An adaptive repositioning method for ABB welding robot tool coordinates

Abstract

This invention relates to the field of robot control technology and discloses an adaptive repositioning method for ABB welding robots, comprising: constructing a digital twin system for the welding robot to achieve real-time mapping and synchronization of virtual and real data; collecting multi-source sensor data to train a long short-term memory neural network to predict tool end wear online; in the repositioning process, subtracting the actual and ideal end positions and deducting the predicted wear, separating the pure spatial position error as the comprehensive coordinate deviation, applying an iterative learning control algorithm combined with a linear decay mechanism to dynamically update the coordinate correction parameters, and issuing the method to the real equipment for execution after verification by digital twin simulation; during normal operation, predicting the wear and calculating the load change rate in real time, and when the rate reaches a trigger threshold, intercepting the normal instructions through a high-priority interrupt and inserting the adaptive repositioning program. This invention achieves effective decoupling of mechanical spatial error and material loss, improving repositioning accuracy and continuous operation efficiency.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to an adaptive repositioning method for tool coordinates of an ABB welding robot. Background Technology

[0002] Currently, industrial robots are widely used in welding lines in county-level manufacturing industries, where there is a high variety of products and small batch production. With the increasing demand for flexible manufacturing, welding torches and other tools need to frequently handle load changes caused by different workpieces. This dynamic production environment places high demands on the real-time accuracy of the robot's tool coordinate system.

[0003] Regarding the aforementioned issues, existing robot relocalization typically relies on manual teaching or fixed calibration blocks. The operator drives the robot's end effector to touch a reference point, and the controller's built-in geometric calculation program determines the coordinate deviation. During work breaks, the system executes a preset static calibration program, which is then manually observed and corrected based on experience, with parameters entered into the system.

[0004] This traditional approach cannot perceive the evolution of the physical state at the end of the tool. The welding torch will experience material loss during continuous operation, and stress changes in the end pipeline will cause positional shifts. Existing solutions cannot accurately decouple mechanical positioning errors from physical wear, lack a mechanism for reusing historical correction data, and each calibration must start from zero. The downtime for testing is too long, production continuity is hindered, and equipment operation is highly dependent on the professional debugging skills of the operators.

[0005] Therefore, this invention provides an adaptive relocation method for tool coordinates of ABB welding robots to address the shortcomings of existing technologies. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an adaptive repositioning method for ABB welding robot tool coordinates. This method solves the problem that existing welding robot tool coordinate repositioning processes cannot dynamically sense and decouple tool wear and load variation errors, leading to frequent machine shutdowns for manual calibration with low efficiency and accuracy.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An ABB welding robot tool coordinate adaptive relocation method includes the following steps: Construct a digital twin system for ABB welding robots, establish a mapping channel between real equipment and virtual simulation models, and realize real-time synchronous interaction of data; Collect multi-source state data during the operation process, train a tool wear prediction model based on a long short-term memory neural network based on the multi-source state data, and output the predicted tool end wear data. Based on the actual end position data and the end wear data of the prediction tool, the comprehensive coordinate deviation is calculated, historical deviation data is extracted, and the coordinate correction parameters are updated by applying an iterative learning control algorithm, and the repositioning process is executed. The system acquires current load data in real time and calculates the load change rate. It then calls the tool wear prediction model to obtain wear increment parameters. When the load change rate or wear increment parameters meet the relocation trigger conditions set based on the system error tolerance range, the system executes a dynamic adaptive compensation process.

[0008] By employing the above technical solution, a high-precision mapping of the physical properties of an entity is established in virtual space through the combination of a digital twin architecture and a long short-term memory neural network. This allows for the extraction of nonlinear correlation features between operation time, electrical parameters, and tool wear. An iterative learning control algorithm is introduced to dynamically adjust the correction step size using historical execution deviations, decoupling pure mechanical spatial errors from material physical losses. Therefore, this approach effectively suppresses end-effector pose drift caused by continuous high-intensity operations, avoids frequent downtime for calibration, and improves system position accuracy and adaptive compensation efficiency under various load conditions.

[0009] Preferably, the specific steps for constructing the ABB welding robot digital twin system and realizing real-time synchronous data interaction are as follows: Collect actual geometric dimension data of the real equipment body and tool end, and build basic component entities one by one in the 3D design software environment to synthesize a 3D geometric assembly model; Assign material properties and mass parameters to each component of the three-dimensional geometric assembly model, set relative kinematic constraints, and construct a dynamic simulation model corresponding to the real equipment. Within the underlying industrial control communication network architecture, data transmission nodes based on an open platform communication unified architecture are deployed to establish a bidirectional data transmission link between real state parameters and virtual simulation models, maintaining real-time updates of various state parameters.

[0010] By adopting the above technical solution, a standard communication link is established in the underlying network to eliminate cross-platform data exchange delays, ensuring that the three-dimensional geometric and dynamic boundary conditions of the virtual simulation benchmark are strictly consistent with the physical entity, and providing a reliable data foundation for subsequent error calculation.

[0011] Preferably, the specific steps for constructing the dynamic simulation model corresponding to the real device are as follows: Define the rules for releasing and locking the degrees of freedom between connecting components, establish a system-level multi-rigid-body topology hierarchy, and introduce a gravitational acceleration vector into the virtual environment to simulate the action of an external constant force field; Numerical integration and algebraic solutions are performed based on the motion equations of the robot system. By combining the robot joint angles, joint angular velocities and joint angular accelerations, inertia matrix, centrifugal force and Coriolis force matrices, gravitational torque vector and joint driving torque vector, various kinematic state parameters of the end effector are derived. The calculated time series parameters are compared with the actual sensor measurement results fed back by the real equipment, and the rotational inertia matrix values ​​of each link are corrected in reverse until the calculated deviation of the derived parameters is not greater than the preset dynamic simulation error threshold. The dynamic simulation error threshold represents the maximum allowable deviation between the calculated and derived parameters and the sensor measurement results.

[0012] By employing the above technical solution, and by comprehensively processing the time-series changes of joint kinematic variables, combined with the system's inertial distribution, centrifugal force, Coriolis force, and gravity, the rotational inertia parameters are calibrated in reverse by comparing them with real sensor feedback data. This eliminates pose calculation errors caused by rigid body dynamic coupling under complex spatial motion conditions, and improves the accuracy of end-effector force state measurement in the virtual environment.

[0013] Preferably, the specific steps for training the tool wear prediction model based on a long short-term memory neural network are as follows: The system continuously collects operation time, welding current, and welding voltage values ​​as input feature parameters, and simultaneously acquires actual tool end wear data measured by a laser rangefinder as the target true value to construct a tool wear database. A long short-term memory neural network topology with three input layers that receive operation time, welding current and welding voltage values ​​respectively is constructed, and an output layer is constructed after the hidden layer to generate data on the wear of the tool tip. Define a model training loss function to guide backpropagation weight updates, perform iterative model training, and solidify and encapsulate it into an independent and callable tool wear prediction model after verifying that the prediction error meets the accuracy standard. The accuracy standard represents the prediction error limit that meets the requirements for online calling of the network model.

[0014] By adopting the above technical solution and using real measurement data to supervise network training, the time series dependency relationship between operation time and arc energy parameters is extracted, enabling the system to independently infer the physical loss size of the end material online.

