Deep learning aided rotation vector optimization method and system oriented to high dynamic environment
By employing a deep learning-assisted rotation vector optimization method, combined with rotation vector differential equations and the fourth-order Runge-Kutta method, the accuracy and complexity issues of attitude calculation in high-dynamic environments were resolved, achieving high-precision attitude calculation.
Patent Information
- Application Number
- CN202511602764.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-04
- Publication Date
- 2026-02-13
AI Technical Summary
In highly dynamic environments, traditional attitude calculation methods suffer from increased non-commutative errors caused by high-speed motion, making it difficult to meet the accuracy requirements of attitude calculation, while also having high computational complexity.
A deep learning-assisted rotation vector optimization method is adopted. By training a network model with a deep learning dataset, and combining the rotation vector differential equation with the fourth-order Runge-Kutta method, the state equation of UKF is constructed to achieve accurate prediction and fusion of rotation vectors.
It improves the accuracy of attitude calculation, reduces computational complexity, and is suitable for attitude calculation in highly dynamic environments.
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention belongs to the field of motion carrier attitude calculation in high dynamic environments, specifically involving a deep learning-assisted rotation vector optimization method and system for high dynamic environments. Background Technology
[0002] High-dynamic environments generally refer to operating conditions where a moving vehicle experiences extreme mechanical conditions such as high overload, high rotational speed, and high initial velocity during operation. Under these extreme conditions, satellite signal rejection is likely to occur, causing the satellite navigation system to malfunction. Sensors such as gyroscopes in the inertial measurement unit are prone to saturation, resulting in significantly increased measurement errors or even malfunctions. Furthermore, due to the non-commutativity of rigid body rotation in differential geometry, non-commutative errors are inevitably introduced during attitude update, severely limiting the stability and accuracy of the solution algorithm. Traditional attitude update algorithms, such as the Euler angle method, direction cosine method, and quaternion method, are mainly suitable for low-dynamic scenarios. However, in high-dynamic application scenarios such as hypersonic aircraft, racing drones, loitering pods, guided projectiles, and aerospace vehicles, the non-commutative errors caused by high-rotational speeds are significantly amplified, making it difficult for these methods to meet the accuracy requirements of attitude calculation for high-dynamic vehicles.
[0003] As the requirements for accuracy and stability in attitude calculation under high dynamic environments continue to increase, traditional attitude calculation methods have obvious limitations under the condition of enhanced non-commutative errors. Although numerical integration methods such as function iteration show high accuracy in compensating for non-commutative errors, their computational complexity is high. Summary of the Invention
[0004] The purpose of this invention is to provide a deep learning-assisted rotation vector optimization method and system for highly dynamic environments, which is suitable for attitude calculation in highly dynamic environments, improves the accuracy of attitude calculation, and reduces computational complexity.
[0005] The technical solution to achieve the purpose of this invention is as follows:
[0006] A deep learning-assisted rotation vector optimization method for highly dynamic environments includes:
[0007] Step 1: Establish a deep learning dataset based on the motion information of the high-dynamic vehicle. Use the gyroscope data recorded by the vehicle motion as the input of the deep learning network model and the rotation vector derived from the attitude information as the output to train the deep learning network model.
[0008] Step 2: Predict the rotation vector using the trained deep learning network model;
[0009] Step 3: Combine the constructed rotation vector differential equation with the fourth-order Runge-Kutta method to construct the state equation of the UKF. Use the rotation vector predicted by the deep learning network model as the observation of the UKF to obtain the fused pose information.
[0010] Furthermore, the deep learning network model includes a 3-layer stacked Transformer encoder structure, with each Transformer encoder structure including 8 attention heads and a hidden dimension of 128.
[0011] Furthermore, the deep learning network model reads data using a sliding window method.
