Low-orbit satellite multi-gnss baseband ai real-time optimization method and system

Abstract

The application provides a low-orbit satellite multi-GNSS baseband AI real-time optimization method and system, and belongs to the technical field of satellite navigation data processing. The method sequentially executes the following at each solution epoch: the first-level arbitration filters out invalid and intermittent fault basebands through multi-dimensional hard veto and sliding window statistics; the second-level arbitration adopts a cascade quality level bit mask rule to score and infer a integer GBDT model in parallel, and outputs a comprehensive score through dynamic weighted fusion; the third-level arbitration performs first-order low-pass time sequence smoothing, and performs hierarchical switching and ping-pong suppression in combination with a multi-condition hysteresis threshold and a sliding window mechanism; and the fourth-level arbitration triggers a three-level degradation keep-alive strategy when all basebands fail. The application completely eliminates the dependence on floating-point operations, significantly reduces the occupation of on-board resources and the frequency of false switching, and guarantees the convergence of orbit determination filtering and the continuity of navigation services.

Description

Technical Field

[0001] This invention belongs to the field of satellite navigation signal processing and spaceborne embedded control technology, and in particular relates to a real-time AI optimization method and system for multiple GNSS basebands on low Earth orbit (LEO) satellites. Specifically, it is applied to the central processor of a LEO GNSS receiver to achieve highly reliable evaluation, smooth switching, and degradation-based keep-alive of multi-baseband navigation data under conditions of limited onboard computing and storage resources. Background Technology

[0002] Low-Earth orbit (LEO) satellite GNSS receivers typically employ multiple independent baseband channels to receive signals from satellite navigation systems such as GPS, BDS, GLONASS, and Galileo in parallel, in order to improve space observation redundancy and orbit determination reliability. Under complex operating conditions such as high-speed motion in LEO, celestial attitude maneuvers, ionospheric scintillation, and space radiation interference, the quality of the output data from each baseband exhibits strong time-varying and asymmetric characteristics. Existing baseband data optimization techniques mainly suffer from the following problems: (1) The single switching logic causes the ping-pong effect: Traditional methods mostly rely on single-epoch instantaneous thresholds or fixed weights, without introducing timing smoothing and hysteresis protection mechanisms. In signal edge conditions or when the observed noise changes abruptly, it is easy to trigger frequent baseband switching, which leads to divergence of downstream real-time orbit determination filtering and drastic fluctuations in orbit determination accuracy.

[0003] (2) Poor adaptability to spaceborne resources: Mainstream AI optimization algorithms rely on floating-point operations and dynamic memory allocation, resulting in large model sizes that cannot be deployed on spaceborne embedded processors with strictly limited clock speeds and RAM. Furthermore, pure rule-based scoring methods are unable to characterize the nonlinear coupling relationships between multidimensional quality indicators, leading to low scoring discrimination.

[0004] (3) Weak fault tolerance in full baseband failure scenarios: When multiple basebands are interfered with or intermittent failures occur at the same time, the existing system lacks a graded degradation strategy and usually directly interrupts the navigation data output, which cannot meet the mission requirements of low-orbit satellites for navigation service continuity and high availability.

[0005] (4) Difficulties in algorithm evolution and on-orbit maintenance: Traditional fixed on-board optimization logic is difficult to be updated through the space-to-ground telemetry and control link, and cannot adapt to changes in orbital segments, load aging or new interference modes.

[0006] Therefore, there is an urgent need in this field for a real-time baseband data optimization scheme that takes into account resource lightweighting, intelligent decision-making, switching stability, and degradation survival, in order to solve the technical bottleneck of multi-source navigation data optimization in spaceborne environments. Summary of the Invention

[0007] In view of this, the present invention aims to propose a real-time optimization method and system for multi-GNSS baseband AI for low-Earth orbit satellites. By constructing a four-level arbitration architecture of hard veto, fusion scoring, smooth switching, and downgrade fallback, it achieves high reliability, low latency, and anti-interference baseband data optimization under the condition of limited on-board resources, improves real-time orbit determination accuracy and navigation service continuity, and meets the requirements of long-term, high-reliability navigation enhancement missions for low-Earth orbit satellites.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A real-time AI optimization method for multiple GNSS basebands of low-Earth orbit satellites is applied to the central processor of a low-Earth orbit satellite GNSS receiver. The method includes the following steps in each calculation epoch: First-level arbitration: Perform multi-dimensional hard rejection judgment on each received GNSS baseband data. If any preset rejection condition is triggered, the baseband is marked as invalid in this epoch. Second-level arbitration: For valid baseband data that has passed the first-level arbitration, rule scoring and AI inference scoring are performed respectively, and the scores are merged to generate a single-epoch comprehensive score; The rule-based scoring employs a multi-dimensional cascaded quality level bitmask mechanism: a cascaded quality level bitmask is preset from low to high bits, with each high-level bit implying that all low-level conditions corresponding to that bit have been met. A basic score is obtained based on the highest-level bit in the quality level bitmask, and continuous increments are added. The AI ​​inference scoring is based on the micro gradient boosting decision tree (GBDT) model, which is stored as an integer constant array. The inference process only performs integer comparisons and addition / subtraction operations. Level 3 arbitration: Perform timing smoothing on the comprehensive score of each baseband, and perform hierarchical switching arbitration based on the smoothing score to output the selected candidate primary baseband; Level 4 Arbitration: Monitor the overall baseband validity status. If the entire baseband fails, execute a graded degradation and keep-alive strategy based on the number of consecutive invalid epochs to determine the baseband data to be output as a fallback. Real-time orbit determination and data output: Real-time orbit determination is performed using the raw observations provided by the selected primary baseband. The orbit determination results and baseband data are framed to generate data packets, which are then output.

[0009] Furthermore, in the first-level arbitration, the preset veto conditions cover the dimensions of data integrity, positioning validity, time reference, geometric configuration, hardware health, observation real-time performance, observation accuracy, and signal stability. The method further includes: for intermittently faulty basebands with fewer than M valid epochs within the last N epochs, forcibly marking them as invalid for the current epoch.