[0015] Preferably, the specific steps for defining the model training loss function and performing iterative model training are as follows: The mean squared error function is used as the model training loss function to calculate the overall deviation between the predicted tool end wear data calculated by the forward propagation of the network model in the training batch and the actual measured tool end wear data, and to generate the mean squared error loss value. An adaptive moment estimation optimizer is used to dynamically adjust the learning step size of each network parameter based on the first-order and second-order moment estimates of the gradient. The data samples in the training set are input into the neural network in batches and the mean squared error loss value is calculated. The connection weights and bias values ​​between network layers are updated iteratively through the backpropagation algorithm.

[0016] By adopting the above technical solution, the overall deviation is quantified using mean square error, and the learning step size is dynamically adjusted according to the first and second moments of the gradient using an optimization algorithm, thereby accelerating the convergence speed of the network's forward and backward propagation and avoiding the model from getting trapped in local optima.

[0017] Preferably, the specific steps for calculating the comprehensive coordinate deviation are as follows: Retrieve the preset welding process program file stored inside the robot system and extract the ideal end position data corresponding to the current process step from it; The actual end position data and the ideal end position data are subjected to coordinate system alignment and filtering processing to ensure that the three-dimensional coordinate values ​​involved in the calculation are in a unified reference coordinate system. The actual end position data is subtracted from the ideal end position data, and the predicted tool end wear data output by the tool wear prediction model is subtracted from the difference result. The pure spatial position error that needs to be eliminated by kinematic compensation is separated as the comprehensive coordinate deviation.

[0018] By adopting the above technical solution, the spatial position difference between the ideal reference point and the actual detection position is calculated under a unified coordinate system, and the end-material loss dimension output by the network model is directly subtracted from this spatial position difference. This calculation method separates the pure spatial position error that needs to be eliminated by mechanical motion compensation, avoiding the misjudgment of material wear error as mechanical positioning error, which would lead to overcompensation of the system.

[0019] Preferably, the specific steps for updating the coordinate correction parameters using the iterative learning control algorithm are as follows: Based on the pre-set initial learning rate and linear decay mechanism, combined with the current iteration round and decay coefficient, the learning rate parameter is gradually reduced to calculate the current adaptive learning rate. The calculated adaptive learning rate is multiplied by the comprehensive coordinate deviation of the current iteration cycle, and the product is added to the current coordinate correction parameters to generate the coordinate correction parameters required for the next iteration. The updated coordinate correction parameters are input into the digital twin model for virtual environment simulation. After confirmation, the parameters are sent to the real device to perform position adjustment and the actual end position data is re-detected to determine whether the deviation has converged to the standard range. The standard range is the allowable repositioning accuracy range that meets the target reference point position requirements.

[0020] By adopting the above technical solution, the learning rate parameter in the iterative control loop is set to decrease linearly with each iteration. The product of the decreasing adaptive learning rate and the comprehensive coordinate deviation of the current iteration is accumulated and added to the existing correction parameter. This strategy suppresses the oscillation overshoot phenomenon during the terminal approach to the target reference point, ensuring that the position error converges smoothly and quickly to the set accuracy range.

[0021] Preferably, the specific steps for obtaining the current load data and calculating the load change rate are as follows: The current load data characterizing the welding torch body and its auxiliary pipelines under the current working posture is extracted through data preprocessing operations. Retrieve the initial load data obtained during the most recent relocation calibration from the stored records; Calculate the absolute difference between the current load data and the initial load data, and then perform a percentage operation on the ratio of the absolute difference to the initial load data to generate a percentage value. Use the percentage value as the load change rate.

[0022] By adopting the above technical solution, the force fluctuation caused by the change in working conditions is quantified by calculating the percentage of the absolute difference between the current load data and the initial load data relative to the initial load data, providing a mechanical index reference for the system to determine whether to perform dynamic compensation.

[0023] Preferably, the specific steps for concurrently calling the tool wear prediction model to obtain wear increment parameters are as follows: The real-time collected operation time, welding current and welding voltage values ​​are used as input features and synchronously fed into the tool wear prediction model running concurrently in the background. The tool wear prediction model calculates and outputs the predicted tool end wear data in real time at the current moment. Retrieve the baseline wear data from the most recent repositioning calibration, calculate the difference between the current predicted tool end wear data and the baseline wear data, and generate a wear increment parameter characterizing the current wear level of the tool.

[0024] By adopting the above technical solution, the real-time wear value is calculated concurrently in the background and the difference between it and the initial reference state is extracted, so as to realize continuous non-interventional monitoring of tool status and ensure the timeliness of wear status early warning.

[0025] Preferably, the specific steps of executing the dynamic adaptive compensation process are as follows: The load change rate and wear increment parameters obtained in real time are compared with the preset relocation trigger threshold in the logic comparison register. The relocation trigger threshold is used to define the allowable range of system error. When the comparison result shows that the wear increment parameter or load change rate is greater than or equal to the relocation trigger threshold, a high-priority interrupt request is sent to the main controller to intercept the currently executing routine welding operation command and control the real equipment to pause its operation. The adaptive repositioning procedure is seamlessly inserted, and the tool end is guided to complete dynamic calibration based on the latest coordinate correction parameters output by the iterative learning control algorithm. After verification, the interruption request status is revoked and the normal welding operation command is restored to maintain continuous operation.

[0026] By adopting the above technical solution, an interception and execution mechanism based on high-priority interruption is established. When the monitoring index exceeds the allowable threshold, the correction instruction closed loop is initiated. After the compensation action is completed, the original operation process is immediately restored, realizing adaptive trajectory correction in non-stop state.

[0027] This invention provides an adaptive relocation method for tool coordinates of an ABB welding robot. It offers the following advantages: 1. This invention achieves real-time data interaction between the real equipment and the virtual simulation model by constructing a digital twin system, and uses a long short-term memory neural network model to predict tool end wear data. When calculating the comprehensive coordinate deviation, the difference between the actual end position and the ideal end position is calculated, and the predicted tool end wear data is directly subtracted from the result, separating the pure spatial position error. This calculation method, which decouples mechanical spatial error from material physical wear, avoids the problem of misjudging material wear error as mechanical positioning error, thus preventing overcompensation of the system and improving the accuracy of tool coordinate repositioning under multi-source error coupling conditions.

[0028] 2. This invention calculates the load change rate by acquiring current load data in real time and concurrently calls a tool wear prediction model to obtain wear increment parameters. When the load change rate or wear increment parameters meet the repositioning trigger conditions set based on the system error tolerance range, the system intercepts regular instructions by sending a high-priority interrupt request and executes a dynamic adaptive compensation process. This enables the welding robot to automatically sense and adapt to continuous tool wear and variable load conditions, automatically switching to a correction closed loop when the error exceeds the threshold, without requiring manual shutdown for measurement and recalibration, thus reducing the dependence of equipment operation on the operator's debugging skills.

