Feedforward compensation method and platform for servo system driven by model prediction

Abstract

The invention discloses a model prediction driven servo system feed-forward compensation method and platform, and relates to the technical field of industrial control, and the method comprises the steps: determining a tail end load mass sequence and a tail end welding posture sequence of a robot tail end according to a preset welding task of a welding robot in a future time period, a tail end load rotational inertia sequence is analyzed and determined; evaluating and determining the servo control complexity of the servo system; and calling a trajectory prediction model and setting an adaptive parameter optimization strategy based on the servo control complexity, and carrying out feed-forward compensation parameter optimization of the servo system according to a tail end load quality sequence, a tail end welding posture sequence and a tail end load rotational inertia sequence. And obtaining the optimal feed-forward compensation parameter to execute the feed-forward compensation of the servo system of the welding robot in the future time period. The technical problem that in the prior art, due to the fact that welding wire consumption causes dynamic change of robot tail end load parameters, the feed-forward compensation precision of a servo system is reduced is solved.

Description

Technical Field

[0001] This invention relates to the field of industrial control technology, specifically to a model prediction-driven servo system feedforward compensation method and platform. Background Technology

[0002] In modern industrial automated production lines, servo systems, as the core execution units of high-end manufacturing equipment, directly determine the operational quality and production efficiency of key equipment such as industrial robots and CNC machine tools through their motion control accuracy and dynamic response performance. Feedforward compensation technology, as an effective means to improve the tracking accuracy of servo systems, can theoretically achieve active suppression of dynamic errors by pre-injecting control signals that match the desired trajectory. However, in continuous operation scenarios such as welding robots, the feedforward compensation parameters used by industrial controllers are usually based on offline tuning and remain unchanged over a long period, making it difficult to cope with time-varying characteristics. This dynamic change makes it difficult for fixed-parameter feedforward compensation strategies to maintain optimal performance in actual operation, thus affecting the accuracy and stability of industrial control. Summary of the Invention

[0003] This application provides a model prediction-driven servo system feedforward compensation method and platform to address the technical problem of decreased servo system feedforward compensation accuracy caused by dynamic changes in robot end-effector load parameters due to welding wire consumption in the prior art.

[0004] In view of the above problems, this application provides a model prediction-driven feedforward compensation method and platform for servo systems.

[0005] In a first aspect, this application provides a model prediction-driven feedforward compensation method for a servo system, the method comprising: Based on the preset welding tasks of the welding robot in the future time period, determine the end load mass sequence and end welding posture sequence of the robot end, and analyze and determine the end load rotational inertia sequence. The servo control complexity of the servo system is evaluated and determined based on the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence. Based on the aforementioned servo control complexity, the trajectory prediction model is invoked and the adaptation parameter optimization strategy is set. The feedforward compensation parameters of the servo system are optimized according to the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence. The optimal feedforward compensation parameters are obtained to execute the feedforward compensation of the welding robot's servo system in the future time period.

[0006] Secondly, this application provides a model prediction-driven servo system feedforward compensation platform, comprising: The load analysis module is used to determine the end-effector load mass sequence and end-effector welding posture sequence of the robot end based on the preset welding tasks of the welding robot in the future time period, and to analyze and determine the end-effector load rotational inertia sequence. The complexity assessment module is used to assess and determine the servo control complexity of the servo system based on the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence. The parameter optimization module is used to call the trajectory prediction model and set the adaptive parameter optimization strategy based on the servo control complexity. It optimizes the feedforward compensation parameters of the servo system according to the end load mass sequence, end welding posture sequence and end load rotational inertia sequence, and obtains the optimal feedforward compensation parameters to execute the servo system feedforward compensation of the welding robot in the future time period.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application proposes a model prediction-driven servo system feedforward compensation method and platform. Before the start of a future welding task, it proactively determines the evolution sequence of the end-load mass, end-welding posture, and end-load rotational inertia inevitably caused by wire consumption in the future time period based on the preset welding task. Based on this, it quantitatively evaluates the dynamic complexity of the servo control process and intelligently invokes predictive resources and configuration parameter optimization strategies to perform precise and adaptive offline optimization of feedforward compensation parameters. This significantly improves the trajectory tracking accuracy and motion control stability of the welding robot under time-varying load conditions. Compared to traditional methods, the method provided in this application achieves the technical effect of enabling the servo system feedforward compensation to evolve synchronously with the dynamic changes of the load, which are strongly coupled with the welding task. Through smooth and adaptive compensation, it suppresses unnecessary torque fluctuations and mechanical vibrations, which is beneficial for extending the mechanical life of the robot body and reducing servo drive energy consumption. This provides a more intelligent, reliable, and adaptable core control guarantee for automated welding production lines. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a flowchart illustrating a model prediction-driven feedforward compensation method for a servo system, as provided in an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the structure of a model prediction-driven servo system feedforward compensation platform provided in an embodiment of this application.

[0011] The components represented by each number in the attached diagram are explained below: Load analysis module 100, complexity assessment module 200, parameter optimization module 300. Detailed Implementation

[0012] This application provides a model prediction-driven servo system feedforward compensation method and platform to address the technical problem of decreased servo system feedforward compensation accuracy caused by dynamic changes in robot end-effector load parameters due to welding wire consumption in the prior art.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, this application provides a model prediction-driven feedforward compensation method for a servo system, wherein the method includes: S10: Determine the end-effector load mass sequence and end-effector welding posture sequence based on the preset welding tasks of the welding robot in the future time period, and analyze and determine the end-effector load rotational inertia sequence.

[0016] In the field of servo control for welding robots, the effectiveness of feedforward compensation technology, used to improve trajectory tracking accuracy, highly depends on accurate modeling of the dynamic characteristics of the robot's end effector load. Traditional methods typically compensate based on a constant load or by obtaining current instantaneous load parameters solely through sensors, which is poorly suited for welding operations where welding wire is continuously consumed.

[0017] Step S10 in the method provided in this application embodiment includes: Based on the preset welding task, determine several end welding postures of the welding robot at several welding points in the future time period, and generate an end welding posture sequence; Monitor and acquire the current end-effector load quality of the welding robot; The amount of welding wire consumed by the welding robot at several welding points in a future time period is determined based on the preset welding task. The end load mass sequence is calculated based on the current end load mass and several welding wire consumptions; Based on the end load mass sequence, by establishing a mass distribution evolution model during the welding wire consumption process, the total mass change is mapped to a continuously changing parameter of the mass geometric distribution, thus obtaining a mass geometric distribution parameter sequence; In each control cycle, based on the current end load mass, mass geometric distribution parameters, and end welding posture, the equivalent rotational inertia tensor of the load about each joint axis is calculated, and the calculated rotational inertia is subjected to time-series filtering and interpolation to generate a smooth end load rotational inertia sequence.

