A high-precision intelligent positioning terminal device for a single Beidou system

Abstract

This invention provides a high-precision intelligent positioning terminal device for a single BeiDou system, including a receiving antenna, a front-end signal processing module, and an MCU chip. The MCU chip integrates a Kalman filter and a positioning enhancement model composed of a positioning enhancement module based on deep reinforcement learning and a multipath signal recognition module based on Transformer. This invention effectively combines multipath signal recognition and positioning correction, thereby improving the accuracy of satellite positioning. It achieves the transformation and simplification of complex models by building the most basic network layer, and constructs a positioning enhancement model that can run on a performance-constrained MCU based on the TensorFlow Lite for Micro library. This invention reduces the impact of perception differences on noise modeling of complex urban environments through a multi-angle environmental perception module, improving the model's perception capability of the current environment and the utilization rate of data.

Description

Technical Field

[0001] This invention relates to the field of BeiDou positioning technology, and in particular to a high-precision intelligent positioning terminal device and method for a single BeiDou system. Background Technology

[0002] In recent years, the BeiDou Global Navigation Satellite System has made comprehensive progress in both hardware and software, achieving centimeter-level positioning accuracy in open areas. However, in complex urban environments, positioning inaccuracies can occur due to complex environmental errors such as multipath propagation and non-line-of-sight issues, with positioning errors potentially reaching tens of meters.

[0003] To achieve high availability and high accuracy GNSS positioning in various scenarios and cities, a series of AI-based positioning correction algorithms have been researched. These include using high-precision maps and inertial navigation to achieve continuous high-precision vehicle positioning. While AI-based methods can learn complex characteristics of environmental noise models using measurement data without strict prior assumptions, this shows great potential for improving GNSS positioning in complex urban areas.

[0004] However, these methods were all verified only in ideal environments, which differ significantly from real-world operating environments. The following problems still exist:

[0005] 1) These algorithms are all developed using deep learning application frameworks such as Tensorflow or PyTorch. Python can be used to easily create, train, and validate deep network models, which can then be deployed to application systems. However, due to insufficient storage resources and computing power in automotive MCU systems, models that can be created and validated within these frameworks cannot be directly deployed.

[0006] 2) During training and validation, the above algorithms obtain input information from existing data samples without considering the real-time nature of data processing. However, in actual operation, the process from receiving the satellite positioning signal to processing it is extremely crucial, as it ensures the real-time nature of high-precision positioning information.

[0007] 3) The above algorithms are not effectively linked together, and the effective combination of multipath signal intelligent recognition algorithm and AI positioning correction algorithm is not considered, so the AI ​​positioning enhancement capability cannot be effectively utilized. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention provides a high-precision intelligent positioning terminal device for a single BeiDou system. This invention effectively combines multipath signal identification and positioning correction, thereby improving the accuracy of satellite positioning.

[0009] The technical solution of this invention is: a high-precision intelligent positioning terminal device for a single BeiDou system, comprising:

[0010] A receiving antenna, used to receive ephemeris information signals from satellites;

[0011] The pre-signal processing module is used to process the received satellite ephemeris information signals and obtain satellite characteristic observation information data;

[0012] The MCU chip integrates a Kalman filter and a localization enhancement model consisting of a localization enhancement module based on deep reinforcement learning and a multipath signal recognition module based on Transformer.

[0013] The Transformer-based multipath signal recognition module is used to remove satellite data affected by multipath from the satellite feature observation information data; and the purified data after removal is passed to the Kalman filter; the deep reinforcement learning-based positioning enhancement module is used to output the correction amount for the current positioning position; the Kalman filter calculates the initial coarse position based on the purified data, and then combines it with the correction amount of the current positioning position output by the positioning enhancement module to output the corrected high-precision positioning position.

[0014] Preferably, the localization enhancement module based on deep reinforcement learning includes a multi-view environment perception module and a localization correction strategy learning module.

[0015] Preferably, the multi-view environment perception module includes a multilayer sensing network and a long short-term memory network (LSTM). The measurement information of the current environment observation and the state information of the historical behavior are respectively input into the multilayer sensing network and the LSTM for processing and then cascaded, and input into the positioning correction strategy learning module to learn the positioning correction strategy.

