Parallel platform pose error prediction and compensation method based on TCN-BiGRU-attention model

Abstract

The application discloses a parallel platform pose error prediction and compensation method based on a TCN-BiGRU-Attention model, selects a target pose of a parallel platform through a grid division mode, controls the platform to realize corresponding fixed-point motion under different loads respectively. The actual pose data of the parallel platform is collected by using a metrological branch chain pose measurement system. The pose error prediction based on the TCN-BiGRU-Attention model, the constructed data set is proportionally divided into a training set and a test set. The data interaction between the host computer and the Beckhoff control system is realized based on the ADS communication protocol, the target pose in the test set and the real-time collected current value are input into the model, the pose error is predicted, and the pose error prediction is completed. The predicted pose error is used as a compensation term and is superimposed into the target pose in advance, and the corrected target pose is input into the parallel platform control system to complete the error compensation. The application can significantly improve the control precision and stability of the platform.

Description

Technical Field

[0001] This invention belongs to the fields of parallel platforms and deep learning technology, specifically relating to a method for predicting and compensating pose errors of a six-degree-of-freedom parallel platform based on the TCN-BiGRU-Attention model, which is used to improve the pose accuracy of the parallel platform under variable load. Background Technology

[0002] Parallel robots are a type of mechanical system with multiple degrees of freedom. Compared to serial robots, they have significant advantages in stiffness, stability, and load-bearing capacity due to their closed-chain structure. Therefore, they are widely used in precision assembly, motion simulation, aerospace, and other fields. In practical applications, when parallel robots are subjected to external loads, their pose errors increase significantly, especially non-geometric errors. Non-geometric errors include structural component assembly errors, friction errors, and local elastic deformation errors of materials. Decreased accuracy under stress not only affects the quality of task completion but may also pose safety hazards. Therefore, effectively suppressing errors caused by load variations and ensuring that robots maintain high-precision operation under load is of great importance.

[0003] Traditional pose error compensation methods primarily rely on kinematic calibration. This involves constructing an error model, identifying geometric error parameters using parameter identification, and then correcting the model accordingly to improve robot accuracy. Some researchers have attempted to replace traditional error models with neural networks. While neural networks can enhance the modeling ability for complex nonlinear errors under load, they are only suitable for stable load conditions. Their generalization ability is insufficient under real-time load variations, making it difficult to achieve high-precision and stable error compensation. To improve the pose accuracy of parallel robots under varying load conditions, this paper proposes a deep learning-based method for pose error prediction and compensation, using the servo motor current as a monitoring parameter for load changes.

[0004] Current kinematic calibration methods suffer from problems such as difficulty in establishing error models and lack of universality. Ordinary neural networks also have problems such as insufficient generalization ability and poor stability under variable loads. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting and compensating pose errors of a six-DOF parallel platform under variable load conditions based on a TCN-BiGRU-Attention model. Addressing the problems of difficulty in establishing error models and lack of universality in current kinematic calibration methods, and the insufficient generalization ability and poor stability of ordinary neural networks under variable load conditions, this invention proposes a method for predicting and compensating pose errors of a six-DOF parallel platform based on a TCN-BiGRU-Attention model. This method significantly improves the overall accuracy and stability of the system, providing theoretical support for precision control of robots under variable load environments. It includes:

[0006] S1: The target pose of the parallel platform is selected through mesh generation. Under various load conditions, the parallel platform is controlled to achieve fixed-point motion of the corresponding pose. The actual pose of the parallel platform is obtained using a pose measurement system, and the pose error between the actual pose and the target pose is calculated. Simultaneously, the current values ​​of six servo motors are collected by the control system to construct a dataset of target pose, current, and actual pose error. For different external load conditions, three loads of 10kg, 30kg, and 50kg are selected, corresponding to light load, medium load, and heavy load conditions, respectively, covering typical load variation ranges. The position and attitude task space is divided into 10×10×10 and 8×8×8 meshes, and 720 positions and attitudes are randomly selected in the mesh as the target pose.

[0007] The pose error between the actual pose and the target pose can be described as follows:

[0008]

[0009] Where ΔQ represents the pose error, Q a Q represents the actual pose. t Indicates the target pose.

[0010] S2: Establish a pose error prediction model for a parallel platform based on TCN, BiGRU and Attention network. The input of the model is the target pose of the parallel platform and its corresponding current value, and the output is the pose error.