[0018] In this embodiment of the application, the end load mass sequence and end welding posture sequence of the robot end are determined according to the preset welding task of the welding robot in the future time period, and the end load rotational inertia sequence is analyzed and determined.

[0019] Specifically, firstly, based on a preset welding task, several end-effector welding postures for several welding points of the welding robot within a future time period are determined, generating an end-effector welding posture sequence. For example, the preset welding task is to weld a complete weld seam. The task is analyzed, and the target position coordinates (X, Y, Z) and target posture angles (e.g., rotation angles around the X, Y, and Z axes) of each welding point required by the task are extracted. The position and posture data of each welding point collectively define an end-effector welding posture. Following the temporal order of the welding task, the end-effector welding postures are integrated in chronological order to obtain the end-effector welding posture sequence.

[0020] Furthermore, the current end-effector load mass of the welding robot is monitored and acquired. For example, a mass measurement is triggered before each new welding task begins to acquire the specific load mass, such as the current end-effector load mass being 5 kg.

[0021] Furthermore, based on the preset welding task, the wire consumption of the welding robot at several welding points within a future time period is determined. Specifically, based on the process parameters of the preset welding task, the wire feed speed, wire density, and welding time are obtained. The wire consumption can be calculated by multiplying the wire feed speed by the welding time and then by the wire density. For example, if the wire feed speed is 10 meters / minute, the welding time at a certain welding point is 6 seconds, and the wire density is 0.05 kg / m, then the wire consumption at that point is calculated as: 10 × (6 / 60) × 0.05 = 0.05 kg.

[0022] Further, an end-load mass sequence is calculated based on the current end-load mass and several welding wire consumption values. Specifically, the welding wire consumption value for each welding point in the future time period is calculated sequentially and arranged in chronological order to obtain a welding wire consumption sequence. The initial mass is used as the first value of the mass sequence. Starting from the second time point, i.e., the first welding point, the load mass at the previous time point is subtracted from the welding wire consumption value at the current time point to obtain the load mass at the current time point. Multiple end-load masses are calculated sequentially and then integrated to obtain the end-load mass sequence.

[0023] Furthermore, based on the aforementioned end-load mass sequence, an evolution model of the mass distribution during the welding wire consumption process is established, mapping the total mass change to continuously varying parameters of the mass geometric distribution, thus obtaining a sequence of mass geometric distribution parameters. For example, the welding wire reel is simplified as a cylinder, and welding wire consumption is considered as a uniform reduction in the height of this cylinder. For instance, the equivalent remaining length of the welding wire reel can be used to describe the mass distribution. Given the total initial mass of the welding wire, the material density, and the reel's cross-sectional area, the relationship between mass and length can be established: Mass = Welding wire density × Welding wire cross-sectional area × Equivalent remaining length. Therefore, knowing the mass at each moment, the equivalent remaining length at each moment and the continuously varying parameters describing the mass geometric distribution can be derived using the formula: Equivalent remaining length = Mass / (Welding wire density × Welding wire cross-sectional area). By sequentially calculating the equivalent remaining length at each time point, the sequence of mass geometric distribution parameters is obtained, reflecting the evolution of the welding wire reel's geometry during the welding process.

[0024] Furthermore, in each control cycle, based on the current end-load mass, mass geometric distribution parameters, and end-welding posture, the equivalent moment of inertia tensor of the load about each joint axis is calculated. The calculated moment of inertia is then subjected to temporal filtering and interpolation to generate a smooth sequence of end-load moment of inertia. Specifically, for each future moment, the end-load is considered a rigid body according to the mass distribution model. Given the current load mass and distribution parameters, the moment of inertia of the load about the three axes in its own center-of-mass coordinate system can be calculated. Further, based on the end-welding posture at that moment, the parallel axis theorem and coordinate rotation formula are used to transform and project the load's moment of inertia tensor in its own coordinate system onto each joint axis of the robot, obtaining the equivalent moment of inertia of the load about that joint axis. Finally, since the above calculations may be discrete at specific points in time, while control requires continuous signals, linear interpolation is performed on the calculated discrete moment of inertia values ​​to obtain the values ​​corresponding to each control cycle. Then, a moving average filter is applied to this continuous sequence, for example, using a sliding window of 5 periods for averaging filtering, to eliminate possible small abrupt changes in the calculation, and finally obtain a time-continuous and smoothly changing end load rotational inertia sequence, which describes the impact of load inertia on each future control cycle.

[0025] By proactively determining the end-load mass sequence and end-welding posture sequence in future time periods based on preset welding tasks, and analyzing the end-load rotational inertia sequence, the serialized prediction and description of the time-varying dynamic parameters of the core load during the welding process is realized.

[0026] S20: Evaluate and determine the servo control complexity of the servo system based on the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence.

[0027] Load parameter changes are not always uniform or gradual; different welding tasks lead to significant differences in the rate of load quality decay, attitude switching frequency, and the amplitude of rotational inertia fluctuations. These differences directly determine the varying levels of control difficulty and challenge faced by the servo system throughout the entire task execution. Existing technologies typically employ a single, fixed strategy to handle all variations, or rely solely on simple thresholds for judgment, failing to quantify and assess the dynamic complexity inherent in the load parameter sequence. This makes it difficult to achieve the optimal balance between control performance and computational efficiency when allocating computational resources, selecting predictive models, or setting optimization parameters.

[0028] Step S20 in the method provided in this application embodiment includes: The terminal load change quality sequence is calculated based on the terminal load quality sequence, wherein the terminal load change quality is the absolute value of the terminal load quality deviation between adjacent nodes; The end load rotational inertia sequence is calculated based on the end load rotational inertia sequence, wherein the end load rotational inertia change sequence is the absolute value of the deviation of the end load rotational inertia between adjacent nodes. Wave behavior analysis was performed on the end load change mass sequence and the end load rotation change inertia sequence, respectively, and the coefficient of variation of the load change mass and the coefficient of variation of the load rotation change inertia were calculated. Based on the aforementioned end-welding posture sequence, the norms of the angular velocity and linear acceleration of the end-tool coordinate system in Cartesian space are calculated to obtain the posture change rate. The servo control complexity of the servo system is determined by evaluating the coefficient of variation of the load change mass, the coefficient of variation of the load rotation change inertia, and the attitude change rate. The servo control complexity of the servo system is determined based on the coefficient of variation of the load change mass, the coefficient of variation of the load rotational inertia, and the attitude change rate, including: The ratio of the load variation quality variation coefficient to the reference load variation quality variation coefficient is used as the first servo control complexity coefficient. The ratio of the coefficient of variation of the load rotational inertia to the coefficient of variation of the reference load rotational inertia is used as the second servo control complexity coefficient. The ratio of the attitude change rate to the reference attitude change rate is used as the third servo control complexity coefficient. The servo control complexity of the servo system is obtained by weighted evaluation based on the first servo control complexity coefficient, the second servo control complexity coefficient, and the third servo control complexity coefficient.

