A high-robustness satellite time fast estimation method based on multi-source fusion

Abstract

The present application relates to the field of satellite navigation and timing technology, and more particularly to a high-robustness satellite time fast estimation method based on multi-source fusion, comprising the following steps: S1, collecting satellite clock historical deviation data, multi-source redundant signals and signal quality indicators, and carrying out denoising, outlier rejection and standardization processing on the data to obtain reliable input for prediction; S2, based on the preprocessed data, using a deep learning model combined with clock physical constraints to predict the nonlinear drift of the satellite clock and generate the next time deviation prediction value. In the present application, through the closed-loop mechanism of multi-source data acquisition and preprocessing, deep learning prediction, credibility evaluation, weighted fusion and robust correction, the fast estimation of satellite clock deviation is realized, thereby improving the problem that the traditional method mostly relies on a single signal source, and due to the lack of discrimination and processing of abnormal signals, the timing result is unstable.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of satellite navigation and timing technology, and particularly relates to a high-robustness satellite time fast estimation method based on multi-source fusion. BACKGROUND

[0002] With the continuous expansion of satellite navigation and timing system applications in key fields such as communication, power, finance, and transportation, the precision and reliability of time synchronization have become the core factors to ensure stable operation of the system. Satellite clock signals are easily affected by ionospheric disturbances, blockage, multipath effects, and malicious interference in complex environments, resulting in deviations or abnormalities in observation data.

[0003] Currently, traditional methods mostly rely on a single signal source, and due to the lack of discrimination and processing of abnormal signals, the timing results are unstable. SUMMARY

[0004] In order to make up for the above shortcomings, the present application provides a high-robustness satellite time fast estimation method based on multi-source fusion, aiming to improve the problem that traditional methods mostly rely on a single signal source, and due to the lack of discrimination and processing of abnormal signals, the timing results are unstable.

[0005] The present application provides the following technical solutions, a high-robustness satellite time fast estimation method based on multi-source fusion, comprising the following steps:

[0006] S1, collect satellite clock historical deviation data, multi-source redundant signals and signal quality indicators, and perform denoising, outlier rejection and standardization processing on the data to obtain reliable input for prediction;

[0007] S2, based on the preprocessed data, use a deep learning model combined with clock physical constraints to predict the nonlinear drift of the satellite clock, and generate a next-time deviation prediction value;

[0008] S3, perform credibility evaluation and fraud detection on the multi-source signals, extract signal features and establish a credibility model to obtain the credibility of each signal source for weighted fusion;

[0009] S4, weight fuse the prediction value and each signal source observation value according to the credibility, and dynamically adjust the weight to obtain a fusion result;

[0010] S5, apply a robust correction method to the fusion result to handle abnormal disturbances and model uncertainties to obtain the final satellite clock deviation estimation value;

[0011] S6, output the satellite clock deviation estimation value corrected by the robustness for timing, wherein each step forms a closed-loop cooperative mechanism, and the output of the previous step is used as the input or depends on the next step for multi-source fusion satellite time fast estimation.

[0012] By adopting the technical scheme, the multi-source data acquisition and preprocessing, deep learning prediction, reliability evaluation, weighted fusion and robust correction closed-loop mechanism are adopted, and then the rapid estimation of satellite clock bias is realized, so that the problem that the traditional method mostly relies on a single signal source, and due to the lack of discrimination and processing of abnormal signals, the problem of unstable time service result is solved.

[0013] Preferably, the data acquisition and preprocessing comprises:

[0014] Acquiring satellite clock historical bias records, multi-source redundant observation signals and corresponding signal quality indicators;

[0015] The collected data is first subjected to wavelet threshold denoising and Kalman smoothing processing, Hampel anomaly detection is performed on the time series data, and observations exceeding the statistical threshold are removed;

[0016] The cleaned time series data is subjected to zero-mean variance standardization and a unified time series sample sequence is constructed and output to the prediction module.

[0017] Preferably, the physical constraint combination comprises:

[0018] The prediction input of the deep learning model is loaded in parallel with the preset satellite clock physical constraint model, and the physical constraint model comprises drift rate and drift acceleration boundary expression;

[0019] The residual of the prediction output and the physical constraint is calculated, and a constrained least squares problem is constructed based on the constraint residual to solve the constraint correction;

[0020] The prediction value optimized by the constraint is output and the value is transmitted to the fusion module.