[0016] Preferably, the measurement information of the current environment observation is a characteristic observation vector composed of the satellite pseudorange residual RES, line-of-sight vector LOS, elevation angle Ao, and carrier-to-noise ratio C / N0 at the current moment.

[0017] Preferably, the Long Short-Term Memory (LSTM) network includes an input gate, a forget gate, hidden units, and an output gate. Inputting observation information into the LSTM network allows for the accurate estimation of the current confidence state h by integrating the hidden states from past time points with the feature information in the current observation. t .

[0018] Preferably, the state information P of the historical behavior t It contains historical behavior information of the historical trajectory for the last N time steps.

[0019] The goal of the localization augmentation module based on deep reinforcement learning at each time step is to find the optimal correction action to perform. This process is based on a sliding window, and the localization enhancement module based on deep reinforcement learning obtains position correction information each time. The sliding window of the multi-view environment perception module moves forward one step, forming a state information of historical behavior with a constant size of N; then, this state information is input into the multilayer perception network to calculate the confidence state based on the behavior information. The calculation expression is as follows:

[0020]

[0021] Among them, w i Let x represent the connection weight of the i-th neuron, b represent the bias term, and x represent the connection weight of the i-th neuron. i It represents the output of the neurons in the previous layer; n represents the total number of neurons; σ represents the sigmoid activation function.

[0022] Finally, the confidence state h of the environmental observation at the current moment is determined. t and confidence state h′ based on behavioral information t The splicing yields confidence states H from multiple angles. t And it serves as the input to the localization correction strategy learning module, that is:

[0023] H t =g[h t ,h′ t ].

[0024] Preferably, the positioning correction strategy learning module is constructed based on a proximal optimization strategy, including a value network and an action network. The strategy value and correction action are learned through the value network and action network. The value network is updated using a value function, and the loss value at time t is calculated using the mean squared return error. Each state H... t The mean squared return error is used to calculate the loss and update the parameters of the value network; in order to effectively quantify the confidence state H t The impact on the output action is assessed by using Generalized Advantage Estimation (GAE) to update the current action network's policy.

[0025] Preferably, the Transformer-based multipath signal recognition module includes a Transformer-based multi-input fusion module, a feature encoding module, and a feature decoding module. By inputting the current epoch observation features and multi-epoch time-series features into the multi-input fusion module, high-dimensional data is mapped to a low-dimensional vector to generate a low-dimensional latent feature representation. Subsequently, the feature encoding module encodes the generated low-dimensional vector and learns its patterns. Then, the feature decoding module classifies and decodes the vector, finally outputting the recognition result of the multipath signal.

[0026] Preferably, the satellite ephemeris information received by the antenna is first processed into satellite feature information by the pre-signal processing module, and then input into the positioning enhancement module based on deep reinforcement learning and the multipath signal recognition module based on Transformer for model training and verification.

[0027] Preferably, the simplification process for the positioning enhancement model includes pruning and quantization.

[0028] As a preferred approach, TensorFlow Lite for Micro is used as a support library for deploying localization enhancement models on MCU chips; the localization enhancement models are then converted into the TensorFlow framework using the basic operators of the localization enhancement models.

[0029] The beneficial effects of this invention are as follows:

[0030] 1. This invention effectively combines multipath signal identification and positioning correction, thereby improving the accuracy of satellite positioning; and based on the characteristics of dual-core MCU and RTOS operating system, it enables parallel computing of the algorithm by splitting the computing modules, and ensures timely output of positioning results through a delay waiting strategy.

