Navigation method capable of adaptively optimizing filtering

Abstract

The invention provides a navigation method for adaptively optimizing filtering. The method comprises the following steps: respectively establishing error models of an inertial navigation system (SINS) and a Doppler velocimeter (DVL) by utilizing Lie group geometric characteristics; transmitting the output data of the SINS and the DVL into an adaptive Kalman filter; a Kalman filter framework is established according to a system model, and a state error and a noise covariance matrix are projected to a Lie group SE2 (3) to be converted into a low-dimensional matrix. When the DVL data is available, a process noise covariance matrix and a measurement noise covariance matrix are explored through DQN training; and when the DVL data is lost, independently optimizing the process noise covariance matrix and the measurement noise covariance matrix by using the trained model. According to the technical scheme provided by the invention, the noise covariance matrix of projection conversion under the Lie group is optimized through the deep reinforcement learning method, Kalman filtering can be adaptively optimized, and high-precision and high-real-time underwater navigation is achieved.

Description

Technical Field

[0001] This invention relates to the field of integrated navigation technology, and in particular to a navigation method with adaptive optimization filtering and its application in underwater sensor navigation. Background Technology

[0002] Autonomous navigation of autonomous underwater vehicles (AUVs) in complex underwater environments is a key aspect driving the development of unmanned underwater robots. With the increasing application of AUVs in intelligent transportation systems, these vehicles are typically deployed in dynamic and uncertain environments to perform critical tasks, making precise navigation one of their core technological requirements. Multi-sensor information fusion technology, especially the combination of strapdown inertial navigation systems (SINS) and Doppler velocimetry (DVL), provides crucial support for reliable estimation of system states. However, traditional methods have limitations in dealing with the time-varying characteristics of process noise covariance matrices and measurement noise covariance matrices, resulting in limited performance in dynamic environments.

[0003] To address these issues, current research focuses on two main approaches. One approach optimizes the derivation of the noise covariance matrix, extending the invariant error on the matrix Lie group to the group affine transformation strapdown inertial navigation system, thereby improving the noise covariance matrix's adaptability to environmental changes. The other approach employs adaptive Kalman filtering to autonomously optimize both the process noise and measurement noise covariance matrices, further enhancing the accuracy and consistency of SINS / DVL fusion. However, the former has limited effectiveness in enhancing filter robustness, while the latter has high computational complexity, limiting its application in scenarios with high real-time requirements.

[0004] In summary, there is an urgent need for a combined navigation method applicable to complex and unknown sea areas to make up for the shortcomings of existing technologies in terms of adaptability, robustness and computational efficiency. Summary of the Invention

[0005] The purpose of this invention is to avoid the shortcomings of existing technologies in terms of insufficient adaptability to noise covariance matrix in dynamic environments, low filter robustness, and limited real-time performance. Therefore, an EKF optimization method based on deep reinforcement learning and SE2(3) is proposed.

[0006] Firstly, this application provides a navigation method for adaptive optimized filtering, comprising: Transform the body coordinate system of the output data of SINS and DVL to the navigation coordinate system; and use the geometric properties of the Lie group SE2(3) to establish the left invariant error measurement model of SINS and the left invariant error observation model of DVL respectively. A Kalman filter framework is constructed, and the process noise covariance matrix and the measurement noise covariance matrix are projected and transformed into a low-dimensional continuous Lie space through SE2(3). When DVL data is available, the state features of SINS are extracted using the noise covariance matrix optimization method of DQN, and the process noise covariance matrix and the measurement noise covariance matrix are optimized and trained. When DVL data is lost, the process noise covariance matrix of SINS is independently optimized using the previously trained model.

[0007] Secondly, this application provides a Kalman filtering method based on deep reinforcement learning optimization and its application in underwater sensor navigation, the steps of which are as follows: Step S01, a coordinate system matching method. The body coordinate system of the SINS and DVL output data is transformed to the navigation coordinate system, and the two are fused under a unified navigation framework through coordinate system transformation.

[0008] Step S02: An error modeling method based on Lie groups and Lie algebras. By utilizing the mathematical structures of Lie groups and Lie algebras, error measurement models for SINS and error observation models for DVL are established respectively.

[0009] Step S03: A method for establishing and initializing a Kalman filter. Based on the system dynamics model and observation equations, a Kalman filter framework is constructed, defining the state vector and its associated state transition equations and observation equations; the state vector and state covariance matrix are initialized, and the process noise covariance matrix and measurement noise covariance matrix are set.

