A Beidou positioning method based on adaptive noise covariance learning and denoising

Abstract

The present invention provides a Beidou positioning method based on adaptive noise covariance learning and denoising, including receiving satellite data in a real city scene, and constructing an adaptive KF satellite positioning environment based on reinforcement learning; constructing a dual-agent reinforcement learning model; training the dual-agent reinforcement learning model; deploying the trained dual-agent reinforcement learning model to a cloud server; and outputting the results to a satellite positioning chip terminal in real time through an end-to-end transmission method. The present invention adaptively denoises the measurement and process noise in the Kalman filter and adjusts the corresponding covariance matrix parameters at the same time through a reinforcement learning method, without relying on cumbersome manual design, and can be effectively applied to deployment and use in different urban environments, and is more adaptable to dynamically changing environments; the measurement and process noise that produces positioning errors are estimated and removed through a dual-agent reinforcement learning method, thereby improving the robustness to random noise that cannot be modeled.

Description

Technical Field

[0001] The present invention relates to the field of satellite positioning technology, and in particular to a Beidou positioning method based on adaptive noise covariance learning and denoising. Background Art

[0002] At present, global navigation satellite system (GNSS) positioning has achieved decimeter or even centimeter-level positioning accuracy in open areas. However, in complex urban areas, such as "urban canyons" and under overpasses, the quality of satellite measurements will be seriously degraded due to multipath effects caused by reflection, diffraction and obstruction caused by high-rise buildings, resulting in random measurement noise in the received satellite signals that is difficult to model. The rapid changes in the vehicle's surrounding environment will further increase the complexity and variability of the measurement noise. In addition, due to the complexity of roads in urban areas, the state of the positioning system will also produce random process noise that is difficult to estimate due to factors such as time and environment. Therefore, in complex urban areas, due to the influence of random noise, the GNSS positioning error can be as high as tens of meters in severe cases.

[0003] Most existing technologies use model-based methods to solve the position, that is, to make a priori assumptions about the degree of environmental noise or the confidence of satellite signal measurement through some prior knowledge or a large number of experimental debugging, such as weighted least squares (WLS) and Kalman filter (KF). In addition, there are also positioning correction methods based on data-driven deep learning, which use a large amount of measurement data collected from the positioning environment to learn the correction strategy for the initial positioning results based on the positioning results of the model-based method, and have a certain adaptability to complex environments.

[0004] However, the model-based satellite positioning method requires a lot of experimental debugging to design and adjust hyperparameters. Its performance depends on the prior assumptions of the environment or model parameters and cannot adapt to the complex dynamic changes in urban environments. The data-driven positioning correction algorithm simply connects the model-based method with the deep learning model in series, and the two are optimized independently. Its final positioning accuracy is still limited by the initial positioning accuracy of the model-based method. In addition, this method only learns a general correction strategy and lacks refined modeling and prediction of positioning errors.

[0005] The above method cannot fundamentally reduce the positioning error caused by random noise, and the final positioning result still contains errors caused by random noise. In addition, it only considers the noise covariance of a single perspective as input observation, ignoring the impact of factors such as multipath effects in the environment and rapid changes in vehicle motion state on positioning. Its observation characteristics cannot effectively characterize the random noise generated by the dynamically changing environment. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention provides a Beidou positioning method based on adaptive noise covariance learning and denoising, which can effectively reduce the impact of random noise on satellite positioning accuracy, and its model parameters can be adaptively adjusted according to the dynamic changes of the current environment.

[0007] The technical solution of the present invention is: a Beidou positioning method based on adaptive noise covariance learning and denoising, comprising the following steps:

[0008] S1), receiving satellite data in real urban scenes and building an adaptive KF satellite positioning environment based on reinforcement learning;

[0009] S2), building a dual-agent reinforcement learning model;

[0010] S3), training the dual-agent reinforcement learning model;

[0011] S4) deploy the trained dual-agent reinforcement learning model to the cloud server; output the results to the satellite positioning chip terminal in real time through end-to-end transmission.