[0031] 2. This invention achieves the transformation and simplification of complex models by building the most basic network layer, and constructs a positioning enhancement model that can run on performance-constrained MCUs based on the TensorFlow Lite for Micro library, effectively reducing the deployment cost of artificial intelligence positioning modules;

[0032] 3. By employing model quantization and pruning-based iterative updates, this invention compresses the model network structure as much as possible while ensuring model calibration accuracy, reducing model storage size and improving the efficiency of model deployment and operation. Attached Figure Description

[0033] Figure 1 This is a structural diagram of the system of the present invention;

[0034] Figure 2 This is a structural framework diagram of the localization enhancement module based on deep reinforcement learning in this invention;

[0035] Figure 3 This is a structural framework diagram of the Transformer-based multipath signal recognition module of the present invention;

[0036] Figure 4 This is a schematic diagram of the pruning process of the positioning enhancement model of the present invention;

[0037] Figure 5This is a schematic diagram of the satellite data processing flow of the MCU chip of the present invention;

[0038] Figure 6 This is a schematic diagram of the process for deploying the model conversion of the present invention. Detailed Implementation

[0039] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0040] like Figure 1 As shown, this embodiment provides a high-precision intelligent positioning terminal device for a single BeiDou system, including:

[0041] A receiving antenna, used to receive ephemeris information signals from satellites;

[0042] The pre-signal processing module is used to process the received satellite ephemeris information signals and obtain satellite characteristic observation information data;

[0043] The MCU chip integrates a Kalman filter and a localization enhancement model consisting of a localization enhancement module based on deep reinforcement learning and a multipath signal recognition module based on Transformer.

[0044] The Transformer-based multipath signal recognition module is used to remove satellite data affected by multipath from the satellite feature observation information data; and the purified data after removal is passed to the Kalman filter; the deep reinforcement learning-based positioning enhancement module is used to output the correction amount for the current positioning position; the Kalman filter calculates the initial coarse position based on the purified data, and then combines it with the correction amount of the current positioning position output by the positioning enhancement module to output the corrected high-precision positioning position.

[0045] As a preferred embodiment, such as Figure 2 As shown, the localization enhancement module based on deep reinforcement learning includes a multi-view environment perception module and a localization correction strategy learning module. This embodiment reduces the impact of perception differences on noise modeling in complex urban environments through the multi-view environment perception module, and also improves the model's ability to perceive the current environment and the utilization rate of data.

[0046] The multi-view environment perception module includes a multilayer sensing network and a long short-term memory network (LSTM). The measurement information of the current environment observation and the state information of the historical behavior are respectively input into the multilayer sensing network and the LSTM for processing and then cascaded, and input into the positioning correction strategy learning module to learn the positioning correction strategy.

[0047] In this preferred embodiment, the measurement information of the current environment observation is a feature observation vector composed of the satellite pseudorange residual RES, line-of-sight vector LOS, elevation angle Ao, and carrier-to-noise ratio C / N0 at the current moment. The measurement information of the current environment observation is represented as follows:

[0048]

[0049] Among them, O t This represents the measurement information of environmental observations at time t, where n is the number of visible satellites in the trajectory; This represents the line-of-sight vector of satellite n at time t. This represents the pseudorange residual of satellite n at time t. This represents the elevation angle of satellite n at time t. This represents the carrier-to-noise ratio of satellite n at time t.

[0050] In this embodiment, the view distance vector and satellite pseudorange residual The expression is as follows:

[0051]

[0052] In the above formula, This represents the position of satellite i at time t. This represents the pseudorange measured by satellite i at time t; information about satellites not visible during the process is filled with 0. This represents the initial approximate position at time t.

[0053] In this preferred embodiment, the Long Short-Term Memory (LSTM) network includes an input gate, a forget gate, hidden units, and an output gate. Inputting observation information into the LSTM network allows for the accurate estimation of the current confidence state h by integrating the hidden states from past time points with the feature information in the current observation. t The calculation expression is as follows:

[0054] f t =σ(W f ·g[h t-1 ,x t ]+b f )

[0055] i t =σ(W i ·g[h t-1 ,x t ]+b i )

[0056]

[0057]

[0058] o t =σ(W o ·g[h t-1 ,x t ]+b o )

[0059] h t =o t ·tanh(C t )

[0060] In the formula, σ represents the sigmoid activation function; W i W f W c W o Let b represent the weight matrices of the LSTM input gate, forget gate, hidden unit, and output gate, respectively. i b f b c b c These represent the bias terms for the LSTM input gate, forget gate, hidden unit, and output gate, respectively. t Let f represent the input information at time t, g[*] represent the vector concatenation operation; t i t o t These represent the outputs of the forget gate, input gate, and output gate, respectively. C t and h t These represent cell state and confidence state, respectively.