[0010] Step S04: A noise covariance matrix optimization method based on Deep Q-Network (DQN). When DVL data is available, the state features of SINS are extracted, and the process noise covariance matrix and measurement noise covariance matrix are optimized and trained. When DVL data is lost, the process noise covariance matrix of SINS is independently optimized using the previously trained model.

[0011] Preferably, before implementing the above scheme, the body coordinate system of the SINS and DVL output data is transformed to the navigation coordinate system, and the two are fused under a unified navigation framework through coordinate system transformation; based on the output characteristics of SINS and DVL, a coordinate transformation relationship from the body coordinate system to the navigation coordinate system is constructed; the attitude angle is transformed using the quaternion method to avoid the singularity problem in Euler angle representation and ensure the continuity and accuracy of attitude transformation; combined with attitude data and velocity information, the efficient computational characteristics of quaternions are used to eliminate dead angle phenomena and the influence of uncertainties in the dynamic environment on coordinate transformation.

[0012] Preferably, the error modeling method based on Lie groups and Lie algebras utilizes the mathematical structures of Lie groups and Lie algebras to establish an error measurement model for SINS and an error observation model for DVL, respectively. This includes: Based on the motion characteristics of the SINS system, and utilizing the geometric properties of the Lie group SE2(3), error measurement models for position, velocity, and attitude are established. The errors are dynamically expressed as continuous variables on the Lie algebra to ensure the mathematical rigor and geometric consistency of the error description. Based on the observation principle of DVL, a velocity error observation model is constructed. Combined with position information in the navigation coordinate system, an observation model including position error is established to describe the deviation relationship between the measurement error and the actual state. The zero-bias error and scaling factor error of the IMU are modeled, and a first-order Gaussian-Markov process is used to describe their dynamic characteristics, thereby capturing the correlation of random errors in the time domain and improving the model's accuracy.

[0013] The SINS error measurement model and the DVL error observation model are integrated to construct a unified error model framework. The system error state is described by 21-dimensional state variables, including position, velocity, attitude, IMU zero bias error and scale factor error. Combined with a 6×21-dimensional observation matrix, the comprehensive observation of velocity error and position error is realized.

[0014] Preferably, the method for establishing and initializing the Kalman filter includes: constructing a Kalman filter framework based on the system dynamics model and observation equations; defining the state vector and its associated state transition equations and observation equations; initializing the state vector and state covariance matrix; and setting the process noise covariance matrix and measurement noise covariance matrix. This includes: Based on system requirements, a state vector is defined that includes state variables such as position, velocity, attitude, IMU zero bias error, and scaling factor error. Each variable is initialized based on initial known conditions to ensure that the initial estimates are consistent with the actual system state. Based on the SE2(3) Lie group manifold, it is projected into a low-dimensional matrix.

[0015] Based on the system dynamics model, a state transition equation is established to describe the evolution of the system state in the time dimension, and process noise is modeled to ensure the accuracy and reliability of the state transition.

[0016] Based on the observation characteristics of the DVL sensor, an observation equation is constructed to describe the mapping relationship between sensor measurements and the actual state of the system. The observation noise is set in combination with the sensor characteristics to ensure the accuracy of the observation model.

[0017] In the early stages of filter operation, the robustness of the filter to initial state estimation errors is enhanced by increasing the initial value of the state covariance matrix; the filter parameters are dynamically adjusted in conjunction with residual analysis to ensure that the filter converges quickly to an accurate state estimate.

[0018] Preferably, the noise covariance matrix optimization method based on Deep Q-Network (DQN) involves: when DVL data is available, extracting the state features of the SINS and optimizing the process noise covariance matrix and measurement noise covariance matrix; when DVL data is lost, independently optimizing the process noise covariance matrix of the SINS using the previously trained model. This includes: Based on the output state of the SINS system, multidimensional feature information including position, velocity, attitude, IMU bias error and scaling factor error is extracted and encoded into the input vector of a deep Q network to describe the current state characteristics of the system.