[0010] Furthermore, in the rule scoring, the setting logic of the cascaded quality level bitmask is as follows: The baseband data quality index is checked bit by bit starting from the least significant bit. If the condition corresponding to the current bit is met and all the less significant bits are set, then the current bit is set. The basic score is obtained by looking up the table based on the bit sequence number of the most significant bit in the quality level bitmask. The continuous quantity bonus includes bonuses for the number of satellites, average carrier-to-noise ratio, shortest continuous lock time, and continuous positioning stability, with upper limits for each bonus.

[0011] Furthermore, in the AI ​​reasoning scoring, the GBDT model, after being trained offline on the ground, converts the decision tree structure into nested judgment statements in C language; The input feature vector of the GBDT model includes: effective number of satellites, average carrier-to-noise ratio, shortest continuous lock time, PDOP value, pseudorange residual, maximum age of observation measurement, baseband self-check health status, and number of effective epochs for continuous positioning.

[0012] Furthermore, the fusion process in the second-level arbitration adopts a weighted fusion method, and the calculation formula is as follows: ; in, For comprehensive scoring, Rate the AI ​​reasoning. Rate the rules The fusion weighting coefficient; Supports uploading and updating via the satellite-to-ground telemetry and control link.

[0013] Furthermore, the third-level arbitration includes: timing smoothing: performing a first-order low-pass filter on the comprehensive score of each baseband to obtain a smoothed score; Tiered handover arbitration: If the current primary baseband fails the first-level arbitration, a forced handover arbitration is executed to switch to the effective baseband with the highest smoothing score; if the current primary baseband is effective, a dominant handover arbitration is executed only when the candidate baseband with the highest smoothing score simultaneously meets the preset score difference threshold, the number of consecutive effective epochs threshold, the shortest consecutive lock time threshold, and the epoch interval threshold from the last baseband handover. Ping-pong handover suppression: Real-time monitoring of the number of baseband handovers within the last P epochs. If the number of handovers exceeds the preset threshold, the minimum handover interval threshold is temporarily increased.

[0014] Furthermore, the graded degradation keep-alive strategy in the fourth-level arbitration includes: Level 1 degradation: when the number of consecutive invalid epochs is less than the first threshold, the primary baseband identifier of the previous epoch is maintained, and the latest historical valid data of the baseband cache is output; Second-level degradation: When the number of consecutive invalid epochs reaches the second threshold, the cached data of the one with the best smoothing score among all baseband historical valid data is forcibly output; Level 3 Degradation: If there is no valid historical data, output the current data of the default baseband and mark the navigation data as invalid.

[0015] Furthermore, the data in the data packet includes: Location results; orbit determination results; raw observations; filtered status fields: including primary baseband ID, smoothing score, switching reason code, data status flag, quality level bitmask, AI inference score, and fusion weight.

[0016] The present invention also provides a low-orbit satellite multi-GNSS baseband AI real-time optimization system for executing the method described in any of the above, deployed in the central processor of a low-orbit satellite GNSS receiver, including: a first-level arbitration module: used to perform multi-dimensional hard veto judgment, and complete the validity marking of baseband data and the filtering of invalid data; The second-level arbitration module has a built-in rule scoring unit and an AI inference unit. The rule scoring unit is configured to execute the cascaded quality level bitmask setting and basic score mapping. The AI ​​inference unit is configured to load the GBDT model in the form of an integer constant array and execute integer inference. The scores of the two units are merged to output a single epoch comprehensive score. The third-level arbitration module is used to perform scoring timing smoothing, hierarchical switching arbitration and ping-pong switching suppression, and output the selected candidate primary baseband ID; The fourth-level arbitration module is used to monitor the overall baseband effectiveness status and execute a graded degradation and keep-alive strategy in the event of a full baseband failure. Real-time orbit determination module: used to receive raw observations from the selected baseband and perform real-time precise orbit determination calculations; Data Packaging Output Module: Used to frame and output the positioning results, orbit determination results, raw observations and filtered status fields of the selected baseband.

[0017] Furthermore, it also includes a model management module, which is connected to the second-level arbitration module, for: solidifying and storing AI model parameters and managing versions; receiving model parameter updates through the satellite-to-ground telemetry and control link, performing integrity verification, and writing them to the backup storage area; switching to the new effective model through atomic pointers; monitoring AI inference time and output range, and automatically switching to pure rule scoring mode when an anomaly occurs.

[0018] Compared with existing technologies, the low-orbit satellite multi-GNSS baseband AI real-time optimization method and system described in this invention have the following beneficial effects: (1) In the second-level arbitration, the rule engine score and the GBDT model inference score are weighted and fused. The rule engine ensures the deterministic evaluation of the basic quality dimensions through concatenated bitmasks, and the GBDT model provides supplementary representation of the nonlinear relationship between features. The combination of the two can improve the rationality of the comprehensive score.

[0019] (2) The structural parameters, ground training process, and lightweight deployment method of the GBDT model were clarified. The model is embedded in Flash in the form of C language integer judgment code. The inference process only requires integer comparison and addition and subtraction operations. An on-orbit update verification and automatic degradation mechanism for anomalies are configured.

[0020] (3) Progressive quality assessment is achieved by cascading quality level bitmasks. The high bit effectively implies that all low bit conditions have been met, ensuring that the high-resolution baseband reaches the preset threshold in each basic dimension, and avoiding the concealment of defects in other dimensions due to the prominence of a single indicator.

[0021] (4) The introduction of first-order low-pass smoothing, hysteresis threshold, minimum hold time and ping-pong switching suppression mechanism in the switching decision reduces the probability of frequent switching caused by single-epoch signal jitter, which is conducive to maintaining the continuity of the input data of the orbit determination filter.

[0022] (5) A graded degradation and keep-alive strategy is set up for the extreme condition of full baseband failure. The degradation process is performed progressively according to the number of consecutive invalid epochs to ensure that each epoch has a valid data frame output and reduce the risk of navigation reference interruption.