[0061] Based on the above calculation expression, the confidence state h of the environmental observation at the current moment is finally obtained. t .

[0062] The state information P of the historical behavior t Historical behavior information, including the historical trajectory of the last N time steps, is represented as:

[0063]

[0064] in, The corrected position at time t-1 is represented by the following calculation process:

[0065]

[0066] in, This represents the initial approximate position at time t. This indicates the correction amount for the position.

[0067] The goal of the localization augmentation module based on deep reinforcement learning at each time step is to find the optimal correction action to perform. This process is based on a sliding window, and the localization enhancement module based on deep reinforcement learning obtains position correction information each time. The sliding window of the multi-view environment perception module moves forward one step, forming a state information of historical behavior with a constant size of N; then, this state information is input into the multilayer perception network to calculate the confidence state h′ based on the behavior information. t The calculation expression is as follows:

[0068]

[0069] Among them, w i Let x represent the connection weight of the i-th neuron, b represent the bias term, and x represent the connection weight of the i-th neuron. i It represents the output of the neurons in the previous layer; n represents the total number of neurons; σ represents the sigmoid activation function.

[0070] Finally, the confidence state h of the environmental observation at the current moment is determined. t and confidence state h′ based on behavioral information t The splicing yields confidence states H from multiple angles. t And it serves as the input to the localization correction strategy learning module, that is:

[0071] H t =g[h t ,h′ t ].

[0072] As a preferred embodiment, such as Figure 2 As shown, the localization correction strategy learning module is constructed based on a proximal optimization strategy, including a value network and an action network. The value network and action network learn the policy value and correction action. In this embodiment, the proximal optimization strategy combines importance sampling and objective function pruning techniques to limit the magnitude of policy updates, thereby reducing the interference of outlier data points on the training of the model policy. This allows the model to learn a more stable correction strategy.

[0073] For value networks, a value function update is used, and the loss value at time t is calculated using the mean squared return error:

[0074]

[0075] Among them, L c This represents the loss function of the value network, used to update the value network parameters; r i+1 It is the reward at time i+1, V π (H t ) is the confidence state value at the current moment. H represents the average value calculated at time t; t This represents the confidence state from multiple angles at time t;

[0076] By using each state H t The mean squared return error is used to calculate the loss and update the parameters of the value network.

[0077] In order to effectively quantify the confidence state H t The impact on the output action is assessed using Generalized Advantage Estimation (GAE) to update the current action network's policy; the calculation formula is as follows:

[0078]

[0079] Among them, L A This represents the loss function of the action network, used to update the value parameters of the action network; θ represents the average value calculated at time t; A The r represents the weights of the action network parameters. t (θ A () represents the ratio of the probability of the new strategy to the probability of the old strategy at time t; denoted as the advantage error function, clip represents the clipping operation, and ∈ represents the clipping ratio of the PPO algorithm.

[0080] As a preferred embodiment, such as Figure 3 As shown, the Transformer-based multipath signal recognition module includes a multi-input fusion module, a feature encoding module, and a feature decoding module based on a Transformer network. By inputting the current epoch observation features and multi-epoch time-series features into the multi-input fusion module, high-dimensional data is mapped to a low-dimensional vector to generate a low-dimensional latent feature representation. Subsequently, the feature encoding module encodes the generated low-dimensional vector and learns its patterns. Then, the feature decoding module classifies and decodes the vector, finally outputting the recognition result of the multipath signal.

[0081] In this embodiment, the current epoch observation features include the elevation angle and azimuth angle of all observable satellites at the current time. The multi-epoch time series features include the elevation angle Ao, azimuth angle Az, and carrier-to-noise ratio C / N0 of each satellite at different times.

[0082] Multipath signal identification falls under the logistic regression binary classification task, and the loss function used is the binary cross-entropy loss L. loss This is used to determine the quality of multipath recognition results, and its calculation formula is as follows:

[0083]

[0084] Among them, y i This represents the binary label of the i-th sample, with a result of either 0 or 1, pr i This indicates that the i-th sample belongs to y. iThe probability of the label; for the above formula, the closer the prediction result of pr is to 1, the closer the value of the loss function is to zero. The binary cross loss reflects the performance of the model to a certain extent.