[0012] Furthermore, the differential equation of the rotating vector mentioned in step 3 is:
[0013] ;
[0014] In the formula, It is the rotation vector of the carrier. It is the angular velocity of the carrier during the attitude update cycle;
[0015] Considering the high sampling rate of the inertial sensor, the carrier angular velocity is a first-order small quantity within a single attitude calculation cycle. The rotation vector differential equation can be rewritten as:
[0016] ;
[0017] In the formula, , is a third-order constant matrix containing the angular velocity of the carrier;
[0018] Multiply both sides of the expression by the left side have to:
[0019] ;
[0020] Summarized as follows:
[0021] ;
[0022] Integrating both sides of the equation with respect to time t yields the rotation vector:
[0023] ;
[0024] In the formula, express and The nonlinear relationship Indicates from arrive The mapping relationship, It is a constant matrix.
[0025] Furthermore, the constructed rotating vector differential equation is combined with the fourth-order Runge-Kutta method to construct the UKF state equations, including:
[0026] The rotating vector differential equation is numerically integrated using the fourth-order Runge-Kutta method, as follows:
[0027]
[0028] In the formula, Let be the rotation vector and angular velocity at the current moment. The state function is based on the Borz equation and the RK4 integral. Let T be the rotation vector at the next moment, and T be the sampling frequency; rearranging, we get:
[0029] ;
[0030] In the formula, It is a nonlinear function combining the rotating vector differential equation and the fourth Runge-Kutta equation. This is process noise.
[0031] A deep learning-assisted rotation vector optimization system for highly dynamic environments includes:
[0032] The model training unit establishes a deep learning dataset based on the motion information of the high-dynamic vehicle, uses the gyroscope data recorded by the vehicle motion as the input of the deep learning network model, and uses the rotation vector derived from the attitude information as the output to train the deep learning network model.
[0033] The prediction unit predicts the rotation vector using a trained deep learning network model.
[0034] The fusion unit combines the constructed rotation vector differential equation with the fourth-order Runge-Kutta method to construct the state equation of the UKF. The rotation vector predicted by the deep learning network model is used as the observation of the UKF to obtain the fused pose information.
[0035] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0036] This invention models the nonlinear dynamics of angular velocity and rotation vector using rotation vector differential equations, uses the three-axis angular velocity output from the gyroscope as input to the deep learning network, uses the rotation vector as output, and utilizes Transformer to capture temporal features to achieve accurate prediction of the rotation vector.
[0037] This invention incorporates the prediction results of Transformer into observation information to construct measurement equations; at the same time, it combines the rotation vector differential equation with the fourth-order Runge-Kutta integral method to form state equations, and uses UKF to achieve the fusion of deep learning models and traditional rotation vectors, which can effectively reduce the influence of non-commutative errors and obtain high-precision attitude calculation results.
[0038] This invention achieves deep integration of deep learning and traditional rotation vectors, which is suitable for attitude calculation in high dynamic environments, while reducing computational complexity. Attached Figure Description
[0039] Figure 1 Flowchart of a deep learning-assisted rotation vector optimization method for highly dynamic environments.
[0040] Figure 2 This is a framework diagram for a deep learning-assisted rotation vector optimization method for highly dynamic environments.
[0041] Figure 3 This is a diagram of a Transformer-based rotation vector prediction model.
[0042] Figure 4 This is a diagram of the data fusion process. Detailed Implementation
[0043] The implementation of the present invention will now be described in detail with reference to the accompanying drawings.
[0044] See Figure 1 and Figure 2 This invention proposes a deep learning-assisted rotation vector optimization method for highly dynamic environments. The framework mainly consists of two parts: 1) training a deep learning network; 2) data fusion. Specifically:
[0045] First, based on the differential equation of the rotation vector, the nonlinear coupling relationship between angular velocity and rotation vector is explored, and a rotation vector prediction model based on Transformer is constructed. That is, a deep learning dataset is established based on the motion information of the high-dynamic vehicle, the gyroscope data recorded by the vehicle motion is used as the input of Transformer, and the rotation vector inverted from the attitude information is used as the output to complete the training of the deep learning network model.