[0023] (6) Using integer arithmetic and bit operations, the AI ​​inference model is stored in the form of integer constants, and the time consumed per epoch does not exceed 1ms. The amount of code and RAM usage are both controlled within a small range, making it suitable for deployment on resource-constrained spaceborne embedded processors. Attached Figure Description

[0024] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall four-level arbitration process described in an embodiment of the present invention; Figure 2 This is a schematic diagram of the multi-dimensional cascaded quality level bitmask progressive setting logic described in an embodiment of the present invention; Figure 3 This is a schematic diagram of the parallel fusion architecture of the rule scoring unit and the AI ​​inference unit in the second-level arbitration according to an embodiment of the present invention; Figure 4 This is a schematic diagram illustrating the data flow from ground training to spaceborne integration and deployment of the GBDT model as described in this embodiment of the invention. Figure 5 This is a schematic diagram of the third-level hierarchical hysteresis switching arbitration and ping-pong suppression logic state machine according to an embodiment of the present invention; Figure 6 This is a schematic diagram of the module structure and data interaction of the system described in an embodiment of the present invention. Detailed Implementation

[0025] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0026] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0027] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0028] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0029] like Figures 1 to 6As shown, this invention addresses the technical bottlenecks of limited onboard computing resources for low-Earth orbit satellites, frequent ping-pong switching in multi-baseband systems, and weak fault tolerance in all-failure scenarios. It proposes an AI-enhanced multi-dimensional hierarchical arbitration method and system for real-time optimization of GNSS baseband data. Its core lies in constructing a four-level progressive architecture: hard veto, fusion scoring, smooth switching, and degradation fallback. The first-level arbitration uses multi-dimensional Boolean thresholds and sliding window statistics to filter out invalid and intermittent fault data. The second-level arbitration innovatively employs cascaded quality bitmask rule scoring and parallel inference with an integer micro GBDT model, using dynamic weight fusion to balance engineering interpretability and nonlinear feature capture, completely eliminating floating-point operations and dynamic memory dependencies. The third-level arbitration introduces first-order low-pass timing smoothing, multi-condition hysteresis comparison, and sliding window ping-pong suppression to effectively suppress instantaneous observation noise and reduce baseband switching frequency. The fourth-level arbitration designs a three-level degradation keep-alive strategy, using cache maintenance, historical best backtracking, and invalid marking as fallback to prevent hard interruptions of navigation data. This invention ensures highly reliable optimization under stringent constraints of RAM < 6KB and single epoch time < 0.9ms, significantly improving the downstream EKF orbit determination convergence accuracy and service continuity, while combining the advantages of lightweight onboard deployment with on-orbit algorithm hot update capabilities. Details of this invention are as follows: A real-time AI-based optimization method for low-Earth orbit (LEO) satellite multi-GNSS basebands is applied to the central processor of an LEO satellite GNSS receiver. This central processor is connected to at least two independent GNSS basebands, and at each epoch, it synchronously receives the positioning results and raw observations output from each baseband. Real-time data optimization is achieved through a four-level arbitration mechanism, where the second level of arbitration incorporates a mechanism that fuses AI inference scoring and rule-based scoring. The method includes the following steps: Level 1 arbitration: Perform multi-dimensional hard veto judgment on each received GNSS baseband data. If any preset veto condition is triggered, mark the baseband epoch as invalid. Second-level arbitration: For valid baseband data that has passed the first-level arbitration, rule scoring and AI inference scoring are performed respectively, and the scores are merged to generate a single-epoch comprehensive score; The rule scoring adopts a multi-dimensional cascaded quality level bitmask mechanism: a cascaded quality level bitmask is preset from low to high bits, and the high bits are valid, implying that all low-bit conditions corresponding to that bit have been met. The basic score is obtained based on the highest valid bit in the quality level bitmask, and continuous increments are added. The AI ​​inference scoring is based on the micro gradient boosting decision tree GBDT model, which is stored in the form of an integer constant array. The inference process only performs integer comparison and addition / subtraction operations. Level 3 arbitration: Perform timing smoothing on the comprehensive score of each baseband, and perform hierarchical switching arbitration based on the smoothing score to output the selected candidate primary baseband; Level 4 Arbitration: Monitor the overall baseband validity status. If the entire baseband fails, execute a graded degradation and keep-alive strategy based on the number of consecutive invalid epochs to determine the baseband data to be output as a fallback. Real-time orbit determination and data output: Real-time orbit determination is performed using the raw observations provided by the selected primary baseband. The orbit determination results and baseband data are framed to generate data packets, which are then output.

[0030] The details of this method are as follows: 1. First-level arbitration: Arbitration based on a hard veto threshold for validity For each baseband's single-epoch data, a multi-dimensional Boolean logic rejection check is performed. If any preset rejection condition is triggered, the baseband is marked as invalid for this epoch and will not proceed to the subsequent arbitration stage. The rejection conditions cover eight dimensions: data integrity, positioning validity, time base, geometric configuration, hardware health, observation real-time performance, observation accuracy, and signal stability. Specific rejection conditions and recommended thresholds are shown in Table 1. Table 1 In addition, for intermittently faulty basebands with fewer than 5 valid epochs within the last 10 epochs, they are forcibly marked as invalid for this epoch.

[0031] 2. Second-level arbitration: AI-enhanced multi-dimensional quality grading and scoring arbitration For valid baseband data that has passed the first level of arbitration, rule scoring and AI inference scoring are performed respectively, and the scores are fused to generate a single-epoch comprehensive score.