[0029] 3. This invention applies an iterative learning control algorithm in the repositioning process. Based on extracted historical deviation data and a linear decay mechanism, an adaptive learning rate is calculated, and coordinate correction parameters are updated sequentially. The updated correction parameters are first verified through simulation in a digital twin model, and then deployed to the real equipment for adjustment. By reusing historical execution data and dynamically adjusting the correction step size, the oscillation overshoot phenomenon during the approach to the target reference point at the end is effectively suppressed, ensuring smooth and rapid convergence of position errors. This strategy accelerates the execution speed of adaptive compensation, reduces equipment downtime during continuous construction operations, and improves overall production efficiency. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the relocation system structure based on digital twin and iterative learning according to the present invention; Figure 2 This is the main flowchart of the tool coordinate adaptive relocation method of the present invention; Figure 3 A flowchart illustrating the construction of the digital twin system of the present invention; Figure 4 This is a flowchart of the wear prediction model training sub-process of the present invention; Figure 5 This is a flowchart of the iterative learning control relocation sub-process of the present invention; Figure 6 This is a flowchart of the dynamic adaptive compensation sub-process of the present invention; Figure 7 This is a curve comparing the convergence of iterative deviations during the tool coordinate repositioning process of the present invention. Figure 8 This is a comparison curve of the dynamic compensation response and positioning error of the present invention over time. Detailed Implementation

[0031] 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.

[0032] See attached document Figure 1This invention provides a system for implementing an adaptive relocation method for ABB welding robot tool coordinates based on digital twins and iterative learning. The system may include a digital twin modeling module, a virtual-real data interaction module, a wear prediction module, an iterative learning control module, a robot execution module, and a data acquisition module. The digital twin modeling module and the virtual-real data interaction module are interconnected for bidirectional communication. The virtual-real data interaction module is connected to the wear prediction module, the iterative learning control module, and the robot execution module, respectively. The iterative learning control module communicates with the robot execution module via a data transmission line using the RAPID language.

[0033] The data acquisition module is used to collect multi-source status data in the working environment. It integrates a laser rangefinder (measurement accuracy ≤ ±0.001mm), current and voltage sensors, and a load sensor. The laser rangefinder measures the dimensions of the tool tip and obtains the actual tool tip wear based on this data. The current and voltage sensors collect welding current and voltage values ​​in real time during the welding process. The load sensor monitors load changes on the tool tip and its associated pipelines. The data acquisition module preprocesses the collected dimensional, current, voltage, and load data before transmitting them to the various calculation modules of the system.

[0034] The digital twin modeling module, deployed on the server side, is used to construct a virtual system model corresponding to the real environment. Based on SolidWorks software, the module creates 3D geometric models of the robot body and the end effector (welding torch). The 3D geometric model is designed strictly according to the dimensional parameters of a real ABB welding robot. The module further uses ANSYS software to assign material properties and motion constraints to the 3D geometric model, constructing a corresponding dynamic simulation model. This dynamic simulation model simulates the force state and trajectory of the end effector during movement, outputting joint torque data and end effector acceleration data, providing a simulation data foundation for tool coordinate analysis.

[0035] The virtual-real data interaction module establishes a communication architecture based on the OPCUA protocol, responsible for building a real-time data transmission channel between the real system and the digital twin modeling module. The data transmission rate of the virtual-real data interaction module is set to be greater than or equal to 100Mbps, and the data synchronization latency is controlled within a range of less than or equal to 10ms. The virtual-real data interaction module realizes real-time mapping and status synchronization of multiple types of parameters. The synchronized data includes twelve types of parameters, including joint angle data, end-effector position data in three-dimensional space (specifically including six dimensions: X, Y, Z, A, B, and C), load data, tool end-effector wear data, welding current values, and welding voltage values.

[0036] The wear prediction module includes a prediction algorithm unit used to calculate the wear trend of the tool's end effector based on historical operation data and real-time status data. The module employs a Long Short-Term Memory (LSTM) neural network to build its model. The prediction model has a 3-dimensional input layer, receiving operation time, welding current, and welding voltage values. After forward propagation through a multi-layer neural network with 2 hidden layers (64 nodes per layer), the output layer (1-dimensional) generates the predicted wear amount of the tool's end effector. The wear prediction module obtains real-time monitored input variables through a virtual-real data interaction module and transmits the calculated predicted wear amount data to the iterative learning control module for compensation calculations.

[0037] The iterative learning control module, as the core control unit of the system, is used to generate and update coordinate correction commands. It receives actual end-effector position data, ideal end-effector position data, and predicted tool end-effector wear data to calculate the current comprehensive coordinate deviation. Based on the iterative learning control algorithm, the module extracts historical deviation data from the historical repositioning process and updates the current coordinate correction parameters. The updated coordinate correction parameters are first input to the digital twin modeling module for verification of the correction effect in a virtual scene. When the verified deviation meets the set accuracy range, the iterative learning control module converts it into the corresponding RAPID language command.

[0038] The robot execution module comprises the ABB welding robot body and its supporting systems, which are responsible for the actual work. The robot execution module receives RAPID language commands from the iterative learning control module, driving the motors of each joint of the robot body to adjust the spatial coordinates of the tool end effector. During welding operations, the robot execution module feeds back its real-time joint angle data, actual end effector position data, and load change data to the virtual-real data interaction module, providing real-time status feedback for the overall system's repositioning process and dynamic adaptive compensation.

[0039] See attached document Figure 2 This invention provides an adaptive relocation method for ABB welding robot tool coordinates based on digital twins and iterative learning. This method relies on the aforementioned system operation, and the overall process encompasses four main operational stages: model construction, algorithm training, deviation iterative optimization, and adaptive trigger compensation. The method may include the following specific steps.

[0040] Step S100: Construct an ABB welding robot digital twin system to achieve real-time synchronous interaction between the real equipment and the virtual model data. This step is executed through the digital twin modeling module and the virtual-real data interaction module. The system establishes a high-precision three-dimensional geometric model based on the actual dimensional parameters and material properties of the real ABB welding robot and tool end effector, and further constructs a multibody dynamics simulation model for stress analysis. The system configures underlying communication transmission protocol nodes to build a bidirectional data interaction architecture, establishing a real-time data mapping channel between the real state parameters and the virtual simulation model, enabling the synchronous transmission of various operating parameters, such as joint angle data, actual end effector position data, and load change data, between the real execution unit and the virtual computing server.

[0041] Step S200 involves collecting multi-source sensor status data and training a tool wear prediction model based on a long short-term memory neural network. This step is executed through the data acquisition module and the wear prediction module. Under set operational boundary conditions, the system continuously collects operational time, welding current, welding voltage, and corresponding actual tool end wear data to construct an initial wear database. Based on this wear database, a long short-term memory neural network topology is constructed, the number of multi-layer neural network nodes and the forward and backward propagation mechanisms are set, and the network weight data of the prediction algorithm is iteratively updated. Finally, the training process of the tool wear prediction model is completed, and the network model is encapsulated to output the predicted tool end wear data.

[0042] Step S300: Execute the tool coordinate repositioning process based on an iterative learning control algorithm to generate coordinate correction parameters and perform accuracy verification and calibration. This step is performed through the iterative learning control module. Upon initial repositioning trigger command, the system generates initial correction parameters and sends them to the lower-level control terminal, subsequently receiving the adjusted actual end-effector position data. The system incorporates the predicted tool end-effector wear data output in real time from the tool wear prediction model to calculate the comprehensive coordinate deviation between the actual and ideal end-effector positions. The system extracts historical deviation data from the historical repositioning process and applies the iterative learning control algorithm to adaptively calculate and update the current coordinate correction parameters. The updated coordinate correction parameters are then input into the digital twin modeling module for closed-loop verification until the verified comprehensive coordinate deviation meets the set convergence accuracy judgment conditions.

