A method and system for real-time compensation of pose error at the end effector of a robotic arm
By using multi-source data fusion and real-time compensation methods with BP neural networks, the problem of distinguishing and compensating for thermal and vibration errors of robotic arms was solved, improving the end-effector trajectory tracking accuracy and vibration error suppression capability, and enhancing the model's adaptability and generalization ability.
Patent Information
- Application Number
- CN202611606067.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-16
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to effectively distinguish and compensate for thermal and vibration errors during robotic arm operation, resulting in poor adaptability and generalization ability of error compensation models. In particular, the robotic arm control system has limited ability to suppress vibration errors under high dynamic conditions.
By acquiring multi-source data, including joint motion, motor drive current, joint torque, end-effector pose, and temperature data, geometric, thermal, and vibration errors are separated using kinematic calibration and state estimation methods. A pre-trained BP neural network is then used to predict end-effector vibration errors, enabling real-time compensation.
It improves the trajectory tracking accuracy of the robotic arm end effector, enhances the control system's ability to suppress vibration errors under high dynamic conditions, and strengthens the adaptability and generalization ability of the error compensation model under complex conditions.
Smart Images

Figure CN123103701A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial robot control technology, and in particular to a method and system for real-time compensation of end-effector pose error of a robotic arm. Background Technology
[0002] Industrial robotic arms commonly exhibit positional errors in their end effectors under conditions of high-speed movement, variable load, and long-term operation. These errors mainly include geometric structural errors, thermal deformation errors, and vibration errors caused by the elasticity of the joint drive system, the periodicity error of the reducer, and the flexibility of the structure.
[0003] In existing technologies, kinematic calibration or PID control methods are commonly used to correct errors in robotic arms. However, existing methods often perform offline calibration for geometric errors, making it difficult to cover dynamic errors during operation. Furthermore, existing methods lack effective means to separate vibration and thermal errors. Thermal errors are mainly affected by material thermal expansion, ambient temperature rise, and continuous operating time, exhibiting low-frequency, slow drift characteristics. Vibration errors, on the other hand, mainly originate from joint drives, motor torque fluctuations, and structural flexibility responses, exhibiting mid-to-high frequency dynamic changes. These two types of errors overlap in time scales and exhibit amplitude coupling during actual operation, making it difficult for traditional simple filtering methods to accurately distinguish between different error sources. Because the low-frequency, slow-changing characteristics of thermal errors are not distinguished from the high-frequency characteristics of jitter errors, signal aliasing cannot be independently modeled and compensated, resulting in poor adaptability and generalization ability of the compensation model. Moreover, under highly dynamic conditions, the robotic arm control system has limited ability to suppress high-frequency vibration errors, and existing methods are usually based on a single information source, lacking a unified modeling mechanism for multi-source data fusion, thus failing to comprehensively reflect the complex dynamic characteristics of the robotic arm during operation. Summary of the Invention
[0004] To address the problems existing in the prior art, this invention provides a method and system for real-time compensation of pose error at the end of a robotic arm. It can utilize multi-source sensor information to separately process and compensate for geometric errors, thermal errors, and vibration errors in real time, thereby improving the trajectory tracking accuracy of the robotic arm's end effector.
[0005] To achieve the above objectives, the present invention provides a method for real-time compensation of pose error at the end effector of a robotic arm, comprising: Acquire multi-source data during the operation of the robotic arm; the multi-source data includes joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data; Based on the desired pose and the end-effector pose data, determine the comprehensive pose error of the end-effector; The geometric error is extracted from the end-effector integrated pose error by kinematic calibration to obtain the first residual error; Based on the temperature data and combined with the temperature change rate constraint, the thermal error is decomposed from the first residual error using the state estimation method to obtain the second residual error. The characteristic parameters are determined based on the multi-source data; the characteristic parameters include velocity, acceleration, jerk, joint torque, motor current, and temperature parameters. The presence of effective jitter is determined based on the jerk and the second residual error. If so, the feature parameters are input into a pre-trained BP neural network to obtain the predicted value of the end-vibration error; otherwise, the predicted value of the end-vibration error is set to zero. The desired trajectory of the robotic arm end effector is corrected in real time based on the predicted values of thermal error and end effector vibration error.
[0006] Optionally, the training samples for the BP neural network are obtained in the following way: Acquire historical multi-source data during the historical operation of the robotic arm, and determine the historical end-effector comprehensive pose error based on the historical expected pose and the historical end-effector pose data in the historical multi-source data; The historical geometric error is extracted from the historical end-point comprehensive pose error by kinematic calibration to obtain the first historical residual error; Based on the historical temperature data in the historical multi-source data, and combined with the temperature change rate constraint, the state estimation method is used to decompose the historical thermal error from the first historical residual error to obtain the second historical residual error. Frequency domain analysis and bandpass filtering are performed on the second historical residual error to extract the historical vibration error; The historical vibration error is used as the training label, and the historical feature parameters determined based on the historical multi-source data are used as the training input to construct training samples.
[0007] Optionally, the input layer of the BP neural network is equipped with a temperature-compensated feature gating mechanism to perform weighted and temperature-compensated processing on the input feature parameters, thereby reducing the impact of thermal errors on the prediction results of end-vibration errors; the expression of the temperature-compensated feature gating mechanism is: X ′= WX + G ( T ) in, X For the input feature parameters, W For feature weights, G ( T ) is the temperature compensation function. X ′ represents the input feature after gating.
[0008] Optionally, the BP neural network uses an adaptive momentum factor for backpropagation weight update during training, and the weight update formula is as follows: in, For the first The weight adjustment amount in the next iteration. For the first The weight adjustment amount in the next iteration. For adaptive momentum factor, For network learning rate, Error function Regarding network weights The gradient.
[0009] Optionally, frequency domain analysis and bandpass filtering are performed on the second historical residual error to extract the historical vibration error, including: Perform a Fourier transform on the second historical residual error to obtain the historical frequency domain signal; The historical vibration error is obtained by performing a fourth-order Butterworth bandpass filter on the components in the historical frequency domain signal whose frequencies are higher than a preset frequency threshold.