[0021] Preferably, the non-linear drift prediction comprises:

[0022] The preprocessed standardized time series sample and environmental feature vector are input into the trained deep learning model;

[0023] The model generates an initial prediction using a recurrent neural network structure and uses a physical regularization term as a loss constraint in the inference process;

[0024] The next moment satellite clock bias prediction value is output and the prediction confidence index is recorded.

[0025] Preferably, the historical reliability and environmental condition combination comprises:

[0026] The historical continuity, packet loss rate and time delay jitter of each signal source are counted as reliability features, and the current ionospheric disturbance index, shielding indication and multipath index are collected as environmental features;

[0027] The historical reliability features and environmental features are input into a credibility modeler, and a Bayesian network structure is used to calculate the posterior reliability probability of each signal source;

[0028] The dynamic credibility index of each signal source at the current time is output for use by a credibility evaluation module.

[0029] Preferably, the signal credibility evaluation and fraud detection includes:

[0030] The amplitude, phase, code correlation, instantaneous signal-to-noise ratio, and timing consistency feature vectors are extracted from multi-source observations;

[0031] The feature vectors are input into a fraud detector and a credibility model, and an anomaly detection algorithm is used to determine abnormal signals and adjust the prior probability of the corresponding signals from the credibility model;

[0032] The detected and adjusted credibility values of each signal source are output, and the signal sources determined to be abnormal are marked.

[0033] Preferably, the weighted fusion includes:

[0034] The deep learning prediction value and the observation value and its credibility value of each signal source are received as input;

[0035] The original weights are assigned to each observation item and prediction item according to the credibility value and normalized, and the weight assignment is adjusted online according to the proximal residual using an exponential decay mechanism;

[0036] The weighted linear combination is performed according to the normalized weights, and the preliminary fusion result and weight update record are output.

[0037] Preferably, the robust correction includes:

[0038] The preliminary fusion result and the corresponding observation residual time sequence are input into a robust correction module;

[0039] An H∞ filter is designed in the robust correction module, and an iterative reweighted M-estimation is combined to perform state update operations on the input;

[0040] The satellite clock bias estimation value after robust correction is output, and the filter internal state parameters are updated.

[0041] Preferably, the closed-loop cooperation includes:

[0042] The final estimation value and residual time sequence are appended to the historical data set to form new training and reference samples;

[0043] The model bias index is calculated based on the new samples, and the bias term of the deep learning model and the prior parameter of the credibility model are adjusted through an online parameter update algorithm;

[0044] The updated model parameters and the updated historical data set are outputted for use in data collection and preprocessing in the next cycle, forming a closed loop cycle.

[0045] The present application has the following advantages:

[0046] 1、In the present application, through the closed loop mechanism of multi-source data collection and preprocessing, deep learning prediction, credibility evaluation, weighted fusion and robust correction, the rapid estimation of satellite clock bias is realized, thereby improving the problem that traditional methods mostly rely on a single signal source, and due to the lack of discrimination and processing of abnormal signals, the problem of unstable time service results is caused.

[0047] 2、In the present application, by denoising, outlier rejection and standardization processing of the collected data, stable prediction input is generated, thereby improving the problem that traditional methods mostly directly use original observation data, and due to the unprocessed noise and outliers, the problem of poor reliability of input data is caused.

[0048] 3、In the present application, by using a deep learning model combined with clock physical constraints for prediction, a bias prediction value at the next time is generated, thereby improving the problem that traditional methods mostly rely on linear or empirical models, and due to the inability to effectively handle nonlinear drift, the problem of insufficient prediction accuracy is caused.

[0049] 4、In the present application, by applying robust correction to the fusion result and forming a closed loop cooperative mechanism, a stable bias estimation value is outputted, thereby improving the problem that traditional methods mostly lack abnormal disturbance processing mechanism, and due to the uncorrected model uncertainty, the problem of unstable time service results is caused. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 A method flowchart of a high-robustness satellite time rapid estimation method based on multi-source fusion is provided for the present application;

[0051] Figure 2 A data collection and preprocessing flowchart of a high-robustness satellite time rapid estimation method based on multi-source fusion is provided for the present application;

[0052] Figure 3 A credibility evaluation and fraud detection flowchart of a high-robustness satellite time rapid estimation method based on multi-source fusion is provided for the present application;

[0053] Figure 4 A weight dynamic adjustment and fusion flowchart of a weighted fusion module of a high-robustness satellite time rapid estimation method based on multi-source fusion is provided for the present application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of the present application.