[0043] Step S400: Real-time monitoring of system status feedback data, calculation of status fluctuation rate, and triggering of dynamic adaptive compensation process. This step is executed by the robot execution module. During the operation of the regular welding procedure, the system continuously reads the current load data transmitted by the load sensor and concurrently calls the aforementioned tool wear prediction model to obtain wear increment parameters. The system calculates and compares the load change rate and wear increment parameters with the preset repositioning trigger threshold in real time. When the monitored status parameter data reaches or exceeds the set repositioning trigger condition, the system intercepts the currently running regular operation command and calls the corresponding repositioning program command to execute dynamic adaptive compensation calculation. After the dynamic adaptive compensation program is completed and the tool coordinates are updated, the regular production operation process resumes.

[0044] See attached document Figure 3 This invention provides a method for constructing a digital twin system of a real device and realizing real-time data interaction, which may include setting and executing operation procedures such as digital twin modeling and data synchronization. Step S100 specifically includes the following sub-steps: Step S101: Collect actual geometric parameters of the real equipment to establish a three-dimensional geometric model.

[0045] The digital twin modeling module receives the actual geometric dimensions of the real equipment body and the tool end from the data acquisition module. The system acquires the positioning reference dimensions of the real equipment base, the length and cross-sectional parameters of each stage of the connecting rod, and records the actual space occupation contour boundary of the working tool assembled on the end flange and the initial installation position deviation data.

[0046] The digital twin modeling module constructs basic component entities one by one in a 3D design software environment based on the acquired geometric dimension data. The system performs 3D contour drawing for components such as the rotary joint shell, drive shaft components, robotic arm linkage body, and end-effector flange of the real equipment, generating independent component models with the same appearance dimensions as the real hardware structure, and creating corresponding 3D solid features for the tool end effector.

[0047] The digital twin modeling module extracts the factory assembly relationship data of the real equipment to assemble and positionally constrain all independent component models. The system defines the overlap, parallelism, and concentric axis relationships between the mating surfaces of each component model, synthesizing a complete 3D geometric assembly model of the equipment. The system performs moving part interference checks and overall size measurement verification on the synthesized 3D geometric assembly model, adjusting the model's detailed features to ensure that the global dimensional deviation of the final 3D geometric model does not exceed a preset geometric accuracy threshold (specifically, geometric accuracy ≤ ±0.003mm).

[0048] Step S102: Construct a dynamic simulation model and calculate the end effector kinematic parameters of the tool.

[0049] After acquiring the assembled 3D geometric model, the digital twin modeling module assigns material properties and mass parameters to each component. The system consults the material configuration database of the real equipment and assigns corresponding material density values ​​to each metal component and non-metallic shell model in the 3D geometric model. Based on the volume and density parameters, the system calculates the absolute mass, spatial centroid coordinates, and rotational inertia matrix distribution data of each component relative to the base coordinate system.

[0050] The digital twin modeling module sets relative kinematic constraints for each connecting node in the three-dimensional geometric assembly model. Based on the mechanical connection form of the real equipment, the system defines the rotation axis vector direction and safe limit rotation angle range of each rotary joint, clarifies the rules for releasing and locking the degrees of freedom between connecting components, establishes a system-level multi-rigid-body topology hierarchy, and introduces a gravitational acceleration vector into the virtual environment to simulate the action of an external constant force field.

[0051] The digital twin modeling module runs a robot dynamics model solver based on the constructed multi-rigid-body topology hierarchy and assigned dynamic parameters. Under pre-defined end-effector trajectory input conditions, the system performs numerical integration and algebraic solutions to the robot system's motion equations to obtain the force and state parameters throughout the process. The fundamental robot dynamics equations involved in the above process are as follows: ; in, For robot joint angles, which represent the actual angular displacement of each rotary joint relative to the mechanical zero point position, in rad; ω is the joint angular velocity, which represents the rate of change of angular displacement of each rotary joint per unit time, and the unit is rad / s; Angular acceleration is the rate of change of the angular velocity of each rotary joint per unit time, expressed in rad / s. 2 ; The robot inertia matrix represents the set of inertial drag coefficients generated by the mass distribution of each link in the real device on the joint rotation motion. is the centrifugal force and Coriolis force matrix, which represents the set of nonlinear coupled forces caused by changes in angular velocity during high-speed motion of each joint; is the gravitational torque vector, which represents the downward tension generated by the Earth's gravitational field on the center of mass of each link and the equivalent torque value at the rotation center of each joint, in N·m. This is the joint driving torque vector, which represents the actual torque value output by the drive motor and applied to the corresponding joint drive shaft through the reducer, in N·m.

[0052] The digital twin modeling module derives the kinematic state parameters of the tool's end effector based on the solution results of the fundamental dynamic equations. The system calculates the linear acceleration, angular acceleration, and combined force vector of the tool's end effector in the three-dimensional workspace, and compares the calculated time-series parameters with the sensor measurements from the actual device executing the same motion trajectory. Based on the comparison results, the system reverse-corrects the rotational inertia matrix values ​​of each link until the calculated parameter deviation is no greater than a preset dynamic simulation error threshold (specifically, dynamic simulation error ≤ 2%).

[0053] Step S103 establishes a virtual-real data interaction channel to achieve real-time mapping and synchronization of multiple parameters.

[0054] The virtual-physical data interaction module is configured with data transmission nodes based on an open platform unified communication architecture within the underlying industrial control communication network architecture. The system establishes a network service host process on the virtual server system responsible for running the digital twin model, while simultaneously deploying compatible client connection configurations on the network module of the electrical control cabinet of the real equipment. The system establishes a low-level bidirectional data transmission link via the Ethernet bus protocol and sets the basic communication frequency parameters for periodic data packet transmission and reception.

[0055] After the underlying communication link is established, the virtual-real data interaction module performs the binding operation between the control system's memory data address identifiers and virtual model variables. The system scans the global register address list within the main control program of the real equipment and extracts the memory block addresses representing operating parameters such as joint angle data, actual end-effector position data, and real-time load change data. The system then matches and logically connects these extracted memory block addresses with the corresponding receive variable identifiers exposed by the digital twin control interface.

[0056] The virtual-real data interaction module initiates a persistent, periodic data synchronization service in the background to maintain real-time updates of various status parameters. When the real device is powered on and running, it packages and sends the updated status parameters in memory to the host process on the virtual server according to the communication frequency set by the system. The system receives data packets and records the system timestamp of the current receiving action. It extracts values ​​through parsing and unpacking operations and refreshes them into the digital twin model, updating the system model's attitude and internal force state in the virtual environment in real time. At the same time, it monitors the congestion status of network transmission nodes to keep the delay time of the data packet synchronization process within a preset delay threshold.

[0057] See attached document Figure 4 This invention provides a method for training a tool wear prediction model based on a long short-term memory neural network, which may include setting and executing procedures such as multi-source sensor data acquisition and iterative training of the network model. Step S200 specifically includes the following sub-steps: Step S201: Collect sensor data under multiple working conditions to construct a tool wear database.