[0010] Optionally, determining whether there is valid jitter based on the jerk and the second residual error includes: The second residual error is analyzed in the frequency domain to calculate the ratio of the energy with a frequency higher than a preset frequency threshold in the second residual error to the total energy of the second residual error, thus obtaining the proportion of high-frequency energy. When the proportion of high-frequency energy is greater than a preset energy proportion threshold and the jerk is greater than a preset jerk threshold, it is determined that there is effective jitter; otherwise, it is determined that there is no effective jitter.
[0011] Optionally, the method further includes: Calculate the ratio of energy with a frequency lower than a preset frequency threshold in the residual error to the total energy; When the ratio is greater than a preset percentage value and the temperature change rate obtained from the temperature data is less than a preset change threshold, the thermal error is determined to be valid.
[0012] Optionally, the method further includes: The actual spatial coordinates of the robotic arm's end effector are obtained in real time using a laser tracker. The actual measurement error is determined based on the actual spatial coordinates and the desired pose. The model prediction error is determined based on the actual measurement error, thermal error, and predicted end-vibration error. The BP neural network is updated online incrementally based on the model prediction error.
[0013] Optionally, the desired trajectory of the robotic arm's end effector can be corrected in real time according to the following correction formula: in, The corrected trajectory. For the desired trajectory, This is the predicted value of the end-vibration error. For thermal error, This refers to the runtime.
[0014] The present invention also provides a real-time compensation system for the pose error of a robotic arm end effector, comprising: The multi-source data acquisition module is used to acquire multi-source data during the operation of the robotic arm; the multi-source data includes joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data. The end-effector integrated pose error determination module is used to determine the end-effector integrated pose error based on the desired pose and the end-effector pose data. A geometric error stripping module is used to strip geometric errors from the end-effector integrated pose error through kinematic calibration to obtain a first residual error. The thermal error decomposition module is used to decompose the thermal error from the first residual error based on the temperature data and combined with the temperature change rate constraint, using a state estimation method to obtain the second residual error. The vibration error determination module is used for: The characteristic parameters are determined based on the multi-source data; the characteristic parameters include velocity, acceleration, jerk, joint torque, motor current, and temperature parameters. The presence of effective jitter is determined based on the jerk and the second residual error. If so, the feature parameters are input into a pre-trained BP neural network to obtain the predicted value of the end-vibration error; otherwise, the predicted value of the end-vibration error is set to zero. The real-time trajectory correction module is used to correct the desired trajectory of the robotic arm end effector in real time based on the predicted values of the thermal error and the end effector vibration error.
[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: The real-time compensation method for end-effector pose error provided by this invention acquires joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data to form a multi-source information fusion error identification basis. This method can comprehensively reflect the motion state, load changes, and temperature changes during the operation of the robotic arm, overcoming the shortcomings of a single information source in fully describing the influencing factors of dynamic errors. This invention first separates geometric errors from the end-effector's overall pose error. Then, based on temperature data and combined with temperature change rate constraints, it uses a state estimation method to decompose thermal errors from the first residual error. Subsequently, when effective jitter is determined based on jerk and the second residual error, feature parameters are input into a pre-trained BP neural network to predict vibration errors. When no effective jitter is determined, the predicted value of the end-effector vibration error is set to zero. This allows geometric errors, thermal errors, and vibration errors to be processed and compensated separately, avoiding the problem of low-frequency, slow-changing characteristics of thermal errors and high-frequency characteristics of vibration errors overlapping and making independent modeling impossible. At the same time, it avoids misjudging high-frequency noise or transient disturbances in a stable operating state as end-effector jitter, avoiding unnecessary vibration compensation, and helping to improve the adaptability and generalization ability of the error compensation model under complex working conditions.
[0016] By real-time correction of the desired trajectory of the robotic arm end effector based on the predicted values of thermal error and end effector vibration error, it is possible to simultaneously compensate for low-frequency thermal errors caused by temperature changes and high-frequency vibration errors caused by dynamic excitation. This is beneficial to improving the end effector trajectory tracking accuracy and enhancing the control system's ability to suppress vibration errors under high dynamic conditions. Attached Figure Description
[0017] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same parts.
[0018] Figure 1 This is a schematic flowchart of a method for real-time compensation of pose error at the end of a robotic arm, as shown in an embodiment of the present invention. Figure 2 This is a flowchart illustrating the multi-source sensor fusion error compensation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the multi-sensor arrangement of a robotic arm according to an embodiment of the present invention; Figure 4 This is a flowchart illustrating jitter identification and frequency domain analysis according to an embodiment of the present invention; Figure 5 This is a logic diagram illustrating error decomposition and compensation in an embodiment of the present invention; Figure 6 This is a schematic diagram of the module structure of a real-time posture error compensation system for the end effector of a robotic arm, as shown in an embodiment of the present invention. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Please see Figure 1 and Figure 2 , Figure 1 This is a flowchart illustrating a method for real-time compensation of pose error at the end effector of a robotic arm. Figure 2 This is a flowchart of error compensation for multi-source sensor fusion.
[0021] Real-time compensation methods for pose error at the end effector of a robotic arm include: S101: Acquire multi-source data during the operation of the robotic arm.
[0022] The multi-source data includes joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data.
[0023] During high-precision tasks performed by robotic arms, the end effector's pose accuracy is affected by a combination of geometric errors, thermal deformation errors, and vibration errors. The physical causes of these various error sources differ, necessitating the collection of information from multiple dimensions to support subsequent error identification and separation. (See also...) Figure 3 Multiple sensors can be deployed on the robotic arm to simultaneously collect multi-dimensional data reflecting its motion state, load status, actual end-effector pose, and structural temperature changes during the robotic arm's operation.
