A beidou navigation chip positioning error intelligent compensation method based on deep learning

Abstract

The application discloses a Beidou navigation chip positioning error intelligent compensation method based on deep learning, which comprises the following steps: acquiring original data sets output by a Beidou navigation chip, organizing the original data sets into continuous time input streams and obtaining reference trajectory data; inputting the continuous time input streams and the reference trajectory data into a continuous time error modeling network to generate a continuous time error vector field; constructing a segmented CfC error compensation model, dividing a closed time substructure; introducing a time constant modulation layer to obtain a segmented CfC error compensation model with a time constant modulation structure; constructing an end-cloud continuous time distillation structure, performing dynamic alignment constraint training, and obtaining a segmented CfC error compensation model trained by end-cloud cooperation; and in a real-time working process, performing error compensation based on real-time continuous time input streams to generate a compensation positioning result. The application improves the positioning precision and environmental adaptability of the Beidou navigation chip.

Description

Technical Field

[0001] This invention relates to the field of satellite navigation and positioning error calibration technology, and in particular to a method for intelligent compensation of positioning errors in BeiDou navigation chips based on deep learning. Background Technology

[0002] Satellite navigation and positioning technology is widely used in transportation, smart terminals, emergency rescue, and high-precision measurement. As the core positioning unit on the terminal side, the BeiDou navigation chip's positioning accuracy is directly affected by ionospheric delay, tropospheric delay, multipath interference, satellite ephemeris errors, satellite clock bias, receiver noise, antenna obstruction, and complex environmental changes. In actual operation, due to the different error compositions and evolution characteristics in different scenarios, traditional error calibration methods based on fixed models or static parameters are difficult to accurately characterize the dynamic laws of error changes with time, environment, and link status.

[0003] Existing error compensation techniques mainly rely on physical error models, differential positioning methods, or linear filtering methods. Physical error models often struggle to effectively describe nonlinear errors in complex scenarios. Differential positioning methods depend on a stable reference station link environment, and their compensation effectiveness decreases when link quality fluctuates. Linear filtering methods lack adaptability to error non-stationarity and cross-scenario variations, making it difficult to achieve refined capture of the error evolution process. Furthermore, existing methods generally lack the ability to model the correlation between the terminal-side operating environment state and positioning errors, and also lack mechanisms for collaborative error knowledge updates between the terminal and platform sides, resulting in insufficient real-time performance and generalization ability of error compensation.

[0004] Meanwhile, existing technologies cannot accurately characterize continuous time errors based on dynamic changes in different scenarios, link states, and error sources, nor can they construct an error compensation mechanism that can adapt to the environment. This makes it difficult to meet the high-precision positioning requirements of Beidou navigation chips in multiple scenarios and states.

[0005] Therefore, how to provide a deep learning-based intelligent compensation method for positioning errors in BeiDou navigation chips is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an intelligent compensation method for positioning errors of BeiDou navigation chips based on deep learning. This invention utilizes continuous-time error vector field modeling, a piecewise CfC time-series structure, and an adaptive time constant modulation method to continuously characterize and finely compensate for positioning errors of BeiDou navigation chips operating in multiple scenarios. This invention constructs an edge-cloud continuous-time distillation structure, enabling collaborative updates between platform-side error knowledge and terminal-side compensation models, thereby improving the model's adaptability to environmental and link state changes. This invention possesses advantages such as high error modeling accuracy, strong cross-scenario generalization ability, and excellent real-time compensation capability, which can improve the positioning accuracy and stability of BeiDou navigation chips in complex environments.

[0007] A method for intelligent compensation of positioning errors in BeiDou navigation chips based on deep learning, according to an embodiment of the present invention, includes the following steps:

[0008] The system acquires raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data output by the BeiDou navigation chip, organizes them into a continuous time input stream, and references trajectory data.

[0009] The continuous-time input stream and reference trajectory data are input into the continuous-time error modeling network, and the continuous-time error vector field is obtained by training based on the position deviation and communication and navigation monitoring status data.

[0010] A piecewise CfC error compensation model is constructed, which divides the error evolution relationship corresponding to the continuous-time error vector field into multiple closed-time substructures.

[0011] A time constant modulation layer is introduced into the piecewise CfC error compensation model to map the conduction monitoring state data into a time scale adjustment factor, thus obtaining a piecewise CfC error compensation model with a time constant modulation structure.

[0012] A continuous-time distillation structure is constructed between the edge and cloud. Based on the continuous-time error vector field, a piecewise CfC error compensation model with a time constant modulation structure, and a continuous-time input stream, dynamic alignment constraint training is performed to obtain a piecewise CfC error compensation model trained by the edge and cloud.

[0013] During real-time operation, the real-time collected data is constructed into a real-time continuous time input stream, which is then input into a segmented CfC error compensation model trained by edge-cloud collaboration. This generates the error compensation amount for the corresponding time point, and error compensation is performed to obtain the compensated positioning result.

[0014] Optionally, the generation of the continuous-time input stream and reference trajectory data includes:

[0015] The raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data output by Beidou navigation chips deployed in different application scenarios are collected with time stamps to form a raw data set;

[0016] The communication and navigation monitoring status data includes communication link quality indicators, monitoring resolution level, and scene labels;

[0017] Time alignment, interpolation and resampling are performed on the original dataset to remove invalid records corresponding to missing time periods, resulting in the original navigation observation data sequence, terminal operation status data sequence and communication and navigation monitoring status data sequence.

[0018] Based on the original navigation observation data sequence, terminal operation status data sequence, and communication and navigation monitoring status data sequence, the data content at each time point is concatenated according to a fixed field order to generate a continuous time input stream;

[0019] Based on the monitoring intensity level and scene label in the communication and navigation monitoring status data sequence, a scene label is generated for each time point. Adjacent time points representing the same monitoring intensity level and the same scene label are combined into a scene time period and encoded into a scene label sequence.

[0020] Based on the positioning reference data collected by the communication and navigation monitoring reference equipment, the corresponding reference position is calculated at each time point, and the reference positions of all time points are stitched together to generate reference trajectory data.