[0085] In this preferred embodiment, the satellite ephemeris information received by the antenna is first processed into satellite feature information by the pre-signal processing module, and then input into the positioning enhancement module based on deep reinforcement learning and the multipath signal recognition module based on Transformer for training and verification of the positioning enhancement model.

[0086] The training and verification process is as follows: (1) Construct a training set and process the satellite feature information into a format suitable for the input processing of the positioning enhancement model; (2) Set training hyperparameters, such as the learning rate, number of iterations, and number of network layers of the positioning enhancement model; (3) Train the positioning enhancement model and complete the training process when the number of iterations reaches the maximum; (4) Save the trained positioning enhancement model and its parameters, measure the accuracy of its output against the real labels, and complete the verification process after testing on multiple test sets and achieving satisfactory results.

[0087] In this preferred embodiment, the localization enhancement model and multipath signal recognition model are simplified, including pruning and quantization. For example... Figure 4 As shown, the pruning process specifically involves:

[0088] First, the pruning ratio of the localization augmentation model parameters is set, and the parameter set size, runtime, and localization correction accuracy of the localization augmentation model are used as rewards in the reinforcement learning process. Then, the pruning of the localization augmentation model parameters is continuously adjusted and optimized based on the compression degree of the localization augmentation model, ultimately achieving the minimum performance loss of the localization augmentation model on the test set after pruning compared to before pruning.

[0089]

[0090] In the formula, W pruned W represents the final pruned localization augmentation model parameter set; W′ represents the pruned parameter set; L is the loss function used to measure the predicted value f of the localization augmentation model. W′ (X) is the difference between the dataset label Y and the input data label; X represents the input data label.

[0091] In a preferred embodiment, the quantization process refers to quantizing the weight parameters of the positioning enhancement model from the original 32-bit float data to 8-bit int data or 16-bit unsigned short / float data; that is:

[0092]

[0093] In the formula, bits represents the number of bits to be quantized, and V x V represents the floating-point input of the original parameters. q This represents the quantized value. Zero_Point represents the zero point after quantization. When Zero_Point is 0, it indicates symmetric quantization.

[0094] In this embodiment, the dequantization process is as follows:

[0095]

[0096] In the formula, V′ x V represents the dequantized value; q This represents the quantized value; bits represents the number of bits to be quantized; V x This represents the floating-point input of the original parameters.

[0097] The output can be converted to a 32-bit output through the above dequantization operation. This embodiment uses different quantization standards according to different MCU performance. Specifically, it is divided into int8 weighted quantization, int16 weighted quantization, and int8 weighted int16 activated int64 biased quantization.

[0098] In this preferred embodiment, TensorFlow Lite for Micro is used as the support library for deploying the localization enhancement model on the MCU chip. The localization enhancement model is converted into the TensorFlow framework using basic operators. For example, for the localization enhancement module based on deep reinforcement learning, this embodiment uses a fully connected layer of basic operators and a long short-term memory structure state estimator to implement the original model's functionality during inference. For the multipath signal recognition module based on Transformer, this embodiment uses a fully connected layer of basic operators to implement the Transformer module's functionality during inference, thereby eliminating network layers that are not used during inference, reducing the model size, and minimizing the accuracy loss caused by quantization compression.

[0099] In this embodiment, as Figure 6 As shown, the conversion process of the positioning enhancement model is as follows:

[0100] 1) First, the parameter weights of the original localization enhancement model are migrated to the new model built with TensorFlow basic operators, realizing the model from .pth to .keras;

[0101] 2) Convert the .keras format to .tflite format;

[0102] 3) In the MCU, call the TFLM library functions to build a program that can load and run the model. This part of the program mainly includes two parts: the model loading function model_init() and the model running function model_run().

