles of the robot's six joint motors and sending them to the joint servo drivers to perform preliminary spatial compensation; planning multiple spatial approach postures with yaw and pitch angle offsets using the estimated contact point as the spatial center, and sequentially re-executing force approximation and fusion calculations to obtain a multivariate observation dataset; constructing a nonlinear least squares objective function with the tool center point coordinates as the independent variable, and introducing the contact force normal vector as a geometric penalty term into the nonlinear least squares objective function to constrain the gradient direction of the coordinate iteration update to be orthogonal to the pre-set calibration plane normal.

[0017] When the coordinate residual vector converges, the repositioning and updating of the robot tool coordinate system is completed. Specifically, this includes: performing numerical iteration using the Levenburg-Marquardt algorithm to dynamically adjust the damping coefficient and calculate the observed residual vector and displacement update step size; calculating the Euclidean norm of the current displacement update step size and the sum of squared residuals of the nonlinear least squares objective function; and determining that the coordinate residual vector has reached convergence when the Euclidean norm is less than a preset step size threshold or the sum of squared residuals is less than a preset residual threshold. The current tool center point coordinate vector is then extracted as the final tool center point calibration coordinates, and these coordinates are written into the system parameter storage area of ​​the robot controller.

[0018] This invention provides a method for precise and rapid tool coordinate repositioning of an ABB welding robot. It offers the following advantages: 1. This invention utilizes a dynamic time warping algorithm to align the visual depth feature sequence and the force control normal force sequence in time phase, and adjusts the equivalent lateral stiffness parameter and dynamic target approach speed in real time based on the calculated optimal warping cost. This technical feature solves the time asynchrony problem in the data acquisition process of heterogeneous visual and force sensors, enabling the robot to smoothly adjust its motion speed and control stiffness according to its state when approaching the calibration object, avoiding rigid collisions at the moment of contact, and improving the smoothness and safety of the calibration approach process.

[0019] 2. This invention extracts the visual spatial surface gradient and the peak value of the force control normal derivative at the moment of contact, constructs a cross-validation logic based on physical consistency, and generates anomaly contact confidence coefficients. This technical feature utilizes the mutual corroboration between visual surface morphology changes and abrupt force changes to accurately identify and filter false contact signals caused by spatter or impurities remaining on the welding workbench surface, eliminating local surface interference in complex industrial environments and improving the accuracy of contact point determination and the system's anti-interference capability.

[0020] 3. This invention calculates a microscopic sideslip prediction compensation vector by combining the instantaneous shear force component and the equivalent lateral compliance matrix, introduces it into an adaptive Kalman filter model, and adds the contact force normal vector as a geometric penalty term in subsequent nonlinear optimization. This technical feature directly compensates for the tiny lateral slip error generated at the moment of force contact at the tool end in the underlying data fusion stage, and constrains the convergence direction of coordinate iteration updates from a geometrical space, effectively eliminating the cumulative influence of sideslip displacement on the calibration results and improving the accuracy of tool coordinate system repositioning. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the system architecture and method of an embodiment of the present invention; Figure 2 This is a comparison chart of calibration residual convergence in an embodiment of the present invention; Figure 3 This is a comparison diagram of contact impact resistance forces in embodiments of the present invention; Figure 4 This is a flowchart illustrating the specific implementation of the coordinate iterative optimization process in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 This invention provides a method for precise and rapid tool coordinate repositioning of an ABB welding robot, which is executed based on a multi-source heterogeneous sensing hardware environment. The hardware environment includes an industrial computer, a robot controller equipped with an external motion guidance interface, and vision sensors and a six-dimensional force sensor mounted on the robot's end flange.

[0024] A vision sensor is used to acquire 3D point cloud data of the robot's working area. A six-dimensional force sensor is used to acquire the spatial force state of the robot's end effector during its contact with the environment. An industrial computer, as the data processing and main control device, establishes a high-frequency communication connection with the robot controller through an external motion guidance interface to receive instructions and drive the robot to perform actions.

[0025] This invention provides a method for accurate and rapid repositioning of tool coordinates on an ABB welding robot, comprising the following steps: S1, Physical parameters and system boundary initialization. The industrial computer obtains the equivalent lateral flexibility matrix and maximum allowable elastic deformation limit of the current end tool, and determines the overall worst response time of the system under the current hardware configuration and communication link as the underlying boundary condition for the system's safety envelope control; S2, Feature Sequence Extraction and Time Warping Cost Evaluation. During the robot's approach to the calibration block, the depth sequence output by the visual sensor and the force sequence output by the six-dimensional force sensor are simultaneously acquired within a set sliding time window. The industrial computer calculates the rate of change of the visual depth feature sequence and the first derivative of the force control normal force sequence. The dynamic time warping algorithm is used to align the time phase of the two sets of sequences, and the optimal warping cost value representing the degree of phase divergence between the two is extracted. S3, Dual Safety Envelope Control. Based on the extracted optimal normalization value, the industrial computer calculates the equivalent lateral stiffness parameters and target approach velocity in real time, and converts them into corresponding pose or velocity correction commands, which are sent to the robot controller via the external guided motion interface, so that the robot end effector exhibits dynamically adjusted lateral compliance and longitudinal speed limiting characteristics when approaching the calibration block; S4, Contact Noise Identification and Cross-validation. When the normal force fed back by the six-dimensional force sensor reaches the preset trigger threshold, the industrial computer extracts the visual spatial surface gradient and the peak value of the force control normal derivative of the contact landing area. Based on the logical relationship between the surface gradient and the peak value of the derivative, it determines whether the current contact state is disturbed by a small obstacle and generates an abnormal contact confidence coefficient. S5, State Prediction Compensation and Data Fusion Solution. An industrial computer constructs an adaptive Kalman filter model to estimate the coordinate deviation of the tool. Combining the instantaneous shear force component output by the six-dimensional force sensor, the equivalent lateral compliance matrix, and the abnormal contact confidence coefficient, a micro-sideslip prediction compensation vector is calculated. This vector is then superimposed as a physical constraint compensation term into the prior state estimation equation of the adaptive Kalman filter to offset the observation error caused by nonlinear lateral slip. S6, Coordinate Iteration Optimization. The industrial computer receives the initial coordinate deviation from the adaptive Kalman filter output, calls a nonlinear optimization algorithm for residual approximation, and introduces the contact force normal vector as a geometric penalty term into the optimization objective function to constrain the gradient direction of the coordinate iteration update to be orthogonal to the normal of the calibration plane. When the coordinate residual vector converges to the set allowable error range, the iteration stops and the repositioning update of the robot tool coordinate system is completed.