[0019] When DVL data is available, the output data of SINS and DVL are fused in real time, and the process noise covariance matrix and measurement noise covariance matrix are optimized and trained using a deep reinforcement learning framework. The parameters of DQN are dynamically adjusted based on the environmental state and reward function to ensure adaptive optimization of the noise covariance matrix, thereby improving the robustness and accuracy of the filter. The trained model parameters are also saved to support separate optimization when DVL data is lost.

[0020] When DVL data is lost, the trained DQN model is invoked to optimize the process noise covariance matrix separately based on the currently extracted SINS state features. The state prediction capability of the Kalman filter is dynamically adjusted by utilizing the optimized process noise covariance matrix, ensuring that the system can maintain high navigation accuracy and optimize cumulative error even without external observation.

[0021] During DQN training, a reward function based on navigation error, state estimation accuracy, and filter stability is designed to guide the model to quickly converge to the optimal solution in complex dynamic environments.

[0022] When DVL data is intermittently available, the DQN model is updated online using new observation data to improve its adaptability to different working conditions; offline batch training is performed by combining historical data and simulation data to further improve the model's generalization performance in unknown environments.

[0023] In the event of long-term loss of DVL data, the deviation of system state estimation is dynamically monitored through residual analysis. Combined with the optimization results of the DQN model, the process noise covariance matrix is ​​dynamically adjusted to ensure the robustness and stability of the filter.

[0024] Thirdly, this application provides a computing device, including a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are to be invoked and executed by the processing component to implement a navigation method of adaptive optimization filtering as described in any of the first aspects.

[0025] This application provides an adaptive optimization filtering navigation method, which includes: transforming the body coordinates of SINS and DVL output data to the navigation coordinate system to achieve fusion processing; based on Lie group and Lie algebra error modeling, establishing error models for SINS and DVL respectively by utilizing the geometric properties of the Lie group SE2(3); establishing a Kalman filter framework based on the system model, projecting the state error and noise covariance matrix onto the Lie group SE2(3) to convert it into a low-dimensional matrix. When DVL data is available, the process noise covariance matrix and measurement noise covariance matrix are optimized by DQN; when DVL data is lost, the process noise covariance matrix is ​​independently optimized using the trained model. The technical solution provided in this application optimizes the Q and R matrices of the projection transformation under the Lie group through deep reinforcement learning, which can adaptively optimize the Kalman filter to achieve high-precision and high-real-time underwater navigation.

[0026] The above technical solution adopted in this application has the following advantages: This application focuses on the invention of optimizing the process noise covariance matrix and measurement noise covariance matrix of SE2(3) dimensionality reduction based on DQN, which enables dynamic adaptive adjustment of noise characteristics in a multi-sensor fusion system, thereby effectively improving the navigation accuracy and reliability of state estimation. Even when DVL data is lost or intermittently unavailable, the process noise covariance matrix can be independently optimized using a previously trained DQN model, ensuring that SINS maintains high estimation accuracy even with missing data, thus enhancing the system's robustness in complex dynamic environments. This invention not only improves the navigation accuracy and robustness of the SINS / DVL sensor fusion system in dynamic environments but also enhances the system's adaptability and real-time processing capabilities through the introduction of deep learning technology, providing more accurate and stable technical support for underwater navigation in intelligent transportation systems.

[0027] These or other aspects of this application will become more apparent in the following description of the embodiments. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of the Kalman filtering algorithm in an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the DQN-enhanced Kalman filter algorithm in an embodiment of the present invention; Figure 3 This is a schematic diagram of the DQN network structure in an embodiment of the present invention; Figure 4 This is a comparison chart of attitude experiment results under the publicly available dataset in the embodiments of the present invention, showing the performance differences of different filtering algorithms in data processing; Figure 5 This is a comparison chart of speed experiment results on the publicly available dataset in this embodiment of the invention, showing the performance differences of different filtering algorithms in data processing; Figure 6 This is a comparison chart of location experiment results under the publicly available dataset in this embodiment of the invention, demonstrating the performance differences of different filtering algorithms in data processing. Detailed Implementation

[0030] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0031] In some of the processes described in the specification, claims, and accompanying drawings of this application, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as S01, S02, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0032] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0033] To make the technical solution of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings. The present invention is implemented in specific steps: See Figure 1 This embodiment discloses a navigation method with adaptive optimization filtering, the steps of which are as follows: Step S01, a coordinate system matching method. The body coordinate system of the SINS and DVL output data is transformed to the navigation coordinate system, and the two are fused under a unified navigation framework through coordinate system transformation.

