Satellite positioning enhancement method based on spatio-temporal information fusion

Abstract

The invention discloses a satellite positioning enhancement method based on spatio-temporal information fusion. The method comprises the following steps: building a reinforcement learning environment by following a Markov process; introducing a self-supervised representation learning module, and constructing a basic model of a deep reinforcement learning network based on Transform space-time fusion representation; and training the basic model based on the reinforcement learning environment. By using the method, the interaction between satellite measurement time characteristics and spatial characteristics is considered, and the dependency relationship between vehicle position sequence points is considered, so that the state information of the current environment can be described more comprehensively, and the positioning precision of the model in the urban complex environment is further improved. The method can be widely applied to the field of satellite positioning.

Description

Technical Field

[0001] The present invention relates to the field of satellite positioning, and in particular to a satellite positioning enhancement method based on spatiotemporal information fusion. Background Art

[0002] Autonomous driving has become a hot topic in the automotive industry. High-precision real-time vehicle positioning in urban environments is one of the keys to ensuring the safe travel of autonomous vehicles. However, multipath effects and signal shielding in urban environments can cause a lot of noise in GNSS measurements, thereby reducing positioning accuracy.

[0003] The deep reinforcement learning-based method can effectively reduce the impact of multipath effects and other problems by modeling environmental noise, and improve the vehicle positioning accuracy of GNSS in complex urban environments. However, existing deep reinforcement learning-based methods often only focus on the temporal characteristics of satellite measurements (such as the time series characteristics of the signal) or spatial characteristics (such as the geometric distribution of satellites), without considering the integration of these two features; at the same time, the existing methods only use a simple fully connected network to process the historical information of vehicle positions, without considering the dependencies between the vehicle position sequence points, making it difficult to obtain a comprehensive representation of the environmental state, resulting in poor decision-making quality of the deep reinforcement learning algorithm, which in turn affects the positioning accuracy. Summary of the invention

[0004] In view of this, in order to solve the technical problem that the existing satellite positioning method cannot achieve a comprehensive representation of the environmental state, thereby resulting in low positioning accuracy, the present invention proposes a satellite positioning enhancement method based on spatiotemporal information fusion, the method comprising the following steps:

[0005] Following the Markov process, a reinforcement learning environment is constructed for positioning correction;

[0006] Build a basic model based on Transformer's spatiotemporal feature extractor and self-supervised auxiliary representation learning module;

[0007] Based on the reinforcement learning environment, the basic model is trained.

[0008] In some embodiments, the base model includes:

[0009] Transformer-based spatiotemporal feature extractor: It includes a spatial graph convolution module and a temporal transformer module, which are used to extract the spatial and temporal features of satellite measurements respectively, while considering the dependencies between the vehicle position sequence points to obtain a comprehensive representation of the environment;

[0010] Self-supervised assisted representation learning module based on data enhancement: Use reasonable data enhancement methods to inject prior knowledge into the training data, simulate the difference in data distribution between the unknown environment and the training environment, and use self-supervised representation learning methods to enable the model to learn a robust representation of the environment, ignoring the influence of different environments, thereby improving the generalization of the model in unknown environments.

[0011] In some embodiments, based on the reinforcement learning environment, the step of training the basic model specifically includes:

[0012] Collect real-time positioning trajectory data of actual vehicles and build a training set;

[0013] Determine whether the update frequency of self-supervised representation learning is reached. If so, perform reinforcement learning training and self-supervised representation learning training at the same time. Otherwise, only perform reinforcement learning training.

[0014] The experience data obtained by the continuous interaction between the agent and the environment is stored in the experience replay pool until the experience replay pool reaches the maximum storage capacity;

[0015] Extract experience from the experience replay pool to train the reinforcement learning model, where the model parameters corresponding to the agent actor network are back-propagated based on stochastic gradient descent (SGD);

[0016] When the maximum number of training times is reached, the training ends.