[0026] In step S1, the specific implementation process of initializing physical parameters and system boundaries includes the following steps: S101, Obtain the equivalent lateral compliance matrix of the current end effector. In this embodiment, considering that the robot's end effector is mounted on the robot's end flange, its overall deformation under stress is affected not only by the elastic modulus of the tool's own material, but also by the inherent stiffness of the tool flange clamp and the microscopic mechanical clearance of the mounting contact surface. To accurately map the end effector force to the actual physical lateral slip displacement, the system obtains this compliance matrix through offline calibration experiments.

[0027] As a preferred method, in offline testing, a known gradient lateral static test force is applied to the robot's end effector. An industrial computer simultaneously records the lateral force components in the flange coordinate system and the actual micro-displacement vector generated at the end effector. The industrial computer performs linear fitting on multiple sets of test data using the least squares method, calculates the mapping coefficient matrix between the lateral force components and the micro-displacement vector, and saves this coefficient matrix as an equivalent lateral compliance matrix. For the specific data acquisition operation of applying the test force and obtaining the micro-displacement data through an external sensor, those skilled in the art can use a standard mechanical loading device in conjunction with a high-precision displacement sensor. The physical testing process is well-known in the field and will not be elaborated upon here.

[0028] S102, Define and input the maximum permissible elastic deformation limit. This limit parameter constitutes the displacement boundary condition of the subsequent safety control envelope, used to limit the maximum geometric intrusion of the robot during blind zone contact. The industrial computer obtains the material yield strength parameters, section moment of inertia, and current assembly extension length of the end effector, and calculates the critical yield displacement value before plastic deformation of the tool using a cantilever beam stress model. To preserve buffer space, the industrial computer multiplies this critical yield displacement value by a set safety margin coefficient to calculate the maximum permissible elastic deformation limit and stores it in system memory. The value of this safety margin coefficient can be set from 0.7 to 0.9. This parameter setting aims to constrain the deformation of the tool body within the elastic deformation range of the material during subsequent collision compensation, thereby facilitating the tool's self-recovery to its initial zero-position state after disengagement from contact.

[0029] S103, Determine the overall worst-case response time of the system under the current hardware configuration and communication link. To enable collision avoidance fallback capability for target approach speed control, the industrial computer needs to obtain the maximum time delay between environmental perception and the actual physical deceleration action of the robotic arm. This overall worst-case response time is the sum of the extreme values ​​of the time overhead across the entire system link.

[0030] The industrial computer, using a built-in clock and high-frequency test command package, calculates and records the following extreme time overhead values ​​based on the system's internal timestamp comparison mechanism: the processing time for the vision sensor to acquire 3D point cloud data and transmit it to the industrial computer; the delay time for the six-dimensional force sensor to feed back force data and perform noise reduction filtering calculations; the maximum delay time due to network communication jitter between the industrial computer and the robot controller; the inherent single-position control cycle time of the robot controller's external guided motion interface; and the electromechanical response time from the robot's underlying servo system receiving control commands to the motor's actual output braking torque. The industrial computer sums these extreme time values ​​to obtain the comprehensive worst-case response time. This time parameter defines the time from algorithm-triggered intervention to the robotic arm's deceleration action under hard contact conditions, providing a low-level time reference for subsequently constructing a dynamic speed-limiting attenuation model.

[0031] In step S2, the specific implementation process of feature sequence extraction and time warping cost evaluation includes the following steps: S201, Construct a sliding time window and synchronously collect heterogeneous data. During the robot's approach to the calibration block at the end effector, the sampling frequencies of the vision sensor and the six-dimensional force sensor differ significantly due to their different physical characteristics. To process the asynchronous data stream, the industrial computer establishes a fixed-length circular queue in memory as a sliding time window. In this embodiment, the vision sensor outputs low-frequency 3D point cloud data, while the six-dimensional force sensor outputs high-frequency force data. The system timestamps each frame of data received according to a unified system clock and stores them concurrently within the sliding time window, thereby forming a visual depth sequence and a force-controlled normal force sequence carrying time tags.

[0032] For the extraction of visual depth sequences, the industrial computer specifically extracts the vertical distance data from the projection of the robot tool's center point onto the calibration block surface from the 3D point cloud data, and combines this data according to timestamps to form the sequence. As a preferred method, the length of this sliding time window can be set between 0.2 seconds and 0.5 seconds to fully cover the transient cycle of the contact event while reducing memory usage.

[0033] S202, calculate feature parameters and normalize the data sequence. To eliminate the differences in physical dimensions between heterogeneous sensor data, the system needs to convert the original displacement and mechanical quantities into a dimensionless sequence characterizing the trend of contact events. The industrial computer performs a first-order difference operation on the visual depth sequence within the sliding time window, calculates the ratio of the vertical distance data between adjacent frames to the time difference, and obtains the rate of change of the visual depth feature sequence. Simultaneously, a first-order difference operation is performed on the force-controlled normal force sequence to calculate the first derivative of the force-controlled normal force sequence.