[0032] 2.1 Rule Scoring Rule-based scoring follows the multi-dimensional cascaded quality level bitmask mechanism, including three sub-steps: 2.1.1 Setting the Cascaded Quality Level Bitmask A 9-bit concatenated quality level bitmask (bit0-bit8) is pre-defined, from least significant bit to most significant bit. The validity of each more significant bit implies that all less significant bit conditions corresponding to that bit have been met. Starting from the least significant bit, the baseband data quality index is checked bit by bit. If the condition corresponding to the current bit is met and all less significant bits are set, the current bit is set, thus generating the quality level bitmask for this epoch of the baseband. The definitions and recommended thresholds for each bit are as follows: bit0: All hard rejection conditions in step S1 have been met; bit1: The deviation between the baseband timestamp and the processor's local timekeeping is ≤10 ms; bit2: Location calculation PDOP < 8; bit3: Pseudo-range residual RMS < 3 m; bit4: Number of valid satellites participating in positioning ≥ 8; bit5: The average carrier-to-noise ratio of all locked satellites is ≥42 dB-Hz; bit6: The shortest consecutive locking time for all locked satellites is ≥10 s; bit7: The time difference between the measurement time and the current epoch is <50 ms; bit8: Baseband self-test status is all healthy and the percentage of satellites that have lost lock and been recaptured is 0%.

[0033] 2.1.2 Basic Hierarchical Mapping The basic score is obtained by using the bit number of the most significant bit in the quality level bitmask through a preset mapping table (as shown in Table 2).

[0034] Table 2 As shown in Table 2, first determine the bit number of the most significant bit in the quality level bit mask (bit0 to bit8) (denoted as...). (Values ​​range from 0 to 8), and then the basic score is obtained through the mapping table. The mapping table is equivalent to a linear formula, and the linear expression of the mapping table is as follows: ; in, The most significant bit is the bit sequence number, ranging from 0 to 8, with bit 0 corresponding to... ,Right now (The most significant bit is bit 0) corresponds to 120 points; (The most significant bit is bit 1) corresponds to 180 points, and so on. Corresponding score: 240; Corresponding to 300 points; Corresponding score: 360; The corresponding score is 420. Corresponding score: 480; The corresponding score is 540. Corresponding to 600 points; if there are no valid bits (theoretically this shouldn't happen, because at least bit 0 would be set), then... deal with.

[0035] 2.1.3 Multi-dimensional continuous quantity bonus In addition to the base score, four consecutive bonus items are added, each with a maximum bonus of 100 points. The calculation formula is as follows: The expression for adding points based on the number of satellites is as follows: Only when Effective immediately; The expression for the average carrier-to-noise ratio (CNR) score is as follows: Only when Effective at dB-Hz; The expression for awarding points for the shortest consecutive locking time is as follows: , The unit is seconds; The expression for the continuous positioning stability bonus is as follows: , This is the number of epochs in which the baseband has continuously passed the first-level arbitration and whose positioning is valid.

[0036] The expression for rule scoring is as follows: If the score exceeds 1000, the score will be clamped down to 1000.

[0037] 2.2 AI Reasoning Scoring In AI inference scoring, the GBDT model, after offline training on the ground, converts its decision tree structure into nested C language conditional statements. The input feature vector of the GBDT model includes the number of effective satellites, average carrier-to-noise ratio, shortest consecutive lock-on time, PDOP value, pseudorange residual, maximum age of observations, baseband self-check health status, and number of consecutive effective epochs of positioning. Details regarding AI inference scoring are as follows: 2.2.1 Model Selection and Deployment Form The central processor's built-in AI inference model is a Gradient Boosting Decision Tree (GBDT). GBDT is a supervised learning model based on decision tree ensembles, which iteratively trains multiple decision trees to gradually approach the target value. After offline training on the ground, the GBDT model converts the decision tree structure into nested conditional statements in C language and stores them in the central processor's Flash memory as an array of integer constants. The model consists of 3 to 5 decision trees, each with a depth of no more than 4 levels, a total of no more than 200 parameters, and a model file size of no more than 4KB. The inference process only requires integer comparisons and addition / subtraction operations.

[0038] 2.2.2 Ground-based offline training The training data comes from historical on-orbit telemetry data of low-Earth orbit satellites of the same type or model, including characteristic fields such as the number of effective satellites reported per epoch for each baseband, average carrier-to-noise ratio, shortest consecutive lock time, PDOP value, pseudorange residual RMS, observation age, baseband self-check status, and number of consecutive effective epochs. The supervision label uses the root mean square (RMS) of the carrier phase residual from post-precision orbit determination, which characterizes the noise level of the original observations. During training, the residual RMS is mapped to a quality score of 0–1000, with a higher score for a smaller residual.

[0039] Training is performed on a ground server using the LightGBM framework. Hyperparameters (number of trees, maximum depth, learning rate, etc.) are determined through grid search and cross-validation. After training, a script is used to convert the model structure into C language integer judgment code, which is then compiled and linked to the central processor software project.

[0040] 2.2.3 On-orbit inference process For each epoch, an 8-dimensional feature vector (all obtained from the baseband's native output fields) is extracted from the baseband that has passed the first-level arbitration, as shown in Table 3: Table 3 The feature vectors are input into the GBDT model code stored in the central processor's Flash memory. After integer comparison and addition / subtraction operations, an AI inference score ranging from 0 to 1000 is output. A single inference attempt takes no more than 200 μs.

[0041] 2.2.4 In-orbit Updates and Safety Degradation GBDT model parameters (decision tree structure and thresholds) can be updated via satellite-to-ground telemetry and control links. The central processor has a model management module that performs integrity verification (CRC32) on newly uploaded model parameters. After successful verification, the parameters are written to the central processor's spare Flash area and activated via atomic pointers. If the AI ​​inference output value exceeds the valid range of 0-1000, or the inference time exceeds a preset threshold, the system automatically switches to pure rule-based scoring mode.

[0042] 2.3 Scoring Integration The overall score uses a weighted fusion method, expressed as follows: ; in, The initial value of the fusion weighting coefficient is set to 0.3. It can be remotely adjusted via satellite-to-ground link based on on-orbit operation data (such as the orbit determination residual statistics of multiple consecutive orbits).