[0012] Preferably, in step S1), the construction of the satellite positioning environment includes the construction of a multi-view observation space, an action space, the setting of a reward function, and the calculation of an adaptive denoising Kalman filter satellite positioning result.

[0013] Preferably, in step S1), the multi-view observation space is defined as:

[0014] O={O g ,O h ,O p ,O m};

[0015] Among them, O g is the satellite observation feature; O h is the historical position sequence; O p is the process noise covariance matrix; O m is the measurement noise covariance matrix.

[0016] As a preference, in step S1), the action space is defined as A={a t,t=0,…,T}; where,

[0017] in, and is the covariance matrix Q of the process and measurement noise at time t t-1 and R t-1 Adjustment of and is an estimate of the process and measurement noise at time t.

[0018] Preferably, in step S1), the reward function is set to correct the advantage error, that is:

[0019]

[0020] In the formula, The reference position is obtained by solving the baseline method; The actual position of the reference; For the predicted location.

[0021] Preferably, in step S2), the dual-agent reinforcement learning model includes two independent agents; the two agents act in the same adaptive KF satellite positioning environment and obtain positioning results, and respectively estimate and adjust the process / measurement noise and the corresponding noise covariance matrix.

[0022] Preferably, in step S2), the dual-agent reinforcement learning model includes an observation feature extractor and an actor-evaluator network.

[0023] Preferably, in step S2), the overall objective function of the dual-agent reinforcement learning model is as follows:

[0024]

[0025] In the formula, is the objective function of the actor network; is the objective function of the evaluator network;

[0026] And the objective function The stochastic gradient descent algorithm is used to iteratively update the weight parameters of the model.

[0027] The beneficial effects of the present invention are:

[0028] 1. The present invention adaptively adjusts the measurement and process noise covariance matrix parameters in the Kalman filter simultaneously through the reinforcement learning method, without relying on cumbersome manual design, and can be effectively applied to deployment and use in different urban environments, and is more adaptable to dynamically changing environments;

[0029] 2. The present invention estimates and removes the measurement and process noise that produces positioning errors through a dual-agent reinforcement learning method, thereby improving the robustness of the positioning model to random noise that cannot be modeled;

[0030] 3. The present invention constructs a reinforcement learning model based on multi-perspective observation, introduces satellite features and historical positioning sequences as observations of the model, thereby effectively modeling the random noise in the environment and the dynamic changes of the vehicle state, which can more effectively reflect the current state of the vehicle agent. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the Beidou positioning method of the present invention;

[0032] Figure 2 A structural diagram of a dual-agent reinforcement learning model constructed by the method of the present invention;

[0033] Figure 3 Schematic diagram of the process flow of the dual-agent reinforcement learning model training and testing phases of the method of the present invention. DETAILED DESCRIPTION

[0034] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings:

[0035] like Figure 1 As shown, this embodiment provides a Beidou positioning method based on adaptive noise covariance learning and denoising, comprising the following steps:

[0036] S1), using the receiver to receive satellite data in a real urban scene and constructing an adaptive KF satellite positioning environment based on reinforcement learning;

[0037] The construction of the satellite positioning environment includes multi-perspective observation space, construction of action space, setting of reward function, and calculation of adaptive denoising Kalman filter satellite positioning results.

[0038] As a preferred embodiment of this invention, in step S1), the observation space is defined as O={o t ,t=0,…,T}, where o t represents the environment observation at time t, T represents the end time of the trajectory; the action space is defined as A = {a t ,t=0,…,T} is the action set of the agent, a t represents the action observation at time t; define the reward function R = {r t ,r=0,…,T}, where r trepresents the reward obtained by the agent at time t; at the same time, the reward discount factor γ∈(0,1) is introduced. The goal of reinforcement learning is to learn an optimal strategy π to maximize the expectation of the cumulative discounted reward obtained in the trajectory τ, that is, Among them, E τ [·] Indicates expectation.