[0024] Specifically, joint motion data can be acquired in real time through a distributed joint encoder, including the angles, angular velocities, and angular accelerations of each joint of the robotic arm. This data characterizes the kinematic state of each joint and is the direct source for subsequent extraction of characteristic parameters such as velocity, acceleration, and jerk. Motor drive current data can be acquired through a motor current acquisition unit, reflecting load changes and torque fluctuations in the joint drive system. Joint torque data can be acquired through joint torque sensors, characterizing the load torque and power changes borne by each joint; joint torque data can also be indirectly calculated based on motor drive current, transmission parameters, and motor torque constant. End-effector pose data can be acquired through an end-effector pose measurement unit installed on the robotic arm's end effector, containing the end effector's position and orientation information in three-dimensional space. This characterizes the actual spatial pose of the end effector, providing a practical pose basis for determining the overall end-effector pose error. Temperature data can be acquired through a distributed temperature sensor array deployed on each joint and end-effector structure of the robotic arm. This data acquires temperature change information for key parts of the robotic arm during operation, including the current temperature value and temperature rise change at each measuring point, characterizing the thermal deformation trend of the robotic arm structure.
[0025] Since the joint encoder, motor current acquisition unit, joint torque sensor, end-effector pose measurement unit, and temperature sensor array belong to different data acquisition channels, and the sampling frequency and data transmission delay of each channel differ, directly using unprocessed data for error modeling will lead to temporal misalignment between the data sources, thus affecting the accuracy of error separation results. Therefore, during the acquisition of multi-source data, microsecond-level timestamps can be uniformly added to the data acquired from each channel. Using these timestamps as a benchmark, the time sequence of the multi-source data is aligned, ensuring that the joint motion state, load information, end-effector pose, and temperature data acquired at the same time can accurately correspond. Afterward, preprocessing can be performed on the time-aligned multi-source data. For example, mean filtering can be used to suppress high-frequency sensor noise, abnormal jump data can be removed based on the 3σ criterion to eliminate invalid measurements introduced by sensor failure or external interference, and the features of each dimension can be normalized to eliminate dimensional differences between different sensor data. The synchronized and preprocessed multi-source data serves as the data foundation for subsequent end-effector integrated pose error determination, error decomposition, and vibration error prediction.
[0026] The aforementioned multi-source data comprehensively covers motion excitation, load excitation, and thermal excitation information during the operation of the robotic arm from different dimensions. Joint motion data and end-effector pose data reflect the motion state and actual trajectory of the robotic arm; motor drive current data and joint torque data reflect the dynamic load state of the joint drive system; and temperature data reflects the main excitation sources of structural thermal deformation. These types of information complement each other, avoiding the problem that a single information source cannot fully reflect various error correlation characteristics.
[0027] S102: Determine the comprehensive pose error of the end effector based on the desired pose and end effector pose data.
[0028] After completing the synchronous acquisition and preprocessing of multi-source data, the actual spatial pose of the robotic arm's end effector at the corresponding moment is extracted from the end-effector pose data. The desired pose is provided by the trajectory planning module in the robotic arm motion control system, representing the theoretically desirable spatial position and orientation of the end effector within the current control cycle. The actual pose and the desired pose are compared moment by moment, and the deviations in position and orientation are calculated to obtain the comprehensive end-effector pose error.
[0029] End-effector integrated pose error This reflects the total deviation of the TCP (Technical Error Track) at the end effector of the robotic arm in three-dimensional space. It is not the result of a single error source, but rather a superposition of geometric errors, thermal deformation errors, and mechanical vibration errors. This superposition relationship can be expressed as: in, Geometric errors include assembly errors and connecting rod deviations; This refers to the low-frequency, slow-temperature deformation error caused by temperature accumulation. This refers to mechanical vibration error, also known as jitter error; the unit for each error component in position is uniformly mm, or the unit for attitude is uniformly rad.
[0030] S103: The geometric error is extracted from the end-effector integrated pose error through kinematic calibration to obtain the first residual error.
[0031] After determining the overall pose error of the end effector, it is necessary to separate the geometric errors. Geometric errors mainly include assembly errors and link deviations, which are static parameter errors of the robotic arm body. Their values remain basically fixed after the robotic arm leaves the factory or is periodically calibrated, and do not change significantly with the change of operating time and dynamic load. They are considered static error components.
[0032] To address the removal of geometric errors, a method based on DH kinematic calibration can be employed. Through DH kinematic calibration, a kinematic model of the robotic arm can be established and the parameter errors of each link can be identified, thereby obtaining the mapping relationship of geometric errors. = f (DH), where The geometric structure error is represented by DH, which denotes the set of static geometric parameter errors consisting of DH parameter deviations. The first residual error is obtained by subtracting the calibrated geometric structure error from the end-effector integrated pose error. This first residual error no longer contains static geometric error components; it mainly consists of two types of dynamic error components: residual thermal error and vibration error.
[0033] S104: Based on the temperature data and combined with the temperature change rate constraint, the thermal error is decomposed from the first residual error using the state estimation method to obtain the second residual error.
[0034] After removing geometric structural errors, the first residual error mainly comprises two dynamic components: thermal deformation error (thermal error) caused by temperature changes and vibration error caused by mechanical dynamic response. The thermal error primarily originates from material thermal expansion, structural temperature rise, and ambient temperature changes during the long-term operation of the robotic arm, exhibiting significant temperature correlation, low-frequency characteristics, and a slow cumulative variation. The amplitude of the thermal error gradually changes with the continuous operation time of the robotic arm and the load heating conditions, showing a significant difference in time scale from the high-frequency dynamic characteristics of the vibration error. Therefore, it is necessary to separate it from the first residual error by incorporating temperature field information.
[0035] To accurately estimate the thermal error component, temperature sensors deployed on the joints and end effector of the robotic arm can be used to acquire temperature changes and form temperature field information. The temperature distribution at each joint and end effector of the robotic arm can be represented as a temperature vector. T ( t )=[ T 1, T 2,..., T n ],in T 1 to T n This represents the temperature value at each temperature measurement point at the current moment. A thermal error mapping relationship is established based on the temperature data. = f ( T , , t ),in T The current temperature. The change in temperature. t This refers to the continuous running time.