[0011] The target pose is input into a multi-layer temporal convolution; the current is input into a multi-layer BiGRU network for bidirectional sequence modeling. Finally, the pose features and current features are adaptively weighted and fused through an attention mechanism. The fused features are then input into a fully connected layer, and a 6-dimensional error prediction value is finally output.

[0012] S3: Divide the target pose-current-pose error dataset into a training set and a test set in a 4:1 ratio, and train the TCN-BiGRU-Attention model using the training set data;

[0013] The TCN-BiGRU-Attention model was trained and tested using the datasets respectively. The number of TCN layers and BiGRU layers was determined by the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The detailed formulas for calculating RMSE, MAE, and R² are as follows:

[0014]

[0015]

[0016]

[0017] in,

[0018]

[0019] in, For predicted values, This is the actual value. Let n represent the average of the actual values, and n be the number of poses in the test set.

[0020] S4: The servo motor current data collected in real time by the Beckhoff control system and the target pose data in the test set are transmitted to the host computer through the ADS communication protocol, and used as the input of the TCN-BiGRU-Attention model to realize pose error prediction.

[0021] S5: The predicted pose error is pre-compensated to the target pose of the parallel platform and used as the input of the control system.

[0022] The predicted pose error is pre-compensated to the target pose of the parallel platform. Specifically, this is achieved by using the predicted pose error ΔQ = [Δx, Δy, Δz, Δα, Δβ, Δγ] and pre-compensating ΔQ to the target pose Q. t The corrected pose Q was then obtained. b Using Q b Replace Q t As the input to the control system, Q b It can be represented as:

[0023]

[0024] The present invention provides a method for predicting and compensating pose errors of a six-DOF parallel platform under variable load based on a TCN-BiGRU-Attention model, which has the following advantages:

[0025] (1) The present invention simultaneously introduces target pose and servo motor current data as model input, fully explores the coupling characteristics between force and pose, and the constructed data-driven model has stronger generalization ability and robustness.

[0026] (2) This invention trains the model by collecting data under various typical loads, and is applicable to parallel platforms under different load changes, thus having high versatility.

[0027] (3) The present invention can effectively improve the pose accuracy of parallel robots under variable load, and the pose error is reduced by 95.64% compared with the uncompensated version, meeting the high precision requirements. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention;

[0029] Figure 2 This is a schematic diagram of the method of the present invention;

[0030] Figure 3 This is a structural diagram of the TCN-BiGRU-Attention model of the method of the present invention;

[0031] Figure 4 This study compares the pose error of the parallel platform under variable load based on this method with the actual pose error. Detailed Implementation

[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0033] Embodiments of the present invention: Addressing the problems of difficulty in establishing error models and weak universality in current kinematic calibration methods, and the insufficient generalization ability and poor stability of ordinary neural networks under variable load conditions, this invention proposes a method for predicting and compensating pose errors of a six-degree-of-freedom parallel platform based on the TCN-BiGRU-Attention model, specifically including:

[0034] refer to Figure 1 The flowchart of the method of the present invention mainly includes the following steps:

[0035] Step S1: Select a large number of target poses of the parallel platform by meshing, control the parallel platform to achieve fixed-point motion of the corresponding poses under various load conditions, use a pose measurement system to obtain the actual pose of the parallel platform, calculate the pose error between the actual pose and the target pose, and construct a dataset of target pose and actual pose error.

[0036] Step S2: Establish a pose error prediction model for a parallel platform based on TCN, BiGRU and Attention. The input of the model is the target pose of the parallel platform and its corresponding current value, and the output is the pose error.

[0037] Step S3: Divide the target pose-current-pose error dataset into a training set and a test set in a 4:1 ratio, and train the TCN-BiGRU-Attention model using the training set data;

[0038] Step S4: Transmit the servo motor current data collected in real time by the Beckhoff control system and the target pose data in the test set to the host computer through the ADS communication protocol, and use them as input to the TCN-BiGRU-Attention model to achieve pose error prediction.

[0039] Step S5: The predicted pose error is pre-compensated to the target pose of the parallel platform as the input of the control system.

[0040] In step S1, three loads of 10kg, 30kg, and 50kg are selected for different external load conditions, corresponding to light, medium, and heavy load conditions, respectively, covering typical load variation ranges. The position and attitude task space is divided into 10×10×10 and 8×8×8 grids, and 720 positions and attitudes are randomly selected in the grid as target poses. The control system realizes the fixed-point motion of the corresponding poses, and the actual pose of the parallel platform is measured using a pose measurement system. The pose error between the actual pose and the target pose is calculated. At the same time, the control system collects the current data of the six servo motors to construct a dataset of the target pose, current, and actual pose error of the parallel platform.