[0021] Optionally, the generation of the continuous-time error vector field includes:

[0022] On the communication and monitoring platform side, the position deviation is calculated based on the positioning results in the continuous time input stream and the reference position data in the reference trajectory data to obtain the positioning error sequence;

[0023] Extract input features that correspond one-to-one with the positioning error sequence from the continuous-time input stream, and concatenate the original data set of each time point into an error modeling feature sequence, which together with the positioning error sequence forms continuous-time error modeling training data.

[0024] A continuous-time error modeling network structure is constructed based on the continuous-time error modeling training data, and the initial continuous-time error modeling network is obtained by initializing all dynamic parameters.

[0025] The training data for continuous time error modeling is input into the initial continuous time error modeling network to obtain the positioning error estimation results at each time point. These results are compared with the positioning error sequence to calculate the position deviation residuals. The training loss value is then aggregated to obtain the training loss value. Based on the training loss value, the dynamic parameters of the continuous time error modeling network are iteratively updated to obtain the converged continuous time error modeling network.

[0026] The converged continuous-time error modeling network is deployed on the communication and monitoring platform. The continuous-time input stream is input into the converged continuous-time error modeling network to obtain the positioning error estimation results. The positioning error estimation results at each time point are arranged in chronological order to form a continuous-time error vector field.

[0027] Optionally, the continuous-time error modeling network structure includes:

[0028] The input unit includes a time series input interface and a preprocessing layer, which converts the error modeling feature vector into an internal feature representation sequence.

[0029] The multi-level continuous-time coding unit receives an internal feature representation sequence, performs state update operations, and generates a hidden state sequence.

[0030] The output unit receives the hidden state sequence and converts it into localization error estimation results at each time point through a fully connected mapping layer.

[0031] Optionally, the construction of the piecewise CfC error compensation model includes:

[0032] On the terminal side where the Beidou navigation chip is located, the continuous time error vector field and scene label sequence are received, and the scene labels and error evolution relationship are associated with a time index to obtain the time-aligned scene label sequence and error evolution relationship sequence.

[0033] Based on the time-aligned scene label sequence, adjacent time points with the same scene label are combined into scene time periods and assigned a unique substructure index to form a set of scene time periods.

[0034] Based on the set of scene time periods and the error evolution relationship sequence, multiple closed time substructures are constructed on the terminal side. A shared network layer is set to share the network structure and parameters. A specific set of scene parameters is set to fit the error evolution relationship corresponding to the scene label in the continuous time error vector field. Each closed time substructure is bound to the corresponding scene time period one by one to form a segmented CfC basic structure with multiple closed time substructures.

[0035] A scene discrimination function is constructed based on the scene label sequence. The scene label corresponding to the current time point and the adjacent scene labels are used as inputs. The output is the weight coefficient of the current time point for each closed time substructure. The weight coefficients are smoothly transitioned near the boundary of the scene time period according to the preset continuous change rule.

[0036] On the terminal side, the continuous time error vector field is input into the piecewise CfC basic structure. At each time point, all closed time substructures output the intermediate error compensation result for that time point. The weight coefficients and the intermediate error compensation results are matched according to the time points, and a weighted sum is performed to obtain the error compensation output for that time point. The output sequence of the piecewise CfC error compensation model is generated by connecting them, realizing continuous differentiable switching between multiple closed time substructures.

[0037] Optionally, the construction of the piecewise CfC error compensation model with time constant modulation structure includes:

[0038] On the terminal side, based on the communication and navigation monitoring status data in the continuous time input stream, the communication link quality index sequence and the monitoring degree level sequence are extracted and bound to the CfC substructure corresponding to each time point in the segmented CfC error compensation model to generate the time constant input sequence.

[0039] A time constant modulation layer is constructed inside the segmented CfC error compensation model. The communication link quality index and the monitoring resolution level are received and combined and mapped into a time scale adjustment factor. At the time boundary of different scenarios, a smooth transition is performed according to a preset continuous change rule to form a time scale adjustment factor sequence.

[0040] The time scale adjustment factor sequence is written into the state update unit of each CfC substructure in the piecewise CfC error compensation model. When the hidden state update is performed, the time scale adjustment factor at the corresponding time point is read to dynamically adjust the time constant inside the CfC substructure.

[0041] After completing the writing of the time scale adjustment factor and the dynamic adjustment of the time constant of the state update unit, the segmented CfC error compensation model with integrated time constant modulation layer is used as the segmented CfC error compensation model with time constant modulation structure.

[0042] Optionally, the construction of the segmented CfC error compensation model trained collaboratively by the edge and cloud includes:

[0043] The continuous-time error vector field is used as the teacher model, and the piecewise CfC error compensation model with time constant modulation structure is used as the student model.

[0044] On the communication and monitoring platform side, the continuous-time input stream is input into the continuous-time error modeling network, and the error increment corresponding to each time point in the continuous-time error vector field is read to form the teacher model output sequence.

[0045] Meanwhile, on the terminal side, the continuous time input stream is input into the segmented CfC error compensation model with a time constant modulation structure, and the error compensation amount corresponding to each time point is read to form the student model output sequence.

[0046] The output sequences of the teacher model and the student model are input into the end-cloud continuous-time distillation structure to construct an end-cloud continuous-time distillation loss function to measure the difference between the teacher error increment and the student error compensation. The end-cloud continuous-time distillation loss function generates time point residuals based on the difference between the error increment and the corresponding error compensation at each time point. The residuals are aggregated to generate a complete time dimension residual sequence, which is used as the end-cloud continuous-time distillation loss.

[0047] Using the continuous-time distillation loss of the end cloud as the training target, joint parameter updates are performed on all CfC substructure parameter sets inside the piecewise CfC error compensation model with time constant modulation structure and all time scale adjustment factors in the time constant modulation layer. This keeps the continuous-time error vector field corresponding to the teacher model from being updated, so that the student model can gradually learn the error evolution relationship represented by the teacher model.