[0055] Embodiment one:

[0056] In the first embodiment of the present application, the present application provides a high-robustness satellite time fast estimation method based on multi-source fusion, as shown in Figures 1-4 The method comprises the following steps:

[0057] S1, collecting satellite clock historical deviation data, multi-source redundant signals and signal quality indicators, and performing denoising, outlier rejection and standardization processing on the data to obtain reliable input for prediction;

[0058] Further, the data collection and preprocessing comprises:

[0059] Collecting satellite clock historical deviation records, multi-source redundant observation signals and corresponding signal quality indicators;

[0060] The collected data is first subjected to wavelet threshold denoising and Kalman smoothing processing, Hampel anomaly detection is performed on the time series data, and observations exceeding the statistical threshold are rejected;

[0061] The cleaned time series data is subjected to zero-mean variance standardization and a unified time series sample sequence is constructed and output to the prediction module.

[0062] Specifically, satellite clock historical deviation records, multi-source redundant observation signals and corresponding signal quality indicators are collected, and the historical deviation records are represented as a sequence: ; wherein is the satellite clock deviation at the th moment. The multi-source redundant observation signals are a set of observations from different signal sources: ; wherein is the observation value of the th signal source at the current moment. The signal quality indicators are a set of quality parameters such as signal-to-noise ratio, packet loss rate, code tracking error corresponding to each observation signal: , and the input data stream is . The data input to the prediction model contains clock historical state, multi-source redundant information and quality constraints, providing a basis for subsequent denoising and modeling.

[0063] Wavelet threshold denoising and Kalman smoothing processing are performed on the collected data. Wavelet threshold denoising: the input sequence is decomposed by wavelet to obtain high-frequency and low-frequency coefficients, and a threshold function is used:

[0064] ;in These are wavelet coefficients. The threshold is used as the threshold, and the output is the denoised coefficient. .

[0065] Kalman smoothing, based on state equations: ;in For state vectors, For observation purposes, , These are process noise and observation noise, respectively. A smoothed estimation sequence is obtained through forward filtering and backward smoothing calculations. High-frequency random noise in the observed signal is removed, and temporal smoothing is used to reduce the impact of abrupt changes on the estimation, thus maintaining signal continuity.

[0066] Perform Hampel anomaly detection and remove observations exceeding the statistical threshold for a length of [length missing]. Using a sliding window, calculate the median within the window. And absolute median: .

[0067] Judgment criteria: If If a point is not found to be an anomaly, it will be removed from further processing. Anomalies caused by signal spoofing, obstruction, or sudden interference should be eliminated to prevent them from entering the standardization and modeling process and causing deviations.

[0068] The cleaned time series data are standardized to zero mean and variance using the following formula: ;in The mean, Standard deviation These are the cleaned sample values. These are the standardized values. This standardizes the numerical magnitude differences between different signal sources, ensuring that the input data is comparable within the model.

[0069] A unified time-series sample sequence is constructed and output to the prediction module. The output is as follows:

[0070] Each element represents a time-series sample after cleaning, denoising, and standardization. The output data stream is generated by... Converted into a standardized time series dataset Generate a unified sequence of input samples to ensure that the subsequent deep learning prediction module can receive a data stream with a consistent format.

[0071] S2. Based on the preprocessed data, use a deep learning model combined with clock physical constraints to predict the nonlinear drift of the satellite clock and generate the predicted deviation value for the next moment.

[0072] Furthermore, the physical constraint combination includes:

[0073] The prediction input of the deep learning model is loaded in parallel with the preset satellite clock physical constraint model, which includes the boundary expressions of drift rate and drift acceleration.

[0074] Calculate the residual between the predicted output and the physical constraints, and construct a constrained least squares problem based on the constraint residuals to solve the constraint correction problem;

[0075] The system outputs the constrained and optimized predicted value and passes it to the fusion module.

[0076] Nonlinear drift prediction includes:

[0077] The preprocessed standardized time series samples and environmental feature vectors are input into the trained deep learning model;

[0078] The model uses a recurrent neural network structure to generate initial predictions and uses physical regularization terms as loss constraints in the inference process.

[0079] Output the predicted satellite clock deviation value for the next moment and record the prediction confidence index.