[0043] 3. Third-level arbitration: Time-sequenced, delayed switching arbitration 3.1 First-order low-pass timing smoothing A first-order low-pass filter is applied to the overall score of each baseband to obtain a smoothed score, as shown in the following expression: ; in, For the current number Smooth scoring of epochs, For comprehensive scoring, Smoothing score for the previous epoch. This indicates rounding down. For basebands with intermittent faults, the smoothing score is halved.

[0044] 3.2, Tiered Arbitration Switching (1) Forced switching arbitration: If the current primary baseband fails the first level arbitration, it will be directly switched to the effective baseband with the highest smoothing score.

[0045] (2) Advantageous handover arbitration: If the current primary baseband is valid, handover will only be performed if the candidate baseband with the highest smoothing score meets all of the following conditions; otherwise, the current primary baseband will be maintained: The difference between the candidate baseband smoothing score and the current primary baseband smoothing score is ≥80 points; The number of valid epochs for continuous positioning of candidate baseband is ≥10; The shortest consecutive locking time for all locked satellites in the candidate baseband is ≥5 seconds; The epoch interval since the last baseband switch is ≥30 epochs (i.e. 3 seconds).

[0046] (3) Ping-pong handover suppression arbitration: Real-time monitoring of the number of baseband handovers within the last 5 epochs. If the number of handovers is ≥2, the minimum handover interval threshold is temporarily increased to 50 epochs.

[0047] 4. Fourth-level arbitration: tiered, downgraded, and fault-tolerant arbitration as a safety net. If all basebands fail the first-level arbitration in this epoch, a three-level graded degradation and survival strategy will be implemented based on the total number of consecutive invalid epochs across all basebands: Level 1 Degradation: When the number of consecutive invalid epochs is less than 3 (e.g., 1 to 2), maintain the primary baseband identifier of the previous epoch, output the latest historical valid data of the baseband cache, and set the status flag of the degradation mode.

[0048] Level 2 Degradation: When the number of consecutive invalid epochs reaches 3 or more, the emergency mode status flag is set, and the cached data of the one with the best smooth score among all baseband historical valid data is forcibly output.

[0049] Level 3 Degradation: If there is no valid historical data, output the current data of the default baseband and mark the navigation data as invalid to prevent the on-board bus data from being idle.

[0050] 5. Real-time orbit determination and data output The central processor utilizes raw observations provided by the selected primary baseband, combined with broadcast ephemeris and dynamic models, to perform real-time orbit determination using the Extended Kalman Filter (EKF) algorithm, generating orbital status (position and velocity). Subsequently, the data is framed according to a protocol to generate data packets, which are then output in real-time to the onboard computer and other onboard payloads via the onboard telemetry bus. The data packets include the following: The system includes: the location results (PVT) of the selected baseband; the orbit determination results generated by the central processor; the raw observations of the selected baseband; and the filtering status fields, including the primary baseband ID, smoothing score, switching reason code, data status flag, bitmask of each baseband quality level, AI inference score, fusion weight, etc.

[0051] A low-Earth orbit (LEO) satellite multi-GNSS baseband AI real-time optimization system, used to execute the aforementioned LEO satellite multi-GNSS baseband AI real-time optimization method, is deployed in the central processor of a LEO satellite GNSS receiver and includes the following modules: Data receiving and parsing module: Receives serial port data packets from each GNSS baseband, performs frame verification, data parsing and core field extraction, and performs buffering according to baseband ID and epoch timestamp.

[0052] The first-level arbitration module connects to the data receiving and parsing module and is used to perform multi-dimensional hard veto judgments, completing the validity marking of baseband data and filtering invalid data.

[0053] The second-level arbitration module is connected to the first-level arbitration module and has a built-in rule scoring unit and an AI inference unit. The rule scoring unit is configured to execute the cascaded quality level bitmask setting and the basic score mapping. The AI ​​inference unit is configured to load the GBDT model in the form of an integer constant array and execute integer inference. The scores of the two units are merged to output a single epoch comprehensive score.

[0054] The third-level arbitration module connects to the second-level arbitration module and is used to perform scoring timing smoothing, hierarchical switching arbitration and ping-pong switching suppression, and output the selected candidate primary baseband ID.

[0055] The fourth-level arbitration module connects to the first-level and third-level arbitration modules respectively. It is used to monitor the overall baseband validity status, execute a graded degradation and keep-alive strategy in the event of overall baseband failure, and determine the baseband data to be output as a fallback.

[0056] Real-time orbit determination module: Connected to the third-level arbitration module, it is used to receive the raw observations of the selected baseband, perform real-time precise orbit determination calculations, and output the orbit determination results.

[0057] Data Packaging Output Module: Connected to the third-level arbitration module, the fourth-level arbitration module, and the real-time orbit determination module, respectively, it is used to frame and output the positioning results, orbit determination results, raw observations, and filtered status fields of the selected baseband.

[0058] Model Management Module: Connects to the second-level arbitration module, permanently stores AI model parameters and manages version; receives model parameter updates via the satellite-to-ground telemetry and control link, performs integrity verification, and writes them to the backup storage area; switches to the new effective model via atomic pointers; monitors AI inference time and output range, and automatically switches to pure rule scoring mode when an anomaly occurs.

[0059] Example 1: This embodiment provides a real-time AI optimization method for multiple GNSS basebands on low-Earth orbit (LEO) satellites, deployed in the central processor of a LEO satellite GNSS receiver. The preferred hardware environment is an ARM Cortex-R5F core (300MHz clock speed, 32KB RAM with ECC verification, and 256KB Flash). It independently connects to GPS, BDS, GLONASS, and Galileo baseband processing chips via four asynchronous RS422 interfaces, with a calculation epoch frequency of 10Hz (period 100ms).