[0048] The continuous-time input stream is continuously fed into the teacher model and the student model, and the parameters are iteratively optimized until the continuous-time distillation loss of the end-cloud meets the preset convergence condition. At this point, the student model is frozen as a piecewise CfC error compensation model trained by the end-cloud collaboration.

[0049] Optionally, the generation of the compensated positioning result includes:

[0050] During the real-time operation of the Beidou navigation chip, raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data are collected in real time to generate a real-time data set with real-time time stamps.

[0051] Perform time alignment processing on the real-time dataset to generate a real-time continuous time input stream, and simultaneously generate a real-time scene label sequence;

[0052] The real-time continuous time input stream and the real-time scene label sequence are input into the segmented CfC error compensation model trained by the edge-cloud collaboration. The CfC substructure at the corresponding time point is selected, the time scale adjustment factor at that time point is generated, the CfC state update is performed after reading the time scale adjustment factor, the real-time error compensation amount is output, and the real-time error compensation sequence is obtained by arranging them.

[0053] The real-time error compensation sequence is matched with the positioning results output in real time by the Beidou navigation chip on a unified time axis. The real-time positioning result at each time point is added with the real-time error compensation amount at the corresponding time point to generate a compensated positioning result.

[0054] The beneficial effects of this invention are:

[0055] First, by constructing a continuous-time error vector field, this invention enables the error variation law of Beidou navigation chips under different operating scenarios to be accurately characterized in a continuous-time form. Compared with traditional error calibration methods that rely on static models, this invention significantly improves the responsiveness of error modeling to complex environmental changes.

[0056] Secondly, this invention introduces a segmented CfC structure on the terminal side and combines it with a time constant modulation mechanism, enabling the model to adaptively adjust the time scale based on the communication link quality index and the monitoring resolution level, thereby maintaining stable compensation performance in both scenarios of rapid error fluctuation and slow error drift.

[0057] Furthermore, this invention constructs an edge-cloud continuous-time distillation structure to realize the synchronous transmission of platform-side error knowledge to the terminal-side compensation model, enabling the error compensation capability to be continuously updated with the operating environment and improving cross-scenario generalization capability.

[0058] Furthermore, this invention forms a feedback data closed loop during real-time operation, enabling the continuous-time error vector field and the collaboratively trained piecewise CfC error compensation model to be automatically corrected based on the latest observations, effectively improving the positioning accuracy and robustness of the BeiDou navigation chip in dynamic environments. Attached Figure Description

[0059] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0060] Figure 1 This is an overall flowchart of a deep learning-based intelligent compensation method for positioning errors in BeiDou navigation chips proposed in this invention.

[0061] Figure 2 This is a schematic diagram of the continuous-time error vector field modeling structure in this invention;

[0062] Figure 3 This is a schematic diagram of the structure in which the segmented CfC error compensation model and the time constant modulation layer work together in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1-3 A method for intelligent compensation of positioning errors in BeiDou navigation chips based on deep learning includes the following steps:

[0065] The system acquires raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data output by BeiDou navigation chips deployed in different application scenarios. The raw navigation observation data includes positioning results, pseudorange observation data, carrier phase observation data, and signal-to-noise ratio data output by the navigation chip. The terminal operation status data includes velocity data and attitude data. The communication and navigation monitoring status data includes communication link quality indicators, monitoring resolution level, and scene labels. Based on a unified time axis, the raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data are organized into a continuous time input stream. A scene label sequence is constructed based on the monitoring resolution level and scene labels in the communication and navigation monitoring status data. Reference trajectory data is generated based on the positioning reference data collected by communication and navigation monitoring reference equipment, including ground reference stations, maritime monitoring platforms, and airborne monitoring equipment.

[0066] On the communication and navigation monitoring platform side, the continuous time input stream and reference trajectory data are input into the continuous time error modeling network. Based on the position deviation between the positioning result and the reference trajectory data at each time point and the communication and navigation monitoring status data in the continuous time input stream, the continuous time error vector field is trained to obtain the continuous time error vector field. The continuous time error vector field is used to characterize the dynamic law of the positioning error changing with time under different scenarios.

[0067] On the terminal side where the Beidou navigation chip is located, a segmented CfC error compensation model is constructed. The error evolution relationship corresponding to the continuous time error vector field is divided into multiple closed time substructures according to the scene label sequence. In each closed time substructure, a shared network layer and a specific scene parameter set are set. A scene discrimination function is constructed based on the scene label sequence. The scene discrimination function continuously weights the output of each closed time substructure at the boundary of adjacent scenes, so as to realize continuous differentiable switching between multiple closed time substructures.

[0068] A time constant modulation layer is introduced into the segmented CfC error compensation model. The communication link quality index and monitoring intensity level in the communication and navigation monitoring status data are mapped to a time scale adjustment factor. The time scale adjustment factor acts on the state update unit in each closed time substructure to dynamically adjust the time constant of the state update unit. This improves the response speed of each closed time substructure to short-term error changes when the communication link quality index is lower than the preset threshold, and enhances the memory ability of long-term error trends when the monitoring intensity level is in the preset high level state. This results in a segmented CfC error compensation model with a time constant modulation structure.

[0069] A continuous-time distillation structure is constructed between the communication and monitoring platform and the terminal. The continuous-time error vector field is used as the teacher model, and the piecewise CfC error compensation model with time constant modulation structure is used as the student model. Based on the continuous-time input stream, dynamic alignment constraint training is performed on the error increment output by the continuous-time error vector field and the error compensation output by the student model in the continuous-time distillation structure of the end cloud. The parameter set in the closed time substructure and the time scale adjustment factor in the time constant modulation layer are updated to obtain the piecewise CfC error compensation model trained by the end cloud.