[0034] Considering the difference in numerical magnitude between the two derivative sequences, the industrial computer uses linear mapping normalization to process the calculated rate of change and first derivative sequences, scaling the data range of both sets of sequences to the interval between 0 and 1, thereby obtaining mutually independent normalized visual feature sequences and normalized force control feature sequences. For the specific algorithm implementation of discrete sequence normalization, those skilled in the art can use conventional numerical processing methods such as max-min normalization or standardized scaling; the data mapping process is well-known in the field and will not be elaborated upon here.

[0035] S203 executes the dynamic time warping algorithm and extracts the optimal warping value. To address the time phase misalignment caused by network jitter and differences in sensor processing time, the system uses the dynamic time warping algorithm to align the time phase of two normalized feature sequences. Its basic principle is to achieve optimal waveform matching by locally stretching or compressing the two signals along the time axis, thereby overcoming the comparison difficulties caused by asynchronous sampling rates of heterogeneous sensors.

[0036] An industrial computer constructs a two-dimensional cumulative distance matrix and uses a dynamic programming algorithm to find the shortest mapping path between two sequences of unequal length to align abrupt changes in depth and force. To avoid logical dead zones, the industrial computer initializes the boundaries of the cumulative distance matrix before recursive calculation, setting the value of the starting node to 0 and setting the values ​​of the remaining nodes in the first row and first column to infinity. Subsequently, the update of the node values ​​in the cumulative distance matrix follows the recursive logic of dynamic programming. The specific expression for the cumulative distance of the current node in the cumulative distance matrix is ​​as follows: ; In the formula, Represents the first normalized visual feature sequence. The data point and the th data point in the normalized force control feature sequence The cumulative distance when aligning data points; The Euclidean distance between the two specific data points is used to characterize local differences. This represents the minimum value function, used to recursively select the minimum cumulative distance value from three adjacent preceding nodes; and These represent the discrete-time indices of the two sets of sequences, respectively.

[0037] The industrial computer traverses the cumulative distance matrix in ascending order of time index until the cumulative distance of the endpoint node is calculated. This cumulative distance is then extracted and stored as the optimal normalization iteration value. This optimal normalization iteration value objectively reflects the degree of phase deviation between the visually predicted contact trend and the actual physical force state. A larger value indicates higher environmental uncertainty and spatiotemporal distortion, thus providing an objective quantitative basis for adjusting the control strategy in the subsequent system.

[0038] In step S3, the specific implementation process of dual safety envelope control includes the following steps: S301, calculate the equivalent lateral stiffness parameter based on the optimal warp algebra value. The basic principle of admittance control is to construct a second-order mathematical model consisting of mass, springs, and damping to dynamically convert the external contact force on the robot end effector into a position or velocity correction deviation, thereby making the originally rigid position control system exhibit mechanical compliance characteristics. In this embodiment, to enable the robot controller to adaptively handle the environmental uncertainties caused by the inconsistency of multi-source sensor signals, the industrial computer calculates the equivalent lateral stiffness parameter based on the optimal warp algebra value extracted in the aforementioned steps. The calculation formula for the equivalent lateral stiffness parameter is as follows: ; In the formula, This represents the equivalent lateral stiffness parameter, which is in diagonal matrix form. This represents the equivalent lateral flexibility matrix, and its inverse matrix represents the basic contact stiffness of the tool under ideal conditions. This represents the value of the optimal regularization generation; The basic stiffness adjustment coefficient can be set from 0.5 to 1.0. The cost weight penalty coefficient is used to adjust the sensitivity of the regularization cost to stiffness attenuation, and its value range can be set from 1.0 to 5.0. Represents the saturation limiting function; This indicates the set lower limit value for stiffness; This indicates the set upper limit value for stiffness. The lower and upper limits for stiffness are specifically obtained through offline force control experiments based on the maximum allowable load-bearing stiffness of the robot's end effector joint and the minimum maintaining stiffness required for the calibration task.

[0039] The physical significance of establishing this mathematical model lies in the fact that a large optimal warp value indicates a significant phase deviation between the visual and force signals, suggesting high uncertainty in the system's contact environment. In this case, the denominator term increases, the equivalent lateral stiffness parameter decreases, and the robot's end effector exhibits more flexible mechanical characteristics, thereby reducing lateral contact reaction forces. To prevent an excessively large optimal warp value from causing the calculated stiffness to approach zero or from negative stiffness due to rounding errors, which could lead to oscillations and divergence in the control system, the industrial computer uses a saturation limiting function to restrict the calculation results between the lower and upper stiffness limits, thus maintaining the positive definiteness of the admittance control law.

[0040] S302, calculate the dynamic target approach speed limited by system latency. Based on the determined lateral compliance parameters, the system needs to further calculate the upper limit of the robot's motion speed along the normal approach to the calibration block to construct a longitudinal safety envelope. An industrial computer constructs a dynamic target approach speed decay model that combines the worst-case response time with the maximum permissible elastic deformation limit. Its physical meaning is that, even after experiencing the longest system communication and servo response delays, when the robot makes hard contact at the current speed, its sliding distance within the blind zone will not exceed the critical deformation threshold for plastic failure of the tool material. The specific calculation formula for the dynamic target approach speed is as follows: ; In the formula, Indicates the dynamic target approach speed; Indicates the maximum permissible elastic deformation limit; Indicates the overall worst-case response time; Represented by natural constant An exponential function with base 0; The velocity decay rate constant can be set from 1.5 to 3.0. This represents the value of the optimal regularization generation.

[0041] Industrial computers utilize this exponential decay logic to ensure that the approach speed of a dynamic target decreases non-linearly and smoothly as the optimal regularization cost increases. Compared to step-type direct braking, this method avoids end-effector mechanical shocks and joint motor torque overloads caused by sudden speed changes, while also limiting the theoretical maximum intrusion displacement within physical safety boundaries.