[0036] Because thermal deformation exhibits slow-changing characteristics, the Kalman state estimation method is used to estimate the thermal error components in real time for the first residual error. The state equation can be expressed as: X k = AX k-1 + BU k + W k The output equation can be expressed as: Y k = CX k + V kBy recursively calculating the state equation and the output equation, the Kalman state estimation method can track the slow changing trend of thermal error in real time under the drive of temperature data, thereby estimating the thermal error component from the first residual error.
[0037] In the process of thermal error estimation, the reasonableness of the estimation results is judged by incorporating the temperature change rate constraint. The slow accumulation characteristic of thermal error determines that it exhibits stable low-frequency drift when the temperature change is relatively gradual, while other low-frequency components that may exist in the first residual error may not be consistent with the temperature change trend. By introducing the temperature change rate constraint, it is possible to determine whether the variation characteristics of the low-frequency components in the first residual error are consistent with the temperature field change trend. This can avoid misestimating low-frequency fluctuations unrelated to thermal deformation as thermal errors and improve the accuracy of thermal error decomposition.
[0038] S105: Determine characteristic parameters based on multi-source data; characteristic parameters include velocity, acceleration, jerk, joint torque, motor current, and temperature parameters.
[0039] Because vibration errors change rapidly and are highly time-varying, a data-driven approach can be used to establish a nonlinear mapping relationship between them and multi-source characteristic parameters, enabling real-time prediction. To this end, characteristic parameters directly related to the vibration error generation mechanism are determined from the synchronized and preprocessed multi-source data. These characteristic parameters include velocity, acceleration, jerk, joint torque, motor current, and temperature parameters.
[0040] Among them, speed The instantaneous velocity of the robotic arm's end effector or joint, measured in m / s; acceleration. This is the first derivative of velocity with respect to time, in m / s. 2 ; jerk Defined as the derivative of acceleration with respect to time, its form is: The unit is m / s 3 Velocity, acceleration, and jerk are used to characterize the rate of change of acceleration during the movement of a robotic arm. In practice, velocity, acceleration, and jerk can be obtained from joint motion data through differentiation or kinematic calculations. Joint torque. The torque output by the joint drive system, measured in N·m, and the motor current. The current value of the joint drive motor is expressed in amperes (A). Both the current and the current value reflect the load state and driving torque fluctuation of the joint drive system. Vibration error is directly related to the dynamic load changes of the joint drive system. Temperature parameter. Temperature data acquired by temperature sensors deployed on the joints and end effector of the robotic arm, although thermal errors have been separated during error decomposition, still serve as auxiliary input features for vibration error prediction. These features reflect the impact of temperature changes on the dynamic characteristics of the structure and help mitigate the interference of thermal errors on the vibration prediction results. The aforementioned six-dimensional feature parameters comprehensively characterize the various excitation factors that cause vibration errors during the robotic arm's operation from kinematic, dynamic, and thermal state perspectives.
[0041] S106: Determine whether there is effective jitter based on the jerk and the second residual error. If so, input the feature parameters into the pre-trained BP neural network to obtain the end vibration error prediction value; otherwise, set the end vibration error prediction value to zero.
[0042] After decomposing the thermal error from the first residual error and obtaining the second residual error, and determining the characteristic parameters based on multi-source data, the second residual error may also contain high-frequency noise or transient disturbances in addition to the thermal error. To avoid misjudging high-frequency noise or transient disturbances during normal operation as end-effector jitter when the robotic arm is in a stable operating state, it is necessary to determine the jitter validity of the second residual error.
[0043] The determination of whether there is valid jitter based on jerk and second residual error specifically includes: Frequency domain analysis is performed on the second residual error to calculate the ratio of the energy with a frequency higher than a preset frequency threshold in the second residual error to the total energy of the second residual error, thus obtaining the proportion of high-frequency energy. When the proportion of high-frequency energy is greater than the preset energy proportion threshold and the jerk is greater than the preset jerk threshold, it is determined that there is effective jitter; otherwise, it is determined that there is no effective jitter.
[0044] Specifically, frequency domain analysis is performed on the second residual error to calculate the ratio of energy with frequencies higher than a preset frequency threshold to the total energy of the second residual error, thus obtaining the high-frequency energy proportion. Simultaneously, the jerk of the robotic arm in its current motion state is determined based on joint motion data from multi-source data; this jerk is a characteristic parameter. The high-frequency energy proportion is compared with a preset energy proportion threshold, and the jerk is compared with a preset jerk threshold. When the high-frequency energy proportion is greater than the preset energy proportion threshold (e.g., 70%), and the jerk is greater than the preset jerk threshold, it is determined that there is effective jitter at the end effector of the robotic arm; when the high-frequency energy proportion is not greater than the preset energy proportion threshold, or the jerk is not greater than the preset jerk threshold, it is determined that there is no effective jitter at the end effector of the robotic arm. Through the joint determination of the high-frequency energy proportion and jerk, it is possible to confirm whether there are compensable vibration components in the second residual error from both frequency domain characteristics and motion excitation perspectives, avoiding misjudging high-frequency noise or transient disturbances in a stable operating state as end effector jitter.
[0045] When a valid jitter is determined to exist, the feature parameters are input into a pre-trained backpropagation (BP) neural network. The BP neural network is used to establish a nonlinear mapping relationship between multi-source features and vibration error; its mapping form is as follows: ,in This provides the predicted vibration error values for the robotic arm's end effector. The BP neural network, through its multi-layer nonlinear mapping capability, can extract dynamic feature patterns related to vibration error from the aforementioned multi-source features, outputting predicted vibration error values for the robotic arm's end effector in the X, Y, and Z directions. The pre-training process of the BP neural network is based on historical operating data. In actual online compensation, only real-time feature parameters need to be input into the network to quickly obtain the predicted vibration error values, eliminating the need for re-performing frequency domain analysis and error decomposition in each compensation cycle. By simultaneously using motion excitation parameters, load parameters, and temperature parameters as prediction inputs, the BP neural network can adaptively identify the variation patterns of vibration error under different operating conditions, handling the complex nonlinear relationships between vibration error and motion state parameters, load parameters, and temperature parameters. This overcomes the limitation of a single information source in comprehensively describing the dynamic characteristics of vibration error, providing a reliable basis for subsequent real-time trajectory correction and vibration error compensation.