[0046] Rotation vector It is a mathematical method for describing the rotation of a carrier relative to inertial space, and can uniquely determine the angular change of the carrier over a given time interval. A commonly used approximate rotation vector differential equation is:
[0047] (1)
[0048] In equation (1), It is the rotation vector of the carrier. It is the angular velocity of the carrier within the attitude update cycle. Considering the high sampling rate of the inertial sensor, the angular velocity of the carrier is a first-order small quantity within a single attitude calculation cycle, and (1) can be further rewritten as:
[0049] (2)
[0050] In the formula, , is a third-order constant matrix containing the angular velocity of the carrier. Multiplying both sides of equation (2) by , have to:
[0051] (3)
[0052] From equation (3), we can obtain:
[0053] (4)
[0054] Integrating both sides of equation (4) with respect to time t, we get:
[0055] (5)
[0056] In the formula, express and The nonlinear relationship Indicates from arrive The mapping relationship, It is a constant matrix.
[0057] As can be seen from (1) and (5), there is not only a strong correlation between angular velocity and rotation vector, but also a significant nonlinear coupling relationship. In order to effectively model this nonlinear coupling relationship, multi-layer nonlinear transformation of Transformer and feature learning are used to implicitly approximate the complex input-output mapping relationship.
[0058] Secondly, the rotation vector differential equation is combined with the fourth-order Runge-Kutta method to construct the state equation of the UKF. The rotation vector predicted by the Transformer is used as the observation of the UKF, thereby realizing the deep integration of deep learning and traditional rotation vector and improving the accuracy of attitude calculation.
[0059] As can be seen from (1) and (5), there is not only a strong correlation between angular velocity and rotation vector, but also a significant nonlinear coupling relationship. In order to effectively model this nonlinear coupling relationship, multi-layer nonlinear transformation of Transformer and feature learning are used to implicitly approximate the complex input-output mapping relationship.
[0060] See appendix Figure 3 Based on the strong coupling between angular velocity and rotation vector, this invention constructs a Transformer model, using angular velocity data as input and rotation vector as network output. The deep learning network adopts a 3-layer stacked Transformer encoder structure, with each layer containing 8 attention heads and a hidden dimension of 128, and adopts a sliding window approach on all readings.
[0061] Using equation (1) as the state model, since this equation cannot be solved analytically directly, it is numerically integrated using the fourth-order Runge-Kutta method, resulting in:
[0062] (6)
[0063] In the formula Let be the rotation vector and angular velocity at the current moment. The state function is based on the Borz equation and the RK4 integral. Let T be the rotation vector at the next moment, and T be the sampling frequency. To prevent the rotation vector state from diverging during long-term integration, equation (6) is used. A coefficient is introduced on the right side of the equation to enhance numerical stability. Equation (6) yields the state equation constructed based on the rotating vector differential equation:
[0064] (7)
[0065] In the formula, It is a nonlinear function combining the rotating vector differential equation and the fourth Runge-Kutta equation. This is process noise.
[0066] The obtained Transformer prediction results are used as the measurement equations for the observation design:
[0067] (8)
[0068] UKF is used to achieve deep fusion of deep learning and traditional rotation vectors, and the resulting rotation vector is used to solve the carrier attitude.
[0069] Participate in the attached Figure 4This invention fuses the Transformer prediction results with traditional rotation vectors. In the state prediction stage, the system uses the angular velocity measured by the gyroscope and the fourth-order Runge-Kutta algorithm based on the rotation vector differential equation to predict the prior estimate of the system state at the next moment. In the measurement update stage, the result predicted by the deep network is used as the measurement. Finally, the UKF filter optimally fuses the state prediction prior value of the dynamic model with the prediction value provided by the Transformer to obtain more accurate attitude information.
[0070] This invention can effectively reduce the impact of non-commutative errors, improve attitude calculation accuracy, and is suitable for attitude calculation in high dynamic environments.
[0071] This embodiment also provides a deep learning-assisted rotation vector optimization system for highly dynamic environments, including:
[0072] The model training unit establishes a deep learning dataset based on the motion information of the high-dynamic vehicle, uses the gyroscope data recorded by the vehicle motion as the input of the deep learning network model, and uses the rotation vector derived from the attitude information as the output to train the deep learning network model.
[0073] The prediction unit predicts the rotation vector using a trained deep learning network model.