[0042] S303 converts and issues external guidance motion commands. After calculating the stiffness and velocity parameters, the industrial computer converts them into low-level control signals. Based on the constructed admittance control law, the industrial computer combines the force state fed back by the current six-dimensional force sensor with the calculated equivalent lateral stiffness parameters, along with the damping coefficient and inertial mass parameters preset based on the critical damping condition, and substitutes them into the damped mass spring model to calculate the target position correction vector for the next control cycle. Simultaneously, the system uses the aforementioned calculated dynamic target approach velocity as a feedforward reference for the robot's normal motion.

[0043] As a preferred approach, the industrial computer packages the target position correction vector and dynamic target approximation velocity into a fixed-length control command data frame according to the communication protocol standard of the external guided motion interface. This command data frame includes a control timestamp and corresponding spatial coordinate information. The industrial computer sends the control command data frame to the robot controller via User Datagram Protocol (UDP) in a local area network environment, with a control communication cycle set to 4 milliseconds. The underlying communication process within the robot controller receives and parses the command data frame, and transmits the position and velocity correction parameters to the servo drivers of each joint via the internal bus, driving the motor rotor to perform corresponding compliant lateral displacement compensation and normal velocity-limited approximation movements. This low-level direct-drive method bypasses the conventional robot advanced trajectory planner, helping to reduce command parsing overhead and meet the high real-time requirements of dual safety envelope control.

[0044] In step S4, the specific implementation process of contact noise identification and cross-validation includes the following steps: S401, Set the normal force trigger threshold and execute data freeze. During the robot's approach to the calibration block under dual safety envelope control, the system monitors the normal force data fed back by the six-dimensional force sensor in real time. To determine whether the tool tip has established stable contact with the environmental surface, and to avoid misjudgments caused by sensor zero-point drift or motion inertial forces, the industrial computer pre-sets the normal force trigger threshold.

[0045] In this embodiment, the normal force trigger threshold is specifically configured offline based on a set multiple of the base noise amplitude of the six-dimensional force sensor in an unloaded, stationary state. Preferably, this set multiple can be 3 to 5 times. When the normal force fed back by the six-dimensional force sensor reaches the preset normal force trigger threshold, the system determines that a contact event has occurred and triggers a data freeze mechanism. The industrial computer extracts the three-dimensional point cloud data and the force-controlled normal force sequence within the current sliding time window and stores them in a static buffer memory. This freeze operation extracts data slices before and after the contact transient, which helps reduce the interference caused by the vibration of the robot's mechanical structure after contact to subsequent data analysis.

[0046] S402, extract the visual spatial surface gradient and the peak value of the force control normal derivative. After acquiring the frozen data, the industrial computer simultaneously extracts the visual morphological features and mechanical response features of the contact point area. For the visual features, the industrial computer extracts a set of local neighborhood point clouds from the frozen 3D point cloud data, centered on the estimated contact point and within a preset radius. The preset radius can be set from 5 mm to 15 mm according to the expected roughness of the calibration block surface.

[0047] The industrial computer uses principal component analysis (PCA) to calculate the variance of the normal vector of the local neighborhood point cloud set, and uses this variance as the surface gradient in visual space. The smaller the gradient value, the more flat the contact area is on a visual scale. For mechanical characteristics, the industrial computer performs a first-order difference operation on the force-controlled normal force sequence within the freeze time window, iterating through the difference sequence to find the maximum absolute value, which is used as the peak value of the force-controlled normal derivative. This peak value reflects the degree of physical impact at the moment of contact. The extraction of these two features provides a cross-sensor modality data foundation for assessing the rationality of the physical contact state.

[0048] S403, construct cross-validation logic and generate anomaly contact confidence coefficients. In actual industrial welding environments, calibration block surfaces often have tiny welding spatters that are difficult for visual sensors to distinguish due to resolution limitations. When the tool tip touches such rigid spatters, the force control system generates pulse contact noise with a non-Gaussian distribution, leading to coordinate calibration deviations. To identify and eliminate such interference, the industrial computer constructs vision and force control cross-validation logic based on extracted features. This logic is based on the principle of physical consistency of heterogeneous data, i.e., low-speed contact on a flat surface should result in a relatively gradual change in normal force; if visual data indicates a smooth surface, but force data shows sharp peaks, then a collision with a tiny foreign object within the blind zone is highly probable. The industrial computer quantifies the above logic and calculates the anomaly contact confidence coefficient, the specific calculation formula of which is as follows: ; In the formula, This represents the confidence coefficient for abnormal contact, and its value is limited to between 0 and 1. Represented by natural constant An exponential function with base 0; This is the sensitivity adjustment coefficient, used to control the model's sensitivity to mutation forces. Its value can be set from 0.1 to 0.5. Indicates the peak value of the force-controlled normal derivative; Represents the surface gradient in visual space; It is a preset minimum positive number, specifically 0.0001, used to prevent calculation errors caused by the denominator being zero due to the zero gradient of the visual space surface.

[0049] According to this formula, when the mechanical impact is significant and the visual surface is smooth, the ratio term increases, and the calculated confidence coefficient for abnormal contact approaches 1. Based on this, the system determines that the current contact has been affected by a minor obstacle. This confidence coefficient is output as a continuous quantitative indicator to the subsequent data fusion stage to guide the filtering algorithm in dynamically adjusting the trust weights for the current observation data.

[0050] In step S5, the specific implementation process of state prediction compensation and data fusion solution includes the following steps: S501, Construct an adaptive Kalman filter model and dynamically adjust the observation noise weights. The industrial computer initializes the state space of the Kalman filter in memory, setting the coordinate deviation of the robot's end effector center point in three-dimensional space as the system state vector. In traditional Kalman filtering, the observation noise variance matrix is ​​usually set to a constant empirical value. In this embodiment, to address non-Gaussian contact noise caused by minor obstacle interference, the system dynamically reduces and adjusts the confidence weights of the current observation data based on the abnormal contact confidence coefficients calculated in the preceding steps.