[0046] When it is determined that there is no valid jitter, the predicted value of the end effector vibration error is set to zero. At this time, subsequent trajectory corrections will no longer add vibration error compensation. By adding jitter validity determination during online real-time compensation, it is possible to avoid misjudging high-frequency noise or transient disturbances during normal movement as end effector jitter when the robotic arm is in a stable operating state, and to avoid unnecessary vibration compensation for the robotic arm trajectory when there is no valid jitter, thereby improving the stability and reliability of real-time compensation results.
[0047] S107: Based on the predicted values of thermal error and end-effector vibration error, the desired trajectory of the robotic arm end effector is corrected in real time.
[0048] After obtaining the predicted values of thermal error and end-effector vibration error, the predicted values of thermal error and vibration error are simultaneously introduced into the trajectory correction stage of the robotic arm end-effector, so that the compensated trajectory can simultaneously suppress the adverse effects of low-frequency thermal drift and high-frequency jitter on the end-effector positioning accuracy.
[0049] Specifically, the desired trajectory of the robotic arm's end effector is corrected in real time according to the following correction formula: in, The corrected trajectory. For the desired trajectory, This is the predicted value of the end-vibration error. For thermal error, This refers to the runtime.
[0050] When the robotic arm's end effector experiences pose deviations due to thermal deformation and mechanical vibration during operation, the actual trajectory is corrected to align with the desired trajectory by pre-subtracting the corresponding predicted error value from the expected trajectory, thereby reducing the deviation between the actual and desired end effector poses. The corrected trajectory command is input to the robotic arm controller in real time, which drives the actuators of each joint to perform motion control according to the corrected trajectory. The controller can employ a PID controller or a Model Predictive Controller (MPC), generating drive commands for each joint based on the corrected trajectory position to achieve closed-loop regulation of the robotic arm's end effector trajectory. Since thermal and vibration errors are obtained online, the corrected trajectory can dynamically adjust with changes in error, thereby achieving online suppression of dynamic errors and improving the end effector trajectory tracking accuracy and dynamic response capability of the robotic arm under high-speed operation and variable load conditions.
[0051] The following section provides a further introduction to the BP neural network. In the vibration error prediction process, the BP neural network is responsible for establishing a nonlinear mapping relationship between multi-source feature parameters and end-effector vibration error. Since the end-effector vibration error of a robotic arm is influenced by multiple factors such as motion state, load changes, drive system characteristics, and temperature conditions, traditional linear regression or fixed-parameter models are insufficient to accurately describe its complex variation patterns. Therefore, this invention employs a BP neural network for online vibration error prediction.
[0052] The BP neural network of this invention adopts a fixed three-layer feedforward structure, including an input layer, a hidden layer, and an output layer. The input layer has six neurons, corresponding to six input features: velocity, acceleration, jerk, joint torque, motor current, and temperature. These six features reflect the excitation sources of vibration error at the robotic arm's end effector from different perspectives, including kinematics, dynamics, the drive system, and the influence of the thermal environment. Because the dimensions and numerical ranges of the various feature parameters differ significantly, the input data needs to be normalized before entering the network, uniformly mapped to the [-1, 1] interval to eliminate the impact of these dimensional differences on the network's training convergence speed and prediction accuracy.
[0053] The input layer of the BP neural network incorporates a temperature-compensated feature gating mechanism to weight and compensate the input feature parameters for temperature variations, thereby reducing the impact of thermal errors on the prediction results of end-vibration errors. The expression for the temperature-compensated feature gating mechanism is as follows: X ′= WX + G ( T ) in, X For the input feature parameters, W For feature weights, X′ represents the input feature after gating; G ( T ) is a temperature compensation function used to reduce the impact of thermal errors on vibration prediction results.
[0054] The hidden layer has 12 hidden neurons. This number was determined through multiple comparative experiments under various working conditions, ensuring the network's fitting ability while avoiding overfitting due to excessive network size. The Tansig hyperbolic tangent function is used as the activation function for the hidden layer, suitable for fitting both positive and negative errors. The output layer has 3 neurons, corresponding to the vibration error components of the robotic arm's end effector in the X, Y, and Z directions, respectively. , , The Purelin linear function is chosen as the activation function for the output layer to ensure that the network output is a continuous real number, which can directly represent the magnitude of the vibration error and is not limited by the output range of nonlinear functions, thus meeting the requirement of output continuity for vibration error prediction.
[0055] The layers of the BP neural network are connected by weights. Input features are passed from the input layer to the hidden layer, undergo a nonlinear transformation using the Tansig activation function, and then passed to the output layer. The Purelin linear function outputs the predicted triaxial vibration error. This three-layer network structure is fixed, establishing a nonlinear mapping relationship between the six-dimensional input features and the triaxial vibration error output through learned weight parameters. During online prediction, the feature parameters determined from multi-source data within the current control cycle are input into the network to quickly obtain the predicted end-efficiency vibration error, meeting the computational efficiency requirements of real-time compensation.
[0056] The training samples for a BP neural network are obtained in the following ways: Acquire historical multi-source data during the robot arm's historical operation, and determine the historical end-effector comprehensive pose error based on the historical expected pose and the historical end-effector pose data in the historical multi-source data; The historical geometric error is extracted from the historical end-point comprehensive pose error by kinematic calibration, and the first historical residual error is obtained. Based on historical temperature data from historical multi-source data and combined with temperature change rate constraints, the state estimation method is used to decompose historical thermal error from the first historical residual error to obtain the second historical residual error. Frequency domain analysis and bandpass filtering were performed on the second historical residual error to extract the historical vibration error; Training samples are constructed by using historical vibration errors as training labels and historical feature parameters determined based on historical multi-source data as training inputs.