[0060] S1: First-level arbitration At the beginning of each epoch, the central processor performs a multi-dimensional Boolean veto check on the navigation messages parsed from each baseband and the original observations. The preset veto conditions and judgment criteria are as follows: Data integrity: CRC32 checksum of data packet passed; Positioning validity: The baseband output positioning status word is a valid identifier (e.g., 0x55); Time reference: The deviation between the baseband timestamp and the spaceborne atomic clock / local clock is ≤50ms; Geometric configuration: ≥4 effective satellites participating in positioning, and position accuracy attenuation factor PDOP≤20; Hardware health: Baseband self-test flag is normal, no hardware faults or functional degradation; Real-time performance of observations: observation age ≤ 200ms; Observation accuracy: pseudorange residual RMS ≤ 30m; Signal stability: The proportion of satellites recaptured after being locked out is ≤50% of the total number of locked satellites.

[0061] If any condition is not met, the current epoch of that baseband is immediately marked as valid=0 and subsequent processing is skipped. To suppress intermittent spatial interference, the processor maintains a sliding window statistician, recording the number of valid epochs for each baseband within the last N (preferably N=10) epochs. If the number of valid epochs for a baseband is less than M (preferably M=5), it is determined to be an intermittent fault, and the current epoch is forcibly marked as valid=0 to prevent degraded data from polluting the scoring pool.

[0062] S2: Second-level arbitration Dual-path parallel scoring is performed only on basebands with valid=1 (e.g.) Figure 3 (as shown) (1) Rule-based scoring: A multi-dimensional concatenated quality level bitmask mechanism is adopted (e.g., Figure 2Define a 9-bit integer mask (Bit0~Bit8) and execute a strict progressive bit-setting logic: starting from Bit0, check each bit sequentially. Bit(k) can only be set to 1 if all bits from Bit0 to Bit(k-1) are 1 and the quality index corresponding to the current Bit(k) meets the standard. For example, Bit4 represents "valid satellite count ≥ 12", and Bit4 is only evaluated after the data has passed the checks of Bits 0~3. After setting the bits, locate the index of the most significant bit and obtain the basic score through a preset ROM mapping table. Then, add consecutive increments to this score. Bonus points for satellite count: 20 points for each additional valid satellite, up to a maximum of 100 points; Average carrier-to-noise ratio (CNR) bonus: 10 points are added for every 42 dB-Hz increase in CNR, up to a maximum of 50 points; Shortest consecutive lockout time bonus: 15 points for every 10 seconds exceeding the limit, up to a maximum of 60 points; Continuous positioning stability bonus: 10 points for every 10 consecutive valid epochs, up to a maximum of 40 points.

[0063] Rule scoring Clamp to [0, 1000].

[0064] (2) AI reasoning scoring: based on a mini GBDT model (e.g. Figure 4 The model is trained on the ground using historical on-orbit telemetry datasets, with the carrier phase residual RMS from post-precision orbit determination serving as the supervision label to optimize the loss function. After training, a dedicated compilation tool extracts the splitting threshold and leaf node values ​​from four decision trees of depth 3, converts them into int16_t constant arrays, generates nested if-else statements in C language, and stores them in the processor's Flash read-only area. During on-orbit inference, an 8-dimensional integer feature vector is input (number of effective satellites, average carrier-to-noise ratio × 10, shortest continuous lock-on time, PDOP × 10, pseudorange residual × 10, maximum age of observations, baseband self-check health status, number of consecutive effective epochs of positioning), and only integer comparison and accumulation operations are performed, outputting a score from 0 to 1000. There are no floating-point instructions or stack allocation throughout the entire process.

[0065] (3) Scoring fusion: according to expression Calculate the overall score. The initial value is preferably 0.3. This coefficient can be updated on-orbit via remote control commands through the satellite-to-ground telemetry and control link to adapt to signal environment migration caused by different orbital altitudes, seasonal ionospheric changes, or load aging.

[0066] S3: Third-level arbitration (1) Time series smoothing: for Perform a first-order low-pass filter: This effectively suppresses single-epoch mutations. If the baseband is judged to be an intermittent fault in S1, its smoothing score is forcibly multiplied by a penalty coefficient of 0.5, reducing its switching priority.

[0067] (2) Tiered switching arbitration (e.g.) Figure 5 ): Forced handover: If the current primary baseband has valid=0, a forced handover is immediately performed, and the valid baseband with the highest smoothing score is designated as the new primary baseband; Advantage handover: If the current primary baseband is valid, the candidate baseband must simultaneously meet the following four hysteresis conditions to trigger a handover: ① Smooth score difference (Score_diff) ≥ 80 points; ② Number of consecutive valid epochs (NavOkCnt) ≥ 10; ③ Shortest consecutive lock time (Lock) ≥ 5 seconds; ④ Episode interval since the last baseband handover ≥ 30 epochs. If any condition is not met, the current primary baseband will remain in use.

[0068] (3) Ping-pong switching suppression: Maintain a sliding window switching counter with a length of P (preferably P=5). If the number of baseband switching times within epoch P is ≥2, it is determined to be a ping-pong condition. The minimum switching interval threshold is temporarily increased to 50 epochs, and the default threshold is automatically restored after the counter is zeroed.

[0069] S4: Level 4 Arbitration When all baseband valid values ​​are 0 in this epoch, the hierarchical degradation state machine is activated: Level 1 Degradation: If the number of consecutive invalid epochs is less than 3, maintain the primary baseband identifier of the previous epoch, output the latest cached historical valid data in the baseband's circular buffer, and set the telemetry status field to the degradation mode. Level 2 Degradation: If the number of consecutive invalid epochs is ≥3, iterate through all baseband historical smoothing score records, force output the cached data of the historical best scorer, and set the status field to emergency mode. Level 3 Degradation: If there is no historical valid data cache after system initialization, output the current raw data of the default baseband and set the navigation data valid flag to 0.

[0070] S5: Real-time Orbit Determination and Data Output The central processor extracts pseudorange, carrier phase, Doppler, and ephemeris data from the selected primary baseband, inputs them into a lightweight extended Kalman filter (EKF) to perform real-time precise orbit determination, and calculates the satellite's position, velocity, and clock bias. Subsequently, the data packaging module frames the baseband positioning results, EKF orbit determination results, raw observations, AI / rule-based scoring details, and degradation status fields according to the CCSDS standard or a spacefarer-defined telemetry frame format, and transmits them to the onboard computer via the CAN / SpaceWire bus.