[0039] As a preferred embodiment of this embodiment, in step S1), the construction of the multi-view observation space specifically includes the following steps:

[0040] S10), roughly positioning the position of the receiver;

[0041] S101), the receiver receives measurement signals from multiple satellites at time t to obtain a set of pseudo-range measurements And get the position coordinates of each satellite based on the broadcast ephemeris Among them, S represents the total number of satellites currently observed, Represents the position coordinates of the sth satellite at time t;

[0042] S102), constructing a pseudorange measurement equation according to the pseudorange measurement and the satellite position coordinates;

[0043] S103), calculate the rough position coordinates of the current receiver by weighted least squares method

[0044] S11) Using satellite observation features g , historical position sequence o h , process noise covariance observation matrix o p , measurement noise covariance observation matrix o m Form a multi-view observation space, that is: O = {o g ,o h ,o p ,o m};

[0045] S111), define satellite observation characteristics o g , using the line-of-sight vector of different satellites S Pseudorange residual Carrier-to-Noise Ratio Altitude angle ε and pseudorange uncertainty The five characteristic quantities are used as satellite observation characteristics; therefore, the satellite observation characteristics at time t are It is expressed as:

[0046]

[0047] Where, the sight vector Represents the rough direction information from the satellite to the receiver, pseudo-range residual Represents the prediction error of the initial position estimate; carrier-to-noise ratio Describes the quality of the current signal; the altitude angle ε represents the vertical angle between the receiver and the satellite; the pseudorange uncertainty Indicates the confidence level of the pseudorange measurement for the current satellite.

[0048] Among them, the line-of-sight vector of satellite i at time t is and pseudorange residual The calculation method is as follows:

[0049]

[0050] In the formula, represents the pseudorange measurement of satellite i at time t; Indicates the current rough position coordinates of the receiver;

[0051] S112) Define the historical position sequence o h , which characterizes the motion state of the current vehicle trajectory and contains the historical predicted positions of L-1 receivers and the receiver's current approximate position Therefore, at time t, the historical position sequence It is expressed as:

[0052]

[0053] S113), define the process noise covariance observation matrix o p and the measurement noise covariance observation matrix o m ; It characterizes the severity of the measurement noise in the environment and the severity of the change in the motion state of the vehicle during motion; The relative ratio of the process noise covariance to the measurement noise covariance determines the contribution of prediction and measurement in the final KF update process of the predicted value; Therefore, the process noise covariance observation matrix o at time t p,t and the measurement noise covariance observation matrix o m,t They are defined as:

[0054]

[0055] Among them, Q t-1 and R t-1 are the noise covariance matrices of the process and measurement at time t-1 respectively.

[0056] As a preferred embodiment of this invention, in step S1), the construction of the action space is as follows:

[0057] The action is defined as a 24-dimensional continuous action setting, namely Action a tEach element a in i,t Output the corresponding mean μ at time t through the actor network i,t and variance σ i,t ; Gaussian distribution is used for sampling, that is,

[0058] The action a t It consists of four parts, namely: in, and is the covariance matrix Q of the process and measurement noise at time t t-1 and R t-1 Adjustment, represents the action scale factor for adjusting the noise covariance matrix; that is:

[0059]

[0060] and is the estimate of process and measurement noise at time t, which is used for the initial control quantity of KF and measurement The correction is:

[0061]

[0062] In the formula, are the action scaling factors for the process and measurement noise estimates, respectively; are the modified KF controlled quantity and measured quantity respectively.

[0063] As a preferred embodiment of this invention, in step S1), the corrected advantage error is used as the reward function, and its expression is:

[0064]

[0065] in, The reference position is obtained by solving the baseline method; The actual position of the reference; For the predicted location.