[0074] The fusion unit combines the constructed rotation vector differential equation with the fourth-order Runge-Kutta method to construct the state equation of the UKF. The rotation vector predicted by the deep learning network model is used as the observation of the UKF to obtain the fused pose information.
[0075] This embodiment also provides a transaction order processing device, including: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the deep learning-assisted rotation vector optimization method.
[0076] This embodiment also provides a computer storage medium storing an executable program, which is executed by a processor to implement the steps of the deep learning-assisted rotation vector optimization method.
[0077] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.
[0078] Obviously, those skilled in the art can make various modifications and variations to the embodiments of the present invention without departing from the spirit and scope of the embodiments of the present invention. Therefore, if these modifications and variations to the embodiments of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention also intends to include these modifications and variations.
Claims
1. A deep learning-assisted rotation vector optimization method for highly dynamic environments, characterized in that, include: Step 1: Establish a deep learning dataset based on the motion information of the high-dynamic vehicle. Use the gyroscope data recorded by the vehicle motion as the input of the deep learning network model and the rotation vector derived from the attitude information as the output to train the deep learning network model. Step 2: Predict the rotation vector using the trained deep learning network model; Step 3: Combine the constructed rotation vector differential equation with the fourth-order Runge-Kutta method to construct the state equation of the UKF. Use the rotation vector predicted by the deep learning network model as the observation of the UKF to obtain the fused pose information.
2. The deep learning-assisted rotation vector optimization method for highly dynamic environments according to claim 1, characterized in that, The deep learning network model includes a 3-layer stacked Transformer encoder structure, with each Transformer encoder structure including 8 attention heads and a hidden dimension of 128.
3. The deep learning-assisted rotation vector optimization method for highly dynamic environments according to claim 2, characterized in that, The deep learning network model reads data using a sliding window method.
4. The deep learning-assisted rotation vector optimization method for highly dynamic environments according to claim 1, characterized in that, The differential equation of the rotating vector mentioned in step 3 is: ； In the formula, It is the rotation vector of the carrier. It is the angular velocity of the carrier during the attitude update cycle; Considering the high sampling rate of the inertial sensor, the carrier angular velocity is a first-order small quantity within a single attitude calculation cycle. The rotation vector differential equation can be rewritten as: ； In the formula, , is a third-order constant matrix containing the angular velocity of the carrier; Multiply both sides of the expression by the left side have to: ； Summarized as follows: ； Integrating both sides of the equation with respect to time t yields the rotation vector: ； In the formula, express and The nonlinear relationship Indicates from arrive The mapping relationship, It is a constant matrix.
5. The deep learning-assisted rotation vector optimization method for highly dynamic environments according to claim 1, characterized in that, By combining the constructed rotating vector differential equations with the fourth-order Runge-Kutta method, the state equations of the UKF are constructed, including: The rotating vector differential equation is numerically integrated using the fourth-order Runge-Kutta method, as follows: ； In the formula, Let be the rotation vector and angular velocity at the current moment. The state function is based on the Borz equation and the RK4 integral. Let T be the rotation vector at the next moment, and T be the sampling frequency; rearranging, we get: ； In the formula, It is a nonlinear function combining the rotating vector differential equation and the fourth Runge-Kutta equation. This is process noise.
6. A deep learning-assisted rotation vector optimization system for highly dynamic environments that implements the method of any one of claims 1-5, characterized in that, include: The model training unit establishes a deep learning dataset based on the motion information of the high-dynamic vehicle, uses the gyroscope data recorded by the vehicle motion as the input of the deep learning network model, and uses the rotation vector derived from the attitude information as the output to train the deep learning network model. The prediction unit predicts the rotation vector using a trained deep learning network model. The fusion unit combines the constructed rotation vector differential equation with the fourth-order Runge-Kutta method to construct the state equation of the UKF. The rotation vector predicted by the deep learning network model is used as the observation of the UKF to obtain the fused pose information.
7. A transaction order processing device, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the deep learning-assisted rotation vector optimization method according to any one of claims 1-5.
8. A computer storage medium, characterized in that, The computer storage medium stores an executable program, which is executed by a processor to implement the steps of the deep learning-assisted rotation vector optimization method according to any one of claims 1-5.