[0058] The wear prediction module receives time and electrical parameters continuously monitored by the data acquisition module when the robot system is under different operating conditions. Under set operating boundary conditions (operating time 1-8 hours, welding current 100-300A, welding voltage 10-30V), the system acquires the cumulative operating time of the robot performing welding operations through the communication network, and simultaneously reads the welding current and welding voltage values ​​in real time during the welding process through current and voltage sensors at a set sampling frequency. The system uses the extracted operating time, welding current, and welding voltage values ​​as a set of input characteristic parameters reflecting the working intensity of the tool end effector.

[0059] The wear prediction module synchronously acquires target label data reflecting the actual material wear state of the tool tip. After the set operation cycle is completed, the system uses a laser rangefinder (measurement accuracy ≤ ±0.001mm) to measure the dimensions of the robot system's tool tip. The system calculates the difference between the measured current tool tip dimension and the initial calibration dimension of the new tool to obtain the actual tool tip wear data, and uses this actual tool tip wear data as the target ground truth during the network model training process.

[0060] The wear prediction module performs timestamp alignment and data cleaning on the acquired set of input feature parameters and corresponding target label data. The system removes missing values ​​and abrupt data jumps caused by sensor malfunctions or communication anomalies. Then, it packages the cleaned operation time, welding current, welding voltage, and corresponding actual tool end wear data into a complete data sample. The system continuously accumulates multiple sets of data samples under different wear levels, ultimately constructing a tool wear database containing 1000 sets of data for algorithm training.

[0061] Step S202: Construct a long short-term memory neural network and set the model training loss function.

[0062] The wear prediction module constructs a long short-term memory neural network topology in system memory to fit the nonlinear relationship between equipment operating status and tool wear. The system sets the input layer of this neural network to have a dimension of 3, corresponding to receiving operation time values, welding current values, and welding voltage values, respectively. After the input layer, the system constructs a hidden layer structure with 2 layers, and sets the number of neurons in each hidden layer to 64 to extract long-term dependency features from the time series data.

[0063] The wear prediction module constructs an output layer after the hidden layer to generate the final prediction result. The system sets the dimension of the output layer to 1, and its single continuous output value represents the wear amount data at the end of the prediction tool under the current input conditions. The system completes the initialization operation of the connection weights between each layer of the network, and configures the network to control the accumulation and forgetting of information at time steps through the forget gate, input gate, and output gate of the network units during the forward propagation process, so as to establish a mapping channel from feature input to wear amount output.

[0064] The wear prediction module defines a model training loss function for the Long Short-Term Memory (LSTM) neural network to guide backpropagation weight updates. The system uses the mean squared error function as a benchmark to measure the accuracy of the network's predicted output, thereby reducing the difference between the predicted and actual values ​​during training iterations. The model training loss function involved in the above process is as follows: ; in, The mean squared error loss value represents the overall deviation between the output of the prediction model and the actual measurement result, and its value ranges from a real number greater than or equal to 0. The total number of samples in the training batch represents the number of data points involved in the computation in a single iteration. For the first The actual tool end wear data for each sample represents the true value measured by a laser rangefinder, in mm. For the first The wear data at the end of the prediction tool for each sample represents the theoretical value calculated by forward propagation of the long short-term memory neural network model, in mm.

[0065] Step S203 executes iterative training of the model and verifies the prediction error to complete model encapsulation.

[0066] The wear prediction module divides the tool wear database into training and validation sets and initiates iterative training of the neural network. The system is configured with the Adam (Adaptive Moment Estimation) optimizer as the driving algorithm for updating network weight parameters, dynamically adjusting the learning step size of each network parameter based on the first and second moment estimates of the gradient. The system inputs data samples from the training set into the neural network in batches and calculates the loss value, then updates the connection weights and biases between network layers using the backpropagation algorithm.

[0067] The wear prediction module executes a set number of iterative cycles during training. The system sets the total number of training iterations to 500, and after each iteration, it uses validation set data to test the generalization ability of the current network model. The system records the absolute difference between the predicted tool end wear data calculated on the validation set under the current network weights and the actual tool end wear data, thereby monitoring the convergence status of the model error.

[0068] After completing the set number of training iterations, the wear prediction module verifies the final output prediction error of the network model. The system evaluates the prediction error value of the model on the validation set and determines whether it meets the accuracy standard of a prediction error not exceeding 0.002mm (i.e., ≤±0.002mm). If the error index meets the set standard, the system locks all weight coefficients between network layers and solidifies and encapsulates the Long Short-Term Memory Neural Network into an independently callable tool wear prediction model for real-time online invocation in subsequent equipment operation processes.

[0069] See attached document Figure 5 This invention provides a method for performing a tool coordinate repositioning process based on iterative learning control to achieve high-precision calibration. Step S300 specifically includes the following sub-steps: Step S301 triggers initial relocation, generates initial correction parameters, and collects status data.

[0070] The iterative learning control module receives a repositioning start signal from the control system to trigger the initial repositioning process. The system calls a pre-built digital twin model and, based on the robot system's fundamental coordinate system definition and the current tool parameter configuration, simulates the reference calibration process of the tool coordinates in a virtual 3D space. The system calculates the spatial pose difference in the initial state by comparing the ideal state with the currently recorded device parameters.

[0071] The system generates initial correction parameters (represented as K1, K2, ..., K6) covering multiple spatial dimensions based on the aforementioned virtual simulation process. These initial correction parameters are then transmitted to the robot execution module via an industrial communication protocol network. The control system drives the robot body to move to the set detection position according to the initial correction parameters and completes the initial adjustment of the tool end effector posture. This initial repositioning time is ≤90s.

[0072] After the robot body reaches the set position and remains stable, the data acquisition module activates the spatial detection mechanism to obtain the actual end-effector position data in three-dimensional space. The system calls the encapsulated tool wear prediction model in parallel, inputting the current operation time value and electrical parameters into the model to calculate and extract the corresponding predicted tool end-effector wear data, providing basic parameters for the subsequent accurate identification of spatial position deviations.

[0073] Step S302 calculates the comprehensive coordinate deviation by combining the wear data at the end of the prediction tool.

[0074] The iterative learning control module receives the actual end-effector position data obtained from measurements and the predicted tool end-effector wear data output by the tool wear prediction model. Simultaneously, the system retrieves the preset welding process program file stored within the robot system and extracts the ideal end-effector position data corresponding to the current process step.

[0075] The system performs coordinate system alignment processing on the actual end-effector position data and the ideal end-effector position data to eliminate system errors caused by inconsistencies in the reference system. The system filters the acquired position data to remove high-frequency noise interference mixed in during sensor detection, thereby ensuring that all 3D coordinate values ​​involved in the calculation are in a unified reference coordinate system and at the same data quality level.

[0076] The system calculates the difference between the actual and ideal end-effector position data according to the established deviation calculation rules, and subtracts the material loss dimension predicted by the model from the result. Through this operation, the system separates the pure spatial position error that needs to be eliminated through kinematic compensation. The comprehensive coordinate deviation calculation equation involved in the above process is as follows: ; in, The overall coordinate deviation represents the pure spatial error between the actual position of the tool tip and the target position after eliminating wear factors, and the unit is mm; This is the actual end-effector position data, which represents the three-dimensional spatial coordinates fed back by the robot body after performing initial corrections, in mm; The ideal end position data represents the coordinates of the target reference point preset in the welding process, in mm. The data for predicting tool tip wear represents the material loss dimension at the welding torch tip predicted by a long short-term memory neural network model, in mm.