[0041] In this implementation example, the metrology branch is used to measure the pose of the parallel platform with a measurement accuracy of ±5μm. However, this method is for the prediction and compensation of pose error of the parallel platform under variable load and is not limited to pose measurement equipment. Other pose measurement equipment can also be used for measurement.

[0042] The pose error between the actual pose and the target pose can be described as follows:

[0043]

[0044] Where ΔQ represents the pose error, Q t Indicates the target pose, Q a Indicates the actual pose.

[0045] In step S2, the target pose is input into a multi-layer temporal convolution; the current is input into a multi-layer BiGRU network for bidirectional sequence modeling; finally, the pose features and current features are adaptively weighted and fused through an attention mechanism, and the fused features are input into a fully connected layer to finally output the error prediction value.

[0046] In step S3, the TCN-BiGRU-Attention model is trained and tested using the dataset. The number of TCN layers and BiGRU layers is determined by the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The detailed calculation formulas for RMSE, MAE, and R² are as follows:

[0047]

[0048]

[0049]

[0050] in,

[0051]

[0052] in, For predicted values, This is the actual value. Let n represent the average of the actual values, and n be the number of poses in the test set.

[0053] In step S4, the servo motor current data collected in real time by the Beckhoff control system and the target pose data in the test set are transmitted to the host computer through the ADS communication protocol as input to the TCN-BiGRU-Attention model to achieve pose error prediction.

[0054] In step S5, the predicted pose error is pre-compensated to the target pose of the parallel platform as the input of the control system.

[0055] Specifically, the implementation involves using the target pose and corresponding current value of the parallel platform in the test set data as input to the TCN-BiGRU-Attention model to obtain the pose error ΔQ = [Δx, Δy, Δz, Δα, Δβ, Δγ]. ΔQ is then pre-compensated to the target pose Q. t The corrected pose Q was then obtained. b Using Q b Replace Q t As the input to the control system, Q b It can be represented as:

[0056]

[0057] The above content can be summarized in the following aspects:

[0058] Construction of the target pose, current, and actual pose error dataset: For different load conditions, three loads of 10kg, 30kg, and 50kg were selected, corresponding to light, medium, and heavy load conditions respectively, covering typical load variation ranges. The position and attitude task space was divided into 10×10×10 and 8×8×8 grids, and 720 positions and attitudes were randomly selected from the grids as the target pose. Fixed-point motion of the corresponding pose was achieved through the control system, and the actual pose of the parallel platform was calculated using a pose measurement system. The pose error between the actual pose and the target pose was calculated. Simultaneously, the current data of the six servo motors were collected using the control system to construct the target pose, current, and actual pose error dataset of the parallel platform.

[0059] Building the TCN-BiGRU-Attention model: The TCN-BiGRU-Attention model for predicting the pose error of a parallel platform is built by combining TCN, BiGRU and Attention networks. The input of the model is the target pose and current of the six-DOF parallel platform, and the output is the pose error.

[0060] Determine suitable TCN-BiGRU-Attention model parameters: Divide the dataset into training and test sets in a 4:1 ratio. Select different numbers of TCN and BiGRU layers, train the model using the training set data, and determine the most suitable number of neural network layers.

[0061] Pose error prediction and compensation for the parallel platform: Servo motor current data acquired in real-time by the Beckhoff control system and target pose data from the test set are transmitted to the host computer via the ADS communication protocol. This data serves as input to the TCN-BiGRU-Attention model to predict pose errors. The predicted pose error is then pre-compensated and incorporated into the target pose of the parallel platform, replacing the target pose as input to the control system.

[0062] refer to This diagram illustrates the prediction and compensation method based on the TCN-BiGRU-Attention model of this invention. The actual pose is acquired in real-time by a pose measurement system, and corresponding current data is obtained from the servo driver to construct a target pose-current-pose error dataset. After training, the TCN-BiGRU-Attention model is deployed to a host computer for predicting pose error. The Beckhoff control system collects servo motor current data, combines it with target pose information, and transmits it to the host computer via the ADS communication protocol. The model takes the target pose and current data as input, outputs the current pose error, and corrects the original target pose command accordingly to form a compensated target pose command.