[0080] Specifically, physical constraints are combined to standardize the input sequence output by S1. Loaded in parallel with the satellite clock physical constraint model. The physical constraint model is defined as the boundary between drift velocity and drift acceleration, expressed as: ;in Due to clock skew, For time intervals, The deviation drift rate, This is the drift acceleration.

[0081] Predicted output generated based on deep learning model Constructing residuals with respect to physical constraints:

[0082] ;in This is an estimate under physical constraints.

[0083] By correcting the residuals, the prediction problem is transformed into a restricted least squares problem:

[0084] After solving, the constrained corrected prediction values ​​are obtained. And output it to the fusion module.

[0085] Nonlinear drift prediction uses preprocessed standardized time series samples. and environmental feature vectors Input a pre-trained deep learning model. This model uses a Long Short-Term Memory (LSTM) network structure, and its state update equation is: ;in The current hidden state. This represents the state of the memory cell from the previous time step. The time dependency of satellite clock changes is preserved through recursive calculation.

[0086] The model constructs its loss function during the training phase using physical constraint regularization terms:

[0087] ;in , For regularization weights, and The desired drift rate and acceleration reference values ​​are defined by minimizing... Ensure that the prediction results remain consistent between the model learning and the physical laws.

[0088] During the inference phase, the model outputs the prediction bias value for the next time step based on the input sequence:

[0089] And calculate the corresponding prediction confidence level. The confidence level is calculated based on the residual distribution of the model output: ;in Output the mean for the model. This represents the historical prediction variance. Confidence level indicates the degree of reliability of the prediction.

[0090] Input and Output Flow: During execution, the system first receives standardized time-series samples, environmental feature vectors, and physical constraint parameters from the preprocessing module, including upper and lower limits for drift velocity and drift acceleration. After the input data is loaded into the model, the neural network generates preliminary predictions. The prediction results are then compared with the output of the physical model to calculate residuals, forming constraint-corrected inputs. Subsequently, a restricted least squares problem is solved based on the boundary conditions to obtain the physically-constrained corrected predictions. The corrected predictions then undergo a confidence calculation stage, ultimately outputting the deviation prediction results and corresponding confidence indices for use by the subsequent weighted fusion module.

[0091] S3. Perform credibility assessment and deception detection on multi-source signals, extract signal features and establish a credibility model to obtain the credibility of each signal source for weighted fusion.

[0092] Furthermore, the combination of historical reliability and environmental conditions includes:

[0093] The historical continuity, packet loss rate, and time delay jitter of each signal source are statistically analyzed as reliability characteristics, and the current ionospheric disturbance index, blockage indication, and multipath index are collected as environmental characteristics.

[0094] Historical reliability features and environmental features are input into the credibility modeler, and a Bayesian network structure is used to calculate the posterior reliability probability of each signal source.

[0095] Output the dynamic credibility index of each signal source at the current moment for use by the credibility assessment module.

[0096] Signal credibility assessment and spoofing detection include:

[0097] Extract feature vectors of amplitude, phase, code correlation, instantaneous signal-to-noise ratio, and temporal consistency from multi-source observations;

[0098] The feature vector is input into the deception detector and the credibility model. Anomaly detection algorithm is used to identify abnormal signals and the prior probability of the corresponding signal is adjusted from the credibility model.

[0099] Output the confidence values ​​of each signal source after detection and adjustment, and mark the signal sources that are judged to be abnormal.

[0100] Specifically, historical reliability is combined with environmental conditions to statistically analyze the continuity indicators of each signal source in past observations. Packet loss rate latency jitter Constructing the historical reliability feature vector:

[0101] ;in Indicates the signal source number.

[0102] Collect current ionospheric disturbance index Obstruction indicator Multipath interference index This forms an environmental feature vector: .

[0103] Will and Input the credibility modeler. The credibility modeler uses a Bayesian network structure to calculate the posterior reliability probability of the signal source based on conditional probability: ;in This indicates that the signal source is reliable. This indicates that the information is unreliable.

[0104] Output the dynamic reliability index of each signal source at the current moment: This indicator is then passed to the credibility assessment module.

[0105] Signal reliability assessment and spoofing detection: Extracting signal features from multi-source observations, including amplitude. Phase Code correlation Instantaneous signal-to-noise ratio Time sequence consistency index This constitutes the signal feature vector: .

