Intelligent prediction and uncertainty analysis method for pose error of six-degree-of-freedom motion platform

Abstract

The invention discloses an intelligent prediction and uncertainty analysis method for pose errors of a six-degree-of-freedom motion platform, and the method comprises the steps: taking an expected pose in a high-precision six-degree-of-freedom motion platform control system as network input data, and obtaining an actual pose through employing a binocular vision measurement system; collecting multiple groups of data, normalizing the data, and dividing the data into a training set and a test set; constructing a pose error prediction network model taking a gated circulation unit neural network as a basic framework, designing a translation and rotation feature coupling attention module, introducing a sparse attention mechanism guided by domain knowledge, and improving the capturing capability of the model for the real mechanical coupling characteristics of the six-degree-of-freedom platform; and inputting a test set into the trained model, and evaluating data of the test set by adopting a Monte Carlo Dropout method to obtain a pose error prediction result and an uncertainty analysis result. According to the method, the static pose precision of the six-degree-of-freedom motion platform can be effectively improved, a visual and reliable network prediction confidence level is provided, and the precision measurement requirement in the metering field is met.

Description

Technical Field

[0001] The present invention relates to the technical field of robot control and metrology, and in particular to an intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, which provides a reliable basis for high-precision measurement and error analysis. Background Art

[0002] Six-degree-of-freedom motion platforms are widely used in flight simulators, ship simulators, robotic control, and other fields. Precisely controlling the position and orientation of these platforms is key to achieving high-precision motion control. However, due to factors such as mechanical structure errors and sensor noise, errors can occur between the actual and desired positions. Traditional error compensation methods rely on precise mathematical models and complex calibration procedures, making them difficult to adapt to dynamically changing environments and mission requirements.

[0003] In recent years, deep learning technology has achieved remarkable results in fields such as image recognition and natural language processing. With the development of neural network technology, its powerful nonlinear mapping capabilities have provided a new approach to solving the problem of pose error prediction for six-degree-of-freedom motion platforms. By learning from large amounts of pose data, neural networks can capture complex pose error patterns. However, research on applying deep learning to pose error prediction for six-degree-of-freedom motion platforms is still in its infancy. Conventional neural networks typically only output a single numerical result when making predictions, which provides no information about the uncertainty of the prediction results. In the field of metrology, effectively assessing the uncertainty of prediction results is key to ensuring data reliability and traceability. Monte Carlo Dropout is an effective technique for introducing randomness into neural networks. By performing multiple forward propagations to obtain different prediction results, the uncertainty of the prediction results can be assessed. This method not only fully utilizes the powerful ability of neural networks to capture complex characteristics, but also provides reliable confidence intervals for measurement results, providing strong support for the accurate prediction and reliable assessment of pose errors for six-degree-of-freedom motion platforms. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, aiming to solve the problems of insufficient accuracy of existing six-degree-of-freedom motion platform posture error prediction methods and inability to effectively evaluate the uncertainty of the prediction results, so as to meet the needs of the metrology field for high-precision measurement and reliable uncertainty assessment.

[0005] The technical solution of the present invention is an intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, which comprises the following steps:

[0006] S1: The desired pose of a high-precision six-degree-of-freedom motion platform control system is used as network input data. A strictly calibrated and calibrated binocular vision measurement system is used to measure the actual pose of the platform. The difference between the two is calculated as error label data. Multiple sets of platform data in the desired pose are collected to form a dataset. The data is normalized and divided into a training set and a test set.

[0007] S2: Construct a network model for predicting the pose error of a six-degree-of-freedom motion platform. The Gated Recurrent Unit (GRU) neural network is selected as the basic architecture. A translation and rotation feature coupling attention module is designed. A sparse attention mechanism guided by domain knowledge is introduced to improve the model's ability to capture the true mechanical coupling characteristics of the six-degree-of-freedom platform. The network output layer outputs the predicted six-degree-of-freedom pose error.

[0008] S3: Use the training set in S1 to train the network model in S2. In each training cycle, the training data is input into the model in batches, the error between the predicted results and the true labels is calculated, and the parameters of the model are accurately updated through the backpropagation algorithm to minimize the loss function.