[0103] Furthermore, this embodiment employs a Real-Time Operating System (RTOS). RTOS has a small kernel and is an operating system that guarantees the completion of specific functions within a certain time limit, making it particularly suitable for fields with high real-time requirements, such as positioning correction. Simultaneously, to address the issue of delayed positioning results output caused by the long processing time of artificial intelligence algorithms due to MCU performance limitations, this embodiment, based on the characteristics of a dual-core MCU, splits the positioning enhancement module into two parts: input processing and output correction. This enables parallel computation between the multipath signal recognition module and the positioning enhancement module, significantly reducing model runtime. In this embodiment, after receiving satellite feature observation information data from the pre-signal processing module, the MCU chip first runs the data processing module in CPU1. This module collects and integrates the received satellite feature observation information data, processing it into the input format required by other modules. Then, it sends the data to the Kalman filter and the input processing part of the positioning enhancement module in CPU1, as well as the Transformer-based multipath signal recognition module in CPU2, respectively, through the message queue in the RTOS and the inter-core communication mechanism based on the message buffer. Upon receiving a message, the input processing sections of the Transformer-based multipath signal identification module and positioning enhancement module read data from the queue and run simultaneously. The multipath signal identification module identifies signals affected by multipath interference and sends the results to the Kalman filter. The input processing section of the positioning enhancement module obtains historical position coordinates from the Kalman filter and sends the calculated confidence state information to the output correction section of the positioning enhancement module in CPU2. After receiving satellite feature data from the data processing module and the identification results from the multipath signal identification module in CPU2, the Kalman filter in CPU1 first removes data affected by multipath interference from the satellite feature data, leaving only pure direct satellite feature data, and immediately performs position calculation. Simultaneously, the output correction section of the positioning enhancement module in CPU2 immediately calculates the position correction amount upon receiving the confidence state information. Finally, the Kalman filter adds the calculated initial coarse position at the current moment to the position correction amount calculated by the output correction section of the positioning enhancement module in CPU2, outputting a corrected high-precision positioning position. Meanwhile, to prevent the computation time on CPU2 from exceeding the maximum acceptable delay for the positioning output, the Kalman filter adopts a delay-waiting strategy at the final output. That is, if the current position correction information is not received within a specified time, the position correction information from the previous moment is used as the correction for the current initial coarse position.

[0104] In the above process, by disassembling the positioning enhancement module, the input processing part and the multipath signal recognition module of the positioning enhancement module are parallelized, as are the Kalman filter and the output correction part of the positioning enhancement module. This improves CPU utilization efficiency, reduces the total computation time, and the delay waiting strategy ensures the timeliness of the positioning module's output. The combined effect of these two factors effectively solves the problem of delayed positioning results output caused by the long processing time of artificial intelligence algorithms due to the limitations of MCU performance. (See also...) Figure 5 As shown.

[0105] The embodiments and descriptions above are merely illustrative of the principles and preferred embodiments of the present invention. Various changes and modifications may be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed.