[0017] The present invention provides a satellite positioning enhancement method based on spatiotemporal information fusion, constructs a Transformer spatiotemporal feature encoder, uses a graph convolutional network to obtain the spatial features of satellite measurements, and then uses the core component of the Transformer, multi-head attention, to extract the temporal features of satellite measurements over a period of time, fuses the spatiotemporal features of satellite measurements, and considers the dependencies between vehicle position sequence points at the same time, so as to obtain a comprehensive representation of the environmental state, thereby achieving accurate positioning correction; further, a self-supervised representation learning based on data enhancement is introduced, which aims to train the model to learn a robust representation of the current environmental state by learning the consistency of the intermediate representation of the enhanced data and the original data, ignoring the differences between the environments, so that the model can effectively adapt to the dynamically changing environment. The method of the present invention can learn a more comprehensive and robust environmental representation to improve the positioning accuracy of the model in complex urban environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the steps of a satellite positioning enhancement method based on spatiotemporal information fusion of the present invention;

[0019] Figure 2It is a structural block diagram of the spatiotemporal feature encoder of the Transformer of a specific embodiment of the present invention;

[0020] Figure 3 It is a data flow diagram of the overall basic model of a specific embodiment of the present invention;

[0021] Figure 4 It is a flowchart of model training according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0023] It should be noted that, for the convenience of description, only the parts related to the relevant invention are shown in the drawings. In the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other.

[0024] It should be understood that the "system", "device", "unit" and / or "module" used in this application is a method for distinguishing different components, elements, parts, portions or assemblies at different levels. However, if other words can achieve the same purpose, the word can be replaced by other expressions.

[0025] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "a kind" and / or "the" do not refer to the singular, but also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of clearly identified steps and elements, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements. The elements defined by the sentence "includes a..." do not exclude the existence of other identical elements in the process, method, commodity or device that includes the elements.

[0026] In the description of the embodiments of the present application, "plurality" means two or more than two. The following terms "first" and "second" are used for descriptive purposes only and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features.

[0027] In addition, flow charts are used in the present application to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the preceding or following operations are not necessarily performed accurately in order. On the contrary, the various steps may be processed in reverse order or simultaneously. At the same time, other operations may also be added to these processes, or a certain step or several steps of operations may be removed from these processes.

[0028] Reference Figure 1 , is a flow chart of an optional example of a satellite positioning enhancement method based on spatiotemporal information fusion proposed in the present invention. The method can be applied to a computer device. The positioning enhancement method proposed in this embodiment may include but is not limited to the following steps:

[0029] Step S1, building a reinforcement learning environment following the Markov process;

[0030] Step S2: introduce a self-supervised representation learning module to build a basic model of a deep reinforcement learning network based on Transformer spatiotemporal fusion representation;

[0031] Step S3: training the basic model based on the reinforcement learning environment.

[0032] The final application stage takes the state of the trajectory as input and outputs the corrected position.

[0033] In some feasible embodiments, in the step S1, it specifically includes:

[0034] The reinforcement learning objective for vehicle positioning correction follows a Markov decision process (MDP), which can be expressed as (s t ,a t ,r t ,s t+1 Specifically, the vehicle agent is based on the state representation s obtained from the environment. t And according to the strategy π(a t |s t ) performs an action a t , the agent completes action a t The environment then returns a reward r t (s t ,a t ) and update the state to the next state s t+1 The goal of reinforcement learning is to optimize the policy function π(a t |s t ) to maximize the expected cumulative return

[0035] (1) Status indication

[0036] The current state input s at time t t Including the vehicle position sequence and the observation feature sequence of different satellites, expressed as:

[0037]

[0038] in, Represents the position sequence of the vehicle within a certain historical time range T, that is, Where P t represents the three-dimensional position coordinates of the vehicle in the Earth-Centered Earth-Fixed (ECEF) coordinate system at time t, that is, P t =(x t ,y t ,z t ); It represents the observation characteristic sequence composed of the line-of-sight vectors and pseudo-range residuals of different satellites within the same historical time range T, that is, Among them, M t Represents the line-of-sight vectors of different satellites at time t and pseudorange residuals The observed characteristics of Represents the satellite measurement characteristics obtained by a single satellite at time t n represents the total number of satellites observed at the current moment.