[0063] refer to The diagram shows the structure of the TCN-BiGRU-Attention model of this invention. The TCN-BiGRU-Attention model first synchronizes and aligns the collected target pose data and current data, and then normalizes them to eliminate the influence of dimensions. Subsequently, the target pose data is constructed into a fixed-length sequence suitable for TCN input, and the current data is constructed into a time-series format suitable for BiGRU input. The model takes the target pose sequence and current sequence as input, extracts features through TCN and BiGRU respectively, and then fuses them through an attention mechanism. Finally, the input is fed to a fully connected layer to output a six-dimensional pose error prediction result. The TCN module consists of three layers of one-dimensional causal convolutions and dilated convolutions. Each convolutional layer expands its receptive field by setting different dilation coefficients, acquiring long-term temporal dependency information without increasing network depth. Simultaneously, the residual connection structure alleviates the gradient vanishing problem during deep network training. This module is used to extract features from the input target pose sequence, mapping the original six-dimensional pose data into a high-dimensional feature representation, thereby extracting the spatial and temporal coupling features during pose change. The BiGRU module consists of two gated recurrent units, one forward and one backward. It iterates bidirectionally through the input sequence, extracting historical and future information respectively, and then fuses them in the time dimension to obtain a complete temporal feature representation. In this invention, the BiGRU module is used to model the servo motor current sequence, extracting the load dynamics and their temporal correlation contained in the current changes. The attention mechanism module is used to perform weighted fusion of features from the TCN and BiGRU. The pose features output by the TCN and the current features output by the BiGRU are concatenated or mapped to a unified feature space. The attention weight calculation unit generates weight coefficients for each feature channel and performs weighted processing on the fused features, thereby highlighting key features that have a significant impact on pose error, suppressing redundant information, and improving the model's prediction accuracy.

[0064] The model parameters of the TCN-BiGRU-Attention neural network include: number of training rounds, learning rate, and decay rate, which are set to 1000, 0.001, and 0.00001, respectively.

[0065] To verify the effectiveness of the method for predicting and compensating pose errors of parallel platforms under varying loads based on the TCN-BiGRU-Attention model, 160 poses for each load (10kg, 30kg, and 50kg) were used as the training set, 40 poses as the test set, and 40 poses as the validation set.

[0066] Table 1 shows the parameters of the 6-UCU parallel platform used in the validation experiment. The model was trained using the training set data, and the number of neural network layers was determined using RMSE, MAE, and R². Smaller RMSE and MAE values ​​indicate better model performance, meaning the predicted values ​​are closer to the true values. R² represents the model's fit to the data; a closer R² is to 1 indicates better performance. Finally, the number of TCN layers was determined to be 3, and the number of BiGRU layers to be 2. Table 2 shows the evaluation metrics for the model's pose prediction error performance, and Table 3 shows the compensation results for the TCN-BiGRU-Attention model on the validation set.

[0067] surface Parallel platform parameters

[0068]

[0069] surface Test set evaluation metrics of the TCN-BiGRU-Attention model

[0070]

[0071] surface Compensation results of the TCN-BiGRU-Attention model validation set

[0072]

[0073] Figure 4 The figure compares the pose error of the parallel platform after compensation using this method with the pose error before compensation. As can be seen from the figure, the pose accuracy of the parallel platform under variable load is greatly improved after compensation using this method.

[0074] The foregoing description provides a detailed account of embodiments of the present invention and is not intended to limit the invention in any way. Those skilled in the art can make various optimizations, improvements, and modifications based on this invention. Therefore, the scope of protection of this invention should be defined by the appended claims.

Claims

1. A method for predicting and compensating pose errors of a parallel platform based on a TCN-BiGRU-Attention model, characterized in that: The method includes the following steps: S1: Select the target pose of the parallel platform by meshing, control the parallel platform under different loads to achieve the fixed-point motion of the target pose, use a pose measurement system to obtain the actual pose of the parallel platform, calculate the pose error between the actual pose and the target pose, and use the control system to collect the current values ​​of the six servo motors to construct a dataset of target pose, current value and actual pose error. S2: Establish a parallel platform pose error prediction model based on TCN, BiGRU and Attention network. The input of the model is the target pose of the parallel platform and its corresponding current value, and the output is the pose error. S3: Divide the target pose-current-pose error dataset into a training set and a test set in a 4:1 ratio, and train the TCN-BiGRU-Attention model using the training set data; S4: The target pose data in the test set and the servo motor current data collected in real time by the Beckhoff control system are transmitted to the host computer through the ADS communication protocol as input to the TCN-BiGRU-Attention model to realize pose error prediction. S5: The predicted pose error is pre-compensated to the target pose of the parallel platform and used as the input of the control system.