Claims

1. A high-precision intelligent positioning terminal device for a single BeiDou system, characterized in that, include: A receiving antenna, used to receive ephemeris information signals from satellites; The front-end signal processing module is used to process the satellite ephemeris information signal received by the receiving antenna to obtain satellite characteristic observation information data; The MCU chip integrates a Kalman filter and a positioning enhancement model consisting of a positioning enhancement module based on deep reinforcement learning and a multipath signal recognition module based on Transformer. The MCU chip has a dual-core structure and achieves parallel computing of the multipath signal recognition module and the positioning enhancement module by splitting the positioning enhancement module into two parts: input processing and output correction. The Transformer-based multipath signal recognition module is used to remove satellite data affected by multipath from satellite feature observation information data; and the purified data after removal is passed to the Kalman filter; the deep reinforcement learning-based positioning enhancement module is used to output the correction amount for the current positioning position; the Kalman filter calculates the initial coarse position based on the purified data, and then combines it with the correction amount of the current positioning position output by the positioning enhancement module to output the corrected high-precision positioning position; The Transformer-based multipath signal recognition module includes a Transformer-based multi-input fusion module, a feature encoding module, and a feature decoding module. By inputting the current epoch observation features and multi-epoch time-series features into the multi-input fusion module, high-dimensional data is mapped to a low-dimensional vector to generate a low-dimensional latent feature representation. Subsequently, the feature encoding module encodes the generated low-dimensional vector and learns its patterns. Then, the feature decoding module classifies and decodes the vector, finally outputting the recognition result of the multipath signal. The localization enhancement module based on deep reinforcement learning includes a multi-view environment perception module and a localization correction strategy learning module; The positioning correction strategy learning module is constructed based on a proximal optimization strategy, including a value network and an action network, and learns the strategy value and correction action through the value network and action network. For value networks, a value function update is used, and the loss value at time t is calculated using the mean squared return error: Among them, L c This represents the loss function of the value network, used to update the value network parameters; r i+1 It is the reward at time i+1, V π (H t ) is the confidence state value at the current moment. H represents the average value calculated at time t; t This represents the confidence state from multiple angles at time t; By using each state H t The mean squared return error is used to calculate the loss and update the parameters of the value network; In order to effectively quantify the confidence state H t The impact on the output action is assessed using Generalized Advantage Estimation (GAE) to update the current action network's policy; the calculation formula is as follows: Among them, L A This represents the loss function of the action network. Used to update action network parameters; θ represents the average value calculated at time t. A The r represents the weights of the action network parameters. t (θ A () represents the ratio of the probability of the new strategy to the probability of the old strategy at time t; This represents the dominance error function, and clip represents the clipping operation. This indicates the pruning ratio of the PPO algorithm; The positioning enhancement model, consisting of a positioning enhancement module based on deep reinforcement learning and a multipath signal recognition module based on Transformer, is trained and validated by receiving satellite ephemeris information signals from the receiving antenna. After simplification and transformation, the model is then simplified and transformed. Finally, the TFLM library function is called in the MCU to build an isomorphic model and run the isomorphic model for positioning inference. The simplification process for the aforementioned localization enhancement model includes pruning and quantization. The conversion process of the aforementioned positioning enhancement model is as follows: 1) First, the parameter weights of the original localization enhancement model are migrated to a new model built with TensorFlow basic operators, realizing the model conversion from .pth to .keras; 2) Convert .keras format to .tflite format; 3) In the MCU, call TFLM library functions to build a program that can load and run the model. This part of the program mainly includes two parts: the model loading function model_init() and the model running function model_run(). The pruning process specifically involves: First, the pruning ratio of the localization augmentation model parameters is set, and the parameter set size, runtime, and localization correction accuracy of the localization augmentation model are used as rewards in the reinforcement learning process. Then, the pruning of the localization augmentation model parameters is continuously adjusted and optimized based on the compression degree of the localization augmentation model, ultimately achieving the minimum performance loss of the localization augmentation model on the test set after pruning compared to before pruning. In the formula, W pruned W represents the final pruned localization augmentation model parameter set; W′ represents the pruned parameter set; L is the loss function used to measure the predicted value f of the localization augmentation model. W′ (X) represents the difference between the input data label and the dataset label Y; X represents the input data label. The quantization process refers to quantizing the weight parameters of the localization enhancement model from the original 32-bit float data to 8-bit int or 16-bit unsigned short / float data; that is: In the formula, bits represents the number of bits to be quantized, and V x V represents the floating-point input of the original parameters. q This represents the quantized value. Zero_Point represents the zero point of the quantized value. When Zero_Point is 0, it indicates symmetric quantization. The dequantization process is as follows: In the formula, V′x represents the dequantized value; V q This represents the quantized value; bits represents the number of bits to be quantized; V x Floating-point input representing the original parameters; The above dequantization operation can convert the output to a 32-bit output. Different quantization standards are used according to different MCU performance, specifically divided into int8 weighted quantization, int16 weighted quantization, and int8 weighted int16 activated int64 biased quantization. The method of using the high-precision intelligent positioning terminal equipment for a single BeiDou system specifically includes the following: Employing a Real-Time Operating System (RTOS), and leveraging the characteristics of a dual-core MCU, the positioning enhancement module is divided into input processing and output correction. After receiving satellite feature observation data from the pre-processing signal calculation module, the MCU chip first runs the data processing module in CPU1. This module collects and integrates the received satellite feature observation data, processing it into the input format required by other modules. Then, it sends the data to the Kalman filter in CPU1 and the input processing section of the positioning enhancement module, as well as the Transformer-based multipath signal identification module in CPU2, via message queues and message buffer-based inter-core communication mechanisms within the ROS. Upon receiving the messages, the multipath signal identification module and the input processing section of the positioning enhancement module read data from the queues and run simultaneously. The multipath signal identification module identifies signals affected by multipath interference and sends the results to the Kalman filter. The input processing section of the positioning enhancement module obtains historical position coordinates from the Kalman filter and sends the calculated confidence state information to the output correction section of the positioning enhancement module in CPU2. The Kalman filter in CPU1, upon receiving satellite feature data from the data processing module and data from the CP... After the multipath signal identification module of U2 identifies the results, it first removes the data affected by multipath from the satellite feature data, leaving the pure direct satellite feature data, and immediately performs position calculation. At the same time, the output correction part of the positioning enhancement module of CPU2 immediately calculates the position correction amount after receiving the confidence state information. Finally, the Kalman filter adds the calculated initial coarse position at the current time to the position correction amount calculated by the output correction part of the positioning enhancement module of CPU2, and outputs the corrected high-precision positioning position. At the same time, in order to prevent the calculation time on CPU2 from exceeding the maximum acceptable delay of the positioning output, the Kalman filter adopts a delay waiting strategy when outputting. That is, if the position correction information at the current time is not received within a specified time, the position correction information at the previous time is used as the correction for the current initial coarse position.