[0070] During the real-time operation of the BeiDou navigation chip, the raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data collected in real time are constructed into a real-time continuous time input stream and a real-time scene label sequence according to a unified time axis. The real-time continuous time input stream and the real-time scene label sequence are input into a segmented CfC error compensation model trained by the end-cloud collaboration to generate the error compensation amount at the corresponding time point. Error compensation is performed on the positioning results output by the BeiDou navigation chip in real time to obtain the compensated positioning result. The real-time continuous time input stream, the error compensation amount, and the compensated positioning result constitute feedback data and are uploaded to the communication and navigation monitoring platform to update the continuous time error vector field and the end-cloud continuous time distillation structure.

[0071] In this embodiment, the generation of the continuous-time input stream and reference trajectory data includes:

[0072] The original navigation observation data, terminal operation status data, and communication and navigation monitoring status data output by Beidou navigation chips deployed in different application scenarios are collected with time stamps. This includes recording the positioning results, pseudorange observation data, carrier phase observation data, and signal-to-noise ratio data output by the navigation chip according to the sampling time, which are recorded as the original navigation observation data; recording the velocity data and attitude data in the terminal operation status data according to the sampling time; and recording the communication link quality indicators, monitoring resolution level, and scene label in the communication and navigation monitoring status data according to the sampling time, forming a set of original data with unified time stamps.

[0073] Time alignment processing is performed on the original dataset with a unified time stamp. Based on the unified time stamp, the original navigation observation data, terminal operation status data and communication and navigation monitoring status data are interpolated and resampled in chronological order. Invalid records corresponding to missing time periods are removed to obtain the original navigation observation data sequence, terminal operation status data sequence and communication and navigation monitoring status data sequence arranged according to a unified time axis.

[0074] Based on the original navigation observation data sequence, terminal operation status data sequence, and communication and navigation monitoring status data sequence arranged according to a unified time axis, the positioning results, pseudorange observation data, carrier phase observation data, signal-to-noise ratio data, velocity data, attitude data, communication link quality indicators, monitoring resolution level, and scene labels at each time point are concatenated into a time-level data vector according to a fixed field order, and the time-level data vector is connected into a continuous time input stream in chronological order.

[0075] Based on the monitoring intensity level and scene label in the communication and navigation monitoring status data sequence, a scene label is generated for each time point. Adjacent time points representing the same monitoring intensity level and the same scene label are combined into a scene time period. The corresponding scene time period is encoded into a scene label sequence and written into the continuous time input stream.

[0076] Based on the positioning reference data collected by the communication and navigation monitoring reference equipment, the positioning reference data output by the ground reference station, the marine monitoring platform and the air surveillance equipment are time-aligned and interpolated according to a unified time mark. At each time point, the corresponding reference position is calculated using the aligned positioning reference data. The reference positions of all time points are connected in chronological order to generate reference trajectory data. The reference trajectory data is then associated with the continuous time input stream and the scene mark sequence by time indexing, so that it can be used by the continuous time error modeling network and the piecewise CfC error compensation model.

[0077] In this embodiment, the generation of the continuous-time error vector field includes:

[0078] On the communication and monitoring platform side, the continuous time input stream and reference trajectory data are paired according to a unified time axis. The positioning result output by the navigation chip is read for each time point in the continuous time input stream, and the reference position data is read for the corresponding time point in the reference trajectory data. The position deviation between the positioning result and the reference position data at each time point is calculated to obtain a positioning error sequence arranged in chronological order. The positioning error sequence is used as the supervision label data of the continuous time error modeling network.

[0079] Extract input features that correspond one-to-one with the positioning error sequence from the continuous time input stream. Concatenate the original navigation observation data, terminal operation status data and communication and navigation monitoring status data at each time point into an error modeling feature vector according to a fixed field order. Connect the error modeling feature vectors at each time point in chronological order to form an error modeling feature sequence. Combine the error modeling feature sequence with the positioning error sequence to form continuous time error modeling training data.

[0080] A continuous-time error modeling network structure is constructed based on the continuous-time error modeling training data, and the initial continuous-time error modeling network is obtained by initializing all dynamic parameters of the continuous-time error modeling network.

[0081] The continuous-time error modeling training data is input into the initial continuous-time error modeling network. In each round of training, the error modeling feature sequence is input into the input unit, and the positioning error estimation results at each time point are obtained in the output unit. The positioning error estimation results are compared with the positioning error sequence, and the position deviation residual is calculated. The position deviation residual is aggregated in the time dimension to obtain the training loss value. The dynamic parameters of the continuous-time error modeling network are iteratively updated based on the training loss value until the training loss value meets the preset convergence condition, and the converged continuous-time error modeling network is obtained.

[0082] The converged continuous-time error modeling network is deployed and run on the communication and navigation monitoring platform. The continuous-time input stream is input into the converged continuous-time error modeling network on the communication and navigation monitoring platform. The positioning error estimation results corresponding one-to-one with the continuous-time input stream are obtained in the output unit. The positioning error estimation results at each time point are arranged in chronological order to form a continuous-time error vector field. The continuous-time error vector field is associated with the scene label sequence by time indexing. This is used to characterize the error evolution relationship of positioning error under different scenarios in the piecewise CfC error compensation model.

[0083] In this embodiment, the continuous-time error modeling network structure includes:

[0084] An input unit is set at the input end of the continuous-time error modeling network. The input unit includes a time series input interface for receiving error modeling feature vectors at each time point in the error modeling feature sequence and a preprocessing layer for performing normalization and dimension mapping processing on the error modeling feature vectors. The preprocessing layer converts the error modeling feature vectors into an internal feature representation sequence.

[0085] In the intermediate layer of the continuous-time error modeling network, multiple levels of continuous-time coding units are set in series. Each level of continuous-time coding unit receives the feature representations of adjacent time points in the internal feature representation sequence, performs state update operation based on time order, and combines the hidden state of the previous time point with the feature representation of the current time point to generate a new hidden state. Residual connection units and normalization units are set between the multiple levels of continuous-time coding units to ensure that the hidden state is stably propagated in the time and depth dimensions. The error evolution relationship of the continuous-time error vector field is represented in the hidden state sequence.