[0071] Example 2: This embodiment provides a real-time AI optimization system for multiple GNSS basebands on low-Earth orbit satellites, deployed as firmware in the central processor of a GNSS receiver. The logical modules are divided as follows (e.g., ...). Figure 6 (as shown) The first-level arbitration module implements the multi-dimensional veto logic and sliding window intermittent fault statistics of S1. The input is the raw baseband parsed data, and the output is the valid flag bits for each channel.

[0072] The second-level arbitration module integrates a rule scoring unit and an AI inference unit. The rule scoring unit has a built-in bitmask status register and lookup table ROM; the AI ​​inference unit loads an integer constant array executor. The fusion calculation unit outputs according to the weighted formula. .

[0073] The third-level arbitration module includes a built-in first-order low-pass filter, a hysteresis comparator group, and a sliding window switching counter. It receives smooth scores and valid status, and outputs a stable primary baseband ID and switching event flag.

[0074] The fourth-level arbitration module implements a fully failed state machine and a three-level buffer scheduler. It maintains historical scoring queues and circular data buffers for each baseband, and outputs degradation instructions and buffer pointers.

[0075] Real-time orbit determination module: Lightweight EKF solution engine, optimized with fixed-point number arithmetic, receives data pointers from the main baseband observations and the output of the fourth-level module, and iterates the orbit state vector in real time.

[0076] Data Packaging Output Module: Responsible for telemetry frame assembly, CRC16 / CRC32 encapsulation, bus driver scheduling, and outputting at a 10Hz cycle.

[0077] Model Management Module: Connected to the second-level arbitration module via an internal bus. Includes primary / backup dual Flash storage areas, a satellite-to-ground link resolver, an SHA-256 integrity checker, and atomic pointer switching logic. A built-in runtime monitor (Watchdog) monitors AI inference time and output boundaries in real time; if a single inference time > 900μs or an output exceeds the limit (<0 or > 1000), a hardware interrupt is immediately triggered, atomically redirecting the execution flow to pure rule scoring mode, ensuring zero system downtime.

[0078] The system code is written according to the C99 standard, with compiler integer optimization enabled and dynamic memory allocation (malloc / free) and recursive calls strictly disabled. Real-world testing on a Cortex-R5F@300MHz platform showed: total processing time per epoch ≤0.9ms, peak RAM usage ≤5.6KB, and Flash model area usage ≤11.8KB, fully meeting the constraints of 10Hz real-time processing and onboard resources.

[0079] The beneficial effects of this invention are: (1) In the second-level arbitration, the rule engine score and the GBDT model inference score are weighted and fused. The rule engine ensures the deterministic evaluation of the basic quality dimensions through concatenated bitmasks, and the GBDT model provides supplementary representation of the nonlinear relationship between features. The combination of the two can improve the rationality of the comprehensive score.

[0080] (2) The structural parameters, ground training process, and lightweight deployment method of the GBDT model were clarified. The model is embedded in Flash in the form of C language integer judgment code. The inference process only requires integer comparison and addition and subtraction operations. An on-orbit update verification and automatic degradation mechanism for anomalies are configured.

[0081] (3) Progressive quality assessment is achieved by cascading quality level bitmasks. The high bit effectively implies that all low bit conditions have been met, ensuring that the high-resolution baseband reaches the preset threshold in each basic dimension, and avoiding the concealment of defects in other dimensions due to the prominence of a single indicator.

[0082] (4) The introduction of first-order low-pass smoothing, hysteresis threshold, minimum hold time and ping-pong switching suppression mechanism in the switching decision reduces the probability of frequent switching caused by single-epoch signal jitter, which is conducive to maintaining the continuity of the input data of the orbit determination filter.

[0083] (5) A graded degradation and keep-alive strategy is set up for the extreme condition of full baseband failure. The degradation process is performed progressively according to the number of consecutive invalid epochs to ensure that each epoch has a valid data frame output and reduce the risk of navigation reference interruption.

[0084] (6) Using integer arithmetic and bit operations, the AI ​​inference model is stored in the form of integer constants, and the time consumed per epoch does not exceed 1ms. The amount of code and RAM usage are both controlled within a small range, making it suitable for deployment on resource-constrained spaceborne embedded processors.

[0085] Those skilled in the art will understand that the threshold parameters in the above embodiments (such as N=10, M=5, score difference 80, epoch interval 30, low-pass coefficient 3 / 2, etc.) are only preferred examples and can be adaptively adjusted according to specific satellite orbit characteristics, baseband chip specifications, mission reliability requirements, and space environment data. The bit width definition of the concatenated mask, the number and depth of trees in the GBDT, and the range of values ​​for the fusion weight α all fall within the scope of technical feature substitution disclosed in this invention. The integer operations can use int16_t or int32_t data types, and the feature input can use fixed-point scaling (such as ×10, ×100) to retain decimal precision.

[0086] This invention deeply couples AI algorithms with the characteristics of spaceborne embedded hardware, utilizing cascaded masks, integer GBDT, timing smoothing hysteresis, and hierarchical degradation mechanisms to construct a complete technical closed loop. All steps can be implemented using general-purpose microcontrollers, DSPs, or FPGA soft cores, and the data flow, state machine logic, memory layout, and communication interfaces are fully disclosed. Those skilled in the art can reproduce this invention and achieve the described technical effects without creative effort based on this specification and the accompanying drawings.