[0029] In this embodiment, the servo control complexity of the servo system is evaluated and determined based on the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence.

[0030] Specifically, firstly, the end load change quality sequence is calculated based on the end load quality sequence, where the end load change quality is the absolute value of the end load quality deviation between adjacent nodes. For example, assuming the quality sequence is [5.0, 4.95, 4.90, 4.80], then the calculated end load change quality sequence is [|4.95-5.0|, |4.90-4.95|, |4.80-4.90|] = [0.05, 0.05, 0.10].

[0031] Further, an end-load rotational inertia sequence is calculated based on the end-load rotational inertia sequence, wherein the end-load rotational inertia change is the absolute value of the deviation of the end-load rotational inertia of adjacent nodes. For example, the same adjacent element difference calculation is performed on this rotational inertia sequence, and the absolute value is taken. For instance, assuming the rotational inertia sequence for a certain joint axis is [1.2, 1.18, 1.15, 1.05], then the calculated end-load rotational inertia change sequence is [|1.18-1.2|, |1.15-1.18|, |1.05-1.15|] = [0.02, 0.03, 0.10].

[0032] Furthermore, volatility analysis is performed on the end-load change mass sequence and the end-load rotational change inertia sequence, respectively, to calculate the coefficient of variation of the load change mass and the coefficient of variation of the load rotational change inertia. Specifically, the following calculations are performed on the "end-load change mass sequence" and the "end-load rotational change inertia sequence": First, the average value of all values ​​in the sequence is calculated; then, the standard deviation of the sequence is calculated; finally, the coefficient of variation is obtained by dividing the standard deviation by the average value. For example, for the change mass sequence [0.05, 0.05, 0.10] kg, its mean is approximately 0.067 kg, and its standard deviation is approximately 0.029 kg, so the coefficient of variation of the change mass sequence is 0.029 / 0.067 = 0.43. The same calculation is performed on the rotational change inertia sequence to obtain the coefficient of variation of the rotational change inertia sequence, which characterizes the relative volatility of the mass change and the rotational change inertia.

[0033] Furthermore, based on the end-effector welding posture sequence, the norms of the angular velocity and linear acceleration of the end-effector coordinate system in Cartesian space are calculated to obtain the attitude change rate. For example, the first derivative of the posture sequence with respect to time is calculated at each time point to approximate the linear velocity vector and angular velocity vector of the end-effector coordinate system at that time point. Then, the norms of these velocity vectors are calculated; for example, the 2-norm is used, i.e., the magnitude of the vector is calculated. The norm of linear velocity reflects the speed of end-effector movement, and the norm of angular velocity reflects the speed of end-effector rotation. Finally, the norms of linear velocity and angular velocity at the same moment are added to obtain the instantaneous attitude change rate at that moment. The average of the instantaneous attitude change rates over all time points in the entire future period is taken to obtain the comprehensive attitude change rate, representing the intensity of the entire task's motion, in a combined unit of millimeters per second and radians per second, which comprehensively reflects the motion agility required by the preset welding task.

[0034] Furthermore, the servo control complexity of the servo system is determined based on the coefficient of variation of the load change mass, the coefficient of variation of the load rotation inertia, and the attitude change rate.

[0035] Specifically, firstly, the ratio of the load variation quality variation coefficient to the baseline load variation quality variation coefficient is used as the first servo control complexity coefficient. For example, the baseline load variation quality variation coefficient can be derived from the average value obtained through statistical analysis of historical typical or high-frequency welding tasks, or it can be set by domain experts. The first servo control complexity coefficient = load variation quality variation coefficient / baseline load variation quality variation coefficient. A first servo control complexity coefficient greater than 1 indicates that the quality fluctuation of the current task is higher than the typical level, with a high contribution to complexity; a coefficient less than 1 indicates that it is lower than the typical level, with a low contribution to complexity.

[0036] Furthermore, the ratio of the load rotational inertia variation coefficient to the reference load rotational inertia variation coefficient is used as the second servo control complexity coefficient. The second servo control complexity coefficient = load rotational inertia variation coefficient / reference load rotational inertia variation coefficient, quantifying the degree of current task load inertia fluctuation relative to a typical level. The reference load rotational inertia variation coefficient is based on the average value obtained from statistical analysis of historical typical or high-frequency welding tasks, or is set by domain experts.

[0037] Furthermore, the ratio of the attitude change rate to the reference attitude change rate is used as the third servo control complexity coefficient. The third servo control complexity coefficient = attitude change rate / reference attitude change rate, reflecting the comparison between the motion intensity of the current preset welding task path and typical tasks. The reference attitude change rate is based on the average value obtained from statistical analysis of historical typical or high-frequency welding tasks, or is set by domain experts.

[0038] Furthermore, the servo control complexity of the servo system is obtained by weighting and evaluating the first, second, and third servo control complexity coefficients. For example, weights are assigned based on the importance of the first, second, and third servo control complexity coefficients. For instance, in a servo system, control due to rotational inertia fluctuations is more complex; therefore, the second servo control complexity coefficient is assigned a weight of 0.4, and the first and third servo control complexity coefficients are assigned a weight of 0.3. Thus, the servo control complexity = 0.3 × first servo control complexity coefficient + 0.4 × second servo control complexity coefficient + 0.3 × third servo control complexity coefficient. A servo control complexity greater than 1 indicates that the overall task difficulty is higher than typical, requiring a more refined compensation strategy; a complexity less than 1 means the task is relatively simple.

[0039] By evaluating and determining the servo control complexity of the servo system based on the end-load mass sequence, attitude sequence, and rotational inertia sequence, the control challenge level caused by the given welding task is accurately quantified. It can clearly distinguish between simple working conditions with stable and gradual load changes and complex working conditions with drastic load changes and agile attitude, providing key decision-making basis for subsequent optimization steps.

[0040] S30: Based on the servo control complexity, call the trajectory prediction model and set the adaptation parameter optimization strategy. According to the end load mass sequence, end welding posture sequence and end load rotational inertia sequence, optimize the feedforward compensation parameters of the servo system, obtain the optimal feedforward compensation parameters, and execute the feedforward compensation of the welding robot's servo system in the future time period.

[0041] Existing feedforward parameter optimization methods require high-fidelity models for refined trajectory prediction and parameter search to ensure compensation accuracy, but this usually comes with huge computational overhead, making it difficult to meet the real-time requirements of online or near-online tasks. On the other hand, if simplified models or fixed step size optimization are used to pursue efficiency, the compensation effect may decline due to inaccurate prediction or insufficient optimization when facing highly complex tasks.