[0057] Historical multi-source data from the robotic arm's operation is acquired. This historical multi-source data is consistent in type with the multi-source data collected during online compensation, including joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data. The historical multi-source data undergoes the same time synchronization and preprocessing procedures as the online data to ensure consistency between the training and online data. Based on the historical desired pose and the historical end-effector pose data from the historical multi-source data, the historical end-effector comprehensive pose error is determined. This historical end-effector comprehensive pose error represents the total deviation between the actual and desired end-effector pose at a given historical moment.
[0058] Historical geometric errors are extracted from the historical end-effector composite pose error through kinematic calibration, yielding the first historical residual error. The extraction method for historical geometric errors is the same as in the online process, using static geometric parameter errors obtained from DH kinematic calibration to subtract them from the historical end-effector composite pose error. The first historical residual error no longer contains geometric structural error components, mainly consisting of two types of dynamic error components: residual thermal error and vibration error. Based on historical temperature data from historical multi-source data, the same state estimation method as in online thermal error estimation is used to decompose historical thermal errors from the first historical residual error, yielding the second historical residual error. The second historical residual error is the residual after removing geometric and thermal errors, primarily containing high-frequency vibration error components caused by the elasticity of the joint drive system, the periodic error of the reducer, and structural flexibility.
[0059] Specifically, the above-mentioned frequency domain analysis and bandpass filtering of the second historical residual error to extract historical vibration errors include: Perform a Fourier transform on the second historical residual error to obtain the historical frequency domain signal; Fourth-order Butterworth bandpass filtering is applied to components in the historical frequency domain signal whose frequencies are higher than a preset frequency threshold to obtain the historical vibration error.
[0060] In application, the second historical residual error ( = The input frequency domain analysis module uses FFT to obtain the frequency response, i.e. F ( )= FFT ( E r For dynamic frequency components above 5Hz, the mechanical vibration error is extracted using a 4th-order Butterworth bandpass filter. E vib = Filter ( E rBy using frequency domain analysis and bandpass filtering, the high-frequency vibration error components with frequencies above 5Hz in the second historical residual error are effectively extracted and used as vibration error labels for training the BP neural network.
[0061] The extracted historical vibration errors are used as training labels, and historical feature parameters determined based on historical multi-source data are used as training inputs to construct training samples. Historical feature parameters include velocity, acceleration, jerk, joint torque, motor current, and temperature parameters, and their determination method is consistent with that used in the online prediction process. Each training sample consists of historical feature parameters and corresponding historical vibration errors at the same time. By constructing a training sample set covering different motion states and load conditions using data collected from a large number of historical operating conditions, the BP neural network can learn the nonlinear mapping relationship between multi-source feature parameters and vibration errors, providing a model foundation for online vibration error prediction.
[0062] During the training of a backpropagation neural network, the traditional standard backpropagation algorithm adjusts the weights only based on the current error gradient direction, resulting in slow convergence and a tendency to get stuck in local optima in complex error surfaces. To address this issue, this invention introduces an adaptive momentum factor during the backpropagation weight update process. This factor utilizes the inertia of historical weight corrections to smooth the gradient descent direction, while dynamically adjusting the momentum intensity based on error trends to accelerate convergence and suppress training oscillations.
[0063] During training, the BP neural network uses an adaptive momentum factor for backpropagation weight updates. The weight update formula is as follows: in, For the first The weight adjustment amount in the next iteration. For the first The weight adjustment amount for the next iteration; This is an adaptive momentum factor used to adjust the degree of influence of the previous iteration's weight correction on the current weight update. The network learning rate is used to adjust the degree of influence of the error gradient on the current weight correction. Error function Regarding network weights The gradient is used to determine the direction and magnitude of the adjustment of network weights.
[0064] Adaptive momentum factor A dynamic adjustment mechanism is adopted, with a value ranging from 0.4 to 0.9. When the error changes relatively smoothly and the gradient direction is consistent, the momentum factor is appropriately increased to maintain greater inertia in the weight updates and accelerate the convergence process. When the error changes drastically or the gradient direction fluctuates, the momentum factor is appropriately decreased to reduce the impact of historical corrections on the current update and avoid training oscillations due to excessive inertia. Through this dynamic adjustment, the training process of the BP neural network has a faster convergence speed and better stability compared to the traditional BP algorithm with a fixed momentum factor, and can more effectively approach the optimal weights in the learning process of the nonlinear mapping relationship between multi-source features and vibration errors.
[0065] During online real-time compensation, when decomposing thermal errors from the first residual error, it is necessary to further determine the validity of the decomposition results to avoid misclassifying low-frequency vibration components as thermal errors. Specifically, the ratio of energy with frequencies below a preset frequency threshold in the first residual error to the total energy is calculated. When this ratio is greater than a preset proportion value, and the temperature change rate obtained from the temperature data is less than a preset change threshold, the thermal error is deemed valid. The proportion of low-frequency energy is... Defined as: This refers to the low-frequency energy in the residual error whose frequency is below a preset frequency threshold. This represents the total energy of the residual error. The temperature change rate is obtained by differentiating temperature data over time. When the low-frequency energy accounts for more than 85%, and the temperature change rate... At that time, it was determined that the currently decomposed component was indeed a thermal error, rather than a low-frequency vibration interference.
[0066] During the construction of BP neural network training samples, when extracting historical vibration errors from the second historical residual error, it is also necessary to determine the validity of the historical vibration errors. Specifically, the ratio of energy with frequencies higher than a preset frequency threshold in the second historical residual error to the total energy is calculated. When this ratio is greater than a preset energy proportion threshold, and the jerk in the feature parameters determined based on historical multi-source data is greater than a preset jerk threshold, the historical vibration error is deemed valid. Among these, the high-frequency energy proportion... Defined as: This refers to high-frequency energy with a frequency higher than a preset frequency threshold within the second historical residual error. This represents the total energy of the second historical residual error. The jerk is calculated based on historical joint motion data from historical multi-source datasets. When the high-frequency energy accounts for... >70%, and acceleration jerk Greater than the preset threshold At that time, it is determined that the high-frequency component extracted is indeed a vibration error, rather than noise or other instantaneous impact interference.