2. The high-precision intelligent positioning terminal device for a single BeiDou system according to claim 1, characterized in that: The multi-view environment perception module includes a multilayer sensing network and a long short-term memory network (LSTM). The measurement information of the current environment observation and the state information of the historical behavior are respectively input into the multilayer sensing network and the LSTM for processing and then cascaded, and input into the positioning correction strategy learning module to learn the positioning correction strategy.

3. A high-precision intelligent positioning terminal device for a single BeiDou system according to claim 2, characterized in that: The measurement information of the current environment observation is a characteristic observation vector composed of the satellite pseudorange residual RES, line-of-sight vector LOS, elevation angle Ao and carrier-to-noise ratio C / N0 at the current moment; The state information of the historical behavior includes the historical behavior information of the historical trajectory of the last N time steps.

4. A high-precision intelligent positioning terminal device for a single BeiDou system according to claim 3, characterized in that: The Long Short-Term Memory (LSTM) network includes an input gate, a forget gate, hidden units, and an output gate. Inputting observation information into the LSTM network allows for the accurate estimation of the current confidence state by integrating the hidden states from past time points with the feature information from the current observation. t The calculation expression is as follows: In the formula, W represents the sigmoid activation function; i W f W c W o Let b represent the weight matrices of the LSTM input gate, forget gate, hidden unit, and output gate, respectively. i b f b c b c These represent the bias terms for the LSTM input gate, forget gate, hidden unit, and output gate, respectively. t This represents the input information at time t. f represents the vector concatenation operation; t i t o t These represent the outputs of the forget gate, input gate, and output gate, respectively. C t and h t These represent cell state and confidence state, respectively. Based on the above calculation expression, the confidence state h of the environmental observation at the current moment is finally obtained. t ; The goal of the localization augmentation module based on deep reinforcement learning at each time step is to find the optimal correction action to perform. This process is based on a sliding window, and the localization enhancement module based on deep reinforcement learning obtains position correction information each time. The sliding window of the multi-view environment perception module moves forward one step to form a state information of historical behavior with a constant size of N; then, this state information is input into the multilayer perception network to calculate the confidence state based on the behavior information. The calculation expression is as follows: Among them, w i Let x represent the connection weight of the i-th neuron, b represent the bias term, and x represent the connection weight of the i-th neuron. i It represents the output of the neurons in the previous layer; n represents the total number of neurons. This represents the sigmoid activation function; Finally, the confidence state of the environmental observations at the current moment is determined. t and confidence state based on behavioral information The splicing yields confidence states H from multiple angles. t And it serves as the input to the localization correction strategy learning module, that is:

Images