[0042] Step S30 in the method provided in this application embodiment includes: A pre-trained trajectory prediction model, wherein the trajectory prediction model comprises P trajectory prediction units, where P is an integer greater than 10; The pre-trained trajectory prediction model includes: Based on the historical operation logs of similar welding robots, several sample end-load mass sequences, several end-welding posture sequences, several end-load rotational inertia sequences, and several historical end-motion trajectories were collected as several sample end-motion trajectories to obtain sample training data. The sample training data is divided into P equal parts, and then P-fold cross-partitioned with replacement is performed to obtain P sample training sets. The deep learning model is trained to convergence using the P sample training sets to obtain P trajectory prediction units; The optimal number of unit calls L is obtained by rounding down the product of the servo control complexity and the initial number of unit selections K, where K equals 5. If the calculated L is less than 2, then L is equal to 2; if the calculated L is greater than P, then L is equal to P. Within the P trajectory prediction units, L trajectory prediction units are randomly called to predict the robot's end-effector trajectory based on the end-effector load mass sequence, end-effector welding posture sequence, and end-effector load rotational inertia sequence. L initial predicted trajectories are output, and the end-effector predicted motion trajectory is obtained by fitting the trajectory mean. The dispersion of the prediction result is obtained by quantizing the trace of the spatial position covariance matrix of the L initial predicted trajectories at each time point in the prediction time domain. Based on the dispersion, the trajectory prediction confidence is calculated by a preset monotonically decreasing mapping function, wherein the trajectory prediction confidence and the dispersion are negatively correlated. Based on the servo control complexity and trajectory prediction confidence, an adaptation parameter optimization strategy is set, and the feedforward compensation parameters of the servo system are optimized according to the predicted motion trajectory of the end effector to obtain the optimal feedforward compensation parameters. The process includes setting an adaptation parameter optimization strategy based on the servo control complexity and trajectory prediction confidence, optimizing the feedforward compensation parameters of the servo system according to the predicted end-effector motion trajectory, and obtaining the optimal feedforward compensation parameters, including: The adaptation parameter optimization step size and the number of adaptation optimization convergences are configured based on the servo control complexity and trajectory prediction confidence. The adaptation parameter optimization step size is negatively correlated with the servo control complexity and positively correlated with the trajectory prediction confidence. The number of adaptation optimization convergences is positively correlated with the servo control complexity and negatively correlated with the trajectory prediction confidence. Obtain the expected end-effector trajectory corresponding to the preset welding task of the welding robot in the future time period; Based on the optimization step size and the number of optimization convergence times of the adaptation parameters, the feedforward compensation parameters of the servo system are optimized according to the expected end motion trajectory and the predicted end motion trajectory to obtain the optimal feedforward compensation parameters. Specifically, based on the adaptation parameter optimization step size and the number of adaptation optimization convergences, the feedforward compensation parameters of the servo system are optimized according to the expected end-effector trajectory and the predicted end-effector trajectory to obtain the optimal feedforward compensation parameters, including: During the task breaks of the welding robot, a fitness evaluation function for offline parameters is constructed with the comprehensive optimization objectives of minimizing the future tracking error between the expected end-effector trajectory and the predicted end-effector trajectory, minimizing control energy consumption, and maximizing motion smoothness. Based on the fitness evaluation function, the feedforward compensation strategy parameters of the servo system are optimized offline according to the fitness parameter optimization step size using a global optimization algorithm. When the fitness optimization convergence number is reached, the feedforward compensation strategy parameters corresponding to the maximum fitness are used as the optimal feedforward compensation parameters and downloaded to the online controller of the welding robot for use.

[0043] In this embodiment, the trajectory prediction model is invoked based on the servo control complexity and the adaptation parameter optimization strategy is set. The feedforward compensation parameters of the servo system are optimized according to the end load mass sequence, end welding posture sequence and end load rotational inertia sequence. The optimal feedforward compensation parameters are obtained to execute the feedforward compensation of the welding robot's servo system in the future time period.

[0044] Specifically, firstly, a pre-trained trajectory prediction model is developed, wherein the trajectory prediction model comprises P trajectory prediction units, where P is an integer greater than 10.

[0045] The pre-trained trajectory prediction model includes: Based on historical operation logs of similar welding robots, several sample end-effector load mass sequences, several end-effector welding posture sequences, several end-effector load rotational inertia sequences, and several historical end-effector motion trajectories were collected as sample end-effector motion trajectories to obtain sample training data. For example, for each historical welding task segment, the same method as described above was used to extract the sample end-effector load mass sequence, sample end-effector welding posture sequence, and sample end-effector load rotational inertia sequence recorded over time during execution. Simultaneously, the spatial position coordinate sequence of the end-effector corresponding to that task segment, fed back by the robot, was extracted as the sample end-effector motion trajectory. The sample training dataset was then integrated to obtain the final dataset.

[0046] Further, the sample training data is divided into P equal parts, and then subjected to P-fold cross-partitioning with replacement to obtain P sample training sets. For example, the total sample training data is randomly shuffled and uniformly divided into P subsets of approximately equal size. Then, P-fold cross-partitioning with replacement is performed: for the i-th subset (i ranges from 1 to P), it is used as the validation set for training the i-th trajectory prediction unit, while the remaining P-1 subsets are merged together as the training set for the i-th trajectory prediction unit. Because the partitioning is done with replacement, the training set for each unit contains most of the original data, but the specific composition is slightly different, enabling the training of differentiated prediction units.

[0047] Further, deep learning models are trained to convergence using the P training sets, resulting in P trajectory prediction units. For example, a deep learning framework is used to construct the trajectory prediction units. The input layer includes three nodes for receiving the end-load mass sequence, the end-welding posture sequence, and the end-load rotational inertia sequence. The hidden layer uses 128 neurons with ReLU activation. The number of output layer nodes corresponds to the position coordinate dimension of the end-load trajectory in the future prediction time domain, and the activation function is linear. During training, mean squared error is used as the loss function, the Adam optimizer is employed, and the initial learning rate is set to 0.001 with a batch size of 32. P trajectory prediction units are trained independently using the P training sets. Each model is trained until its loss on the corresponding validation set no longer significantly decreases, i.e., convergence. This yields P trained trajectory prediction units capable of independent trajectory prediction, constituting the trajectory prediction model.