[0051] In principle, the core mechanism of Kalman filtering lies in allocating confidence levels by balancing the variance between model predictions and actual observations. When the variance of observation noise is amplified, the filter reduces its dependence on the current observation value and instead tends to trust the theoretical prediction value more. When the confidence coefficient for anomalous contact increases, it indicates a higher probability that the current observation data is interfered with by foreign objects such as small splashes. Industrial computers multiply the basic observation noise matrix by a dynamic scaling factor to amplify the observation noise variance setpoint for the current observation period in real time.

[0052] As a preferred approach, the dynamic scaling factor can be set to 1 plus the product of the abnormal contact confidence coefficient and the preset scaling gain, where the scaling gain can be set to a value between 10 and 50. This operation reduces the trust weight of contaminated observation data in data fusion calculations, which is beneficial for maintaining the stability of state estimation. For the specific algorithm steps of the iterative update of the covariance matrix in the adaptive Kalman filter model, those skilled in the art can refer to the standard optimal estimation algorithm theory for implementation; its recursive process is a well-known technique in this field and will not be elaborated here.

[0053] S502, combining heterogeneous characteristics to calculate the micro-sideslip prediction compensation vector. When the robot's end effector makes physical contact with the calibration block in a set posture, due to the tool flange assembly clearance and the inherent flexibility of the material, the end effector will experience micro-lateral slippage while under normal compression. Without intervention, this nonlinear sideslip displacement can easily be introduced into the observation data, leading to coordinate calibration errors. To compensate for this physical error, the industrial computer extracts the instantaneous shear force component output in real time from the six-dimensional force sensor. Combining the obtained equivalent lateral compliance matrix and the generated abnormal contact confidence coefficient, the industrial computer calculates the micro-sideslip prediction compensation vector. The specific calculation formula is as follows: ; In the formula, This represents the microscopic sideslip prediction compensation vector, whose physical dimension is spatial displacement. Indicates the confidence coefficient for abnormal contact; Represents the equivalent lateral compliance matrix; It represents the instantaneous shear force component, which includes the lateral force vectors corresponding to the X and Y axes of the flange coordinate system.

[0054] The physical basis for introducing the abnormal contact confidence coefficient as an adjustment factor is that when contact occurs on a normal smooth surface, the coefficient approaches zero, and the system trusts the theoretical lateral slip amount calculated based on the compliance matrix. When the tool impacts a small protrusion, causing the abnormal contact confidence coefficient to increase, the original contact force state is disrupted, and the compliance mechanics mapping model based on the linear assumption is no longer accurate. Accordingly, the system adaptively weakens the intervention amplitude of the sideslip compensation to reduce the risk of filter divergence caused by overcompensation.

[0055] S503, the micro-sideslip prediction compensation vector is superimposed as a physical constraint compensation term onto the prior state estimate. After solving for the compensation vector, the industrial computer introduces it into the Kalman filter prediction update stage. The industrial computer establishes the prior state estimation equation including the physical constraint compensation term, the specific expression of which is as follows: ; In the formula, Indicates the current discrete time step The prior state estimation vector, i.e., the coordinate deviation prediction value that does not include the correction of the current observation data; Indicates the previous discrete time step The posterior state estimation vector is the optimal coordinate deviation estimate output from the previous iteration. This represents the microscopic sideslip prediction compensation vector.

[0056] Through this prior state estimation equation, the system directly transforms the lateral slip derived from physical forces into displacement compensation for coordinate deviation. This process injects deterministic physical and mechanical prior knowledge into a purely mathematical data filtering framework, allowing the filter to compensate for the impact of slip errors in the state prediction stage before receiving observation data containing slip errors. Subsequently, the industrial computer combines the dynamically adjusted observation noise variance matrix with the sensor observation vector of the current period to perform posterior state update calculations. Specifically, the sensor observation vector is the end-effector position deviation observation value derived synchronously from the aforementioned extracted vertical distance data and the forward kinematics solution of the robot. After the update is completed, the system outputs the initial coordinate deviation after compensation and fusion calculation and stores it in memory for subsequent coordinate iteration optimization stages.

[0057] See attached document Figure 4 In step S6, the specific implementation process of coordinate iterative optimization includes the following steps: S601, Initial spatial compensation is performed using the initial coordinate deviation. After obtaining the initial coordinate deviation output in step S5, the industrial computer uses it as a spatial displacement compensation term and adds it to the robot's current tool coordinate system parameter matrix. The system calls the robot's inverse kinematics algorithm to map this spatial displacement compensation amount into the target compensation angles of the six joint motors and sends them to each joint servo driver to perform position fine-tuning. This operation helps to make the physical position of the robot's end effector approach the true calibration reference point, thereby providing a more accurate initial reference point for subsequent multi-pose observation.

[0058] S602, planning multi-pose repositioning and acquiring multiple sets of observation features. Since single-pose contact measurements are easily affected by local surface topography and it is difficult to completely decouple the three-axis coordinate errors in three-dimensional space mathematically, the system needs to perform repeated data acquisition in multiple poses. The industrial computer plans at least three different spatial approach poses above the calibration block, using the current estimated contact point as the spatial center. As a preferred approach, the yaw and pitch angle offsets of these spatial approach poses relative to the surface normal of the calibration block can be set between 5 and 15 degrees. The industrial computer drives the robot to sequentially re-execute the aforementioned force approximation and fusion calculation process in these planned poses, obtaining multiple sets of independent observation residual vectors to form a multivariate observation dataset.