[0034] Step S02: An error modeling method based on Lie groups and Lie algebras. By utilizing the mathematical structures of Lie groups and Lie algebras, error measurement models for SINS and error observation models for DVL are established respectively.

[0035] Step S03, a method for establishing and initializing a Kalman filter. Based on the system dynamics model and observation equations, construct the Kalman filter framework, define the state vector and its associated state transition equations and observation equations; initialize the state vector and state covariance matrix, and set the process noise covariance matrix and measurement noise covariance matrix through SE2(3).

[0036] Step S04: A noise covariance matrix optimization method based on Deep Q-Network (DQN). When DVL data is available, the state features of SINS are extracted, and the process noise covariance matrix and measurement noise covariance matrix are optimized and trained. When DVL data is lost, the process noise covariance matrix of SINS is independently optimized using the previously trained model.

[0037] See Figure 2 The overall algorithm flow and its specific implementation method are as follows. Specifically, in step S01, the machine system and navigation system are defined as follows: Navigation coordinate system (N system): This invention uses the NED (North-East-Down) coordinate system as the reference coordinate system for navigation calculation, in which the X-axis points due east, the Y-axis points due north, and the Z-axis is perpendicular to the plane and points to the ground, thus serving as the standard axial definition of the navigation coordinate system.

[0038] Body coordinate system (B-frame): The origin of the body coordinate system is set at the center of mass of the carrier. The X-axis points to the right side of the carrier, the Y-axis points to the front of the carrier, and the Z-axis points to the top of the carrier, forming a local coordinate system for describing the motion state of the carrier and its relative relationship with the external environment.

[0039] Both SINS and DVL directly output state variables based on the sensor's B-frame. Therefore, to achieve multi-sensor data fusion processing, they need to be uniformly converted to the N-frame. This coordinate transformation process is achieved through the stepwise accumulation of quaternions to ensure the accuracy and continuity of rotation calculation. For the state variables of SINS, their attitude variables need to be updated first. The attitude update is based on the angular velocity data collected by the IMU and is calculated through equation (1) to ensure that the attitude solution in the navigation coordinate system matches the actual motion, providing accurate attitude information for the subsequent fusion algorithm; for the state variables of DVL, the attitude information is obtained through the IMU and the state information in the N-frame is obtained through the accumulation of quaternions.

[0040]

[0041] In the above formula,

[0042] in, and The modulus can be calculated from the angular increment of the gyroscope. Quaternions This represents the transformation related to the attitude change of the N-system.

[0043]

[0044] Where, vector rotation angle For, vector The attitude is determined by multiple factors, including the Earth's rotation constant. Using the formula above, all the quaternions needed to update the attitude can be obtained, and then the attitude update can be calculated by updating the quaternions.

[0045] Next, the velocity update equation within the navigation coordinate system is proposed. Velocity update needs to consider gravity, the Coriolis force, and the IMU-specific force. In practical engineering applications, the velocity update equation using dual sampling is obtained by approximating the gravity / Coriolis integral term and the specific force integral term, as shown in the following formula:

[0046]

[0047]

[0048] in, and These values ​​are derived from the Earth's rotation parameters and the angular increments output by the gyroscope. Using the above formulas, the velocity update equation can be obtained.

[0049] Finally, the position update equation is as follows:

[0050] Furthermore, in step S02, the error states of SINS and DVL observations are modeled using Lie groups and Lie algebras. First, the error state of SINS is defined:

[0051] In the formula, r represents the position error, and v represents the velocity error. The attitude error is represented by bg, which refers to the bias vector related to the three-axis gyroscope; ba represents the bias vector of the three-axis accelerometer. The gyroscope scale factor error is represented by sg, and sa represents the accelerometer scaling factor error vector. All of the above factors are time-dependent. To derive the error propagation state equation of the system, the differential equation of δx in continuous time needs to be established first, expressed as follows:

[0052] Differential equations for position and velocity are derived using Lie groups and Lie algebras.

[0053] The results of the SE2(3) error propagation algorithm during the process of proposing left-invariant error propagation are as follows:

[0054] Furthermore, in step S03, the F matrix can be obtained through S02:

[0055] in,

[0056] Based on this, the noise matrix G(t) of a continuous-time system can also be easily derived.

