Hydraulic mechanical arm cross load position compensation method and system

Abstract

The application discloses a hydraulic mechanical arm cross-load end position compensation method and system. A kinematics model corrected by geometric parameters is acquired, the mechanical arm state and the measured position of an end reference point under source and target load working conditions are collected, a difference between the measured position and a theoretical position is used to construct source domain and target domain training data sets, and TrAdaBoost.R2 migration regression models are trained respectively for three-axis position errors. During compensation, first inverse kinematics solving is performed on an expected end position, three-dimensional position error prediction values are obtained by using three position error component prediction models and the expected end position is reversely corrected, second inverse kinematics solving is performed, the obtained corrected mechanical arm state is converted into hydraulic cylinder target displacement and closed loop execution is performed. The system comprises a hydraulic mechanical arm, a controller, an end position measurement component and a data processing unit. The application can realize cross-load compensation by using a small amount of target load samples and existing calibration data, and reduces the workload of re-calibration.

Description

Technical Field

[0001] This invention relates to the fields of hydraulic robot positioning control, error compensation, and transfer learning, and in particular to a method and system for establishing a three-dimensional position error prediction model based on calibrated load condition data and a small number of samples of target load condition, and realizing end-effector position compensation of a hydraulic robotic arm through target position pre-correction, secondary inverse kinematics solution, and hydraulic cylinder displacement closed-loop control. Background Technology

[0002] Hydraulic robotic arms are characterized by high output force, high power density, and strong impact resistance. They can achieve large load capacity with a relatively compact structure, making them suitable for scenarios such as handling large-weight components, engineering construction, hazardous environment operations, and high-load assembly. When the end-effector load changes, the connecting rods, pins, hydraulic cylinders, and connecting structures will undergo varying degrees of elastic deformation. Hydraulic oil compressibility, valve nonlinearity, off-center friction, system internal leakage, and joint hysteresis will also introduce positional deviations related to the load and configuration.

[0003] The positioning errors of hydraulic robotic arms can be divided into geometric errors and non-geometric errors. Geometric errors, such as manufacturing and assembly errors, joint zero-position errors, and linkage parameter deviations, can be corrected by identifying kinematic parameters. Non-geometric errors, such as load deformation, friction, hydraulic hysteresis, oil temperature changes, and pressure changes, have obvious nonlinear and mechanical-hydraulic coupling characteristics, and are difficult to accurately characterize with a set of fixed geometric parameters in all load conditions and workspaces.

[0004] Existing data-driven compensation methods typically require collecting a large amount of end-effector position calibration data under the load being studied. When the load on the robotic arm's end effector changes, high-density sampling needs to be performed again, resulting in long measurement time and limitations imposed by obstructions, vibrations, and measurement accessibility during the experimental process. Therefore, the key to reducing calibration costs lies in how to utilize existing calibration data under existing load conditions and establish a position error model by combining only a small number of samples under the target load condition.

[0005] Transfer learning enables the use of calibrated source load data for modeling target load conditions. However, simply mixing different load data can lead to interference from samples in the source domain that do not conform to the target domain's patterns, thus hindering target load prediction. Transfer regression methods such as TrAdaBoost.R2 iteratively adjust the weights of samples in the source and target domains, gradually reducing the weight of source domain samples that are unsuitable for the target load condition while giving more attention to samples in the target domain that are difficult to fit. This approach is suitable for modeling cross-load position errors in hydraulic robotic arms.

[0006] Furthermore, the position error output by the transfer model cannot be directly used as a hydraulic actuator command. The position error needs to be converted into a corrected target position according to a unified error direction, and then converted into a target quantity for the joint or hydraulic cylinder through inverse kinematics and hydraulic cylinder geometry. Under measurement conditions with only a single spherical reflective target, the laser tracker can directly obtain the three-dimensional position of the target's center, but the end-effector posture cannot be uniquely determined from the measurement result of a single target. Therefore, the target's center can be defined as the end-effector measurement reference point, and a three-dimensional position error model can be established around this reference point to ensure that the model output is consistent with the actual available measurement data.