[0086] An output unit is set at the output end of the continuous-time error modeling network. The output unit receives the hidden state sequence output by the multi-level continuous-time coding unit, converts the hidden state sequence into the positioning error estimation result at each time point through a fully connected mapping layer, and outputs the positioning error estimation result as the error component in the continuous-time error vector field, establishing a correspondence with the continuous-time input stream on the time axis.

[0087] In this embodiment, the construction of the piecewise CfC error compensation model includes:

[0088] On the terminal side where the Beidou navigation chip is located, the continuous time error vector field and the scene label sequence are received. The scene label corresponding to each time point in the scene label sequence is associated with the error evolution relationship corresponding to each time point in the continuous time error vector field by establishing a time index association, so as to obtain the scene label sequence and the error evolution relationship sequence aligned by time.

[0089] Based on the time-aligned scene label sequence, adjacent time points with the same scene label are combined into scene time periods. A unique substructure index is assigned to each scene time period to form a set of scene time periods containing multiple scene time periods. Each scene time period in the set of scene time periods is bound to the corresponding substructure index.

[0090] Based on the set of scene time periods and the sequence of error evolution, multiple closed-time substructures are constructed on the terminal side. In each closed-time substructure, a shared network layer and a specific scene parameter set are set. The shared network layer is used to share the network structure and shared parameters among all closed-time substructures. The specific scene parameter set is used to fit the error evolution relationship with the scene label in the continuous time error vector field within the corresponding scene time period. Each closed-time substructure is bound to the corresponding scene time period in the set of scene time periods to form a segmented CfC basic structure with multiple closed-time substructures.

[0091] A scene discrimination function is constructed based on the scene label sequence. The scene discrimination function takes the scene label corresponding to the current time point and the scene label adjacent to the current time point as input, and outputs the weight coefficient of the current time point for each closed time substructure. The weight coefficient remains stable within the scene time period. Near the boundary of the scene time period, it smoothly transitions between the closed time substructures corresponding to adjacent scene time periods according to the preset continuous change rule, so that the weight coefficient corresponding to each time point changes continuously in the time dimension.

[0092] On the terminal side, the continuous-time error vector field is input into a segmented CfC basic structure with multiple closed-time substructures. At each time point, all closed-time substructures output the intermediate error compensation results for that time point. The weight coefficients output by the scene discrimination function are matched with the intermediate error compensation results according to the time points. The error compensation output for that time point is obtained by performing a weighted summation on the intermediate error compensation results of each closed-time substructure at that time point. The error compensation outputs of all time points are connected in chronological order to generate the output sequence of the segmented CfC error compensation model, realizing continuous differentiable switching between multiple closed-time substructures and providing a scene-specific error compensation basis for time constant modulation layer injection and end-to-cloud continuous-time distillation structure.

[0093] In this embodiment, the construction of the piecewise CfC error compensation model with a time constant modulation structure includes:

[0094] On the terminal side, based on the communication and navigation monitoring status data in the continuous time input stream, the communication link quality index sequence and the monitoring intensity level sequence are extracted according to a unified time axis. The communication link quality index sequence and the monitoring intensity level sequence are bound to the CfC substructure corresponding to each time point in the segmented CfC error compensation model according to the time index, and a time constant input sequence consistent with the time span of each CfC substructure is generated.

[0095] A time constant modulation layer is constructed inside the segmented CfC error compensation model. The time constant modulation layer receives the communication link quality index and monitoring resolution level at each time point in the time constant input sequence. The communication link quality index and monitoring resolution level are combined and mapped into a time scale adjustment factor through the time constant mapping unit. The time scale adjustment factor keeps changing continuously within the same scene time period and smoothly transitions at the boundaries of different scene time periods according to a preset continuous change rule, forming a time scale adjustment factor sequence corresponding to each time point.

[0096] The time scale adjustment factor sequence is written into the state update unit of each CfC substructure in the segmented CfC error compensation model. When the state update unit performs hidden state update, it reads the time scale adjustment factor at the corresponding time point and dynamically adjusts the time constant inside the CfC substructure. The time constant is shortened during the time period when the communication link quality index is lower than the preset threshold to improve the response speed to short-term error changes. The time constant is extended during the time period when the monitoring intensity is at the preset high level to enhance the memory ability of long-term error evolution relationship. This enables each CfC substructure to have time scale adaptive characteristics that match the communication and navigation monitoring status data in different scenario time periods.

[0097] After completing the writing of the time scale adjustment factor and the dynamic adjustment of the time constant of the state update unit, the piecewise CfC error compensation model with integrated time constant modulation layer is output as a piecewise CfC error compensation model with time constant modulation structure, so that the end-cloud continuous time distillation structure can jointly optimize the time scale adjustment factor and CfC substructure parameters during the training process.

[0098] In this embodiment, the construction of the segmented CfC error compensation model trained collaboratively by the edge and cloud includes:

[0099] The continuous-time error vector field is used as the teacher model, and the piecewise CfC error compensation model with time constant modulation structure is used as the student model.

[0100] The continuous-time input stream and scene label sequence are received on the communication and monitoring platform side. The continuous-time input stream is input into the continuous-time error modeling network, and the error increment corresponding to each time point in the continuous-time error vector field is read to form the teacher model output sequence.

[0101] Simultaneously, the terminal receives a continuous-time input stream and a scene marker sequence, inputs the continuous-time input stream into a segmented CfC error compensation model with a time constant modulation structure, reads the error compensation amount corresponding to each time point, and forms the student model output sequence.

[0102] The output sequences of the teacher model and the output sequences of the student model are time aligned according to a unified time axis, so that each time point of the two sequences forms a one-to-one pairing relationship between the error increment and the error compensation amount.

[0103] The time-aligned output sequences of the teacher model and the student model are input into the edge cloud continuous time distillation structure. Inside the edge cloud continuous time distillation structure, an edge cloud continuous time distillation loss function is constructed to measure the difference between the teacher error increment and the student error compensation. The edge cloud continuous time distillation loss function generates time point residuals based on the difference between the error increment and the corresponding error compensation at each time point. The time point residuals are aggregated in time order to generate a complete time dimension residual sequence, and the aggregated value of the time dimension residual sequence is used as the edge cloud continuous time distillation loss.