[0087] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time AI-based optimization method for multi-GNSS baseband low-orbit satellites, characterized in that, The method, applied to the central processor of a low-Earth orbit satellite GNSS receiver, includes the following steps in each solution epoch: Level 1 arbitration: Perform multi-dimensional hard veto judgment on each received GNSS baseband data. If any preset veto condition is triggered, mark the baseband epoch as invalid. Second-level arbitration: For valid baseband data that has passed the first-level arbitration, rule scoring and AI inference scoring are performed respectively, and the scores are merged to generate a single-epoch comprehensive score; The rule scoring adopts a multi-dimensional cascaded quality level bitmask mechanism: a cascaded quality level bitmask is preset from low to high bits, and the high bits are valid, implying that all low-bit conditions corresponding to that bit have been met. The basic score is obtained based on the highest valid bit in the quality level bitmask, and continuous increments are added. The AI ​​inference scoring is based on the micro gradient boosting decision tree GBDT model, which is stored in the form of an integer constant array. The inference process only performs integer comparison and addition / subtraction operations. Level 3 arbitration: Perform timing smoothing on the comprehensive score of each baseband, and perform hierarchical switching arbitration based on the smoothing score to output the selected candidate primary baseband; Level 4 Arbitration: Monitor the overall baseband validity status. If the entire baseband fails, execute a graded degradation and keep-alive strategy based on the number of consecutive invalid epochs to determine the baseband data to be output as a fallback. Real-time orbit determination and data output: Real-time orbit determination is performed using the raw observations provided by the selected primary baseband. The orbit determination results and baseband data are framed to generate data packets, which are then output.

2. The method according to claim 1, characterized in that, In the first-level arbitration, the preset veto conditions cover the dimensions of data integrity, positioning validity, time reference, geometric configuration, hardware health, observation real-time performance, observation accuracy, and signal stability. The method further includes: for intermittently faulty basebands with fewer than M valid epochs within the last N epochs, forcibly marking them as invalid for the current epoch.

3. The method according to claim 1, characterized in that, In the rule-based scoring, the bit-setting logic of the cascaded quality level bitmask is as follows: The baseband data quality index is checked bit by bit starting from the least significant bit. If the condition corresponding to the current bit is met and all the less significant bits are set, then the current bit is set. The basic score is obtained by looking up the table based on the bit sequence number of the most significant bit in the quality level bitmask. The continuous quantity bonus includes bonuses for the number of satellites, average carrier-to-noise ratio, shortest continuous lock time, and continuous positioning stability, with upper limits for each bonus.

4. The method according to claim 1, characterized in that, In the AI ​​reasoning scoring, the GBDT model, after being trained offline on the ground, converts the decision tree structure into nested judgment statements in C language; The input feature vector of the GBDT model includes: effective number of satellites, average carrier-to-noise ratio, shortest continuous lock time, PDOP value, pseudorange residual, maximum age of observation measurement, baseband self-check health status, and number of effective epochs for continuous positioning.

5. The method according to claim 1, characterized in that, The fusion process in the second-level arbitration adopts a weighted fusion method, and the calculation formula is as follows: ； in, For comprehensive scoring, Rate the AI ​​reasoning. Rate the rules For fusion weighting coefficients; The fusion weight coefficient Supports uploading and updating via the satellite-to-ground telemetry and control link.

6. The method according to claim 1, characterized in that, The third level of arbitration includes: Timing smoothing: A first-order low-pass filter is applied to the overall score of each baseband to obtain a smoothed score; Tiered Arbitration Switching: If the current primary baseband fails the first-level arbitration, a forced handover arbitration will be executed to switch to the effective baseband with the highest smoothing score. If the current primary baseband is valid, the dominant handover arbitration will be performed only if the candidate baseband with the highest smoothing score simultaneously meets the preset score difference threshold, the number of consecutive valid epochs threshold, the shortest consecutive lock time threshold, and the epoch interval threshold from the last baseband switch. Ping-pong handover suppression: Real-time monitoring of the number of baseband handovers within the last P epochs. If the number of handovers exceeds the preset threshold, the minimum handover interval threshold is temporarily increased.

7. The method according to claim 1, characterized in that, The graded downgrade and survival strategy in the fourth-level arbitration includes: Level 1 Degradation: When the number of consecutive invalid epochs is less than the set threshold, maintain the primary baseband identifier of the previous epoch and output the latest historical valid data of the baseband cache; Secondary degradation: When the number of consecutive invalid epochs reaches a set threshold, the cached data of the one with the best smoothing score among all baseband historical valid data is forcibly output; Level 3 Degradation: If there is no valid historical data, output the current data of the default baseband and mark the navigation data as invalid.

8. The method according to claim 1, characterized in that, The data in the data packet includes: Location results; orbit determination results; raw observations; filtered status fields: including primary baseband ID, smoothing score, switching reason code, data status flag, quality level bitmask, AI inference score, and fusion weight.

9. A low-orbit satellite multi-GNSS baseband AI real-time optimization system, used to execute the method according to any one of claims 1-8, characterized in that, The central processor deployed in a low-Earth orbit satellite GNSS receiver includes: The first-level arbitration module is used to perform multi-dimensional hard veto judgments and complete the validity marking and invalid data filtering of baseband data. The second-level arbitration module has a built-in rule scoring unit and an AI inference unit. The rule scoring unit is configured to execute the cascaded quality level bitmask setting and basic score mapping. The AI ​​inference unit is configured to load the GBDT model in the form of an integer constant array and execute integer inference. The scores of the two units are merged to output a single epoch comprehensive score. The third-level arbitration module is used to perform scoring timing smoothing, hierarchical switching arbitration and ping-pong switching suppression, and output the selected candidate primary baseband ID; The fourth-level arbitration module is used to monitor the overall baseband effectiveness status and execute a graded degradation and keep-alive strategy in the event of a full baseband failure. Real-time orbit determination module: used to receive raw observations from the selected baseband and perform real-time precise orbit determination calculations; Data Packaging Output Module: Used to frame and output the positioning results, orbit determination results, raw observations and filtered status fields of the selected baseband.

10. The system according to claim 9, characterized in that, It also includes a model management module, which connects to the second-level arbitration module, for: Fixed storage AI model parameter and version management; The model parameter updates are received via the satellite-to-ground telemetry and control link, and after integrity verification, they are written to the backup storage area. Switch the effective new model using atomic pointers; Monitor AI inference time and output range, and automatically switch to pure rule scoring mode when an anomaly occurs.

Images