[0048] Furthermore, the optimal number of unit calls, L, is obtained by rounding down the product of the servo control complexity and the initial number of selected units, K, where K equals 5. If the calculated L is less than 2, then L is set to 2; if the calculated L is greater than P, then L is set to P. For example, if the calculated servo control complexity is 1.2, then the optimal number of unit calls = 1.2 × 5 = 6. When the result is not an integer, it is rounded down. If the servo control complexity is small and the calculated L is less than 2, then L is set to 2 to ensure that at least two trajectory prediction units participate in the prediction. If the servo control complexity is large and the calculated L is greater than P, then L = P is set to call all trajectory prediction units for prediction to improve prediction accuracy.

[0049] Further, within the P trajectory prediction units, L trajectory prediction units are randomly selected to predict the robot's end effector trajectory based on the end effector load mass sequence, end effector welding posture sequence, and end effector load rotational inertia sequence. L initial predicted trajectories are output, and the mean of these trajectories is fitted to obtain the final predicted end effector trajectory. For example, from all P trajectory prediction units, L units are randomly selected without repetition using a random number generator. The end effector load mass sequence, end effector welding posture sequence, and end effector load rotational inertia sequence are respectively input into the L selected trajectory prediction units. Each unit independently performs forward calculation and prediction, outputting a prediction of the robot's end effector trajectory in Cartesian space within the same future time period—that is, an initial predicted trajectory. A total of L predicted trajectories are obtained. Finally, the mean of these L trajectories is fitted: at each future prediction time point, the arithmetic mean of the spatial coordinates (X, Y, Z) of the L trajectories at that time point is taken, and these averaged coordinate points are reconnected in chronological order to form a single trajectory. The result is a single final effector predicted trajectory, representing the comprehensive prediction result based on multi-model ensemble.

[0050] Further, the dispersion of the prediction result is obtained by calculating the trace quantization of the spatial position covariance matrix of the L initial predicted trajectories at each time point in the prediction time domain. Based on this dispersion, the trajectory prediction confidence is calculated using a preset monotonically decreasing mapping function, where the trajectory prediction confidence and the dispersion are negatively correlated. Specifically, first, the dispersion is calculated. For each time point in the future prediction time domain, the three-dimensional spatial positions of the L initial predicted trajectories at that time point are collected, forming a set containing L sample points. The three-dimensional spatial covariance matrix of this point set is calculated. Then, the trace of this covariance matrix is ​​calculated. This trace value quantifies the dispersion of the L predictions at that time point. The average of the trace values ​​at all time points in the prediction time domain is used to obtain the dispersion of the prediction result. Further, the trajectory prediction confidence is calculated. A preset monotonically decreasing mapping function is used, such as defining a linear function: Trajectory prediction confidence = 1 - Dispersion of prediction result / Dispersion threshold, where the dispersion threshold can be set based on the specific scenario. When the dispersion of the prediction result is less than the dispersion threshold, the trajectory prediction confidence is between 0 and 1; if the dispersion of the prediction result is greater than or equal to the dispersion threshold, the trajectory prediction confidence is 0. The obtained trajectory prediction confidence is used to represent the reliability of this trajectory prediction.

[0051] Furthermore, an adaptation parameter optimization strategy is set based on the servo control complexity and trajectory prediction confidence, and the feedforward compensation parameters of the servo system are optimized according to the predicted motion trajectory of the end effector to obtain the optimal feedforward compensation parameters.

[0052] Specifically, the optimization step size and the number of optimization convergence iterations are configured based on the servo control complexity and trajectory prediction confidence. The optimization step size is negatively correlated with the servo control complexity and positively correlated with the trajectory prediction confidence. Similarly, the number of optimization convergence iterations is positively correlated with the servo control complexity and negatively correlated with the trajectory prediction confidence. For example, a baseline optimization step size and a baseline number of optimization convergence iterations are set, such as setting the baseline optimization step size to 0.5% of the original control parameters. The baseline number of optimization convergence iterations is 50. These two parameters are further dynamically adjusted using the servo control complexity and trajectory prediction confidence. The optimization step size is calculated as: Optimization step size = Baseline optimization step size × (1 / Servo control complexity) × Trajectory prediction confidence. This means that when the servo control complexity is higher or the trajectory prediction confidence is lower, a smaller optimization step size is used for a more refined and cautious search. The number of optimization convergence iterations is equal to the number of baseline optimization convergence iterations multiplied by the servo control complexity multiplied by (1 / trajectory prediction confidence). When the servo control complexity is higher or the trajectory prediction confidence is lower, the maximum number of optimization iterations is set to allow for a more thorough search.

[0053] Furthermore, the expected end-effector trajectory corresponding to the preset welding task of the welding robot within a future time period is obtained. For example, the preset welding task within the future time period is analyzed, including the target position coordinate sequence of the robot's end effector at each moment, to obtain the expected end-effector trajectory. The expected end-effector trajectory represents the ultimate goal of control.

[0054] Furthermore, based on the optimization step size and the number of optimization convergence times of the adaptation parameters, the feedforward compensation parameters of the servo system are optimized according to the expected end motion trajectory and the predicted end motion trajectory to obtain the optimal feedforward compensation parameters.

[0055] Specifically, during the task breaks of the welding robot, a fitness evaluation function for offline parameters is constructed with the comprehensive optimization objectives of minimizing the future tracking error between the expected end-effector trajectory and the predicted end-effector trajectory, minimizing control energy consumption, and maximizing motion smoothness. First, the future tracking error term is calculated. This term is obtained by calculating the difference between the three-dimensional spatial coordinates of the expected end-effector trajectory and the predicted end-effector trajectory at each same time point, and then summing the squared differences at all time points. The smaller the future tracking error term, the higher the trajectory tracking accuracy. Second, the control energy consumption term is calculated. This term is approximated by squaring the predicted torque required by each joint in the future time period and then summing it over all joints and all time points. The smaller the control energy consumption term, the lower the energy consumption. Finally, the motion smoothness term is calculated. This term is represented by squaring the predicted acceleration of each joint in the future time period and then summing it over all joints and all time points. The smaller the motion smoothness term, the smoother the motion. Furthermore, importance weight coefficients are assigned to these three terms respectively, with the future tracking error term and control energy consumption term multiplied by a positive weight coefficient, and the motion smoothness term multiplied by a negative weight coefficient. Then, the three weighted terms are summed, and their negative values ​​are taken to obtain the fitness evaluation value. The higher the fitness evaluation value, the better a set of feedforward compensation parameters results in smaller future tracking errors, lower control energy consumption, and smoother motion.