[0039] (2) Action indication

[0040] In this embodiment, the action is defined as a correction operation on the vehicle position. In order to avoid the problem of dimensionality disaster caused by the increase of the dimension of the action space as a discrete action, we use a continuous action setting that obeys the Gaussian distribution. The position correction values ​​in the three directions of X, Y, and Z in the coordinate system are all sampled from the Gaussian distribution. Therefore, we define the decision action a t Set to:

[0041]

[0042] in So the position correction is ΔP = (Δx t ,Δy t ,Δz t ), the corrected position of the final model output It can be expressed as:

[0043]

[0044] (3) Rewards

[0045] This example uses the corrected advantage error as the reward function of reinforcement learning to evaluate the effectiveness of the action, which is defined as follows:

[0046]

[0047] Among them, P t represents the initial position, It represents the real position obtained by the satellite-inertial combined positioning system. Represents the corrected position predicted by the model. Intuitively, the closer the corrected position output by the model is to the true position, the higher the reward.

[0048] In some feasible embodiments, the basic model in step S2 specifically includes:

[0049] (1) Transformer-based spatiotemporal feature encoder

[0050] In the existing positioning correction method based on deep reinforcement learning, only the temporal characteristics or spatial characteristics of satellite measurements are considered, and the dependency between the vehicle position sequence points is not considered, which cannot fully describe the environmental state information, resulting in limited positioning correction performance based on deep reinforcement learning methods. Therefore, this embodiment proposes a Transformer-based spatiotemporal feature encoder, which aims to combine the spatiotemporal characteristics of satellite measurements and the dependency between vehicle position sequence points to more comprehensively describe the environmental state information and achieve accurate positioning correction. The network structure of the Transformer-based spatiotemporal feature encoder is as follows: Figure 2 shown.

[0051] Extract satellite measurement spatial features through the spatial graph convolutional network module:

[0052] Satellite receivers can receive signals from multiple constellations at the same time, including the Global Positioning System GPS, the Beidou Satellite Navigation System BDS and the Galileo Satellite Navigation System GAL. First, considering the spatial relationship between multi-constellation satellites, we construct a multi-constellation graph G of all visible satellites at the current time t t =(V t ,E t ), V t represents a node, E t We regard satellites as nodes of the graph, and the node features include satellite measurement features composed of line-of-sight vectors and pseudo-range residuals. The edge connection between two satellites is determined based on whether the two satellites are from the same constellation or whether the pseudorange residual is less than a certain threshold. We represent the constellation of satellite i at time t as The construction based on this regular graph can be expressed as:

[0053]

[0054] Among them, i and j represent the two satellites observed at the current time t, represents the constellation type of satellites i and j, and η is an artificially set threshold.

[0055] Next, we introduce the core components for extracting satellite measurement spatial features. We only construct a multi-constellation graph G for all visible satellites at the current time t. t = ( V t ,E t) As input, the graph convolutional network (GCN) is used to extract the spatial features, which can be expressed as:

[0056]

[0057] in, and represents the hidden features of the l+1th layer and the lth output in the graph convolutional network at time t. Note that the 0th layer The propagation rule for each convolutional layer in the graph convolutional network is formulated as:

[0058]

[0059] D is E t The diagonal matrix of is the trainable weight of the graph convolution layer, and σ(·) is the ELU nonlinear activation function. We set the number of GCN layers to L and get the input G t Hidden features at layer L At the same time, we use a multi-layer perceptron (MLP) to extract satellite measurement features from tT to t-1 and map them as Finally, we concatenate the hidden features from tT to t, expressed as:

[0060]

[0061] The temporal features of satellite measurements and vehicle position sequences are extracted through the time series Transformer module:

[0062] Transformer has shown remarkable effectiveness in various tasks in the field of time series. It relies on the core component multi-head attention mechanism to model long-term sequences and capture the dependencies between different positions in the sequence, making it very attractive in time series modeling. Below, we specifically introduce the multi-head attention mechanism and the application of time series transformer in extracting satellite measurement time features and the dependencies between vehicle position sequence points.

[0063] Multi-head attention mechanism, which is the core mechanism in Transformer, learns the temporal features of the input sequence from different aspects. This mechanism enhances the model's ability to capture complex dependencies by projecting the query, key, and value into multiple subspaces, performing attention operations, and finally merging the results. The calculation process of the multi-head attention mechanism is expressed as:

[0064] MHA(Q,K,V)=concat(head1,head2,...,head h )W O

[0065] The calculation for each head is:

[0066]

[0067] Among them, Q, K, and V are query, key, and value matrices respectively. The learnable weights corresponding to each head are used to project Q, K, and V into different subspaces. The Attention(·) function is the scaled dot-product attention (ScaledDot-ProductAttention), and its attention score calculation formula is:

[0068]

[0069] Among them, d k represents the dimension of the key vector, Used to scale the dot product to stabilize the gradient of the softmax(·) function. The outputs of all heads are concatenated and passed through another learnable linear transformation W O are fused to obtain the final output.