[0106] Will Input the deception detector and the credibility model. The deception detector is based on anomaly detection algorithms (statistical threshold detection or distribution-fit-based detection) and calculates feature bias: ;in This represents the mean feature vector of the training samples. When If the signal is deemed abnormal, its prior probability is adjusted in the credibility model. ;in This is the inhibition coefficient.

[0107] Based on the detected and corrected model, output the confidence values ​​of each signal source: The signal sources that are determined to be abnormal are marked for use by the subsequent weighted fusion module.

[0108] Input / output flow, Input: Historical reliability characteristics Environmental characteristics Multi-source signal feature vector Model processing: The credibility modeler calculates the initial credibility through a Bayesian network. The deception detector identifies anomalies and adjusts the prior probabilities accordingly, updating the confidence level. Output: The dynamic confidence level of each signal source at the current moment. and the results of abnormal signal marking.

[0109] S4. The predicted values ​​and the observed values ​​of each signal source are weighted according to their confidence level and the weights are dynamically adjusted to obtain the fusion result.

[0110] Furthermore, weighted fusion includes:

[0111] It receives deep learning predictions and observations from each signal source, along with their confidence values, as inputs.

[0112] The original weights of each observation and prediction item are assigned and normalized based on the confidence level values. The weight allocation is then adjusted online using an exponential decay mechanism based on the near-end residuals.

[0113] Perform a weighted linear combination based on normalized weights and output the preliminary fusion results and weight update records.

[0114] Specifically, weighted fusion receives deep learning predictions. Observations of each signal source and the corresponding credibility value. ,in The input set is: .

[0115] Based on credibility values Assign original weights to each observation and prediction term. The weights of the prediction values ​​are denoted as... The weights are typically set to a fixed baseline or dynamically based on historical performance. The original weights for observed and predicted values ​​are: .

[0116] The original weights are normalized to obtain the standardized weights: .

[0117] For each time step, calculate the residuals between the fused output and each input. The initial estimate of the fused output is: The residual is: .

[0118] The weights are adjusted based on the residuals using an exponential decay mechanism, and the update formula is as follows: ;in This is the attenuation factor.

[0119] Normalize the updated weights again: .

[0120] The final weighted linear combination yields the fusion result: .

[0121] The output includes the fusion result. And the weight update record for this period. This is for use by the subsequent robust correction module.

[0122] Input / output flow: Input: Predicted value Observations of each signal source Credibility index Model processing: Initial weights are generated based on confidence level, preliminary fusion results are calculated, the weights are adjusted using the residual exponential decay mechanism, normalized again, and the final weighted combination value is output. Output: Fusion result Weight update record.

[0123] S5. Apply a robust correction method to the fusion results to handle abnormal disturbances and model uncertainties, and obtain the final satellite clock deviation estimate.

[0124] Furthermore, robustness modifications include:

[0125] The preliminary fusion results and the corresponding observation residual time series are input into the robust correction module;

[0126] In the robust correction module, an H∞ filter is designed and combined with iterative reweighted M-estimation to perform state update operations on the input;

[0127] Output a robustly corrected estimate of the satellite clock offset and update the internal state parameters of the filter.

[0128] Specifically, robust correction will be applied to the preliminary results obtained through weighted fusion. The residual sequence between each observation item Also input the robust correction module. The input dataset is: .

[0129] Establish a state-space model in the robust correction module: ;in Indicates the system status. Indicates observation, and These are process noise and observation noise, respectively.

[0130] The gain matrix is ​​designed using the H∞ filtering method. The objective function for filtering is: ;in To estimate the output, For performance metrics. Through constraints Calculate the upper bound of the filter gain that satisfies the condition. .

[0131] Iterative reweighted M-estimation is introduced during the filtering update process to suppress anomalous perturbations. The M-estimation employs a loss function... The weights are updated as follows: ;in , The residuals are used to reduce the impact of large residuals on the estimation by iteratively calculating weights and re-estimating the state.

[0132] By combining the results of H∞ filtering and M estimation, the state estimate is updated: ;in The weights are adjusted based on the M estimate.

[0133] Output the final estimated satellite clock offset: And update the internal state parameters of the filter. With weight function Save to the next cycle.

[0134] Input / output flow: Preliminary fusion results Observation residual sequence Module processing: State updates are performed within the H∞ filter framework, and residual weights are dynamically adjusted using iterative reweighted M-estimation. Output: The final satellite clock offset estimate. , and the updated internal state parameters of the filter.