[0067] By jointly determining the proportion of low-frequency energy and the rate of temperature change, and jointly determining the proportion of high-frequency energy and jerk, reliable decomposition of thermal errors in the online compensation phase and accurate extraction of vibration error labels in the historical training phase are achieved, respectively. Both methods utilize the triple constraints of temperature information, motion state, and frequency domain characteristics to solve the signal aliasing problem caused by the overlap of time scales and amplitude coupling between thermal and vibration errors.
[0068] In one embodiment, the above method further includes: The actual spatial coordinates of the robotic arm's end effector are obtained in real time using a laser tracker. The actual measurement error is determined based on the actual spatial coordinates and the desired pose. The model prediction error is determined based on the actual measurement error, thermal error, and predicted end-vibration error. The BP neural network is updated online incrementally based on the model prediction error.
[0069] After the BP neural network is pre-trained and put into online prediction, the prediction accuracy of the pre-trained model for vibration error may gradually decrease during long-term operation due to factors such as changes in working conditions, wear of mechanical structures, and slow changes in joint drive characteristics. To suppress the cumulative error and system drift during long-term operation, a laser tracker can be used to obtain the actual spatial coordinates of the robotic arm's end effector in real time, and use them as an external reference to perform online incremental weight updates for the BP neural network, forming a closed-loop feedback mechanism.
[0070] A laser tracker measures the spatial coordinates (TCP) of the robotic arm's end effector with high precision. This measurement is independent of the robotic arm's own joint encoders and end-effector pose sensors, providing an external reference unaffected by internal robotic arm errors. The actual measurement error is determined based on the actual spatial coordinates measured by the laser tracker and the desired pose. This actual measurement error represents the true pose deviation of the robotic arm's end effector under an external high-precision measurement reference. Since the actual measurement error includes geometric, thermal, and vibration errors, and the geometric error has been pre-exposed through kinematic calibration, while the thermal error has been decomposed in real-time using state estimation methods, determining the model prediction error requires a comprehensive calculation based on the actual measurement error, the estimated thermal error, and the predicted end-effector vibration error. Model Prediction Error e The calculation relationship is as follows: e=E measured E predictd in, Emeasured This represents the actual measurement error, i.e., the true deviation at the end measured by the laser tracker. E predictd This represents the prediction error of the neural network.
[0071] Online incremental weight updates are performed on the BP neural network based on the model's prediction error. This online incremental update relies on real-time measurement data from a laser tracker, performing a local weight update every 200ms, eliminating the need for offline retraining of the entire machine and adapting to long-term operational drift. During the incremental update process, only the network weights are fine-tuned, without requiring the collection of large amounts of historical data for complete retraining. Through incremental weight updates based on error feedback, the BP neural network can continuously adapt to the slowly changing dynamic characteristics of the robotic arm during operation, improving the model's generalization ability and robustness under complex conditions, and suppressing accumulated errors and system drift during long-term operation.
[0072] Through the aforementioned closed-loop feedback mechanism, the high-precision external measurement information provided by the laser tracker is continuously introduced into the parameter update process of the BP neural network, forming a closed-loop process of acquisition, feature extraction, error decomposition, prediction compensation, and measurement correction, which takes into account both the dynamic real-time performance and long-term operational stability of the robotic arm's end-effector error compensation.
[0073] Corresponding to the aforementioned application function implementation method embodiments, the present invention also provides a real-time compensation system for the pose error of the robotic arm end effector and corresponding embodiments.
[0074] Please see Figure 6 , Figure 6 This is a schematic diagram of the module structure of a real-time posture error compensation system for the end effector of a robotic arm.
[0075] A real-time posture error compensation system for the robotic arm's end effector includes: The multi-source data acquisition module 61 is used to acquire multi-source data during the operation of the robotic arm; the multi-source data includes joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data; The end-effector integrated pose error determination module 62 is used to determine the end-effector integrated pose error based on the desired pose and the end-effector pose data. The geometric error stripping module 63 is used to strip the geometric error from the end-effector integrated pose error through kinematic calibration to obtain the first residual error. The thermal error decomposition module 64 is used to decompose the thermal error from the first residual error based on the temperature data and combined with the temperature change rate constraint, and to obtain the second residual error by using the state estimation method. Vibration error determination module 65 is used for: Characteristic parameters are determined based on multi-source data; these parameters include velocity, acceleration, jerk, joint torque, motor current, and temperature. The presence of effective jitter is determined based on the jerk and the second residual error. If it is, the feature parameters are input into the pre-trained BP neural network to obtain the predicted value of the end vibration error; otherwise, the predicted value of the end vibration error is set to zero. The real-time trajectory correction module 66 is used to correct the desired trajectory of the robotic arm end effector in real time based on the predicted values of thermal error and end effector vibration error.
[0076] In one embodiment, when determining whether there is valid jitter based on jerk and second residual error, the vibration error determination module 65 is specifically used for: Frequency domain analysis is performed on the second residual error to calculate the ratio of the energy with a frequency higher than a preset frequency threshold in the second residual error to the total energy of the second residual error, thus obtaining the proportion of high-frequency energy. When the proportion of high-frequency energy is greater than the preset energy proportion threshold and the jerk is greater than the preset jerk threshold, it is determined that there is effective jitter; otherwise, it is determined that there is no effective jitter.
[0077] In one embodiment, the thermal error decomposition module 64 is further configured to: Calculate the ratio of energy with a frequency lower than a preset frequency threshold in the residual error to the total energy; When the ratio is greater than the preset percentage value and the temperature change rate obtained from the temperature data is less than the preset change threshold, the thermal error is deemed valid.
[0078] In one embodiment, the system further includes: Incremental update unit, used for: The actual spatial coordinates of the robotic arm's end effector are obtained in real time using a laser tracker. The actual measurement error is determined based on the actual spatial coordinates and the desired pose. The model prediction error is determined based on the actual measurement error, thermal error, and predicted end-vibration error. The BP neural network is updated online incrementally based on the model prediction error.