[0059] S603, Constructing a Nonlinear Objective Function and Performing Iterative Solving. Based on the acquired multivariate observation dataset, the industrial computer constructs a nonlinear least-squares objective function with the tool center point coordinates as the independent variable. To prevent non-physical surface penetration errors during iterative optimization, the system introduces the contact force normal vector as a geometric penalty term into the objective function. This objective function aims to minimize the sum of squared spatial residuals under all observed poses, as well as the projected inner product of the coordinate iteration update direction and the contact force normal vector, thereby constraining the gradient direction of the coordinate iteration update to tend to be orthogonal to the calibration plane normal. To solve this nonlinear optimization problem, the industrial computer uses the Levenberg-Marquardt algorithm for numerical iteration. In principle, this algorithm is essentially an optimization method based on the trust region. When the current solution is far from the optimal solution, it behaves as a gradient descent method, while when it is close to the optimal solution, it behaves as a Gauss-Newton method. By dynamically adjusting the damping factor, this algorithm can ensure a relatively fast convergence speed while avoiding the singularity of the Jacobian matrix. Its core coordinate iteration update formula is as follows: ; In the formula, Indicates the first The coordinate vector of the tool center point in the next iteration; Indicates the first The coordinate vector of the tool center point in the next iteration; The Jacobian matrix representing the residual function with respect to the coordinates of the tool's center point; The matrix representing the transpose of the Jacobian matrix; This represents the damping coefficient, used to adjust the iteration step size and search direction. Its value can be set from 0.01 to 0.1. Represents the identity matrix; This represents the observation residual vector under each current attitude. In this embodiment, the vector is specifically derived from the basic residual obtained by subtracting the actual observation position of the sensor from the theoretical position in the current coordinate system, and is jointly extended by the orthogonal penalty residual formed by the dot product of the contact force normal vector and the displacement update step size. This represents the index of the current discrete iteration count.

[0060] For the specific partial derivative calculation of the Jacobian matrix and the dynamic update strategy of the damping coefficient, those skilled in the art can use conventional nonlinear optimization algorithm libraries to implement them. The underlying numerical calculation logic is a well-known technology in this field and will not be elaborated here.

[0061] S604, Evaluate the convergence state of the iteration and output the final calibration parameters. After each iteration calculation, the industrial computer needs to perform a convergence evaluation on the current system state to avoid the algorithm getting stuck in an infinite loop. The industrial computer calculates the Euclidean norm of the current iteration step size and the sum of squared residuals of the objective function. When the Euclidean norm of the iteration step size is less than a preset step size threshold, or the sum of squared residuals is less than a preset residual threshold, the system determines that the iteration process has reached a convergence state. In this embodiment, the step size threshold can be set to 0.001 mm to 0.005 mm, and the residual threshold can be set to 0.01 square millimeters to 0.05 square millimeters. In addition, the system sets a maximum number of iterations as the termination boundary condition for the algorithm to exit, and the maximum number of iterations can be set to 30 to 50.

[0062] If any of the above convergence conditions are met, the industrial computer extracts the current tool center point coordinate vector and uses it as the final tool center point calibration coordinates. If the convergence conditions are not met and the maximum number of iterations has not been reached, the system will increment the discrete iteration count index and return to the nonlinear objective function update stage to continue execution. Finally, the system writes the final tool center point calibration coordinates into the system parameter storage area of ​​the robot controller via local area network communication, so that they can be called upon in subsequent actual production and processing tasks to complete the automatic calibration task of the tool center point.

[0063] Specific application examples: Application scenarios and physical parameter initialization: An automated welding production line for automotive chassis uses an ABB IRB 2600 industrial robot to perform MIG / MAG gas shielded welding. After continuous operation, the tool center point of the welding torch at the robot's end effector experienced an initial physical offset of approximately 2.5 mm. The industrial computer then executes the ABB welding robot tool coordinate precise and rapid repositioning method of this invention.

[0064] In step S1, the industrial computer obtains the equivalent lateral compliance matrix of the current end-effector. The worst-case response time of the industrial computer measurement system under current hardware configuration and communication link conditions. The time is 0.012 seconds. The industrial computer defines and inputs the maximum permissible elastic deformation limit based on the tool material properties. It is 1.5 mm.

[0065] Heterogeneous data alignment and dual security envelope control: In steps S2 and S3, the robot approaches the calibration block in a preset posture. The vision sensor and the six-dimensional force sensor simultaneously acquire data. The industrial computer executes a dynamic time warping algorithm to align the sequence in time phase. The specific expression for the cumulative distance of the current node in the cumulative distance matrix is ​​as follows: ; The industrial computer iterates through the computations to the endpoint and extracts the cumulative distance to that endpoint as the optimal normalization value. .

[0066] The industrial computer calculates the equivalent lateral stiffness parameters based on the extracted optimal normalization time value: ; Meanwhile, the industrial computer constructs a dynamic target approach velocity decay model to calculate the dynamic target approach velocity: ; The industrial computer converts the equivalent lateral stiffness parameters and the dynamic target approach speed into instructions and sends them to the robot controller.

[0067] Combined with appendix Figure 3 illustrate: See attached document Figure 3 The horizontal axis represents the approximation and contact time (seconds), and the vertical axis represents the end-contact normal force (Newtons). Under conventional calibration control without the present invention (as shown in the attached...), Figure 3 The traditional contact force curve (shown as a solid hollow square) is limited by system delay. When physical contact occurs at approximately 0.4 seconds, the robot cannot brake in time, resulting in a peak normal force of approximately 145 Newtons, which poses a risk of damaging the equipment.

[0068] After adopting the dual security envelope control of the present invention (as shown in the appendix) Figure 3 The double safety envelope force curve (shown as a dashed hollow triangle) demonstrates that, because the system outputs the reduced equivalent lateral stiffness parameter and dynamic target approach velocity in real time based on the optimal regularization iteration value, the robot exhibits dynamically adjusted lateral compliance and longitudinal velocity limiting characteristics when approaching the calibration block. During the contact phase, the force curve transitions smoothly, and the peak normal force is limited to approximately 38 Newtons, achieving safe and compliant contact.