[0009] S4: Input the S1 test set into the network model trained in S3, use the Monte Carlo Dropout method to test and evaluate the test set data, obtain the test sample pose error prediction results, and output the corresponding uncertainty analysis results.

[0010] In the aforementioned intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, in step S1, the load of the six-degree-of-freedom motion platform is 9.9 kg, and the posture is obtained by uniform random sampling within the platform workspace, with 40 test set samples and 257 training set samples.

[0011] In the aforementioned method for intelligent prediction and uncertainty analysis of the pose error of a six-degree-of-freedom motion platform, the six-degree-of-freedom motion platform pose error prediction network model structure in step S2 includes an input layer, a coupled attention module, a GRU layer, a dropout layer, a fully connected layer, and an output layer. The input layer size is set to 6, corresponding to the platform's six degrees of freedom. The GRU layer performs in-depth feature extraction and processing on the input data. Its hidden layer size is set to 64, with a layer number of 1, and the output layer size is set to 6.

[0012] In the aforementioned intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, in step S3, the network model parameters are set as follows: the dropout ratio is set to 0.005, the Adam optimizer is used for training, the sample batch size is 16, the initial learning rate is set to 0.01, and after 100 training times, it is multiplied by 0.75 to reduce the learning rate, and the mean square error loss function is used for training.

[0013] In the aforementioned intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, in step S4, the number of Monte Carlo experiments is set to 50, and the root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators to evaluate the prediction performance of the model.

[0014] Beneficial effects of the present invention: Compared with the existing technology, the present invention can significantly improve the static posture accuracy of the platform by constructing a six-degree-of-freedom motion platform posture error prediction network model, while providing an intuitive and reliable confidence level, which can be flexibly adapted to different working conditions and scenarios to meet the precision measurement needs in the metrology field.

[0015] The present invention has the following advantages:

[0016] (1) The present invention can realize the posture error prediction of a six-degree-of-freedom motion platform and can significantly improve the static posture accuracy of the platform compared with traditional calibration methods.

[0017] (2) The coupling attention module designed in this invention improves the model's ability to capture the real mechanical coupling characteristics of the six-degree-of-freedom platform by introducing domain knowledge.

[0018] (3) The present invention uses the Monte Carlo Dropout method to perform uncertainty analysis on the test results, provide an intuitive confidence level, and enhance the reliability of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Attachment Figure 1 This is a flow chart of the intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform of the present invention;

[0020] Attachment Figure 2a Schematic diagram of the spatial distribution of measuring point locations;

[0021] Attachment Figure 2b Schematic diagram of the spatial distribution of measurement point postures;

[0022] Attachment Figure 3 Schematic diagram of the Monte Carlo Dropout method;

[0023] Attachment Figure 4 This is the deviation result diagram of the posture error;

[0024] Attachment Figure 5a Predict the x-axis pose error and uncertainty results for the network;

[0025] Attachment Figure 5b Predict the y-axis pose error and uncertainty results for the network;

[0026] Attachment Figure 5cThe network predicts the z-axis pose error and uncertainty results.

[0027] Attachment Figure 5d Predict the α-axis pose error and uncertainty result map for the network;

[0028] Attachment Figure 5e Predict the β-axis pose error and uncertainty results for the network;

[0029] Attachment Figure 5f Graph of the network's predicted γ-axis pose error and uncertainty. DETAILED DESCRIPTION

[0030] The present application will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to illustrate the relevant invention and are not intended to limit the invention. The present application will be described in detail below with reference to the accompanying drawings and examples.

[0031] Figure 1 The present invention shows an intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, which includes the following steps:

[0032] S1: The desired posture of a high-precision six-degree-of-freedom motion platform control system is used as the network input data. A strictly calibrated and calibrated binocular vision measurement system is used to measure the actual posture of the platform. The difference between the two is calculated as the error label data. Multiple sets of platform data in the desired posture are collected to form a data set. The data is normalized and divided into a training set and a test set.