[0070] The time series transformer extracts the satellite measurement time characteristics and the vehicle position sequence point dependencies: given a vehicle position sequence And the hidden feature B extracted by the spatial graph convolutional network module t , the temporal Transformer encodes them respectively and extracts the temporal features, which are expressed as:

[0071]

[0072] After being processed by the multi-head attention mechanism, the output features are input into an addition and normalization layer (Add&Norm) and a feed forward neural network layer (Feed Forward Layer).

[0073] Multi-observation input feature fusion through cross-attention module: In the above process, temporal features and spatial features are extracted from satellite measurements respectively, and the dependencies between vehicle position sequence points are considered. The subsequent key step is to effectively fuse multi-observation input features for positioning correction. Some simple feature fusion methods include concatenation or feature summation, but simple concatenation or summation is difficult to capture the complex interactions between multiple input features. Cross-attention has shown significant effectiveness in feature fusion, so we use the cross-attention module to fuse and Features, the cross attention module uses the multi-head attention mechanism mentioned above, in which we As a query, As keys and values, output fused features z t , the specific operations are as follows:

[0074]

[0075] Finally, the fusion feature z is obtained t , which obtains a more comprehensive representation of the environmental state by fusing the spatiotemporal features of satellite measurements with the dependencies between the vehicle position sequence points, and is used for the positioning decision of the reinforcement learning model.

[0076] (2) Self-supervised auxiliary representation learning module based on data enhancement

[0077] Since the surrounding environment of the vehicle changes dynamically during driving, the data distribution between different environments may be inconsistent, which makes it difficult for the model obtained in the training environment to generalize to other unknown environments. In order to improve the generalization of the model, we introduce the SODA model to construct a self-supervised auxiliary representation learning module based on data enhancement, aiming to train the model to learn a robust representation of the current environment state and ignore the differences between data distributions in different environments. SODA is a self-supervised auxiliary task that shares a common encoder f(·) with the reinforcement learning strategy. SODA consists of a common online encoder f θ (·), online projector g θ (·) and a predictor h θ (·) consists of three parts. In addition, SODA also considers a target encoder with the same network structure and target projector The workflow of the self-supervised assisted representation learning module based on data enhancement is as follows:

[0078] First, given the input state representation s t We use the data augmentation method of adding Gaussian noise to the training data to inject prior knowledge into the original state representation to obtain the enhanced state representation s't , used to simulate the data distribution that may exist in an unknown environment, the enhanced state representation is defined as:

[0079] s′ t =s t +N

[0080] in, N means that the mean is 0 and the variance is σ 2 Gaussian noise. t With s' t It can be regarded as different perspectives of the same state.

[0081] Then input both into the target encoder and the online encoder f θ (·) and then input to the target projector and online projectorg θ (·) outputs two potential representations z t and z' t Here, we set the network structure of the two encoders to the Transformer-based spatiotemporal feature extractor defined in (1), and the network structure of the two projectors to the MLP layer, which are expressed as:

[0082]

[0083] z′ t =g θ (f θ (s′ t ))

[0084] where the latent representation z' obtained by the online projector t As the predictor h θ The input and output prediction of (·) are expressed as

[0085]

[0086] We will potentially represent z t Defined as the target representation The goal of SODA is to make the predictor h θ (·) Predicted representation With target representation Since the initial inputs of the two simulate the differences in data distribution in different environments, the consistency between the learned representations can be used to promote the public online encoder f θ (·) Learn a robust representation that adapts to different environments. We use similarity loss to calculate the predicted representation With target representation The distance between them is expressed as:

[0087]

[0088] We use this similarity loss to calculate the reverse gradient to update the online encoder f θ (·), online projector g θ (·) and the predictor h θ (·) parameter θ, and at the same time, the target encoder and target projector Update parameters using the exponential moving average (EMA) method It is expressed as:

[0089]

[0090] Among them, τ is a manually set hyperparameter, are the model parameters at the current moment and the previous moment.