[0066] As a preferred embodiment of this invention, in step S1), the adaptive denoising Kalman filter satellite positioning process is as follows:

[0067] S121), define the state equation and measurement equation of the vehicle positioning of KF at time t, namely:

[0068] x t =Fx t-1 +Bu t +w t

[0069] zt =Hx t +v t

[0070] In the formula, x t and x t-1 is the predicted position of the vehicle at time t and time t-1 by KF; u t is the control quantity of KF at time t, i.e., the vehicle speed; is the state transfer matrix; it is defined as the identity matrix here; is the control variable matrix, which is defined as the identity matrix here; represents random measurement noise, which is assumed to follow a Gaussian distribution w t ~N(0,R), where R represents the measurement noise covariance matrix; represents random process noise, assuming it follows a Gaussian distribution v t ~N(0,Q), Q represents the process noise covariance matrix; z t is the measured value at time t, i.e., the coarse positioning result solved by weighted least squares method; represents the state-to-measurement transformation matrix, which is defined here as the identity matrix;

[0071] S122), in the prediction stage, the position at the current moment is predicted according to the position at the previous moment through the dynamic model, that is, the state equation, wherein the control amount is corrected by the process noise estimated by the reinforcement learning model, and the process noise covariance matrix is ​​adjusted by the output action of the reinforcement learning model, and then the a priori estimated covariance matrix is ​​predicted; that is:

[0072]

[0073]

[0074] in, represents the posterior estimated covariance at time t-1, represents the prior estimated covariance at time t, P0 is initialized to a unit diagonal matrix at time t = 0; Q0 is initialized to the covariance matrix of the speed control amount u0 calculated using the weighted least squares method at time t = 0, and the initialization x0 of the KF state is the position solved using the weighted least squares method;

[0075] S123), in the update phase, upon receiving the measured value z t After that, the KF state is updated, and the measurement noise covariance matrix is ​​adjusted through the output action of the reinforcement learning model, and then the covariance matrix of the prior estimate is predicted.

[0076] As a preferred embodiment of this invention, step S123) specifically includes the following steps:

[0077] S1231), calculate the Kalman gain K, which is used to measure the weight of prediction and measurement:

[0078]

[0079] In the formula, represents the state-to-measurement transformation matrix, which is defined here as the identity matrix; represents the prior estimated covariance at time t; R t is the measurement noise covariance matrix at time t;

[0080] S1232) Calculate the optimal estimated state The Kalman gain K is used to measure the weight and combine the predicted state and the measured data to calculate the optimal state estimate at the current moment:

[0081]

[0082] S1233), covariance matrix P t Update: is the identity matrix;

[0083] Therefore, the final satellite positioning result of the adaptive denoising Kalman filter based on reinforcement learning is

[0084] S2), building a dual-agent reinforcement learning model;

[0085] like Figure 1 As shown, in this embodiment, the dual-agent reinforcement learning model includes two independent agents π m and π p The two intelligent agents act in the same adaptive KF satellite positioning environment and obtain positioning results, and estimate and adjust the process / measurement noise and the corresponding noise covariance matrix respectively.

[0086] In this embodiment, the dual-agent reinforcement learning model includes an observation feature extractor and an actor-evaluator network. The observation feature extractor includes an attention module and an LSTM network module; the attention module can adaptively learn the corresponding importance weights of observation features from different perspectives; and the LSTM network module is used to extract important time series information from historical observation data.

[0087] The actor-evaluator network includes an actor network and an evaluator network, wherein the actor network is used to sample and output the action of the intelligent agent from the action space, and the evaluator network is used to output the value of the current observation state and guide the learning of the actor network.