[0007] Therefore, it is necessary to provide an end-effector position compensation method and system that combines the actual position measurement data of the hydraulic manipulator, the cross-load migration regression algorithm, and the hydraulic execution constraints. Summary of the Invention

[0008] The purpose of this invention is to provide a method and system for end-effector position compensation of a hydraulic robotic arm based on cross-load migration prediction, in order to solve the problems of requiring a large amount of recalibration of the target load condition, the easy interference caused by direct mixing of source load data, and the difficulty in converting model prediction results into executable hydraulic control commands.

[0009] To achieve the above objectives, the present invention adopts the following technical solution.

[0010] In a first aspect, the present invention provides a method for end-effector position compensation of a hydraulic robotic arm, including position data construction, cross-load migration training, position error prediction, target position reverse correction, quadratic inverse kinematics solution, and hydraulic cylinder displacement closed-loop execution.

[0011] A fixed relationship is established between the base coordinate system of the robotic arm and the coordinate system of the position measuring device, and a kinematic model corrected by geometric parameters is obtained. The center of a single spherical reflective target rigidly mounted at the end of the robotic arm is directly defined as the end-effector reference point.

[0012] Under one or more calibrated source load conditions and target load conditions, the robotic arm is driven to a preset sampling state. After the hydraulic cylinder position stabilizes, the robotic arm state, load condition identifier, and measured position of the end effector reference point are recorded simultaneously. The robotic arm state is then input into a kinematic model corrected for geometric parameters to obtain the theoretical position of the end effector reference point.

[0013] Let the forward kinematics model after geometric parameter correction be: , No. The state of the robotic arm for each sample is as follows: The corresponding theoretical position is The actual measured location is Define three-dimensional position error : (1) The source load condition samples constitute the source domain dataset. The target load condition samples are divided into target domain training datasets. and independent test dataset: (2) (3) in, is the state feature vector of the robotic arm; y is the currently trained position error component. For and Establish independent transfer regression models separately and : (4) (5) (6) The three models use the same source and target domain training sets and independent test sets, but calculate sample loss and update sample weights based on the corresponding error components. After training, they are combined to obtain the predicted 3D position error value. (7) In the preferred embodiment, the TrAdaBoost.R2 transfer regression model is used. Source domain samples are arranged before target domain training samples to form a joint training dataset. Let the number of source domain samples be... The number of training samples in the target domain is The preset maximum number of iterations is Set the same initial weights for all joint training samples: (8) In the In each iteration, weighted sampling with replacement is performed from the joint training dataset according to the current sample weights. The number of samples is the same as the number of joint training samples, and the sampled samples are used to train the current weak regressor. Apply the current weak regressor to all joint training samples and calculate the... Absolute prediction error of joint training samples The maximum absolute prediction error across all joint training samples. This serves as the baseline for error normalization in this round. The normalized prediction error for all jointly trained samples is calculated. : (9) (10) (11) when When the prediction error of the current weak regressor for all joint training samples is zero, the current weak regressor is determined as the final prediction model for the corresponding position error component and the training ends.

[0014] when At that time, based on the current sample weights and normalized prediction errors of all joint training samples, the weighted error of the current weak regressor on the target domain training samples is calculated: (12) when When the current weak regressor is determined as the final prediction model for the corresponding position error component, training ends; when Discard the current weak regressor and stop iterative training; if at least one effective weak regressor has been retained before stopping iterative training, then perform subsequent fusion based on the retained effective weak regressors; if no effective weak regressors have been retained, then train the preset backtracking regressor using only the target domain training data; when When calculating the error coefficient of the weak regressor and fusion weight : (13) Based on the number of source domain samples and preset maximum number of iterations Calculate the source domain attenuation coefficient : (14) Calculate the unnormalized weights for the next round of samples for both the source domain and the target domain training samples. For the source domain samples, the weights are updated as follows: (15) For the training samples in the target domain, the weights are updated as follows: (16) Calculate the weight normalization factor for this round: (17) And obtain the normalized sample weights for the next round: (18) because and For source domain samples with larger normalized prediction errors, the weight decay is more significant; for target domain training samples with larger normalized prediction errors, the relative weight is higher. Repeatedly perform weighted sampling, weak regressor training, error calculation, and weight updates until the preset maximum number of iterations is reached or the early stopping condition is met.