[0077] Step S303: Apply an iterative learning algorithm to update the coordinate correction parameters until the accuracy requirements are met.

[0078] The iterative learning control module introduces the calculated comprehensive coordinate deviation into the closed-loop control cycle, and uses an iterative learning control algorithm to progressively optimize the system's feedforward correction mechanism using historical deviation data. The system dynamically calculates the adaptive learning rate for adjusting the control parameters based on the current iteration number.

[0079] The system progressively reduces the learning rate parameter based on pre-configured baseline learning weights (initial learning rate of 0.01) and a linear decay mechanism during iteration to prevent overshoot and oscillation when the robot's end effector approaches the target position. The adaptive learning rate decay equation involved in the above process is as follows: ; in, For the first The adaptive learning rate of the next iteration represents the actual learning step size weight used in the current iteration step, and its value ranges from 0 to 1. The initial learning rate (set to 0.01) represents the maximum corrected weights when the algorithm starts. is the decay coefficient, which represents the linear decrease in the learning rate after each iteration; This represents the current iteration number, indicating the number of loops the algorithm has executed. The maximum number of iterations is limited to ≤8.

[0080] The system multiplies the calculated adaptive learning rate by the comprehensive coordinate deviation of the current iteration cycle, and then adds the product to the existing coordinate correction parameters to generate the coordinate compensation values ​​required for the next iteration. The system inputs the updated coordinate correction parameters into the digital twin modeling module for virtual environment simulation. After confirming the absence of motion interference and singularities, the parameters are sent to the robot body to perform a new round of position adjustments. The iterative correction parameter update equations involved in the above process are as follows: ; in, For the next correction parameter, it represents the value at the 1st... The coordinate compensation values ​​sent to the robot during each iteration, in mm; The current correction parameter represents the value at the th . The coordinate compensation values ​​used in the next iteration are in mm. For the first The adaptive learning rate of the next iteration represents the weight coefficient that controls the current correction step size in the iterative learning algorithm, and its value ranges from 0 to 1. For the first The comprehensive coordinate deviation of the next iteration represents the spatial position error calculated under the current correction parameters, in mm.

[0081] After each parameter update and robot action execution, the system re-detects the actual end-effector position data and calculates the new comprehensive coordinate deviation through the data acquisition module. The system determines whether the absolute value of this new deviation meets the requirement of a corrected comprehensive coordinate deviation ≤ ±0.018mm. If the determination result indicates that the current position error is within the set standard, the system terminates the iteration loop and determines that the repositioning process is complete; if the determination result is negative, the system increments the current iteration number and repeats the deviation calculation and parameter update operation until the deviation converges within the standard range or the set maximum iteration limit is reached.

[0082] See attached document Figure 6 This invention provides a method for real-time monitoring of job status triggering and dynamic adaptive compensation, wherein step S400 specifically includes the following sub-steps: Step S401 involves concurrently calling the prediction model and reading sensor data in real time.

[0083] The iterative learning control module establishes a continuous background monitoring mechanism during the robot's routine welding operations. The system maintains a continuous connection between the virtual and real data interaction channels to ensure that it can acquire various operational parameters and end-effector status feedback information of the robot body in real time without interrupting the main control program.

[0084] The system continuously receives real-time electrical signals from the sensor network via a data acquisition module. The system synchronously feeds the real-time acquired operation time, welding current, and welding voltage values ​​into the tool wear prediction model. This model runs concurrently in the background and calculates and outputs the predicted tool end wear data for the current moment in real time.

[0085] The system simultaneously reads the real-time force data measured by the load sensor via the industrial communication bus. The data preprocessing unit filters and performs analog-to-digital conversion on the received load analog signal to extract the current load data characterizing the welding torch body and its auxiliary pipelines in the current working posture, providing real-time data support for subsequent system status assessment.

[0086] Step S402 calculates the load change rate and wear increment to determine the trigger threshold.

[0087] After receiving real-time updated end-of-pipe wear data and current load data from the predictive tool, the iterative learning control module initiates the state assessment program. The system automatically retrieves the baseline wear data and initial load data recorded in the storage unit from the most recent repositioning calibration, establishing a unified reference benchmark for deviation calculation.

[0088] The system calculates the difference between the current predicted tool end wear data and the baseline wear data, generating a wear increment parameter characterizing the current degree of tool wear. Simultaneously, the system calculates the absolute difference between the current load data and the initial load data, and then proportionally calculates this absolute difference to the initial load data. The dynamic load change rate calculation equation involved in the above process is as follows: ; in, The load change rate represents the percentage fluctuation of the load currently borne by the robot end relative to the initial calibrated load. This is the current load data, which represents the total mass of the device end as collected in real time by the load sensor, in kg. This is the initial load data, representing the baseline mass recorded when the robot last completed repositioning and calibration, in kg.

[0089] The system transmits the calculated wear increment parameters and load change rate to the logic comparison register. The comparison register is pre-programmed with trigger thresholds to define the system error tolerance range. The system compares the real-time calculated parameters with the preset trigger thresholds through hard connections or real-time computing tasks to determine whether the current system operating state deviates from the set accuracy maintenance range.

[0090] Step S403 intercepts routine operation instructions and inserts a program to perform dynamic compensation.

[0091] When the comparison result of the logic comparison register indicates that the wear increment parameter is ≥0.05mm or the load change rate is ≥5%, the repositioning process is automatically triggered (dynamic compensation response time ≤30s). The iterative learning control module immediately sends a high-priority interrupt request to the main controller. Upon receiving the interrupt request, the main controller responds and intercepts the currently executing routine welding operation command, controlling the robot body to pause its movement in the current safe pose.

[0092] After routine operation commands are suspended, the system seamlessly inserts an adaptive repositioning program written in a dedicated RAPID control language, enabling dynamic calibration of the tool coordinates without manual intervention. The control system generates drive electrical signals based on the latest coordinate correction parameters output by the iterative learning control algorithm, controlling each joint motor to run along the compensation trajectory, thereby guiding the tool end effector to complete dynamic calibration and deviation compensation of the spatial coordinate system.

[0093] After the system completes the dynamic adaptive compensation program (with a dynamic compensation time ≤ 60s) and verifies the coordinate deviation through the digital twin modeling module, it sends a compensation end signal. Upon receiving this signal, the main controller cancels the previous interruption request and releases the suspended and intercepted routine welding operation commands. The robot body continues to execute the remaining welding processes according to the restored control commands, maintaining continuous operation throughout without the need for manual intervention.