[0056] Furthermore, based on the fitness evaluation function, a global optimization algorithm is used to offline optimize the feedforward compensation strategy parameters of the servo system according to the optimization step size of the adaptation parameters. When the optimization convergence count is reached, the feedforward compensation strategy parameters corresponding to the maximum fitness are used as the optimal feedforward compensation parameters and downloaded to the online controller of the welding robot. For example, a particle swarm optimization algorithm can be used as the global optimization algorithm. The feedforward compensation strategy parameters to be optimized are encoded as the position vectors of each particle in the particle swarm. These parameters may include, but are not limited to: adaptive mapping function coefficients, such as the slope and intercept of a linear function that maps servo control complexity to the predicted time domain length; dynamic weight baseline values ​​of the cost function, such as the initial scaling factor of the tracking error weight matrix. The parameters of the particle swarm algorithm are set, where the maximum number of iterations is set as the optimization convergence count. In each iteration of the algorithm, the velocity term of particle position updates is constrained by the optimization step size of the adaptation parameters to achieve fine control of the search range. During algorithm execution, each particle is substituted into the fitness evaluation function F to calculate its fitness value. The particle swarm continuously updates its parameters based on the historical best positions of individuals and the swarm, searching for better parameters. The search stops when the algorithm reaches the required number of iterations for optimal convergence. From all iteration histories, the parameter combination corresponding to the particle positions that maximizes the fitness function F is selected, obtaining the optimal feedforward compensation parameters. These optimal feedforward compensation parameters are transmitted and written into the corresponding parameter register of the online feedforward compensation module in the welding robot's servo drive. The online controller will then directly use these optimized parameters for feedforward compensation calculations when executing the preset welding task, equipping the welding robot's servo system with an optimal feedforward compensation strategy tailored to the upcoming specific load change task.

[0057] By dynamically invoking the trajectory prediction model and setting an adaptive parameter optimization strategy based on servo control complexity, the optimization and execution of feedforward compensation parameters are ultimately completed, achieving intelligent and adaptive compensation strategy generation. Specifically, based on complexity scores, the system intelligently determines how many and what set of trajectory prediction units to invoke for collaborative prediction of future motion trajectories, ensuring that the computational resource investment in the prediction stage is positively correlated with the task difficulty. Simultaneously, combined with the confidence level of the prediction results, the step size and convergence count for parameter optimization are further dynamically configured: for high-complexity, low-confidence tasks, a cautious optimization with smaller step sizes and more convergence counts is adopted; for low-complexity, high-confidence tasks, a more efficient optimization with larger step sizes and faster convergence can be used. This dynamic adaptation mechanism enables the system to efficiently and accurately optimize and obtain the optimal feedforward compensation parameter sequence that best matches the load change sequence for a specific future time period, guided by minimizing tracking errors and energy consumption and ensuring smoothness, both offline and during task intervals.

[0058] Example 2, as Figure 2As shown, based on the same inventive concept as the model prediction-driven servo system feedforward compensation method provided in Embodiment 1, this embodiment of the invention also provides a model prediction-driven servo system feedforward compensation platform, including: The load analysis module 100 is used to determine the end load mass sequence and end welding posture sequence of the robot end based on the preset welding tasks of the welding robot in the future time period, and to analyze and determine the end load rotational inertia sequence. The complexity assessment module 200 is used to assess and determine the servo control complexity of the servo system based on the end load mass sequence, end welding posture sequence and end load rotational inertia sequence. The parameter optimization module 300 is used to call the trajectory prediction model and set the adaptive parameter optimization strategy based on the servo control complexity, optimize the feedforward compensation parameters of the servo system according to the end load mass sequence, end welding posture sequence and end load rotational inertia sequence, obtain the optimal feedforward compensation parameters and execute the feedforward compensation of the welding robot's servo system in the future time period.

[0059] In one embodiment, the load analysis module 100 is further configured to: Based on the preset welding task, determine several end welding postures of the welding robot at several welding points in the future time period, and generate an end welding posture sequence; Monitor and acquire the current end-effector load quality of the welding robot; The amount of welding wire consumed by the welding robot at several welding points in a future time period is determined based on the preset welding task. The end load mass sequence is calculated based on the current end load mass and several welding wire consumptions; Based on the end load mass sequence, by establishing a mass distribution evolution model during the welding wire consumption process, the total mass change is mapped to a continuously changing parameter of the mass geometric distribution, thus obtaining a mass geometric distribution parameter sequence; In each control cycle, based on the current end load mass, mass geometric distribution parameters, and end welding posture, the equivalent rotational inertia tensor of the load about each joint axis is calculated, and the calculated rotational inertia is subjected to time-series filtering and interpolation to generate a smooth end load rotational inertia sequence.

[0060] In one embodiment, the complexity evaluation module 200 is further configured to: The terminal load change quality sequence is calculated based on the terminal load quality sequence, wherein the terminal load change quality is the absolute value of the terminal load quality deviation between adjacent nodes; The end load rotational inertia sequence is calculated based on the end load rotational inertia sequence, wherein the end load rotational inertia change sequence is the absolute value of the deviation of the end load rotational inertia between adjacent nodes. Wave behavior analysis was performed on the end load change mass sequence and the end load rotation change inertia sequence, respectively, and the coefficient of variation of the load change mass and the coefficient of variation of the load rotation change inertia were calculated. Based on the aforementioned end-welding posture sequence, the norms of the angular velocity and linear acceleration of the end-tool coordinate system in Cartesian space are calculated to obtain the posture change rate. The servo control complexity of the servo system is determined by evaluating the coefficient of variation of the load change mass, the coefficient of variation of the load rotation change inertia, and the attitude change rate. The servo control complexity of the servo system is determined based on the coefficient of variation of the load change mass, the coefficient of variation of the load rotational inertia, and the attitude change rate, including: The ratio of the load variation quality variation coefficient to the reference load variation quality variation coefficient is used as the first servo control complexity coefficient. The ratio of the coefficient of variation of the load rotational inertia to the coefficient of variation of the reference load rotational inertia is used as the second servo control complexity coefficient. The ratio of the attitude change rate to the reference attitude change rate is used as the third servo control complexity coefficient. The servo control complexity of the servo system is obtained by weighted evaluation based on the first servo control complexity coefficient, the second servo control complexity coefficient, and the third servo control complexity coefficient.