[0015] Let the number of effective weak regressors retained at the end of training be . .when When, the predicted value of the effective weak regressor is directly used. When At that time, the effective weak regressors are arranged in order of iteration, and the effective weak regressors in the latter half are selected to participate in the prediction. For the state of the robotic arm to be predicted... The predicted values ​​of the selected effective weak regressors are arranged in ascending order, and then fused according to the corresponding weights. Calculate the cumulative weight; take the predicted value corresponding to the first time the cumulative weight reaches half of the total weight of the selected effective weak regressor as the final predicted value of the corresponding position error component, that is, use the weighted median to fuse the weak regressors.

[0016] During compensation, the desired end position of the target ball's center is received. And the reference manipulator state used for inverse kinematics selection. With the sixth joint maintaining a preset fixed state, the desired end effector position is achieved using the end effector reference point. The first inverse kinematics solution is performed as a position constraint; when multiple feasible inverse solutions satisfying the position constraint exist, the candidate robot arm state is determined based on the reference robot arm state, joint limits, and hydraulic cylinder stroke constraints. The candidate robotic arm state Enter them separately and The predicted value of the three-dimensional position error is obtained. Calculate the correction command position according to the error definition direction: (19) While maintaining the preset fixed state at the sixth joint, the correction command position is reached using the end reference point. The second inverse kinematics solution is performed as a position constraint. When multiple feasible inverse solutions satisfying the position constraint exist, the one that best matches the candidate robot arm state is selected. The corrected robotic arm state is obtained by solving the inverse equation of equations that have small changes and satisfy joint limits and hydraulic cylinder stroke. The first and second inverse kinematics solutions do not take the complete 3D end-effector pose as a necessary objective constraint; this method only compensates for the 3D position of the end-effector reference point.

[0017] The corrected robotic arm state obtained from the second inverse kinematics solution is converted into the target rotation angle of the first joint and the target displacement of the hydraulic cylinders corresponding to the second to fifth joints, and the target displacement of the hydraulic cylinders is sent to the hydraulic robotic arm controller. The hydraulic robotic arm controller performs closed-loop control based on the feedback from the displacement sensors of each hydraulic cylinder, and drives the hydraulic robotic arm to complete the end position compensation.

[0018] Secondly, the present invention provides an end-effector position compensation system, comprising a hydraulic robotic arm, a hydraulic robotic arm controller, a rotary encoder for a rotary hydraulic support, a hydraulic cylinder displacement sensor, a rotary encoder for a swing hydraulic cylinder, an end-effector position measurement component, and a data processing unit. The end-effector position measurement component includes a single spherical reflective target rigidly mounted at the end of the robotic arm and a laser tracker; the data processing unit performs dataset construction, transfer model training, two inverse kinematics operations, target position correction, and feasibility checks; and the hydraulic robotic arm controller performs closed-loop control of the hydraulic cylinder displacement.

[0019] Thirdly, the present invention provides a computer device in which a program for performing the dataset construction, transfer model training and location compensation methods is stored in its memory, and the processor executes the program to implement the above method steps.

[0020] This invention uses the location error data of the calibrated source load condition to train a small number of samples of the target load condition, so that the existing calibration data can participate in the modeling of the new load condition, thereby reducing the workload of re-collecting high-precision location tags for the target load condition.

[0021] This invention establishes a single-output migration regression model for each of the three position error components, which is compatible with the data structure and regressor interface of existing programs, thus avoiding the direct mixing of different error components into a single loss that lacks clear physical meaning.

[0022] This invention reduces the weight of high-error source domain samples and increases the weight of high-error target domain samples during training. It also uses a weighted median fusion with a weak regressor in the second half of the training process. This helps to reduce the impact of source samples that are inconsistent with the target load pattern and outliers of a single weak regressor on the final prediction.

[0023] This invention obtains the candidate robotic arm state corresponding to the target operation point through the first inverse kinematics, then corrects the target position in reverse according to the prediction error and performs the second inverse kinematics solution, so that the data-driven prediction result can be converted into the actual executable joint state and hydraulic cylinder target displacement.

[0024] This invention defines the center of a single spherical reflective target rigidly mounted at the end of a robotic arm as the end reference point, so that the output of the three-dimensional position error model is consistent with the measurement directly obtained by the laser tracker, without making inferences about the end attitude error that has not been measured by the single target.