[0104] In the continuous-time distillation structure of the end cloud, the continuous-time distillation loss of the end cloud is used as the training target. Joint parameter updates are performed on all CfC substructure parameter sets inside the piecewise CfC error compensation model with time constant modulation structure and all time scale adjustment factors in the time constant modulation layer. During the joint parameter update process, the continuous-time error vector field corresponding to the teacher model is kept unupdated, so that the student model gradually learns the error evolution relationship represented by the teacher model in the time scale adjustment structure and CfC substructure, so that the error compensation amount output by the student model approximates the error increment output by the teacher model in the continuous-time dimension.

[0105] During the continuous-time distillation training process between the edge and cloud, the continuous-time input stream is continuously fed into the teacher model and the student model. The continuous-time distillation loss of the edge is calculated in batches and the parameters are iteratively optimized until the continuous-time distillation loss of the edge meets the preset convergence condition. At this time, the parameter sets of all CfC substructures inside the student model and all time scale adjustment factors in the time constant modulation layer are frozen. The frozen student model is determined as the piecewise CfC error compensation model trained by the edge and cloud. The piecewise CfC error compensation model trained by the edge and cloud is used as a dedicated model for the real-time error compensation stage of the Beidou navigation chip to generate the continuous-time error compensation amount.

[0106] In this embodiment, the generation of the compensation positioning result includes:

[0107] During the real-time operation of the Beidou navigation chip, raw navigation observation data is collected in real time. Real-time positioning results, real-time pseudorange observation data, real-time carrier phase observation data, and real-time signal-to-noise ratio data are recorded according to a unified time stamp. Real-time speed data and real-time attitude data in the terminal operation status data are recorded according to a unified time stamp. Real-time communication link quality indicators, real-time monitoring resolution level, and real-time scene labels in the communication and navigation monitoring status data are recorded according to a unified time stamp, forming a real-time data set with real-time time stamps.

[0108] Time alignment processing is performed on the real-time dataset. Real-time navigation observation data, real-time terminal operation status data, and real-time communication and navigation monitoring status data are resampled according to a unified time axis and concatenated by field into a time-level real-time data vector. All time-level real-time data vectors are connected in chronological order to generate a real-time continuous time input stream. At the same time, a real-time scene label sequence is generated based on the real-time monitoring video resolution level and real-time scene label, so that the real-time continuous time input stream and the real-time scene label sequence are indexed and associated on a unified time axis.

[0109] The real-time continuous time input stream and the real-time scene label sequence are input into the segmented CfC error compensation model trained by edge-cloud collaboration. Inside the segmented CfC error compensation model, the CfC substructure corresponding to the real-time scene label sequence is selected, and the time scale adjustment factor of the time point is generated by the time constant modulation layer. After reading the time scale adjustment factor, the segmented CfC error compensation model trained by edge-cloud collaboration performs CfC state update and outputs the real-time error compensation amount corresponding to each time point in the real-time continuous time input stream. The real-time error compensation amounts output at each time point are arranged in chronological order to obtain the real-time error compensation sequence.

[0110] The real-time error compensation sequence is matched with the positioning results output by the Beidou navigation chip in real time on a unified time axis. The real-time positioning result at each time point is added with the real-time error compensation amount at the corresponding time point to generate the compensated positioning result arranged in time order. The compensated positioning result is then used as the real-time positioning output of the Beidou navigation chip.

[0111] The real-time continuous-time input stream, real-time error compensation sequence, and compensation positioning results are combined in chronological order to form feedback data. This feedback data is then uploaded to the communication and navigation monitoring platform. On the platform side, the continuous-time error vector field is updated with the feedback data to ensure that the error evolution relationship in the continuous-time error vector field is consistent with the real-time working environment. The edge-cloud continuous-time distillation structure is also updated with the feedback data, enabling the edge-cloud collaborative training process to receive the latest scene observation data and error compensation performance in subsequent training rounds. This achieves synchronous updates of the continuous-time error modeling network and the piecewise CfC error compensation model trained by the edge-cloud collaboration during actual operation.

[0112] Example 1:

[0113] To verify the feasibility of this invention in practice, it was applied to the task of compensating for positioning errors of BeiDou navigation chips in a complex urban canyon environment. This environment is characterized by dense high-rise buildings, rapidly changing road conditions, and significant fluctuations in communication link quality. In this environment, BeiDou navigation chips are affected by multiple path reflections, occlusion losses, link attenuation, and abrupt attitude changes, resulting in positioning errors exhibiting nonlinear, non-stationary, and strongly scene-dependent time-varying characteristics. Existing physical error models, differential enhancement methods, and linear filtering methods struggle to accurately characterize the time-varying nature of errors in this environment, and cannot achieve piecewise adaptive compensation for the error evolution process. The continuous-time error vector field modeling structure, piecewise CfC error compensation model, and time constant modulation layer provided by this invention can provide continuous-time, cross-scene, and adaptive fine-grained compensation for positioning errors in such complex environments.

[0114] In the specific implementation process, the raw navigation observation data generated by the BeiDou navigation chip, terminal operation status data, and communication link monitoring data are uploaded to the communication and navigation monitoring platform to construct a continuous-time input stream. The platform then generates reference trajectory data as a monitoring benchmark. Based on the continuous-time error vector field modeling structure, the platform generates the positioning error evolution relationship for each time point. Subsequently, the error modeling feature sequence and positioning error sequence are input into the continuous-time error modeling network, forming a continuous-time error vector field through input units, continuous-time coding units, and output units. On the terminal side, the continuous-time error vector field is divided into multiple closed-time substructures, and a segmented CfC error compensation model is constructed. Each closed-time substructure contains a shared network layer and a specific scenario parameter set. Communication link quality indicators and monitoring resolution levels are mapped as time-scale adjustment factors, acting within the state update unit. This enables the CfC substructure to accelerate short-time error response during link weakening phases and enhance the ability to maintain long-term error trends during monitoring resolution improvement phases.