[0135] S6. Output the robustly corrected satellite clock deviation estimate for time synchronization. Each step forms a closed-loop collaborative mechanism, with the output of the previous step serving as the input or dependency for the next step, to perform rapid satellite time estimation through multi-source fusion.

[0136] Furthermore, closed-loop collaboration includes:

[0137] The final estimate and residual time series are appended to the historical dataset to form new training and reference samples;

[0138] The model bias index is calculated based on the new samples, and the bias terms of the deep learning model and the prior parameters of the credibility model are adjusted through an online parameter update algorithm.

[0139] The updated model parameters and the updated historical dataset are output for use in the next cycle of data collection and preprocessing, forming a closed loop.

[0140] Specifically, the final estimated satellite clock offset is denoted as... ,in This represents the observation period at the current moment. This estimate is compared with the corresponding residual time series. Add to historical dataset Updated to: This historical dataset serves as input for subsequent model training and reference.

[0141] Calculate the model bias index based on the new samples. For example, mean squared deviation can be used:

[0142] ;in For reference clock value, This represents the number of samples. This metric is used to measure the degree of bias in a deep learning model at its current stage.

[0143] The bias terms of the deep learning model are updated using an online parameter update algorithm. Prior parameters of the credibility model Adjustments will be made: ;in For learning rate, This represents the gradient with respect to the parameters.

[0144] After the parameters are updated, the system outputs the updated model parameters. With new historical datasets This data is then input into the data acquisition and preprocessing module for the next cycle. The input-output relationship is: input, final estimated value. residual Historical datasets ; Operations: add data, calculate deviation indicators, update model parameters; Output: updated parameter set, updated historical dataset.

[0145] This closed-loop collaborative mechanism forms a periodic cycle of data acquisition, bias estimation, robust correction, and model updating, enabling the satellite clock bias estimation to continuously and adaptively approach the true value.

[0146] Example 2:

[0147] In applications involving national time service centers and ground reference station networks, clock signals from multiple navigation satellites are needed to provide unified and continuous time services for power dispatching systems, financial settlement systems, and critical infrastructure. Satellite signals collected by base stations in various locations are affected by ionospheric disturbances, multipath effects, and signal blockage. Observations from a single satellite are prone to abnormal noise, making it difficult to guarantee stability and reliability. While multi-source observations can enhance redundancy, significant differences in signal quality among sources during data fusion mean that credibility modeling and weight allocation are insufficient to eliminate accumulated biases, easily affecting the accuracy of the final clock bias estimate. Existing robust correction methods have high computational complexity under large-scale data processing, and filters are at risk of divergence under strong abnormal disturbances, making it difficult to guarantee real-time performance. During closed-loop collaborative updates, the accumulation of historical data and online adjustment of model parameters lag, and the prior knowledge of deep learning predictions and credibility models cannot adapt to dynamic environmental changes in a timely manner. This leads to shifts in time service results during continuous operation, failing to meet the high accuracy and reliability requirements of power dispatching and financial settlement systems. To address the aforementioned problems, this invention provides a robust and fast satellite time estimation method based on multi-source fusion, the structure of which is as follows: Figure 1 As shown. The specific implementation process of this method is as follows:

[0148] Specifically, data acquisition and preprocessing involves receiving historical satellite clock deviation data, multi-source redundant signals, and signal quality indicators. Noise reduction is performed on the input data to remove outliers, and standardization operations are used to unify data from different sources and with different dimensions, resulting in effective input samples that can be used for subsequent predictions.

[0149] Nonlinear drift prediction is used to invoke a trained deep learning model based on preprocessed input data. During the prediction process, a physical constraint model for the satellite clock is introduced in parallel, including boundary conditions for drift rate and drift acceleration. After the deep learning model outputs initial predictions, the residual between the prediction results and the physical constraints is calculated, and the predictions are corrected to output a predicted satellite clock deviation for the next time step that meets the physical boundary conditions.

[0150] Signal reliability assessment and spoofing detection are used to analyze multi-source signals. Features such as amplitude, phase, instantaneous signal-to-noise ratio, code correlation, and timing consistency are extracted and combined with historical continuity, packet loss rate, latency jitter, ionospheric disturbance, and multipath indicators. These are input into a reliability modeler, which calculates the posterior reliability probability of each signal source using a Bayesian network. The spoofing detector adjusts the prior probabilities of the reliability model based on anomalies in the features, outputting dynamic reliability indicators for each signal source and marking anomalous signal sources.