[0061] In one embodiment, the parameter optimization module 300 is further configured to: A pre-trained trajectory prediction model, wherein the trajectory prediction model comprises P trajectory prediction units, where P is an integer greater than 10; The pre-trained trajectory prediction model includes: Based on the historical operation logs of similar welding robots, several sample end-load mass sequences, several end-welding posture sequences, several end-load rotational inertia sequences, and several historical end-motion trajectories were collected as several sample end-motion trajectories to obtain sample training data. The sample training data is divided into P equal parts, and then P-fold cross-partitioned with replacement is performed to obtain P sample training sets. The deep learning model is trained to convergence using the P sample training sets to obtain P trajectory prediction units; The optimal number of unit calls L is obtained by rounding down the product of the servo control complexity and the initial number of unit selections K, where K equals 5. If the calculated L is less than 2, then L is equal to 2; if the calculated L is greater than P, then L is equal to P. Within the P trajectory prediction units, L trajectory prediction units are randomly called to predict the robot's end-effector trajectory based on the end-effector load mass sequence, end-effector welding posture sequence, and end-effector load rotational inertia sequence. L initial predicted trajectories are output, and the end-effector predicted motion trajectory is obtained by fitting the trajectory mean. The dispersion of the prediction result is obtained by quantizing the trace of the spatial position covariance matrix of the L initial predicted trajectories at each time point in the prediction time domain. Based on the dispersion, the trajectory prediction confidence is calculated by a preset monotonically decreasing mapping function, wherein the trajectory prediction confidence and the dispersion are negatively correlated. Based on the servo control complexity and trajectory prediction confidence, an adaptation parameter optimization strategy is set, and the feedforward compensation parameters of the servo system are optimized according to the predicted motion trajectory of the end effector to obtain the optimal feedforward compensation parameters. The process includes setting an adaptation parameter optimization strategy based on the servo control complexity and trajectory prediction confidence, optimizing the feedforward compensation parameters of the servo system according to the predicted end-effector motion trajectory, and obtaining the optimal feedforward compensation parameters, including: The adaptation parameter optimization step size and the number of adaptation optimization convergences are configured based on the servo control complexity and trajectory prediction confidence. The adaptation parameter optimization step size is negatively correlated with the servo control complexity and positively correlated with the trajectory prediction confidence. The number of adaptation optimization convergences is positively correlated with the servo control complexity and negatively correlated with the trajectory prediction confidence. Obtain the expected end-effector trajectory corresponding to the preset welding task of the welding robot in the future time period; Based on the optimization step size and the number of optimization convergence times of the adaptation parameters, the feedforward compensation parameters of the servo system are optimized according to the expected end motion trajectory and the predicted end motion trajectory to obtain the optimal feedforward compensation parameters. Specifically, based on the adaptation parameter optimization step size and the number of adaptation optimization convergences, the feedforward compensation parameters of the servo system are optimized according to the expected end-effector trajectory and the predicted end-effector trajectory to obtain the optimal feedforward compensation parameters, including: During the task breaks of the welding robot, a fitness evaluation function for offline parameters is constructed with the comprehensive optimization objectives of minimizing the future tracking error between the expected end-effector trajectory and the predicted end-effector trajectory, minimizing control energy consumption, and maximizing motion smoothness. Based on the fitness evaluation function, the feedforward compensation strategy parameters of the servo system are optimized offline according to the fitness parameter optimization step size using a global optimization algorithm. When the fitness optimization convergence number is reached, the feedforward compensation strategy parameters corresponding to the maximum fitness are used as the optimal feedforward compensation parameters and downloaded to the online controller of the welding robot for use.

[0062] In summary, the embodiments of this application have at least the following technical effects: This application proposes a model prediction-driven servo system feedforward compensation method and platform. Before the start of a future welding task, it proactively determines the evolution sequence of the end-load mass, end-welding posture, and end-load rotational inertia that will inevitably result from the consumption of welding wire in the future time period based on the preset welding task. Based on this, it quantitatively evaluates the dynamic complexity of the servo control process and then intelligently calls predictive resources and configuration parameter optimization strategies to perform accurate and adaptive offline optimization of feedforward compensation parameters. This significantly improves the trajectory tracking accuracy and motion control stability of the welding robot under time-varying load conditions. Specifically, by sequentially predicting load parameters, the compensator can anticipate and adapt to future load change trends, thereby injecting a more suitable compensation amount at the control source. By introducing servo control complexity assessment, a quantitative perception of task difficulty is achieved, enabling dynamic adjustment of the model set size used for trajectory prediction and strategies such as search step size and convergence depth for parameter optimization based on the actual challenge, optimizing the allocation efficiency of computational resources while ensuring prediction accuracy. Furthermore, combining confidence assessment based on multi-model prediction results provides reliability guidance for the optimization process, ensuring more prudent parameter adjustment strategies are adopted when prediction uncertainty is high, thus enhancing the system's robustness. Compared to traditional methods, the method provided in this application achieves the technical effect of enabling the servo system's feedforward compensation to evolve synchronously with the dynamic load changes strongly coupled with the welding task. Through smooth and adaptive compensation, unnecessary torque fluctuations and mechanical vibrations are suppressed, which is beneficial for extending the mechanical life of the robot and reducing servo drive energy consumption, providing a more intelligent, reliable, and adaptable core control guarantee for automated welding production lines.

[0063] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0064] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0065] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A model prediction-driven feedforward compensation method for a servo system, characterized in that, The methods include: Based on the preset welding tasks of the welding robot in the future time period, determine the end load mass sequence and end welding posture sequence of the robot end, and analyze and determine the end load rotational inertia sequence. The servo control complexity of the servo system is evaluated and determined based on the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence. Based on the aforementioned servo control complexity, the trajectory prediction model is invoked and the adaptation parameter optimization strategy is set. The feedforward compensation parameters of the servo system are optimized according to the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence. The optimal feedforward compensation parameters are obtained to execute the feedforward compensation of the welding robot's servo system in the future time period.

2. The model prediction-driven feedforward compensation method for a servo system according to claim 1, characterized in that, The robot's end-effector load mass sequence and end-effector welding posture sequence are determined based on the preset welding task, including: Based on the preset welding task, determine several end welding postures of the welding robot at several welding points in the future time period, and generate an end welding posture sequence; Monitor and acquire the current end-effector load quality of the welding robot; The amount of welding wire consumed by the welding robot at several welding points in a future time period is determined based on the preset welding task. The end load quality sequence is calculated based on the current end load quality and the consumption of several welding wires.

3. The model prediction-driven feedforward compensation method for a servo system according to claim 1, characterized in that, The analysis determines the sequence of rotational inertia of the end load, including: Based on the end load mass sequence, by establishing a mass distribution evolution model during the welding wire consumption process, the total mass change is mapped to a continuous change parameter of the mass geometric distribution, thus obtaining a mass geometric distribution parameter sequence; In each control cycle, based on the current end load mass, mass geometric distribution parameters, and end welding posture, the equivalent rotational inertia tensor of the load about each joint axis is calculated, and the calculated rotational inertia is subjected to time-series filtering and interpolation to generate a smooth end load rotational inertia sequence.