[0033] In this embodiment, the high-precision six-degree-of-freedom motion platform consists of six legs, a fixed base and a moving platform. One end of each leg is connected to the fixed base through a Hooke's hinge, and the other end is connected to the moving platform through a ball joint. The length of the legs is adjusted by a servo motor, and the platform can achieve six-degree-of-freedom motion, including translation (along the X, Y, and Z axes) and rotation (around the X, Y, and Z axes) in three-dimensional space. The position and posture of the moving platform are expressed as (x, y, z, α, β, γ), where x, y, z represent the position of the platform and α, β, γ represent its posture. The origin of the coordinate system is set at the initial position of the moving platform to ensure consistent representation. A binocular vision measurement system is used to measure the actual position and posture of the platform. To ensure the reliability of the measurement, random and uniform sampling is performed within the working space of the platform. Figure 2a is the spatial distribution of the measurement point locations, Figure 2b The spatial distribution of the measured point postures, the diamond points represent 40 randomly selected test set samples, and the remaining 257 circles represent training set samples.

[0034] S2: Construct a network model for predicting the pose error of a six-degree-of-freedom motion platform. Select the gated recurrent unit (GRU) neural network as the basic architecture, design a translation and rotation feature coupling attention module, and introduce a sparse attention mechanism guided by domain knowledge to improve the model's ability to capture the real mechanical coupling characteristics of the six-degree-of-freedom platform. The network output layer outputs the predicted six-degree-of-freedom pose error.

[0035] In this embodiment, the six-degree-of-freedom motion platform posture error prediction network model structure includes an input layer, a coupled attention module, a GRU layer, a dropout layer, a fully connected layer, and an output layer. Among them, the input layer size is set to 6, corresponding to the six degrees of freedom of the platform, and can fully receive the platform's expected posture input data. Furthermore, a coupled attention module is designed to perform feature embedding operations on translation features (x, y, z) and rotation features (α, β, γ) respectively, and a multi-head self-attention mechanism is applied to translation features and rotation features respectively to extract the dependencies within each degree of freedom; through the cross-attention mechanism constrained by the physical mask matrix, the dynamic coupling relationship between the translational degree of freedom and the rotational degree of freedom is modeled, and a physical mask matrix basis is designed to only allow feature interaction between the degrees of freedom that actually have mechanical coupling. The cross-attention weight is calculated by the following formula:

[0036]

[0037] Among them, Q, K, V are translation and rotation characteristics; ⊙ is the Hadamard product; is a scaling factor to prevent gradient vanishing; M is a physical mask matrix, which is 1 when the translational and rotational degrees of freedom are coupled, and 0 otherwise. Furthermore, the physical coupling characteristics of the mechanical system are encoded into the attention mechanism, enabling sparse interaction guided by domain knowledge.

[0038] On this basis, the GRU layer extracts and processes features from the data passed by the module. Specifically, the hidden layer size is set to 64, and the number of layers is set to 1, ensuring the model has sufficient feature learning capabilities. To prevent overfitting during training, a dropout layer is introduced to improve the model's generalization ability by randomly discarding some neurons. The fully connected layer integrates the features processed by the GRU layer to further explore the potential connections between features. Finally, the output layer size is set to 6, and its output is the predicted six-degree-of-freedom pose error, which directly provides key reference information for subsequent error compensation and platform control.

[0039] S3: Use the training set in S1 to train the network model in S2. In each training cycle, the training data is input into the model in batches, the error between the predicted results and the true labels is calculated, and the parameters of the model are accurately updated through the backpropagation algorithm to minimize the loss function.

[0040] In this embodiment, the network model parameters are set as follows: the dropout ratio is set to 0.005, the Adam optimizer is used for training, the sample batch size is 16, the initial learning rate is set to 0.01, and after 100 training times, it is multiplied by 0.75 to reduce the learning rate, and the mean square error loss function is used for training.

[0041] S4: Input the S1 test set into the network model trained in S3, use the Monte Carlo Dropout method to test and evaluate the test set data, obtain the test sample pose error prediction results, and output the corresponding uncertainty analysis results.