[0069] Contact noise identification and condition prediction compensation: In steps S4 and S5, when the normal force fed back by the six-dimensional force sensor reaches a preset trigger threshold, the tool tip contacts a 0.5 mm rigid spatter on the calibration surface. An industrial computer extracts the visual spatial surface gradient of the contact point area. and the peak value of force control normal coefficient Generate anomaly contact confidence coefficients based on logical relationships: ; The current abnormal contact confidence coefficient is calculated. The industrial computer determines that the current contact is being interfered with by a minor obstacle. This coefficient is then incorporated into the adaptive Kalman filter model to amplify the observation noise variance, thereby reducing the weight of trust in the current observation data.

[0070] Meanwhile, to address the physical sideslip caused by contact with foreign objects, industrial applications incorporate an equivalent lateral flexibility matrix. and the instantaneous shear force component output by the six-dimensional force sensor. Calculate the micro-scale sideslip prediction compensation vector: ; The industrial computer superimposes the micro-slip prediction compensation vector as a physical constraint compensation term into the prior state estimation equation: ; This is done in advance to offset the observation error caused by nonlinear lateral slip and output the initial coordinate deviation.

[0071] Coordinate Iteration Optimization and Verification: In step S6, the industrial computer performs multi-pose repositioning based on the initial coordinate deviation, and collects multiple sets of observation features to form the observation residual vector. Subsequently, a nonlinear objective function is constructed, and the Levenberg-Marquardt algorithm is used for residual approximation. The formula for coordinate iterative update is as follows: ; Combined with appendix Figure 2 illustrate: See attached document Figure 2 The x-axis represents the number of iterations (times), and the y-axis represents the calibration residual (millimeters) of the tool center point coordinates. In traditional calibration methods that do not eliminate splash interference and ignore microscopic sideslip (as shown in the appendix)... Figure 2 The traditional calibration method curve (shown as a solid hollow circle) introduces physical errors into the Jacobian matrix operation, causing the calibration residual of the tool's center point coordinates to fail to converge after decreasing to approximately 0.85 mm.

[0072] Applying the method of the present invention (as shown in the appendix) Figure 2 The repositioning method curve of this invention (shown by the dashed asterisk) effectively identifies contact noise and superimposes a microscopic sideslip prediction compensation vector during the data fusion and calculation step, eliminating the interference of observation errors. The curve shows that by the 4th to 5th iteration, the tool center point coordinate calibration residual rapidly converges to an allowable error range of approximately 0.04 mm. The industrial computer then stops iterating and outputs the final tool center point calibration coordinates to complete the repositioning update.

Claims

1. A method for precise and rapid tool coordinate repositioning of an ABB welding robot, applied to a hardware environment including an industrial computer, robot controller, vision sensor, and six-dimensional force sensor, characterized in that... include: Obtain the equivalent lateral flexibility matrix, maximum allowable elastic deformation limit, and estimated contact point of the robot's current end-effector, and determine the system's overall worst-case response time. Within a set sliding time window, visual depth feature sequences and force-controlled normal force sequences are simultaneously acquired. The dynamic time warping algorithm is used to align the visual depth feature sequences and the force-controlled normal force sequences in time phase and extract the optimal warping generation value. Based on the optimal normalization generation value, the equivalent lateral stiffness parameter and the dynamic target approximation velocity are calculated in real time, and control commands are sent to the robot controller to execute the action. When the normal force fed back by the six-dimensional force sensor reaches the preset normal force trigger threshold, the visual space surface gradient and the peak value of the force control normal derivative are extracted to generate an abnormal contact confidence coefficient. An adaptive Kalman filter model is constructed, and a micro-sideslip prediction compensation vector is calculated by combining the instantaneous shear force component output by the six-dimensional force sensor, the equivalent lateral compliance matrix, and the abnormal contact confidence coefficient. The micro-sideslip prediction compensation vector is then superimposed on the prior state estimation equation of the adaptive Kalman filter model to solve for the initial coordinate deviation. A nonlinear optimization algorithm is called to approximate the coordinate residual vector. When the coordinate residual vector converges, the repositioning and update of the robot tool coordinate system is completed.

2. The method for precise and rapid repositioning of ABB welding robot tool coordinates according to claim 1, characterized in that, The process of obtaining the equivalent lateral flexibility matrix and maximum permissible elastic deformation limit of the robot's current end-effector, and determining the system's overall worst-case response time, includes: A known gradient lateral static test force is applied to the end effector, and the lateral force components in the flange coordinate system of the robot and the actual micro displacement vector generated by the end effector are recorded simultaneously. The equivalent lateral compliance matrix is ​​obtained by linear fitting calculation using the least squares method. The material yield strength parameters, cross-sectional moment of inertia and assembly extension length of the end tool are obtained. The critical yield displacement value is calculated using the cantilever beam stress model. The maximum allowable elastic deformation limit is calculated by multiplying the critical yield displacement value by the set safety margin factor. Based on the system's internal timestamp comparison mechanism, the processing time of the visual sensor acquiring three-dimensional point cloud data and transmitting it to the industrial computer, the delay time of the six-dimensional force sensor feeding back force data and being calculated after noise reduction filtering, the maximum delay time of network communication jitter, the single position control cycle time, and the electromechanical response time are calculated and recorded respectively. The summation of these values ​​yields the system's overall worst-case response time.

3. The method for precise and rapid repositioning of ABB welding robot tool coordinates according to claim 1, characterized in that, The process of extracting the optimal normalization generation value includes: Calculate the rate of change of the visual depth feature sequence and the first derivative of the force control normal force sequence, and use linear mapping normalization to obtain a normalized visual feature sequence and a normalized force control feature sequence with a unified value range. Construct a two-dimensional cumulative distance matrix, initialize the starting node of the two-dimensional cumulative distance matrix to zero, and initialize the remaining nodes in the first row and first column to infinity; The two-dimensional cumulative distance matrix is ​​traversed in ascending order of time index. Dynamic programming is used to recursively calculate the cumulative distance of the endpoint node of the two-dimensional cumulative distance matrix by combining the Euclidean distance of local differences with the minimum cumulative distance value among the adjacent predecessor nodes. The cumulative distance of the endpoint node of the two-dimensional cumulative distance matrix is ​​extracted as the optimal normalization generation value.