[0025] This invention performs inverse kinematics existence checks, joint limit checks, and checks on the slewing bearing and hydraulic cylinder stroke before output. Attached Figure Description

[0026] Figure 1 This is a flowchart of the end-position compensation method of the present invention; Figure 2This is a schematic diagram of the end target ball position measuring device and system composition of the present invention; Figure 3 This is a schematic diagram of the three-axis position error migration prediction and evaluation process of the present invention; Figure 4 This is a flowchart of the training process for the TrAdaBoost.R2 transfer regression model of this invention. Figure 5 This is a flowchart of the end-position error prediction and compensation process of the present invention; Figure 6 This is a functional block diagram of the end-position compensation system. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. These embodiments are used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0028] Please see Figure 1 As shown, this embodiment provides a method for end-effector position compensation across loads in a hydraulic robotic arm. As a preferred embodiment of the present invention, taking an end-effector position compensation experiment of a six-degree-of-freedom tandem hydraulic robotic arm under different load conditions as an example, the method includes the following steps: Step S1: Build an experimental platform consisting of a six-degree-of-freedom serial hydraulic manipulator, a laser tracker, and a host computer. Install a single spherical reflective target at the end of the manipulator, establish the communication and measurement relationship between the manipulator controller, the host computer, and the laser tracker, and calibrate the manipulator's base coordinate system. Step S2: A kinematic model of a six-degree-of-freedom hydraulic manipulator is established using the improved DH method, the conversion relationship between hydraulic cylinder displacement and corresponding joint angle is established, and the kinematic model is corrected using the geometric parameter identification results. Step S3: Collect the state variables of the robotic arm and the actual three-dimensional position coordinates of the end effector reference point under multiple preset source load conditions and target load conditions. Calculate the theoretical three-dimensional position coordinates of the end effector reference point based on the corrected kinematic model, and obtain the end effector reference point along... Position error in three coordinate directions; Step S4: The data collected under the multiple source load conditions are used to form a source domain dataset, a portion of the data under the target load condition is used to form a target domain training dataset, and the remaining data under the target load condition that did not participate in model training are used to form an independent test dataset. Step S5, respectively with The position errors in the three coordinate directions are used as output labels to train the transfer regression model. Through weighted sampling, weak regressor training, sample weight updating, and weak regressor fusion, the following results are obtained: The predicted position errors in the three coordinate directions are combined to form the predicted three-dimensional load position error. Step S6: Perform the first inverse kinematics solution on the desired end position, input the obtained robotic arm state into three position error prediction models, use the predicted three-dimensional position error to correct the desired end position, and then perform the second inverse kinematics solution on the corrected desired end position to obtain the final joint command value and drive the robotic arm to move. Step S7: Under the target load condition, for the same verification target point, control commands are generated using the initial kinematic model before parameter identification, the kinematic model after geometric parameter correction, and the joint compensation method composed of the kinematic model after geometric parameter correction and the load position error prediction model, respectively; the actual three-dimensional position coordinates of the end reference point of the robotic arm are measured respectively, and the end position errors corresponding to the three methods are compared.

[0029] Specifically, step S1 includes: Please see Figure 2 As shown, the experimental platform mainly consists of a six-degree-of-freedom tandem hydraulic manipulator, an API R-20RADIAN series laser tracker, and a host computer. The host computer is equipped with SA measurement software and manipulator motion control software; the SA measurement software is used to control the laser tracker and collect measurement data, while the manipulator motion control software is used to issue manipulator motion commands and record the actual three-dimensional position coordinates of the end-effector reference point measured by the laser tracker.

[0030] A stud connector is installed on the flange at the output end of the robotic arm's end-effector cylinder. The stud is used to install end loads of different specifications. An M6 threaded hole is machined at the center of the stud end for fixing the target ball holder. The target ball holder fixes a 1.5-inch SMR target ball using a strong magnet.

[0031] The hydraulic robotic arm controller communicates with a host computer via a CAN bus. Researchers input desired joint variables, such as the linear displacement of the hydraulic cylinder and the rotation angle of the hydraulic motor, into the robotic arm motion control software. The host computer encapsulates these desired joint variables into data frames with corresponding CAN IDs according to a preset communication protocol and sends them to the controller. Upon receiving the motion command, the controller uses the desired joint variables as setpoints, collects the actual motion values ​​of each joint in real time, and runs a closed-loop control algorithm internally. It calculates the control quantity based on the deviation between the desired and actual joint variables and outputs control signals to the corresponding proportional valves to drive the robotic arm's movement. During the motion, the controller transmits the actual motion values ​​of each joint back to the host computer via the CAN bus for display and recording. All joint closed-loop control calculations are completed internally by the controller; the host computer is used to send desired joint variables and display and record joint motion data.