[0057]

[0058] Define the position observation vector as SINS and calculate the difference between the position and the DVL record position. This is a readily available quantity, and its specific expression is as follows:

[0059] In the above formula, with The superscript variable represents the calculated value, where The position vector of the DVL phase center obtained through inertial navigation calculation, with The superscript variable represents the observed value, where These are DVL location observations. This represents the DVL position error. Typically, to simplify processing, the error in the DVL position measurement is modeled as a Gaussian white noise sequence, i.e.:

[0060] Therefore, the DVL position observation equation can be written in the following form:

[0061] Among them, the observation matrix for:

[0062] In actual calculations, the observation matrix uses the latest inertial navigation attitude at the current moment. The above is the DVL position observation equation corresponding to the system state equation generated by the SINS mechanical orchestration algorithm.

[0063] This allows us to substitute the above matrix into the Kalman filter formula. The Kalman filter process is as follows: The formula for discrete Kalman filtering is given. Kalman filtering is a recursive filter used to estimate the system state from a set of incomplete and imprecise observation data. In SINS and DVL joint navigation, the working principle of the Kalman filter is as follows: First, define the system state vector and observation vector. The state equation and observation equation are as follows:

[0064]

[0065] Among them, is State transition matrix, It's process noise. It is the observation matrix. It is observation noise. This is achieved by predicting the current state. Covariance Matrix :

[0066] Where is the covariance matrix of the process noise. Then, the Kalman gain is calculated:

[0067] Where is the covariance matrix of the observation noise. Finally, update the state estimate and the covariance matrix:

[0068]

[0069] The prediction and update steps are repeated continuously until all observation data has been processed. In this way, Kalman filtering can effectively fuse SINS and DVL data, reducing systematic errors and improving navigation accuracy.

[0070] Further, in step S04, see Figure 3 This embodiment applies the DQN (Deep Q-Network) network structure to EKF data fusion control, transforming the EKF data fusion control problem into a Markov decision process. Through agent exploration and learning in the environment, the optimal fusion control behavior is obtained. The algorithm converts the covariance matrix of the current system state into control actions, achieving the desired control objective through the mapping of the reward function and control variables, thereby optimizing data fusion performance. The process noise covariance matrix is ​​initialized as follows:

[0071] in:

[0072] gry represents the gyroscope output on each axis; acc represents the accelerometer output on each axis; arw represents the angular random walk (ARW) parameter on each axis; bias_std represents the standard deviation of the sensor bias; scale_std represents the standard deviation of the sensor scaling factor. The Hadamard product (element-wise product) represents the element-wise multiplication of a vector with itself. After projecting the system using SE2(3):

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] The first and second formulas are exponential and logarithmic mappings of SE2(3), and the third formula is its Jacobian matrix format. Left and right multiplication matrices are obtained based on Lie group projections, and a simplified version of the matrix can be obtained by multiplying Qj with it. In engineering, the Joseph form is typically used to calculate the covariance matrix, which ensures the symmetry and positive definiteness of the matrix. Therefore, the system state noise variance matrix and measurement noise variance matrix are taken as the control objects, and the state space is a vector composed of the projections of the current matrices Q and R.

[0079] In this example, the action space is defined as modifying the diagonal elements of matrices Q and R. To simplify the problem, we focus only on adjusting the diagonal elements of the two matrices, which represent independent noise sources in the system and measurement process. For each diagonal element, the action space consists of a discrete set of adjustments within a predetermined range. Specifically, each element can undergo a discrete scaling operation corresponding to the current diagonal element value. The adjustment is given by , where i is a discrete variation factor, and Therefore, for each diagonal element, the action space is defined as:

[0080] Here, represents the j-th diagonal element of matrix Q. Similarly, the diagonal elements of matrix R can be adjusted using the same equation. The adjustment range is set to the current element value, meaning each diagonal element can vary within the following range:

[0081] or:

[0082] This design allows for adaptive control of the noise covariance matrix while maintaining filter stability and accuracy. The reward function is designed to minimize the estimation error. In this example, the reward is set to the negative mean squared error:

[0083] Where represents the observed value, H represents the matrix mapping the system state to the observed variables, and represents the state prediction value of the EKF. Logarithmic scaling of the reward function can increase the negative reward variance. Specifically, the negative mean squared error is logarithmically processed:

[0084] in, It is a very small positive number, used to prevent log(0), set to 0.01. The logarithmic transformation maps smaller error values ​​to a larger range of negative values, thus preventing the reward from changing too slowly when the mean square error is close to 0. At each step, the EKF generates updated state estimates and associated errors based on the current state and noise covariance matrix settings (i.e., the actions determined by the DQN). These estimates and errors are then fed back to the DQN agent for further decision-making. To further illustrate the advantages and objectives of this embodiment, the following will compare the performance of Kalman filtering (KF), SE2(3) enhanced Kalman filtering (SEKF), DQN enhanced Kalman filtering, and the Kalman filtering algorithm (DQN-SEKF) disclosed in this embodiment in terms of navigation effects using experimental results on a public dataset. Figure 4 As shown in Figures 5 and 6, in terms of attitude, velocity, and position accuracy, DQN-SEKF has a mean square error for navigation that is more than 10 times higher than that of EKF, which leads to the conclusion that it is more feasible.

[0085] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0086] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0087] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0088] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A navigation method with adaptive optimization filtering, characterized in that, include: Establish the left-invariant error measurement model for SINS and the left-invariant error observation model for DVL; An adaptive Kalman filter framework is constructed, which transforms the process noise covariance matrix and the measurement noise covariance matrix into a low-dimensional continuous Lie space by SE2(3) projection; and optimizes the process noise covariance matrix and the measurement noise covariance matrix by using a deep reinforcement learning noise covariance matrix optimization method when DVL data is available, extracting the state features of SINS and DVL; When DVL data is lost, the process noise covariance matrix and measurement noise covariance matrix of SINS are independently optimized using the previously trained model.

2. The navigation method for adaptive optimization filtering according to claim 1, characterized in that, The error models for SINS and DVL are constructed based on the SE2(3) left-invariant error modeling method, including: Based on the SINS and DVL mechanical arrangement equations, and combined with the SE2(3) geometric invariance, an error measurement model for position, velocity and attitude is established; Based on the group multiplication rule and differential properties of SE2(3), and combining the IMU error model in SINS and the speed measurement error model of DVL, the transformation relationship of the error state vector is constructed, and the error sources of IMU and DVL are integrated into the dynamic equation of the error state vector. Based on the exponential and logarithmic mappings from Lie groups to their corresponding Lie algebras, a projection relationship is established from the high-dimensional error covariance matrix to the low-dimensional continuous Lie space. Through the adjoint representation operator of SE2 (3), a geometric consistency constraint is applied to the projected low-dimensional covariance matrix to ensure that the transformed process noise covariance matrix and the measurement noise covariance matrix maintain their geometric manifold characteristics under coordinate system transformation. Error threshold detection is performed on the low-dimensional projection matrix. When the deviation between the high-dimensional covariance matrix and the low-dimensional projection matrix exceeds the threshold, the low-dimensional matrix is ​​iteratively corrected by the differential correction method of the Lie group SE2 (3) until the error convergence condition is met. Finally, the low-dimensional covariance matrix data that can be directly used for DQN training is obtained.

3. The navigation method for adaptive optimization filtering according to claim 1, characterized in that, The noise covariance matrix optimization methods based on DQN include: Based on the output status of the SINS system, extract the position, velocity, attitude, IMU zero bias error and scaling factor error. Based on the output status of the DVL sensor, extract the beam measurement velocity vector and encode the two into the DQN input vector. When DVL data is available, the output data of SINS and DVL are fused in real time, and the dimensionality-reduced process noise covariance matrix and measurement noise covariance matrix are optimized and trained using a deep reinforcement learning framework. When DVL data is lost, the trained DQN model is invoked to optimize the process noise covariance matrix separately based on the currently extracted SINS state features.

4. The navigation method for adaptive optimization filtering according to claim 1, characterized in that, The reward function design described in claim 3 is based on navigation error, state estimation accuracy, and filter stability to guide the model to quickly converge to the optimal solution in complex dynamic environments.

5. The navigation method for adaptive optimization filtering according to claim 1, characterized in that, The online update method of claim 3 updates the DQN model online using new observation data when DVL data is intermittently available, thereby improving the model's adaptability to different operating conditions. In the event of prolonged DVL data loss, residual analysis is used to dynamically monitor the deviation in system state estimation, and the PNCM is dynamically adjusted based on the optimization results of the DQN model to ensure the robustness and stability of the filter.

Images