[0091] (3) Actor-Critic Network

[0092] This embodiment uses the most advanced policy gradient algorithm based on the Actor-Critic structure, the Proximal Policy Optimization (PPO) algorithm, to train and update the parameters of the model. We use two deep neural network (DNN) modules to form the Actor network and the Critic network. a and L c layer neural network, so the parameters of the Actor network are The parameters of the Critic network are Based on the above designed Transformer-based spatiotemporal feature encoder, the output fusion feature z is obtained t After that, we put the Actor and Critic in L a , L c The output of the layer network is represented as

[0093]

[0094] Where ψ(·) represents a nonlinear activation function. At time t, the Actor network outputs the action space, and the Critic network outputs the estimated state value, expressed as:

[0095]

[0096]

[0097] Where μ and σ are the mean and standard deviation of each dimension, and Respectively represent the L a , L c For the continuous positioning and correction environment, the Actor network outputs the action Gaussian distribution and determines the action by sampling from the continuous action space. The Critic needs to predict the value of different states to guide the training of the Actor network.

[0098] To ensure the effectiveness of the Actor-Critic learning function, we use PPO for Actor-Critic training, which ensures that the trust region is updated with an advantage clipping strategy and is efficient in continuous action environments. We use generalized advantage estimation to guide Actor training, defined as follows:

[0099]

[0100] in It takes into account the time difference reward residual of the Critic in adjacent time steps to help the agent consider the temporal association. The goal of the policy network is to update the parameters of the Actor network so that the output value is constantly close to the optimal policy. We define the loss function of the Actor network by clipping the PPO objective function:

[0101]

[0102] in, represents the generalized advantage estimate, Represents the probability ratio between the new and old policy networks, clip ( · represents the pruning function, which modifies the objective function of the Actor network by pruning the probability ratio, eliminating r i( θ a) In the range [ The part outside 1-∈,1+∈].

[0103] On the other hand, in order to achieve accurate value estimation and long-term positioning accuracy, the critic loss function L c (θ c ) is defined as the mean-squared return error (MSRE), taking into account the cumulative expected return reward, as follows:

[0104]

[0105] in, Consider L v The cumulative discounted reward for the vehicle trajectory, represents the Critic network model, θ cDenotes the parameters of the Critic network, and γ denotes the discount factor. Different from the common single-step error used in most learning-based methods, MSRE helps the agent consider the corrected accuracy of sequence positioning.

[0106] Next, this embodiment applies the entropy loss function To ensure that the agent has sufficient exploratory power when interacting with the environment, it is defined as:

[0107]

[0108] Finally, the overall objective function is obtained:

[0109]

[0110] Where β1 and β2 are loss coefficients. Through interaction with the positioning correction environment, the agent can collect experience and use back propagation based on stochastic gradient descent (SGD) to train the overall parameter set.

[0111] The overall data flow of the basic model of the deep reinforcement learning network based on Transformer spatiotemporal fusion representation is referenced Figure 3 .

[0112] In some feasible embodiments, the step S3 specifically includes:

[0113] S3.1. The vehicle drives in a real urban scene and uses the on-board receiver to collect real-time positioning trajectory data.

[0114] S3.2. Determine whether the update frequency of self-supervised representation learning is reached. If so, perform reinforcement learning training and self-supervised representation learning training at the same time. Otherwise, only perform reinforcement learning training.

[0115] S3.3, initialize the experience replay pool, and store the experience data (s t ,a t ,r t ,s t+1 ) quadruplets are stored in the experience replay pool until the experience replay pool reaches its maximum storage capacity.

[0116] S3.4. Extract experience from the experience replay pool to train the reinforcement learning model, where the model parameters corresponding to the agent Actor network are back-propagated using stochastic gradient descent (SGD).

[0117] S3.5. When the maximum number of training times is reached, the training ends.

[0118] The overall training process reference Figure 4 .

[0119] A satellite positioning enhancement system based on spatiotemporal information fusion, comprising:

[0120] Environmental modeling unit, building a reinforcement learning environment based on Markov processes;

[0121] The model building unit introduces a self-supervised representation learning module to build a basic model of a deep reinforcement learning network based on Transformer spatiotemporal fusion representation;

[0122] A training unit trains the basic model based on the reinforcement learning environment.