[0032] Before collecting end-effector position data, the robot arm's base coordinate system is first calibrated. The target ball is rigidly mounted on the target ball mount at the end of the robot arm. Joint 1 is rotated, and the trajectory of the target ball's center is measured sequentially using a laser tracker. The obtained trajectory points are then fitted with a spatial circle to obtain the axis of the fitted circle. And use it as the base coordinate system of the robotic arm. Axial direction. Control the rotation of joint 2 and obtain the fitted circular axis using the same method. And use it as the base coordinate system of the robotic arm. The initial orientation of the axis. Then, the target ball is placed sequentially at at least three non-collinear positions on the robotic arm's rotary platform, and the coordinates of the target ball's center are measured and fitted to the spatial plane. Using this plane as the base coordinate system of the robotic arm Plane, with axis With plane The intersection point is taken as the origin of the coordinate system. The coordinate system is then used along the axis. The vertical upward direction is The positive direction of the axis; based on the positive rotation of joint 2, the axis is determined according to the right-hand rule. In the positive direction, that is, when viewed from the axis, the positive rotation of joint 2 is counterclockwise. To ensure that the coordinate axes are orthogonal, according to... The initial positive direction of the axis is Determining the cross product in the positive direction of the axis A base coordinate system for the robotic arm, conforming to the right-hand rule, is thus established along the positive axis.

[0033] Specifically, step S2 includes: A modified DH method was used to establish the kinematic model of a six-degree-of-freedom hydraulic manipulator. For the hydraulic cylinder-driven joints, the transformation relationship between the hydraulic cylinder displacement and the joint angle was established using triangular geometric relationships based on the hydraulic cylinder installation dimensions and the link length. By multiplying the transformation matrices of adjacent link coordinate systems, the theoretical three-dimensional position of the end effector reference point relative to the manipulator's base coordinate system was obtained.

[0034] Before collecting load error data, the kinematic model is corrected using the geometric parameter identification results. The geometric error compensation process includes establishing the relationship between end-position error and structural parameter error based on linear differentiation, eliminating redundant parameters through QR decomposition, and identifying structural parameter errors using a weighted LM iteration method. The identified structural parameter errors are then compensated to the original kinematic model to obtain the corrected kinematic model.

[0035] Specifically, step S3 includes: Multiple preset source loads and target loads are sequentially applied to the end effector of the robotic arm. For each load condition, randomly generated joint configuration commands are sequentially input into the robotic arm controller. After the robotic arm moves into position, the robotic arm state and the actual three-dimensional position coordinates of the end effector reference point are recorded. The robotic arm state consists of a five-dimensional state vector composed of the rotation angle of the first joint and the hydraulic cylinder displacements corresponding to the second to fifth joints. The sixth joint maintains a preset fixed state during data acquisition and compensation, and the variables of the sixth joint are not used as input features for the position error migration regression model.

[0036] The acquired five-dimensional robotic arm state is input into the kinematic model corrected for geometric parameters to calculate the theoretical three-dimensional position coordinates of the end effector reference point. Based on the error definition of subtracting the theoretical position coordinates from the actual position coordinates, the position coordinates of the end effector reference point along... Position errors in three coordinate directions. Each data sample includes a five-dimensional robotic arm state vector and its corresponding three-dimensional position error vector.

[0037] Specifically, step S4 includes: See Figure 3 As shown, the data from each source load condition are merged to form the source domain dataset. The data for the target load condition is divided into a target domain training dataset. and independent test dataset .in, and Used together for model training, It is only used to evaluate the predictive performance of the model and is not used in model training or sample weight updates.

[0038] Specifically, step S5 includes: Please see Figure 4 As shown, respectively with The positional errors in the three directions are used as output labels to train the transfer regression model. The same robotic arm state input and dataset partitioning are used in the three directions, and weak regressor training, sample weight update and final prediction are completed respectively.

[0039] For position error prediction in one direction, let the amount of calibrated data in the source domain be... The target domain has been marked with a data volume of [data volume]. The total dataset is constructed by arranging the source domain samples first and the target domain samples last. ; The first in the total dataset The sample at the th The weight of the wheel is denoted as The initial weights are all set to .