[0094] Specific application examples: This embodiment uses a multi-variety, small-batch welding production line in a county-level manufacturing industry as an application scenario, specifically addressing the problems of load variations caused by frequent workpiece changes and welding torch wear due to continuous operation in this scenario. The hardware configuration of this system includes: one physical device (i.e., an ABB IRB 2600 model, with a repositioning accuracy of ≤±0.015mm) equipped with an ABB Robotics IRBP 664 welding torch at its end effector; a virtual server equipped with an Intel XeonGold 6330 CPU, an NVIDIA RTX A6000 graphics card, and 64GB of memory, used to deploy the digital twin modeling module; a data acquisition module equipped with a laser rangefinder with a measurement accuracy of ≤±0.001mm, current and voltage sensors with an accuracy of ±1%, and a load sensor with an accuracy of ±0.5%; the communication architecture of the virtual-physical data interaction module relies on an industrial Ethernet switch supporting the OPC UA protocol, with a data transmission rate set to 1Gbps. In terms of software implementation, SolidWorks 2023 and ANSYS 2023 are used to complete high-precision digital twin modeling, KEPServerEX OPC UA server is used to realize virtual and real data interaction, tool wear prediction model is built and run based on TensorFlow framework, and iterative learning control algorithm is compiled by MATLAB and integrated into server control software. The robot execution module executes automatic repositioning program through RAPID language.

[0095] When the system performs a specific relocation task, the digital twin modeling module and the iterative learning control module perform real-time numerical calculations on the aforementioned formulas. Taking the third iteration optimization process as an example, the system first calculates the current adaptive learning rate according to the adaptive learning rate decay equation: Let the initial learning rate be... attenuation coefficient Current iteration number Then, the adaptive learning rate for the third iteration is calculated as follows: .

[0096] Subsequently, the system executes the comprehensive coordinate deviation calculation equation based on feedback from the laser rangefinder and the model. Let's assume the actual end-point position data fed back by the data acquisition module is located in a certain dimension. Ideal end position data preset by the process Predicted tool end wear data output by the tool wear prediction model Then the overall coordinate deviation of this dimension is: .

[0097] After obtaining the deviation, the system uses the iterative correction parameter update equation to calculate and update the current coordinate correction parameters. Let the current correction parameters be those at the third iteration. Combined with the adaptive learning rate obtained above With comprehensive coordinate deviation The next correction parameter (i.e., the 4th correction) is calculated as follows: The updated coordinate correction parameters are immediately sent to the digital twin modeling module for simulation verification, and are iterated until the deviation approaches zero.

[0098] Meanwhile, during operation, the system continuously monitors the load status to perform calculations using a dynamic load change rate calculation equation. After a fixture change, the system reads the initial load data. The load sensor provides real-time feedback of the current load data. The load change rate was calculated as follows: Because the detected load change rate of 6.0% reaches the preset relocation trigger threshold of ≥5%, the system automatically intercepts the currently running routine operation instructions and triggers the execution of the dynamic adaptive compensation process.

[0099] See attached document Figure 7 The graph shows the number of iterations (from 1 to 10) on the x-axis and the overall coordinate deviation (mm, ranging from 0 to 0.12) on the y-axis. A horizontal dashed line at 0.018 mm is set as a reference line for the repositioning accuracy threshold. The graph illustrates the convergence characteristics of the two methods when the initial coordinate deviation is 0.12 mm. The solid lines marked with filled circles in the figure represent the method of this invention. This method employs an iterative learning control strategy based on an adaptive learning rate (combining an iterative correction parameter update equation and an adaptive learning rate decay equation). Its overall coordinate deviation decreases rapidly and smoothly with increasing iteration count. The deviation drops to 0.05 mm in the third iteration and to approximately 0.016 mm in the sixth iteration, successfully crossing the 0.018 mm accuracy threshold from top to bottom, and finally stabilizing at around 0.012 mm in the tenth iteration. The dashed lines marked with hollow squares in the figure represent traditional relocation methods that do not use adaptive learning rates and historical bias data from digital twins. Their bias convergence speed is significantly lagging. In the 6th iteration, the bias is still as high as about 0.055mm. After a complete 10 iterations, the bias is still around 0.040mm, failing to reach the specified accuracy requirement of 0.018mm, and the convergence efficiency is low.

[0100] Experimental verification and effect comparison section: Systematic experimental verification was conducted in the aforementioned application scenarios.

[0101] First, a three-dimensional geometric model and a dynamic simulation model of the ABB IRB 2600 robot were constructed. After communication debugging, the data synchronization delay of the virtual and real data interaction module was stabilized at 8ms, and the dynamic simulation error was reduced to 1.8%.

[0102] After the model was built, multi-source data was collected under the conditions of operation time of 0-8 hours, welding current of 150-250A, and welding voltage of 15-25V to train the long short-term memory neural network model. The results showed that the output prediction error of the tool wear prediction model was ±0.0018mm.

[0103] In the practical application testing phase, after replacing the welding torch with a brand new ABB Robotics IRBP 664, the initial repositioning process was initiated. The system uses historical deviation data from the digital twin to generate initial correction parameters and iterates. After 6 iterations, the overall coordinate deviation stabilized at ±0.016mm, and the overall initial repositioning time was only 75s.

[0104] Subsequently, the robot was controlled to continuously perform routine welding operations. When the 4th hour was reached, the predicted wear data of the tool end of the tool output by the tool wear prediction model reached 0.06mm. Since this value is greater than or equal to the trigger threshold of 0.05mm, the system automatically triggered and executed the dynamic adaptive compensation process, which took 45 seconds to complete.

[0105] When continuous operation reached the 7th hour, the load on the end effector of the tool was manually changed by adding a 5kg counterweight. The system, through the load sensor, calculated that the load change rate was 6.2%, which met the repositioning trigger condition of greater than or equal to 5%. The system then automatically intercepted the normal operation command and executed the dynamic adaptive compensation process, which took 42 seconds to complete the compensation.

[0106] See attached document Figure 8 The horizontal axis of the graph represents the operation time (h, from 0 to 8 hours), and the vertical axis represents the positioning error (mm, ranging from 0 to 0.12). This is combined with the aforementioned experimental data... Figure 8It can be seen that when facing dynamic operating scenarios such as tool wear and load changes, the traditional method, represented by the hollow diamond mark in the figure, lacks dynamic monitoring and early compensation mechanisms. Its positioning error, starting from approximately 0.015mm initially, shows a clear linear cumulative increase over time. By the fourth hour of operation, the error had risen to approximately 0.058mm, and by the end of the eighth hour, it had reached approximately 0.108mm, completely exceeding the allowable range for precision welding.

[0107] In contrast, the solid line marked with a solid inverted triangle in the figure represents the method of the present invention. By introducing the advance prediction and compensation of digital twin virtual-real linkage and tool wear prediction model, the positioning error always presents a stable broken line trend throughout the entire continuous operation cycle of 0 to 8 hours, and is strictly controlled below the accuracy threshold of 0.018mm (maintained between approximately 0.015mm and 0.018mm).

[0108] Figure 8 The method curve of this invention clearly marks two dynamic compensation trigger points with hollow large circles: the first large circle is located at the 4th hour of the horizontal axis operation time (corresponding to an error of approximately 0.016 mm), confirming the compensation mechanism automatically triggered due to excessive wear caused by continuous operation in the aforementioned experiment; the second large circle is located at the 7th hour of the horizontal axis operation time (corresponding to an error of approximately 0.015 mm), confirming the compensation mechanism automatically intercepted and triggered due to sudden load changes caused by human intervention in the end effector in the aforementioned experiment. This invention shortens the initial repositioning time (measured at 75 s) and the dynamic adaptive compensation process time (measured within 45 s), enabling preventative measures to be taken even with no downtime or very short downtime. It achieves unmanned intervention for high-precision repositioning, improving the production continuity of county-level manufacturing and the level of intelligent robotic operations.