[0088] As a preferred embodiment of this invention, in order to extract effective features and generate accurate potential representations, this embodiment learns the attention weights of different observation features through the attention module; for the agent π m ; The input observation feature is the satellite observation feature o g,t , historical position sequence o h,t and the measurement and process noise covariance observation matrix o m,t ,o p,t ; By having parameter set Attention module Learn the attention weights for multiple input observation sets, specifically:

[0089] aw g =σ(w g ×o g,t );

[0090] aw h =σ(w h ×o h,t );

[0091] aw m =σ(w m ×o m,t );

[0092] aw p =σ(w p ×o p,t );

[0093] Among them, aw g ,aw h ,aw p and aw m is the attention weight of different input observation sets, σ ​​is the sigmoid activation function;

[0094] Then, the weighted satellite observation features, weighted historical position sequence and weighted measurement / process noise covariance matrix are concatenated to obtain the agent π m Attention-weighted multi-input observations Right now:

[0095]

[0096] In the formula, concat represents the concatenation operation between vectors.

[0097] As a preferred embodiment of the present invention, the LSTM network module can effectively capture the temporal relationship between data samples; in this embodiment, the intelligent agents π m and π p Construct the corresponding LSTM network module; for the intelligent agent π m and π p The attention-weighted multi-input observations are passed through the forget gate to remove some useless information in the previous moment; then the important feature information is extracted through the input gate; finally, the output gate outputs the potential representation containing the effective information in the observation features; that is, the corresponding potential representation of the extracted observation features is obtained through the corresponding LSTM network module and The expression is:

[0098]

[0099] in, and Denotes the agent π p and π m The weight parameters of the LSTM network module; For the intelligent agent π p The attention-weighted multi-input observations, For the intelligent agent π m The attention-weighted multiple input observations.

[0100] Therefore, the parameters in the observation feature extractor of the dual-agent reinforcement learning model are:

[0101]

[0102] In the formula, They are respectively p and π m The weight parameters of the attention module; and Denotes the agent π p and π m The weight parameters of the LSTM network module.

[0103] The expression of the potential representation of a multi-input observation set through the observation feature extractor Φ is:

[0104]

[0105] As a preferred embodiment of this embodiment, for the intelligent agent π m and π p Construct the corresponding actor network respectively and in and is the corresponding actor network weight parameter; the output of the observation feature extractor Actor Network and Input; Actor Network and The outputs of are used for process (measurement) noise estimation And the corresponding covariance matrix adjustment The mean μ corresponding to each element in i,t and variance σ i,t .

[0106] In this embodiment, the intelligent agent π m and π p Construct the corresponding evaluator network respectively and in, and is the corresponding evaluator network weight parameter; its input observation feature extractor output potential representation h t , the output is the value function V(o t )’s estimate.

[0107] S3), training the dual-agent reinforcement learning model;

[0108] like Figure 2 As shown, the training of the dual-agent reinforcement learning model includes the following steps:

[0109] S31), the pseudo-range measurement obtained in step S1) Get the position coordinates of each satellite based on the broadcast ephemeris The current approximate position coordinates of the receiver Baseline position The actual location of the reference As a training set;

[0110] S32), initialize an experience replay pool, agent π m and π p Continue to interact with the environment, observing o at time t t Next, by observing the feature extractor and the actor network and Get action a t ;

[0111] The positioning result is obtained by Kalman filter positioning method And use and Calculate the current reward r t, and reaches the observation o at time t+1 t+1 , and then collect the empirical data (o t ,a t ,r t ,o t+1 ) is stored in the playback experience pool;

[0112] S33), when the amount of stored data in the experience replay pool reaches the set size, the parameters of the dual-agent reinforcement learning model are updated; generalized advantage estimation is used Calculate the strategic advantage, namely:

[0113]

[0114] Among them, V π represents the value function under the current observation; γ is the reward discount factor; r t represents the reward obtained by the agent at time t; the value function is estimated by the mean of the output of the evaluator network of each of the two agents, that is:

[0115]

[0116] In the formula, They are respectively m and π p The network of evaluators; and is the corresponding evaluator network weight parameter;