2. The method for predicting and compensating pose errors of a parallel platform based on the TCN-BiGRU-Attention model according to claim 1, characterized in that: In step S1, different load conditions are set, and the target pose of the parallel platform is selected by dividing the grid: three different loads of 10kg, 30kg and 50kg are selected, which correspond to light load, medium load and heavy load conditions respectively; the position and attitude task space is divided into 10×10×10 grid and 8×8×8 grid, and 720 positions and attitudes are randomly selected in the grid as the target pose; The pose error between the actual pose and the target pose is described as follows: ； Where ΔQ represents the pose error, Q a Q represents the actual pose. t Indicates the target pose.

3. The method for predicting and compensating pose errors of a parallel platform based on the TCN-BiGRU-Attention model according to claim 1, characterized in that: In S2, the target pose is input into a multi-layer temporal convolution; the current is input into a multi-layer BiGRU network for bidirectional sequence modeling. The pose features and current features are adaptively weighted and fused through a feature fusion attention mechanism. The fused features are then input into a fully connected layer, and the final error prediction value is output.

4. The method for predicting and compensating pose errors of a parallel platform based on the TCN-BiGRU-Attention model according to claim 1, characterized in that: In step S3, the TCN-BiGRU-Attention model is trained and tested using the dataset. The number of TCN layers and BiGRU layers is determined by the root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). The detailed calculation formulas for RMSE, MAE, and R² are as follows: ； ； ； in, ； in, For predicted values, This is the actual value. Let be the average of the actual values, n be the number of poses in the test set, and j be the pose index.

5. The method for predicting and compensating pose error of a parallel platform based on the TCN-BiGRU-Attention model according to claim 1, characterized in that: In S4, the Beckhoff control system collects servo motor current data, combines it with target pose information, and transmits it to the host computer via the ADS communication protocol; the model takes the target pose and current data as input and outputs the current pose error.

6. The method for predicting and compensating pose errors of a parallel platform based on the TCN-BiGRU-Attention model according to claim 1, characterized in that: In step S5, the predicted pose error is pre-compensated to the target pose of the parallel platform: using the predicted pose error ΔQ = [Δx, Δy, Δz, Δα, Δβ, Δγ], where Δx, Δy, and Δz are the translation errors along the X, Y, and Z axes, and Δα, Δβ, and Δγ are the attitude angle errors around the X, Y, and Z axes; ΔQ is pre-compensated to the target pose Q. t The corrected pose Q was then obtained. b Using Q b As the input to the control system, Q b Represented as: 。 7. The method for predicting and compensating pose errors of a parallel platform based on the TCN-BiGRU-Attention model according to claim 1, characterized in that: The TCN-BiGRU-Attention model first synchronizes and aligns the collected target pose data and current data, and then performs normalization to eliminate the influence of dimensions. Subsequently, the target pose data is constructed into a fixed-length sequence suitable for TCN input, and the current data is constructed into a time series suitable for BiGRU input. Using the target pose sequence and current sequence as input, features are extracted by TCN and BiGRU respectively, then fused through an attention mechanism, and finally input to a fully connected layer to output a six-dimensional pose error prediction result. The TCN module consists of three layers of one-dimensional causal convolutions and dilated convolutions. Each convolutional layer expands its receptive field by setting different dilation coefficients, acquiring long-term temporal dependency information without increasing network depth. The residual connection structure alleviates the gradient vanishing problem during deep network training. The TCN module is used to extract features from the input target pose sequence, mapping the original six-dimensional pose data into a high-dimensional feature representation, and extracting the spatial representation of the pose change process. The BiGRU module, composed of two gated recurrent units (forward and reverse), performs bidirectional traversal of the input sequence to extract historical and future information, which are then fused in the time dimension to obtain a complete temporal feature representation. It is used to model the servo motor current sequence, extracting the load dynamics and temporal correlation contained in the current changes. The attention mechanism module performs weighted fusion of features from the TCN and BiGRU. It concatenates or maps the pose features output by the TCN with the current features output by the BiGRU to a unified feature space. The attention weight calculation unit generates weight coefficients for each feature channel and weights the fused features, thereby highlighting key features that significantly affect pose error, suppressing redundant information, and improving the model's prediction accuracy.

Images