[0042] In this embodiment, the number of Monte Carlo trials N is set to 50. Assume that the network model with parameter θ is f(Φ,θ). In Monte Carlo Dropout, by performing N forward propagations on the network, multiple prediction means and variances are obtained.

[0043]

[0044] The forecast mean gives the most likely estimate of the forecast, while the forecast variance quantifies the uncertainty in that forecast. Figure 3 Schematic diagram of the Monte Carlo Dropout method.

[0045] The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators to evaluate the prediction performance of the model. Table 1 shows the error prediction results of this embodiment.

[0046] Table 1 Error prediction results

[0047]

[0048] Figure 4 Figure 5 shows the deviation of the predicted pose error output by the network on 40 test samples. The results demonstrate that this method achieves high accuracy in all dimensions. Figure 5 shows the pose error and uncertainty for each axis of the test sample. The shaded area around the predicted mean represents the uncertainty of the network output obtained using the Monte Carlo Dropout method. This visualization provides insight into the reliability of the predictions, enabling the identification of potential outliers or areas for further improvement of model performance.

[0049] The above description is a detailed description of an embodiment of the present invention and is not intended to limit the present invention in any form. Those skilled in the art may make a series of optimizations, improvements, and modifications based on the present invention. Therefore, the scope of protection of the present invention shall be defined by the appended claims.

Claims

1. An intelligent prediction and uncertainty analysis method for the posture error of a six-degree-of-freedom motion platform, characterized by: It includes the following steps: S1: The desired pose of a high-precision six-degree-of-freedom motion platform control system is used as network input data. A strictly calibrated and calibrated binocular vision measurement system is used to measure the actual pose of the platform. The difference between the two is calculated as error label data. Multiple sets of platform data in the desired pose are collected to form a dataset. The data is normalized and divided into a training set and a test set. S2: Construct a six-degree-of-freedom motion platform pose error prediction network model. The Gated Recurrent Unit (GRU) neural network is selected as the basic architecture. A translation and rotation feature coupling attention module is designed. Domain knowledge-guided sparse attention is introduced to improve the model's ability to capture real-world mechanical coupling characteristics. The network output layer outputs the predicted six-degree-of-freedom pose error. S3: Use the training set in S1 to train the network model in S2. In each training cycle, the training data is input into the model in batches, the error between the predicted results and the true labels is calculated, and the parameters of the model are accurately updated through the backpropagation algorithm to minimize the loss function. S4: Input the S1 test set into the network model trained in S3, use the Monte Carlo Dropout method to test and evaluate the test set data, obtain the test sample pose error prediction results, and output the corresponding uncertainty analysis results.

2. The method for intelligent prediction and uncertainty analysis of posture errors of a six-degree-of-freedom motion platform according to claim 1, characterized in that: In step S1, the load of the six-degree-of-freedom motion platform is 9.9 kg, and poses are obtained by uniform random sampling in the platform workspace, including 40 test set samples and 257 training set samples.

3. The method for intelligent prediction and uncertainty analysis of posture errors of a six-degree-of-freedom motion platform according to claim 1, characterized in that: The six-degree-of-freedom motion platform posture error prediction network model structure in step S2 includes an input layer, a coupled attention module, a GRU layer, a dropout layer, a fully connected layer and an output layer; the input layer size is set to 6, corresponding to the six degrees of freedom of the platform, the GRU layer performs in-depth feature extraction and processing on the input data, its hidden layer size is set to 64, the number of layers is 1, and the output layer size is set to 6.

4. The method for intelligent prediction and uncertainty analysis of posture errors of a six-degree-of-freedom motion platform according to claim 1, characterized in that: In step S3, the network model parameters are set as follows: the dropout ratio is set to 0.005, the Adam optimizer is used for training, the sample batch size is 16, the initial learning rate is set to 0.01, and the learning rate is reduced by multiplying it by 0.75 after 100 training cycles, and the mean square error loss function is used for training.

5. The method for intelligent prediction and uncertainty analysis of posture errors of a six-degree-of-freedom motion platform according to claim 1, characterized in that: In step S4, the number of Monte Carlo experiments is set to 50, and the root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators to evaluate the prediction performance of the model.

Images