Claims

1. An adaptive repositioning method for tool coordinates of an ABB welding robot, characterized in that, Includes the following steps: Construct a digital twin system for ABB welding robots, establish a mapping channel between real equipment and virtual simulation models, and realize real-time synchronous interaction of data; Collect multi-source state data during the operation process, train a tool wear prediction model based on a long short-term memory neural network based on the multi-source state data, and output the predicted tool end wear data. Based on the actual end position data and the end wear data of the prediction tool, the comprehensive coordinate deviation is calculated, historical deviation data is extracted, and the coordinate correction parameters are updated by applying an iterative learning control algorithm, and the repositioning process is executed. The system acquires current load data in real time and calculates the load change rate. It then calls the tool wear prediction model to obtain wear increment parameters. When the load change rate or wear increment parameters meet the relocation trigger conditions set based on the system error tolerance range, the system executes a dynamic adaptive compensation process.

2. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for constructing the ABB welding robot digital twin system and achieving real-time data synchronization and interaction are as follows: Collect actual geometric dimension data of the real equipment body and tool end, and build basic component entities one by one in the 3D design software environment to synthesize a 3D geometric assembly model; Assign material properties and mass parameters to each component of the three-dimensional geometric assembly model, set relative kinematic constraints, and construct a dynamic simulation model corresponding to the real equipment. Within the underlying industrial control communication network architecture, data transmission nodes based on an open platform communication unified architecture are deployed to establish a bidirectional data transmission link between real state parameters and virtual simulation models, maintaining real-time updates of various state parameters.

3. The ABB welding robot tool coordinate adaptive repositioning method according to claim 2, characterized in that, The specific steps for constructing the dynamic simulation model corresponding to the real device are as follows: Define the rules for releasing and locking the degrees of freedom between connecting components, establish a system-level multi-rigid-body topology hierarchy, and introduce a gravitational acceleration vector into the virtual environment to simulate the action of an external constant force field; Numerical integration and algebraic solutions are performed based on the motion equations of the robot system. By combining the robot joint angles, joint angular velocities and joint angular accelerations, inertia matrix, centrifugal force and Coriolis force matrices, gravitational torque vector and joint driving torque vector, various kinematic state parameters of the end effector are derived. The calculated time series parameters are compared with the actual sensor measurement results fed back by the real equipment, and the rotational inertia matrix values ​​of each link are corrected in reverse until the calculated deviation of the derived parameters is not greater than the preset dynamic simulation error threshold. The dynamic simulation error threshold represents the maximum allowable deviation between the calculated and derived parameters and the sensor measurement results.

4. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for training the tool wear prediction model based on a long short-term memory neural network are as follows: The system continuously collects operation time, welding current, and welding voltage values ​​as input feature parameters, and simultaneously acquires actual tool end wear data measured by a laser rangefinder as the target true value to construct a tool wear database. A long short-term memory neural network topology with three input layers that receive operation time, welding current and welding voltage values ​​respectively is constructed, and an output layer is constructed after the hidden layer to generate data on the wear of the tool tip. Define a model training loss function to guide backpropagation weight updates, perform iterative model training, and solidify and encapsulate it into an independent and callable tool wear prediction model after verifying that the prediction error meets the accuracy standard. The accuracy standard represents the prediction error limit that meets the requirements for online calling of the network model.

5. The ABB welding robot tool coordinate adaptive repositioning method according to claim 4, characterized in that, The specific steps for defining the model training loss function and performing iterative model training are as follows: The mean squared error function is used as the model training loss function to calculate the overall deviation between the predicted tool end wear data calculated by the forward propagation of the network model in the training batch and the actual measured tool end wear data, and to generate the mean squared error loss value. An adaptive moment estimation optimizer is used to dynamically adjust the learning step size of each network parameter based on the first-order and second-order moment estimates of the gradient. The data samples in the training set are input into the neural network in batches and the mean squared error loss value is calculated. The connection weights and bias values ​​between network layers are updated iteratively through the backpropagation algorithm.

6. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for calculating the comprehensive coordinate deviation are as follows: Retrieve the preset welding process program file stored inside the robot system and extract the ideal end position data corresponding to the current process step from it; The actual end position data and the ideal end position data are subjected to coordinate system alignment and filtering processing to ensure that the three-dimensional coordinate values ​​involved in the calculation are in a unified reference coordinate system. The actual end position data is subtracted from the ideal end position data, and the predicted tool end wear data output by the tool wear prediction model is subtracted from the difference result. The pure spatial position error that needs to be eliminated by kinematic compensation is separated as the comprehensive coordinate deviation.

7. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for updating the coordinate correction parameters using the applied iterative learning control algorithm are as follows: Based on the pre-set initial learning rate and linear decay mechanism, combined with the current iteration round and decay coefficient, the learning rate parameter is gradually reduced to calculate the current adaptive learning rate. The calculated adaptive learning rate is multiplied by the comprehensive coordinate deviation of the current iteration cycle, and the product is added to the current coordinate correction parameters to generate the coordinate correction parameters required for the next iteration. The updated coordinate correction parameters are input into the digital twin model for virtual environment simulation. After confirmation, the parameters are sent to the real device to perform position adjustment and the actual end position data is re-detected to determine whether the deviation has converged to the standard range. The standard range is the allowable repositioning accuracy range that meets the target reference point position requirements.

8. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for obtaining the current load data and calculating the load change rate are as follows: The current load data characterizing the welding torch body and its auxiliary pipelines under the current working posture is extracted through data preprocessing operations. Retrieve the initial load data obtained during the most recent relocation calibration from the stored records; Calculate the absolute difference between the current load data and the initial load data, and then perform a percentage operation on the ratio of the absolute difference to the initial load data to generate a percentage value. Use the percentage value as the load change rate.

9. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for concurrently calling the tool wear prediction model to obtain wear increment parameters are as follows: The real-time collected operation time, welding current and welding voltage values ​​are used as input features and synchronously fed into the tool wear prediction model running concurrently in the background. The tool wear prediction model calculates and outputs the predicted tool end wear data in real time at the current moment. Retrieve the baseline wear data from the most recent repositioning calibration, calculate the difference between the current predicted tool end wear data and the baseline wear data, and generate a wear increment parameter characterizing the current wear level of the tool.

10. The ABB welding robot tool coordinate adaptive repositioning method according to claim 1, characterized in that, The specific steps for executing the dynamic adaptive compensation process are as follows: The load change rate and wear increment parameters obtained in real time are compared with the preset relocation trigger threshold in the logic comparison register. The relocation trigger threshold is used to define the allowable range of system error. When the comparison result shows that the wear increment parameter or load change rate is greater than or equal to the relocation trigger threshold, a high-priority interrupt request is sent to the main controller to intercept the currently executing routine welding operation command and control the real equipment to pause its operation. The adaptive repositioning procedure is seamlessly inserted, and the tool end is guided to complete dynamic calibration based on the latest coordinate correction parameters output by the iterative learning control algorithm. After verification, the interruption request status is revoked and the normal welding operation command is restored to maintain continuous operation.

Images