[0117] Define agent π m and π p The corresponding parameters of the dual-agent reinforcement learning model are m and θ p , where includes the observation feature extractor parameters and the actor and evaluator network parameters; therefore, the actor network parameters corresponding to the two agents are optimized by maximizing the following objective function:

[0118]

[0119] In the formula, represents the computational expectation; θ actor,k+1 and θ actor,k represents the parameters of the actor network during the k+1th and kth rounds of updates; clip(·) represents the truncation operation used to clip the probability ratio to a specified interval; ∈=0.2 is the preset hyperparameter; ratio represents the output action probability ratio of the new and old strategies; represents the generalized advantage estimate; represents the objective function of the actor network;

[0120] For the evaluator network, the mean squared return error is used to define the objective function to update the parameters of the evaluator network;

[0121]

[0122] In the formula, for; represents the computational expectation;

[0123] Finally, the overall objective function of the dual-agent reinforcement learning model is as follows:

[0124]

[0125] And the objective function The stochastic gradient descent algorithm is used to iteratively update the weight parameters of the model.

[0126] S4) deploy the trained dual-agent reinforcement learning model to the cloud server; output the results to the satellite positioning chip terminal in real time through end-to-end transmission.

[0127] This embodiment adopts the observation o at time t t ={o g,t ,o h,t ,o p,t ,o m,t}, Rough position coordinates of the receiver Test the trained dual-agent reinforcement learning model by transforming the observation o t Input into the observation feature extractor and actor network corresponding to the dual agents to obtain the output action Combined with the estimated denoising of the measurement / process noise and the corresponding covariance adjustment, the satellite positioning results are solved in real time through the adaptive denoising Kalman filter satellite positioning method based on reinforcement learning After the test is completed, the dual-agent reinforcement learning model will be deployed to the cloud server; the results will be output to the satellite positioning chip terminal in real time through end-to-end transmission.

[0128] The above embodiments and descriptions are only for illustrating the principles and best embodiments of the present invention. Without departing from the spirit and scope of the present invention, the present invention may be subject to various changes and improvements, all of which fall within the scope of the present invention to be protected.