4. The model prediction-driven feedforward compensation method for a servo system according to claim 1, characterized in that, The servo control complexity of the servo system is evaluated and determined based on the end-load mass sequence, end-welding posture sequence, and end-load rotational inertia sequence, including: The terminal load change quality sequence is calculated based on the terminal load quality sequence, wherein the terminal load change quality is the absolute value of the terminal load quality deviation between adjacent nodes; The end load rotational inertia sequence is calculated based on the end load rotational inertia sequence, wherein the end load rotational inertia change sequence is the absolute value of the deviation of the end load rotational inertia between adjacent nodes. Wave behavior analysis was performed on the end load change mass sequence and the end load rotation change inertia sequence, respectively, and the coefficient of variation of the load change mass and the coefficient of variation of the load rotation change inertia were calculated. Based on the end-welding posture sequence, the norms of the angular velocity and linear acceleration of the end-tool coordinate system in Cartesian space are calculated to obtain the posture change rate. The servo control complexity of the servo system is determined based on the coefficient of variation of the load change mass, the coefficient of variation of the load rotation inertia, and the attitude change rate.

5. The model prediction-driven feedforward compensation method for a servo system according to claim 4, characterized in that, The servo control complexity of the servo system is determined based on the coefficient of variation of the load change mass, the coefficient of variation of the load rotational inertia, and the attitude change rate, including: The ratio of the load variation quality variation coefficient to the reference load variation quality variation coefficient is used as the first servo control complexity coefficient. The ratio of the coefficient of variation of the load rotational inertia to the coefficient of variation of the reference load rotational inertia is used as the second servo control complexity coefficient. The ratio of the attitude change rate to the reference attitude change rate is used as the third servo control complexity coefficient. The servo control complexity of the servo system is obtained by weighted evaluation based on the first servo control complexity coefficient, the second servo control complexity coefficient, and the third servo control complexity coefficient.

6. The model prediction-driven feedforward compensation method for a servo system according to claim 1, characterized in that, Based on the aforementioned servo control complexity, a trajectory prediction model is invoked and an adaptation parameter optimization strategy is set. The feedforward compensation parameters of the servo system are optimized according to the end-load mass sequence, end-welding posture sequence, and end-load rotational inertia sequence to obtain the optimal feedforward compensation parameters, including: A pre-trained trajectory prediction model, wherein the trajectory prediction model comprises P trajectory prediction units, where P is an integer greater than 10; The optimal number of unit calls L is obtained by rounding down the product of the servo control complexity and the initial number of unit selections K, where K equals 5. If the calculated L is less than 2, then L is equal to 2; if the calculated L is greater than P, then L is equal to P. Within the P trajectory prediction units, L trajectory prediction units are randomly called to predict the robot's end-effector trajectory based on the end-effector load mass sequence, end-effector welding posture sequence, and end-effector load rotational inertia sequence. L initial predicted trajectories are output, and the end-effector predicted motion trajectory is obtained by fitting the trajectory mean. The dispersion of the prediction result is obtained by quantizing the trace of the spatial position covariance matrix of the L initial predicted trajectories at each time point in the prediction time domain. Based on the dispersion, the trajectory prediction confidence is calculated by a preset monotonically decreasing mapping function, wherein the trajectory prediction confidence and the dispersion are negatively correlated. Based on the servo control complexity and trajectory prediction confidence, an adaptation parameter optimization strategy is set, and the feedforward compensation parameters of the servo system are optimized according to the predicted motion trajectory of the end effector to obtain the optimal feedforward compensation parameters.

7. The model prediction-driven feedforward compensation method for a servo system according to claim 6, characterized in that, Pre-trained trajectory prediction models include: Based on the historical operation logs of similar welding robots, several sample end-load mass sequences, several end-welding posture sequences, several end-load rotational inertia sequences, and several historical end-motion trajectories were collected as several sample end-motion trajectories to obtain sample training data. The sample training data is divided into P equal parts, and then P-fold cross-partitioned with replacement is performed to obtain P sample training sets. The deep learning model is trained to convergence using the P sample training sets to obtain P trajectory prediction units.

8. The model prediction-driven feedforward compensation method for a servo system according to claim 6, characterized in that, Based on the servo control complexity and trajectory prediction confidence, an adaptation parameter optimization strategy is set. The feedforward compensation parameters of the servo system are optimized according to the predicted end-effector trajectory to obtain the optimal feedforward compensation parameters, including: The adaptation parameter optimization step size and the number of adaptation optimization convergences are configured based on the servo control complexity and trajectory prediction confidence. The adaptation parameter optimization step size is negatively correlated with the servo control complexity and positively correlated with the trajectory prediction confidence. The number of adaptation optimization convergences is positively correlated with the servo control complexity and negatively correlated with the trajectory prediction confidence. Obtain the expected end-effector trajectory corresponding to the preset welding task of the welding robot in the future time period; Based on the optimization step size and the number of optimization convergence times of the adaptation parameters, the feedforward compensation parameters of the servo system are optimized according to the expected end motion trajectory and the predicted end motion trajectory to obtain the optimal feedforward compensation parameters.

9. A model prediction-driven feedforward compensation method for a servo system according to claim 8, characterized in that, Based on the optimization step size and the number of optimization convergence times according to the adaptation parameters, the feedforward compensation parameters of the servo system are optimized according to the expected end-effector trajectory and the predicted end-effector trajectory to obtain the optimal feedforward compensation parameters, including: During the task breaks of the welding robot, a fitness evaluation function for offline parameters is constructed with the comprehensive optimization objectives of minimizing the future tracking error between the expected end-effector trajectory and the predicted end-effector trajectory, minimizing control energy consumption, and maximizing motion smoothness. Based on the fitness evaluation function, the feedforward compensation strategy parameters of the servo system are optimized offline according to the fitness parameter optimization step size using a global optimization algorithm. When the fitness optimization convergence number is reached, the feedforward compensation strategy parameters corresponding to the maximum fitness are used as the optimal feedforward compensation parameters and downloaded to the online controller of the welding robot for use.

10. A model prediction-driven feedforward compensation platform for a servo system, characterized in that, For implementing the model prediction-driven feedforward compensation method for a servo system according to any one of claims 1-9, the platform comprises: The load analysis module is used to determine the end-effector load mass sequence and end-effector welding posture sequence of the robot end based on the preset welding tasks of the welding robot in the future time period, and to analyze and determine the end-effector load rotational inertia sequence. The complexity assessment module is used to assess and determine the servo control complexity of the servo system based on the end load mass sequence, end welding posture sequence, and end load rotational inertia sequence. The parameter optimization module is used to call the trajectory prediction model and set the adaptive parameter optimization strategy based on the servo control complexity. It optimizes the feedforward compensation parameters of the servo system according to the end load mass sequence, end welding posture sequence and end load rotational inertia sequence, and obtains the optimal feedforward compensation parameters to execute the feedforward compensation of the welding robot's servo system in the future time period.

Images