[0123] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0124] A satellite positioning enhancement device based on spatiotemporal information fusion:

[0125] at least one processor;

[0126] at least one memory for storing at least one program;

[0127] When the at least one program is executed by the at least one processor, the at least one processor implements the satellite positioning enhancement method based on spatiotemporal information fusion as described above.

[0128] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0129] A storage medium stores processor-executable instructions, which are used to implement a satellite positioning enhancement method based on spatiotemporal information fusion as described above when executed by the processor.

[0130] The contents of the above method embodiments are all applicable to the present storage medium embodiments. The functions specifically implemented by the present storage medium embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0131] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art may make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A satellite positioning enhancement method based on spatiotemporal information fusion, characterized in that: The following steps are involved: Building a reinforcement learning environment following the Markov process; Introducing a self-supervised representation learning module to build a basic model of a deep reinforcement learning network based on Transformer spatiotemporal fusion representation; Based on the reinforcement learning environment, the basic model is trained.

2. The satellite positioning enhancement method based on spatiotemporal information fusion according to claim 1, characterized in that: The state space in the reinforcement learning environment is represented as follows: Among them, s t represents the state input at time t, represents the position sequence of the vehicle within a certain historical time range T, Represents the observation feature sequence composed of the line-of-sight vectors and pseudorange residuals of different satellites within the historical time range T.

3. The satellite positioning enhancement method based on spatiotemporal information fusion according to claim 2, characterized in that: The basic model includes a Transformer spatiotemporal feature extractor and a self-supervised auxiliary representation learning module, where: Transformer spatiotemporal feature extractor, which extracts satellite measurement spatial features through the spatial graph convolutional network module, extracts temporal features of satellite measurement and vehicle position sequence through the temporal transformer module, and performs multi-observation input feature fusion through the cross attention module; The self-supervised auxiliary representation learning module simulates the data distribution of an unknown environment through data augmentation and performs reverse gradient update using similarity loss.

4. The satellite positioning enhancement method based on spatiotemporal information fusion according to claim 2, characterized in that: The step of training the basic model based on the reinforcement learning environment specifically includes: Obtain real-time positioning trajectory data and build a training set; Controlling the intelligent agent to interact with the environment and storing the experience data after the interaction into the experience replay pool until the experience replay pool reaches the maximum capacity; Experience is extracted from the experience replay pool to train the basic model.

5. The satellite positioning enhancement method based on spatiotemporal information fusion according to claim 4, characterized in that: The step of training the basic model based on the reinforcement learning environment further includes: When the update frequency of self-supervised representation learning is reached; Injecting prior knowledge into the original state in the training set to obtain an enhanced state; Processing the original state and the enhanced state through an encoder and a projector respectively, outputting an original latent representation and an enhanced latent representation; Taking the enhanced latent representation as input during a prediction period, and outputting a predicted representation; The similarity loss between the predicted representation and the target representation is calculated, and the encoder, projector, and predictor are updated with the reverse gradient using the similarity loss.

6. The satellite positioning enhancement method based on spatiotemporal information fusion according to claim 3, characterized in that: The step of extracting satellite measurement spatial features through the spatial graph convolutional network module specifically includes: Using satellites as graph nodes and satellite measurement features as node features, a multi-constellation graph is constructed; Taking the multi-constellation graph as input, a graph convolutional network module is used to extract spatial features to obtain satellite measurement spatial features.

7. The satellite positioning enhancement method based on spatiotemporal information fusion according to claim 3, characterized in that: The step of extracting the temporal features of the satellite measurement and vehicle position sequence by the temporal Transformer module specifically includes: Use multi-head attention mechanism to learn the temporal characteristics of input; The vehicle position sequence and satellite measurement spatial features are encoded respectively through the temporal Transformer to extract the temporal features.

8. A satellite positioning enhancement system based on spatiotemporal information fusion, characterized in that: include: Environmental modeling unit, building a reinforcement learning environment based on Markov processes; The model building unit introduces a self-supervised representation learning module to build a basic model of a deep reinforcement learning network based on Transformer spatiotemporal fusion representation; A training unit trains the basic model based on the reinforcement learning environment.

9. A satellite positioning enhancement device based on spatiotemporal information fusion, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the satellite positioning enhancement method based on spatiotemporal information fusion as described in any one of claims 1 to 7.

Images