[0040] In the In each iteration, the data is processed according to the current sample weight from the total dataset. Sampling is performed to give samples with larger weights a higher probability of being selected; the sampled dataset is used to train a base regressor to obtain the first... A weak regressor .

[0041] weak regressor For each sample in the total training dataset, the predicted value for each sample is obtained. The maximum absolute prediction error across all joint training samples. As the normalization benchmark for this round of errors: (20) Get the first The sample at the th Normalized error in round iteration: (twenty one) The error of the weak regressor is calculated based on the normalized error of all joint training samples on the weak regressor. : (twenty two) No. The error coefficient of the weak regressor : (twenty three) Source attenuation coefficient : (twenty four) Unnormalized next-round sample weights: (25) Weight normalization factor: (26) Next round of normalized sample weights: (27) satisfy: (28) because The larger the normalization error of the source domain samples, the greater the degree of weight decay, thereby reducing the impact of source domain samples that are inconsistent with the target load condition distribution on subsequent training; for target domain samples, the larger the normalization error, the higher the updated relative weight, thus making the subsequent weak regressor pay more attention to the target domain samples that have not yet been accurately fitted.

[0042] Through the aforementioned weight updates, the weights of source domain samples with large prediction errors are reduced in the next iteration to mitigate the impact of source domain samples with inconsistent distributions with the target load condition data on model training. Conversely, the weights of target domain samples with large prediction errors are increased in the next iteration, causing the subsequent weak regressor to pay more attention to target domain samples that have not yet been adequately fitted. This process of weighted sampling, weak regressor training, and sample weight updates is repeated until the preset target load condition is met. The iteration continues until a preset termination condition is met.

[0043] After the model training is completed, select the effective weak regressor in the second half of the iteration, and use... As the first The fusion weights of the selected weak regressors are used to calculate the weighted median of their regression outputs, yielding the predicted position error in the current coordinate direction. This is then applied to... Independent load position error prediction models are established in three coordinate directions, and the prediction results of the three models are combined to obtain the three-dimensional load position error prediction value of the reference point at the end of the robotic arm.

[0044] Specifically, step S6 includes: Please see Figure 5 As shown, during joint compensation, the desired end-effector position is first input into the corrected kinematic model for the first inverse kinematics solution, yielding the initial robotic arm state corresponding to the desired end-effector position. Subsequently, the initial robotic arm state is input into three load position error prediction models respectively, resulting in... Predicted values ​​of end load position error in three coordinate directions.

[0045] According to the definition of position error as the difference between the actual position and the theoretical position of the end reference point, the expected position after compensation is obtained by subtracting the predicted value of the three-dimensional load position error from the expected position of the end.

[0046] The compensated desired position is then input into the corrected kinematic model for inverse kinematics solution to obtain the final corrected joint command value; the joint command value is then input into the robotic arm controller to drive the robotic arm to move to the target position.

[0047] Specifically, step S7 includes: A target load is applied to the end effector of the robotic arm, and multiple target points are selected within the end effector's Cartesian workspace for joint compensation verification. The corresponding joint command values ​​are solved using the initial kinematic model before parameter identification, the corrected kinematic model, and a joint compensation method combining the corrected kinematic model with a load position error prediction model. These joint command values ​​are then input into the robotic arm controller.

[0048] Using a laser tracker, the actual three-dimensional position coordinates of the end effector reference point were measured when the robotic arm reached each target point under different compensation methods. The distance between the actual position and the desired position of the end effector reference point was then calculated. The position deviations in the three coordinate directions are calculated, and the end position accuracy of different compensation methods is compared based on the position deviations.

[0049] In this embodiment, the data labels, transfer learning model output, and compensation effect evaluation are all based on the reference point at the robotic arm's end effector. The position error is calculated in three coordinate directions; the target ball's center position measured by the laser tracker is used to obtain the actual three-dimensional position coordinates of the end effector reference point. The transfer learning model does not establish an end effector attitude error prediction model, nor does it perform end effector attitude compensation.