4. The method for precise and rapid repositioning of ABB welding robot tool coordinates according to claim 1, characterized in that, The real-time calculation of equivalent lateral stiffness parameters and dynamic target approximation velocity based on the optimal normalization value includes: A basic stiffness adjustment coefficient and a cost weight penalty coefficient are introduced. The inverse matrix of the equivalent lateral flexibility matrix is ​​adjusted based on the optimal normalization generation value. The equivalent lateral stiffness parameter is output by combining the lower stiffness limit and the upper stiffness limit value through a saturation limiting function constraint. Introducing a velocity decay rate constant, the optimal regularization value is substituted into the exponential decay function, and combined with the ratio of the maximum allowable elastic deformation limit to the worst-case response time of the system, the smoothly decreasing dynamic target approximation velocity is calculated and output.

5. The method for precise and rapid repositioning of ABB welding robot tool coordinates according to claim 4, characterized in that, Sending control commands to the robot controller to execute actions includes: The force state currently fed back by the six-dimensional force sensor, the calculated equivalent lateral stiffness parameter, the preset damping coefficient and the inertial mass parameter are substituted into the damping mass spring model to calculate the target position correction vector for the next control cycle. The target position correction vector and the dynamic target approximation velocity are packaged and encapsulated into a fixed-length control command data frame containing control timestamps and spatial coordinate information; The fixed-length control command data frame is sent to the underlying communication process within the robot controller via the User Datagram Protocol (UDP) in a local area network environment, and then transmitted to the servo drivers of each joint via the internal bus.

6. The method for precise and rapid repositioning of ABB welding robot tool coordinates according to claim 1, characterized in that, When the normal force fed back by the six-dimensional force sensor reaches a preset normal force trigger threshold, the peak values ​​of the visual space surface gradient and the force control normal derivative are extracted, including: The preset normal force trigger threshold is set in advance based on the amplitude of the base noise under no-load static state; When the normal force reaches the preset normal force trigger threshold, the data freezing mechanism is triggered, and the three-dimensional point cloud data collected by the visual sensor and the force-controlled normal force sequence within the currently set sliding time window are extracted and stored in the static buffer memory. Extract a set of local neighborhood point clouds centered on the estimated contact point from the frozen 3D point cloud data, and call the principal component analysis algorithm to calculate the normal vector variance as the visual space surface gradient; Perform a first-order difference operation on the frozen force-controlled normal force sequence, iterate through the difference sequence to find the maximum absolute value, and use the maximum absolute value as the peak value of the force-controlled normal derivative.

7. The method for precise and rapid repositioning of ABB welding robot tool coordinates according to claim 6, characterized in that, The generation of the abnormal contact confidence coefficient includes: We construct a cross-validation logic based on the physical consistency of heterogeneous data, and introduce a sensitivity adjustment coefficient and a very small positive number to prevent abnormal denominator calculation. The ratio term is calculated by multiplying the peak value of the force control normal derivative by the sensitivity adjustment coefficient as the numerator, and summing the visual spatial surface gradient with the minimum positive number as the denominator. The ratio term is subjected to a negative exponentiation with the natural constant as the base, and the difference between 1 and the result of the negative exponentiation is calculated to generate the abnormal contact confidence coefficient.

8. The method for precise and rapid repositioning of tool coordinates of an ABB welding robot according to claim 1, characterized in that, The calculation of the micro-sideslip prediction compensation vector by combining the instantaneous shear force component output by the six-dimensional force sensor, the equivalent lateral compliance matrix, and the abnormal contact confidence coefficient includes: The coordinate deviation of the end-effector center point in three-dimensional space is set as the system state vector. The basic observation noise matrix of the adaptive Kalman filter model is multiplied by a dynamic scaling factor that includes the abnormal contact confidence coefficient and the preset scaling gain, thereby amplifying the observation noise variance setting value of the current observation period. The instantaneous shear force component is the component of the lateral force vector in the X and Y directions of the flange coordinate system of the robot, which is output in real time by the six-dimensional force sensor. Calculate the difference between 1 and the confidence coefficient of the abnormal contact, and multiply the obtained difference, the equivalent lateral flexibility matrix and the instantaneous shear force component to obtain the microscopic sideslip prediction compensation vector.

9. The method for precise and rapid repositioning of tool coordinates of an ABB welding robot according to claim 1, characterized in that, The step of using a nonlinear optimization algorithm to approximate the coordinate residual vector includes: The initial coordinate deviation is mapped to the target compensation angle of the robot's six joint motors and sent to the joint servo driver to perform preliminary spatial compensation. Using the estimated contact point as the center of the space sphere, plan multiple spatial approach attitudes with offsets in yaw and pitch angles, and sequentially re-execute the force approximation and fusion calculation to obtain a multivariate observation dataset; A nonlinear least squares objective function is constructed with the coordinates of the tool's center point as the independent variable. The contact force normal vector is introduced into the nonlinear least squares objective function as a geometric penalty term, and the gradient direction of the coordinate iteration update is constrained to be orthogonal to the normal of the preset calibration plane.

10. A method for precise and rapid repositioning of tool coordinates of an ABB welding robot according to claim 9, characterized in that, The step of completing the repositioning update of the robot tool coordinate system when the coordinate residual vector converges includes: The Levenburg Marquardt algorithm is used for numerical iteration, and the damping coefficient is dynamically adjusted to calculate the observation residual vector and displacement update step size. Calculate the sum of squared residuals between the Euclidean norm of the current displacement update step and the nonlinear least squares objective function; When the Euclidean norm is less than a preset step size threshold or the sum of squared residuals is less than a preset residual threshold, it is determined that the coordinate residual vector has reached a convergence state. The current tool center point coordinate vector is extracted as the final tool center point calibration coordinate, and the final tool center point calibration coordinate is written into the system parameter storage area of ​​the robot controller.

Images