[0115] In real-time applications, the model's ability to handle complex scenarios was effectively verified. Test data came from several urban main roads and elevated sections, with continuous operation exceeding 120 minutes. The total number of raw data entries collected from the BeiDou navigation chip exceeded 300,000. The proportion of communication link quality dropping from the stable range to the low-quality range reached 42%, and the positioning scenario changed 63 times. After compensation using the method of this invention, the average chip positioning error decreased from 3.87 meters to 2.81 meters, the error variance decreased from 6.92 to 4.31, and the maximum instantaneous error decreased from 11.54 meters to 7.32 meters. During 150 time periods with poor communication link quality, the method of this invention, relying on a time constant modulation structure, improved the error compensation response speed by approximately 23%, effectively suppressing error amplification caused by link fluctuations. For road sections with high monitoring density, the method of this invention, by enhancing the ability to maintain long-term error trends, reduced the amplitude of cross-scenario error drift by approximately 19%. In densely built-up areas where multipath enhancement is significant, the segmented CfC structure of the present invention achieves segmented compensation for the error subspace, thereby significantly reducing the degree of deviation of the positioning trajectory.

[0116] The table below shows the error performance of the method of this invention and the comparative method:

[0117] Table 1. Performance Comparison of Different Error Compensation Methods in Complex Urban Environments

[0118]

[0119] Table 1 shows that the experimental data fully demonstrates that the method of the present invention has higher positioning error control capability under different environmental conditions. The average positioning error is reduced by about 27% compared with the original data and by about 11% compared with the differential positioning method. The main reason is that the continuous-time error vector field constructed by the present invention can characterize the non-stationary changes of the error in a continuous-time form, so that error compensation no longer depends on a fixed model structure. The reduction in error variance indicates that the error fluctuation of the present invention is more stable in various scenarios, thanks to the continuous and differentiable smooth switching between different closed-time substructures achieved by the segmented CfC structure. The maximum instantaneous error decreases significantly because the time constant modulation layer can quickly compress the response time of the model when the link quality deteriorates, making the instantaneous error easier to suppress. In the multipath enhancement area and the low-quality communication link area, the reduction in the mean error of the present invention is further increased, thanks to the sensitive response capability of the time scale adjustment factor to the rapid change segment of error. The reduced error drift amplitude in the scene switching segment indicates that the present invention has stronger cross-scene generalization capability.

[0120] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent compensation of positioning errors in BeiDou navigation chips based on deep learning, characterized in that, Includes the following steps: The system acquires raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data output by the BeiDou navigation chip, organizes them into a continuous time input stream, and references trajectory data. The continuous-time input stream and reference trajectory data are input into the continuous-time error modeling network, and the continuous-time error vector field is obtained by training based on the position deviation and communication and navigation monitoring status data. A piecewise CfC error compensation model is constructed, which divides the error evolution relationship corresponding to the continuous-time error vector field into multiple closed-time substructures. A time constant modulation layer is introduced into the piecewise CfC error compensation model to map the conduction monitoring state data into a time scale adjustment factor, thus obtaining a piecewise CfC error compensation model with a time constant modulation structure. A continuous-time distillation structure is constructed between the edge and cloud. Based on the continuous-time error vector field, a piecewise CfC error compensation model with a time constant modulation structure, and a continuous-time input stream, dynamic alignment constraint training is performed to obtain a piecewise CfC error compensation model trained by the edge and cloud. During real-time operation, the real-time collected data is constructed into a real-time continuous time input stream, which is input into the segmented CfC error compensation model trained by the edge-cloud collaboration, generating the error compensation amount at the corresponding time point, and performing error compensation to obtain the compensated positioning result. The construction of the piecewise CfC error compensation model includes: On the terminal side where the Beidou navigation chip is located, the continuous time error vector field and scene label sequence are received, and the scene labels and error evolution relationship are associated with a time index to obtain the time-aligned scene label sequence and error evolution relationship sequence. Based on the time-aligned scene label sequence, adjacent time points with the same scene label are combined into scene time periods and assigned a unique substructure index to form a set of scene time periods. Based on the set of scene time periods and the error evolution relationship sequence, multiple closed time substructures are constructed on the terminal side. A shared network layer is set to share the network structure and parameters. A specific set of scene parameters is set to fit the error evolution relationship corresponding to the scene label in the continuous time error vector field. Each closed time substructure is bound to the corresponding scene time period one by one to form a segmented CfC basic structure with multiple closed time substructures. A scene discrimination function is constructed based on the scene label sequence. The scene label corresponding to the current time point and the adjacent scene labels are used as inputs. The output is the weight coefficient of the current time point for each closed time substructure. The weight coefficients are smoothly transitioned near the boundary of the scene time period according to the preset continuous change rule. On the terminal side, the continuous time error vector field is input into the piecewise CfC basic structure. At each time point, all closed time substructures output the intermediate error compensation result at that time point. The weight coefficients and the intermediate error compensation results are matched according to the time point, and the weighted summation is performed to obtain the error compensation output at that time point. The output sequence of the piecewise CfC error compensation model is generated by connecting them, so as to realize the continuous differentiable switching between multiple closed time substructures. The construction of the piecewise CfC error compensation model with time constant modulation structure includes: On the terminal side, based on the communication and navigation monitoring status data in the continuous time input stream, the communication link quality index sequence and the monitoring degree level sequence are extracted and bound to the CfC substructure corresponding to each time point in the segmented CfC error compensation model to generate the time constant input sequence. A time constant modulation layer is constructed inside the segmented CfC error compensation model. The communication link quality index and the monitoring resolution level are received and combined and mapped into a time scale adjustment factor. At the time boundary of different scenarios, a smooth transition is performed according to a preset continuous change rule to form a time scale adjustment factor sequence. The time scale adjustment factor sequence is written into the state update unit of each CfC substructure in the piecewise CfC error compensation model. When the hidden state update is performed, the time scale adjustment factor at the corresponding time point is read to dynamically adjust the time constant inside the CfC substructure. After completing the writing of the time scale adjustment factor and the dynamic adjustment of the time constant of the state update unit, the segmented CfC error compensation model with integrated time constant modulation layer is used as the segmented CfC error compensation model with time constant modulation structure.