[0079] Regarding the system in the above embodiments, the specific manner in which each unit module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated further here.
[0080] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.
Claims
1. A method for real-time compensation of end-effector pose error of a robot arm, characterized in that, include: Acquire multi-source data during the operation of the robotic arm; The multi-source data includes joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data. Based on the desired pose and the end-effector pose data, determine the comprehensive pose error of the end-effector; The geometric error is extracted from the end-effector integrated pose error by kinematic calibration to obtain the first residual error; Based on the temperature data and combined with the temperature change rate constraint, the thermal error is decomposed from the first residual error using the state estimation method to obtain the second residual error. Determine feature parameters based on the multi-source data; The characteristic parameters include speed, acceleration, jerk, joint torque, motor current, and temperature parameters; The presence of effective jitter is determined based on the jerk and the second residual error. If so, the feature parameters are input into a pre-trained BP neural network to obtain the predicted value of the end vibration error. If not, then set the predicted value of the end vibration error to zero; The desired trajectory of the robotic arm end effector is corrected in real time based on the predicted values of thermal error and end effector vibration error.
2. The method of claim 1, wherein, The training samples for the BP neural network are obtained in the following way: Acquire historical multi-source data during the historical operation of the robotic arm, and determine the historical end-effector comprehensive pose error based on the historical expected pose and the historical end-effector pose data in the historical multi-source data; The historical geometric error is extracted from the historical end-point comprehensive pose error by kinematic calibration to obtain the first historical residual error; Based on the historical temperature data in the historical multi-source data, and combined with the temperature change rate constraint, the state estimation method is used to decompose the historical thermal error from the first historical residual error to obtain the second historical residual error. Frequency domain analysis and bandpass filtering are performed on the second historical residual error to extract the historical vibration error; The historical vibration error is used as the training label, and the historical feature parameters determined based on the historical multi-source data are used as the training input to construct training samples.
3. The real-time compensation method for end-effector pose error of a robotic arm according to claim 1, characterized in that, The input layer of the BP neural network is equipped with a temperature-compensated feature gating mechanism, which is used to weight and compensate the input feature parameters for temperature, thereby reducing the impact of thermal errors on the prediction results of end-vibration errors; the expression of the temperature-compensated feature gating mechanism is as follows: X ′= WX + G ( T ) in, X For the input feature parameters, W For feature weights, G ( T ) is the temperature compensation function. X ′ represents the input feature after gating.
4. The real-time compensation method for end-effector pose error of a robotic arm according to claim 1, characterized in that, The BP neural network uses an adaptive momentum factor for backpropagation weight update during training. The weight update formula is as follows: in, For the first The weight adjustment amount in the next iteration. For the first The weight adjustment amount in the next iteration. For adaptive momentum factor, For network learning rate, Error function Regarding network weights The gradient.
5. The real-time compensation method for the pose error of the robotic arm end effector according to claim 2, characterized in that, Frequency domain analysis and bandpass filtering are performed on the second historical residual error to extract the historical vibration error, including: Perform a Fourier transform on the second historical residual error to obtain the historical frequency domain signal; The historical vibration error is obtained by performing a fourth-order Butterworth bandpass filter on the components in the historical frequency domain signal whose frequencies are higher than a preset frequency threshold.
6. The real-time compensation method for end-effector pose error according to claim 1, characterized in that, Determining whether there is valid jitter based on the jerk and the second residual error includes: The second residual error is analyzed in the frequency domain to calculate the ratio of the energy with a frequency higher than a preset frequency threshold in the second residual error to the total energy of the second residual error, thus obtaining the proportion of high-frequency energy. When the proportion of high-frequency energy is greater than a preset energy proportion threshold and the jerk is greater than a preset jerk threshold, it is determined that there is effective jitter; otherwise, it is determined that there is no effective jitter.
7. The real-time compensation method for end-effector pose error of a robotic arm according to claim 1, characterized in that, The method further includes: Calculate the ratio of energy with a frequency lower than a preset frequency threshold in the residual error to the total energy; When the ratio is greater than a preset percentage value and the temperature change rate obtained from the temperature data is less than a preset change threshold, the thermal error is determined to be valid.
8. The real-time compensation method for end-effector pose error of a robotic arm according to claim 1, characterized in that, The method further includes: The actual spatial coordinates of the robotic arm's end effector are obtained in real time using a laser tracker. The actual measurement error is determined based on the actual spatial coordinates and the desired pose. The model prediction error is determined based on the actual measurement error, thermal error, and predicted end-vibration error. The BP neural network is updated online incrementally based on the model prediction error.
9. The real-time compensation method for end-effector pose error of a robotic arm according to claim 1, characterized in that, The desired trajectory of the robotic arm's end effector is corrected in real time according to the following correction formula: in, The corrected trajectory. For the desired trajectory, This is the predicted value of the end-vibration error. For thermal error, This refers to the runtime.
10. A real-time compensation system for the pose error of a robotic arm's end effector, characterized in that, include: The multi-source data acquisition module is used to acquire multi-source data during the operation of the robotic arm; The multi-source data includes joint motion data, motor drive current data, joint torque data, end-effector pose data, and temperature data. The end-effector integrated pose error determination module is used to determine the end-effector integrated pose error based on the desired pose and the end-effector pose data. A geometric error stripping module is used to strip geometric errors from the end-effector integrated pose error through kinematic calibration to obtain a first residual error. The thermal error decomposition module is used to decompose the thermal error from the first residual error based on the temperature data and combined with the temperature change rate constraint, using a state estimation method to obtain the second residual error. The vibration error determination module is used for: Determine feature parameters based on the multi-source data; The characteristic parameters include speed, acceleration, jerk, joint torque, motor current, and temperature parameters; The presence of effective jitter is determined based on the jerk and the second residual error. If so, the feature parameters are input into a pre-trained BP neural network to obtain the predicted value of the end vibration error. If not, then set the predicted value of the end vibration error to zero; The real-time trajectory correction module is used to correct the desired trajectory of the robotic arm end effector in real time based on the predicted values of the thermal error and the end effector vibration error.