Claims

1. A method for compensating the end-effector position of a hydraulic robotic arm across a load, characterized in that, include: S1. Obtain the kinematic model of the hydraulic manipulator after geometric parameter correction; Under at least one calibrated source load condition and a target load condition different from the source load condition, respectively collect the manipulator state and the measured position of the end reference point in the manipulator base coordinate system, and use the kinematic model of the hydraulic manipulator to calculate the theoretical position of the end reference point corresponding to the manipulator state. S2, calculate the three-dimensional position error according to the direction of the measured position minus the theoretical position, construct the robot arm state and three-dimensional position error corresponding to the source load condition as the source domain dataset, and construct the robot arm state and three-dimensional position error corresponding to the target load condition as the target domain training dataset. S3, merge the source domain dataset and the target domain training dataset into a joint training dataset, and train transfer regression models for the position error components along the three coordinate axes of the robot arm base coordinate system respectively; during the iterative training process, based on the prediction error of the weak regressors on the joint training samples, reduce the relative weight of the source domain samples with larger prediction errors, increase the relative weight of the target domain samples with larger prediction errors, and fuse multiple weak regressors to obtain three position error component prediction models; S4, Receive the desired end position, and under the condition that the sixth joint is kept in a preset fixed state, perform the first inverse kinematics solution with the desired end position as the position constraint of the end reference point, and determine the candidate robot arm state from the feasible inverse solution according to the preset joint constraints; input the candidate robot arm state into the three position error component prediction models respectively, and combine the outputs of the three position error component prediction models into a three-dimensional position error prediction value. S5, according to the three-dimensional position error defined direction in step S2, subtract the predicted three-dimensional position error from the expected end position to obtain the corrected command position; under the condition that the sixth joint maintains the preset fixed state, use the corrected command position as the position constraint of the end reference point to perform the second inverse kinematics solution, and determine the corrected robot arm state from the feasible inverse solution according to the preset joint constraints. S6, the corrected robotic arm state is converted into the target rotation angle of the first joint and the target displacement of the hydraulic cylinders corresponding to the second to fifth joints, and the robotic arm controller performs closed-loop control based on the feedback of the corresponding joint states.

2. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 1, characterized in that, The end reference point is the center of a single spherical reflective target ball that is rigidly mounted to the end of the hydraulic manipulator via a target ball seat; The three-dimensional coordinates of the center of the spherical reflective target in the laser tracker's measurement coordinate system are measured using a laser tracker. When the laser tracker's measurement coordinate system does not coincide with the robot arm's base coordinate system, the three-dimensional coordinates are converted into the measured position of the end reference point in the robot arm's base coordinate system using a pre-established fixed coordinate transformation between the two.

3. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 1, characterized in that, The hydraulic manipulator is a six-degree-of-freedom cascaded hydraulic manipulator; the manipulator state is a five-dimensional state vector composed of the rotation angle of the first joint and the hydraulic cylinder displacements corresponding to the second to fifth joints; the sixth joint maintains a preset fixed state during data acquisition, inverse kinematics solution, and end-effector position compensation, and the sixth joint variable is not used as a component of the five-dimensional state vector; the inverse kinematics solution uses the three-dimensional position of the end-effector reference point as the target constraint, and determines the corresponding manipulator state from feasible inverse solutions that satisfy the target constraint according to the preset joint constraints.

4. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 3, characterized in that, All three position error component prediction models use the five-dimensional state vector as numerical input; the load quality of the source load condition and the target load condition is only used to determine the domain affiliation of the corresponding samples and is not used as numerical input to the three position error component prediction models.

5. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 1, characterized in that, The source domain dataset consists of multiple samples with calibrated source load conditions, and the number of samples in the target domain training dataset is less than the number of samples in the source domain dataset. The three position error component prediction models use the same source domain samples and target domain training samples, but calculate the prediction error and update the sample weights according to the corresponding position error components. The data corresponding to the target load condition also includes an independent test dataset, which is not involved in the training of the transfer regression model, the update of sample weights, and the selection of weak regressors.

6. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 1, characterized in that, The transfer regression model in step S3 uses the TrAdaBoost.R2 transfer regression model. The training of the TrAdaBoost.R2 transfer regression model for each location error component includes: Set initial sample weights for the source domain samples and target domain training samples in the joint training dataset; perform sampling with replacement from the joint training dataset according to the current sample weights, and train the current weak regressor using the sampled samples; The current weak regressor is applied to all joint training samples, and the absolute prediction error of each joint training sample is calculated. The maximum absolute prediction error among all joint training samples is used as the normalization benchmark. When the maximum absolute prediction error is greater than zero, the absolute prediction error of each joint training sample is divided by the maximum absolute prediction error to obtain the normalized prediction error of each joint training sample. Based on the current sample weights and corresponding normalized prediction errors of all joint training samples, the weighted error of the current weak regressor on the joint training dataset is calculated. When the maximum absolute prediction error is zero or the weighted error is zero, the current weak regressor is determined as the final prediction model for the corresponding position error component and training ends; when the weighted error is greater than or equal to 0.5, the current weak regressor is discarded and iterative training stops; when the weighted error is greater than zero and less than 0.5, the error coefficient of the current weak regressor is calculated based on the weighted error, the source domain decay coefficient is calculated based on the number of source domain samples and the preset maximum number of iterations, the source domain sample weights are reduced according to the decay magnitude positively correlated with the normalized prediction error, the target domain training sample weights are increased according to the increase magnitude positively correlated with the normalized prediction error, and the sample weights of all joint training samples are normalized before the next round of training.

7. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 6, characterized in that, The weak regressor is a regression decision tree with a maximum depth of a preset value, and the transfer regression model is iteratively trained according to a preset maximum number of iterations; let the number of effective weak regressors retained at the end of training be... ;when When the effective weak regressor's predicted value is used as the final predicted value of the corresponding position error component, then... When selecting a serial number, the following conditions must be met: ； An effective weak regressor is used for prediction, in order to: ； As the fusion weight of the t-th effective weak regressor, the predicted values ​​of the selected effective weak regressors are sorted and weighted, and the predicted value corresponding to the first time the accumulated weight reaches half of the total weight of the selected effective weak regressors is taken as the final predicted value of the corresponding position error component.

8. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 1, characterized in that, The second inverse kinematics solution in step S5 includes determining whether there is an inverse solution that satisfies the preset joint limit and hydraulic cylinder stroke range at the position of the correction command; when the inverse solution exists, the state of the corrected robotic arm is determined from the inverse solution that satisfies the preset joint limit and hydraulic cylinder stroke range; step S6 only converts the state of the corrected robotic arm that satisfies the preset joint limit and hydraulic cylinder stroke range into the target displacement of each hydraulic cylinder.

9. The method for compensating the end position of a hydraulic robotic arm across a load according to claim 3, characterized in that, The output of the three position error component prediction models only includes the position error of the end reference point along the three coordinate axes of the manipulator base coordinate system, and does not include the end posture error of the hydraulic manipulator. When performing the first and second inverse kinematics solutions, the three-dimensional position of the end reference point is used as the target constraint, and the sixth joint is kept in the preset fixed state. The complete three-dimensional end posture of the hydraulic manipulator is not used as the target constraint for the inverse kinematics solution.

10. A hydraulic robotic arm end-effector position compensation system across load, characterized in that, Includes a hydraulic robotic arm, a hydraulic robotic arm controller, a rotary hydraulic bearing rotary encoder, a hydraulic cylinder displacement sensor, a swing hydraulic cylinder rotary encoder, an end-effector position measurement component, and a data processing unit; The end-effector position measurement component includes a single spherical reflective target ball rigidly mounted on the end of the hydraulic manipulator via a target ball mount, and a laser tracker for measuring the three-dimensional coordinates of the center of the spherical reflective target ball. The center of the spherical reflective target ball is defined as the end-effector reference point of the hydraulic manipulator. The end-effector position measurement component is used to obtain the measured position of the end-effector reference point in the manipulator base coordinate system. The data processing unit includes a processor and a memory. The memory stores a computer program. When the processor executes the computer program, it calculates the three-dimensional position error based on the measured position of the end reference point and the theoretical position output by the kinematic model of the hydraulic manipulator after geometric parameter correction. Based on this, it constructs a source domain dataset and a target domain training dataset. It trains a transfer regression model for each of the three position error components of the three-dimensional position error, and performs the first inverse kinematics solution, three-dimensional position error prediction, reverse correction of the expected end position, the second inverse kinematics solution, and a feasibility check to obtain the target displacement of the hydraulic cylinder and output it to the hydraulic manipulator controller. The hydraulic robotic arm controller is used to perform closed-loop control of hydraulic cylinder displacement based on the feedback from the hydraulic cylinder displacement sensor and the target displacement of the hydraulic cylinder.

Images