2. The intelligent compensation method for positioning errors of Beidou navigation chips based on deep learning according to claim 1, characterized in that, The generation of the continuous-time input stream and reference trajectory data includes: The raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data output by Beidou navigation chips deployed in different application scenarios are collected with time stamps to form a raw data set; The communication and navigation monitoring status data includes communication link quality indicators, monitoring resolution level, and scene labels; Time alignment, interpolation and resampling are performed on the original dataset to remove invalid records corresponding to missing time periods, resulting in the original navigation observation data sequence, terminal operation status data sequence and communication and navigation monitoring status data sequence. Based on the original navigation observation data sequence, terminal operation status data sequence, and communication and navigation monitoring status data sequence, the data content at each time point is concatenated according to a fixed field order to generate a continuous time input stream; Based on the monitoring intensity level and scene label in the communication and navigation monitoring status data sequence, a scene label is generated for each time point. Adjacent time points representing the same monitoring intensity level and the same scene label are combined into a scene time period and encoded into a scene label sequence. Based on the positioning reference data collected by the communication and navigation monitoring reference equipment, the corresponding reference position is calculated at each time point, and the reference positions of all time points are stitched together to generate reference trajectory data.

3. The intelligent compensation method for positioning errors of Beidou navigation chips based on deep learning according to claim 1, characterized in that, The generation of the continuous-time error vector field includes: On the communication and monitoring platform side, the position deviation is calculated based on the positioning results in the continuous time input stream and the reference position data in the reference trajectory data to obtain the positioning error sequence; Extract input features that correspond one-to-one with the positioning error sequence from the continuous-time input stream, and concatenate the original data set of each time point into an error modeling feature sequence, which together with the positioning error sequence forms continuous-time error modeling training data. A continuous-time error modeling network structure is constructed based on the continuous-time error modeling training data, and the initial continuous-time error modeling network is obtained by initializing all dynamic parameters. The training data for continuous time error modeling is input into the initial continuous time error modeling network to obtain the positioning error estimation results at each time point. These results are compared with the positioning error sequence to calculate the position deviation residuals. The training loss value is then aggregated to obtain the training loss value. Based on the training loss value, the dynamic parameters of the continuous time error modeling network are iteratively updated to obtain the converged continuous time error modeling network. The converged continuous-time error modeling network is deployed on the communication and monitoring platform. The continuous-time input stream is input into the converged continuous-time error modeling network to obtain the positioning error estimation results. The positioning error estimation results at each time point are arranged in chronological order to form a continuous-time error vector field.

4. The intelligent compensation method for positioning errors of Beidou navigation chips based on deep learning according to claim 3, characterized in that, The continuous-time error modeling network structure includes: The input unit includes a time series input interface and a preprocessing layer, which converts the error modeling feature vector into an internal feature representation sequence. The multi-level continuous-time coding unit receives an internal feature representation sequence, performs state update operations, and generates a hidden state sequence. The output unit receives the hidden state sequence and converts it into localization error estimation results at each time point through a fully connected mapping layer.

5. The intelligent compensation method for positioning errors of Beidou navigation chips based on deep learning according to claim 1, characterized in that, The construction of the segmented CfC error compensation model trained collaboratively by the cloud and the edge includes: The continuous-time error vector field is used as the teacher model, and the piecewise CfC error compensation model with time constant modulation structure is used as the student model. On the communication and monitoring platform side, the continuous-time input stream is input into the continuous-time error modeling network, and the error increment corresponding to each time point in the continuous-time error vector field is read to form the teacher model output sequence. Meanwhile, on the terminal side, the continuous time input stream is input into the segmented CfC error compensation model with a time constant modulation structure, and the error compensation amount corresponding to each time point is read to form the student model output sequence. The output sequences of the teacher model and the student model are input into the end-cloud continuous-time distillation structure to construct an end-cloud continuous-time distillation loss function to measure the difference between the teacher error increment and the student error compensation. The end-cloud continuous-time distillation loss function generates time point residuals based on the difference between the error increment and the corresponding error compensation at each time point. The residuals are aggregated to generate a complete time dimension residual sequence, which is used as the end-cloud continuous-time distillation loss. Using the continuous-time distillation loss of the end cloud as the training target, joint parameter updates are performed on all CfC substructure parameter sets inside the piecewise CfC error compensation model with time constant modulation structure and all time scale adjustment factors in the time constant modulation layer. This keeps the continuous-time error vector field corresponding to the teacher model from being updated, so that the student model can gradually learn the error evolution relationship represented by the teacher model. The continuous-time input stream is continuously fed into the teacher model and the student model, and the parameters are iteratively optimized until the continuous-time distillation loss of the end-cloud meets the preset convergence condition. At this point, the student model is frozen as a piecewise CfC error compensation model trained by the end-cloud collaboration.

6. The intelligent compensation method for positioning errors of Beidou navigation chips based on deep learning according to claim 1, characterized in that, The generation of the compensated positioning result includes: During the real-time operation of the Beidou navigation chip, raw navigation observation data, terminal operation status data, and communication and navigation monitoring status data are collected in real time to generate a real-time data set with real-time time stamps. Perform time alignment processing on the real-time dataset to generate a real-time continuous time input stream, and simultaneously generate a real-time scene label sequence; The real-time continuous time input stream and the real-time scene label sequence are input into the segmented CfC error compensation model trained by the edge-cloud collaboration. The CfC substructure at the corresponding time point is selected, the time scale adjustment factor at that time point is generated, the CfC state update is performed after reading the time scale adjustment factor, the real-time error compensation amount is output, and the real-time error compensation sequence is obtained by arranging them. The real-time error compensation sequence is matched with the positioning results output in real time by the Beidou navigation chip on a unified time axis. The real-time positioning result at each time point is added with the real-time error compensation amount at the corresponding time point to generate a compensated positioning result.

Images