Claims

1. A Beidou positioning method based on adaptive noise covariance learning and denoising, characterized in that: The following steps are involved: S1), receiving satellite data in real urban scenes and building an adaptive KF satellite positioning environment based on reinforcement learning; S2), building a dual-agent reinforcement learning model; The dual-agent reinforcement learning model includes two independent agents; the two agents act in the same adaptive KF satellite positioning environment and obtain positioning results, and respectively estimate and adjust process noise, measurement noise and corresponding covariance matrix; The dual-agent reinforcement learning model includes an observation feature extractor and an actor-evaluator network; wherein the observation feature extractor includes an attention module and an LSTM network module; the attention module adaptively learns the corresponding importance weights of observation features from different perspectives; the LSTM network module is used to extract important time series information from historical observation data; The actor-evaluator network includes an actor network and an evaluator network, wherein the actor network is used to sample and output the action of the agent from the action space, and the evaluator network is used to output the value of the current observation state and guide the learning of the actor network; In order to extract effective features and generate accurate potential representations, the attention weights of different observation features are learned through the attention module; for the agent π m , the input observation feature is the satellite observation feature o g,t , historical position sequence o h,t , and the measurement and process noise covariance observation matrix o m,t ,o p,t ; By having parameter set Attention module Learn the attention weights for multiple input observation sets, specifically: oh g =σ(in g ×o g,t ); oh h =σ(in h ×o h,t ); oh m =σ(in m ×o m,t ); oh p =σ(in p ×o p,t ); Among them, aw g ,aw h ,aw p and aw m is the attention weight of different input observation sets, σ ​​is the sigmoid activation function; Then, the weighted satellite observation features, weighted historical position sequence, weighted measurement noise covariance matrix, and process noise covariance matrix are concatenated to obtain the agent π m Attention-weighted multi-input observations Right now: In the formula, concat represents the concatenation operation between vectors; The LSTM network module can effectively capture the temporal relationship between data samples; they are respectively the agent π m and π p Construct the corresponding LSTM network module; for the intelligent agent π m and π p The attention-weighted multi-input observations are passed through the forget gate to remove some useless information in the previous moment; then the important feature information is extracted through the input gate; finally, the output gate outputs the potential representation containing the effective information in the observation features; that is, the corresponding potential representation of the extracted observation features is obtained through the corresponding LSTM network module and The expression is: in, and Denotes the agent π p and π m The weight parameters of the LSTM network module; For the intelligent agent π p The attention-weighted multi-input observations, For the intelligent agent π m The attention-weighted multi-input observations; Therefore, the parameters in the observation feature extractor of the dual-agent reinforcement learning model are: In the formula, They are respectively p and π m The weight parameters of the attention module; and Denotes the agent π p and π m The weight parameters of the LSTM network module; The expression of the potential representation of a multi-input observation set through the observation feature extractor Φ is: in, They are respectively p , π m The corresponding observation feature extractor; They are respectively p , π m The corresponding attention module; They are respectively p , π m The weight parameters of the attention module; For the agent π m and π p Construct the corresponding actor network respectively and in and is the corresponding actor network weight parameter; Observe the output of the feature extractor Actor Network and Input; Actor Network and The outputs of Measurement noise estimation And the corresponding covariance matrix adjustment The mean μ corresponding to each element in i,t and variance σ i,t ; For the agent π m and π p Construct the corresponding evaluator network respectively and in, and is the corresponding evaluator network weight parameter; its input observation feature extractor output potential representation h t , the output is the value function V(o t ) S3), training the dual-agent reinforcement learning model; S4) deploy the trained dual-agent reinforcement learning model to the cloud server; output the results to the satellite positioning chip terminal in real time through end-to-end transmission.

2. A Beidou positioning method based on adaptive noise covariance learning and denoising according to claim 1, characterized in that: In step S1), the construction of the satellite positioning environment includes the construction of a multi-view observation space, an action space, the setting of a reward function, and the calculation of an adaptive denoising Kalman filter satellite positioning result.

3. A Beidou positioning method based on adaptive noise covariance learning and denoising according to claim 2, characterized in that: In step S1), the construction of the multi-view observation space is defined as: O={O g ,O h ,O p ,O m }; Among them, O g is the satellite observation feature; O h is the historical position sequence; O p is the process noise covariance matrix; O m is the measurement noise covariance matrix.

4. The Beidou positioning method based on adaptive noise covariance learning and denoising according to claim 1, characterized in that: In step S3), the overall objective function of the dual-agent reinforcement learning model is as follows: In the formula, is the objective function of the actor network; is the objective function of the evaluator network; by Use the stochastic gradient descent algorithm to iteratively update the model’s weight parameters, θ m and θ p They are respectively n and π p Parameters of the corresponding two-agent reinforcement learning model.

5. A Beidou positioning method based on adaptive noise covariance learning and denoising according to claim 4, characterized in that: The actor network parameters corresponding to the two agents are optimized by maximizing the following objective function: In the formula, represents the computational expectation; θ actor,k+1 and θ actor,k represents the parameters of the actor network during the k+1th and kth rounds of updates; clip(·) represents the truncation operation used to clip the probability ratio to a specified interval; ∈ is a preset hyperparameter; ratio represents the output action probability ratio of the new and old strategies; represents the generalized advantage estimate; represents the objective function of the actor network; θ m and θ p For the intelligent agent π m and π p Parameters of the corresponding dual-agent reinforcement learning model; For the evaluator network, the mean squared return error is used to define the objective function to update the parameters of the evaluator network; In the formula, is the objective function of the evaluator network; represents the computational expectation; They are respectively m and π p The evaluator network; γ is the reward discount factor; r l represents the reward obtained by the agent at time l.

Images