[0151] Weighted fusion receives deep learning predictions, observations from each signal source, and their corresponding confidence scores. It assigns original weights based on the confidence scores and normalizes them, then adjusts them online based on the residual sequence. Each observation and prediction is weighted and linearly combined, outputting preliminary fusion results and weight update records.

[0152] Robust correction is used to receive the preliminary fusion results and residual timing. It uses an H∞ filter combined with an iterative reweighted M algorithm to update and correct the state of the input data, outputs the robustly processed satellite clock deviation estimate, and updates the internal parameters of the filter to enhance the system's ability to suppress disturbances and uncertainties.

[0153] The output is used to provide the robustly corrected satellite clock offset estimate to the time synchronization system, and to append this estimate and residual time series to the historical dataset to form new training and reference samples. Based on the new samples, the parameters of the deep learning model and the reliability model are updated online, so that the data acquisition and prediction in subsequent cycles can be gradually corrected and form a closed loop, realizing rapid time synchronization under multi-source fusion conditions.

[0154] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. 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 robust satellite time estimation method based on multi-source fusion, characterized in that, Includes the following steps: S1. Collect historical satellite clock deviation data, multi-source redundant signals and signal quality indicators, and perform noise reduction, outlier removal and standardization on the data to obtain reliable input for prediction. S2. Based on the preprocessed data, use a deep learning model combined with clock physical constraints to predict the nonlinear drift of the satellite clock and generate the predicted deviation value for the next moment. S3. Perform credibility assessment and deception detection on multi-source signals, extract signal features and establish a credibility model to obtain the credibility of each signal source for weighted fusion. S4. The predicted values ​​and the observed values ​​of each signal source are weighted according to their confidence level and the weights are dynamically adjusted to obtain the fusion result. S5. Apply a robust correction method to the fusion results to handle abnormal disturbances and model uncertainties, and obtain the final satellite clock deviation estimate. S6. Output the robustly corrected satellite clock deviation estimate for time synchronization. Each step forms a closed-loop collaborative mechanism, with the output of the previous step serving as the input or dependency for the next step, to perform rapid satellite time estimation through multi-source fusion. The weighted fusion includes: It receives deep learning predictions and observations from each signal source, along with their confidence values, as inputs. The original weights of each observation and prediction item are assigned and normalized based on the confidence level values. The weight allocation is then adjusted online using an exponential decay mechanism based on the near-end residuals. Perform a weighted linear combination based on normalized weights and output the preliminary fusion results and weight update records; The robustness correction includes: The preliminary fusion results and the corresponding observation residual time series are input into the robust correction module; In the robust correction module, an H∞ filter is designed and combined with iterative reweighted M-estimation to perform state update operations on the input; Output a robustly corrected estimate of the satellite clock offset and update the internal state parameters of the filter; The closed-loop collaboration includes: The final estimate and residual time series are appended to the historical dataset to form new training and reference samples; The model bias index is calculated based on the new samples, and the bias terms of the deep learning model and the prior parameters of the credibility model are adjusted through an online parameter update algorithm. The updated model parameters and the updated historical dataset are output for use in the next cycle of data collection and preprocessing, forming a closed loop.

2. The robust satellite time estimation method based on multi-source fusion according to claim 1, characterized in that, The process of collecting historical satellite clock deviation data, multi-source redundant signals, and signal quality indicators, and performing noise reduction, outlier removal, and standardization on the data includes: Collect historical deviation records of satellite clocks, multi-source redundant observation signals, and corresponding signal quality indicators; The collected data were processed by wavelet thresholding and Kalman smoothing, and Hampel anomaly detection was performed on the time series data to remove observations that exceeded the statistical threshold. The cleaned time series data is standardized with zero mean and variance, and a unified time series sample sequence is constructed and output to the prediction module.

3. The robust satellite time estimation method based on multi-source fusion according to claim 1, characterized in that, The use of deep learning models combined with clock physical constraints includes: The prediction input of the deep learning model is loaded in parallel with the preset satellite clock physical constraint model, which includes the boundary expressions of drift rate and drift acceleration. Calculate the residual between the predicted output and the physical constraints, and construct a constrained least squares problem based on the constraint residuals to solve the constraint correction problem; The system outputs the constrained and optimized predicted value and passes it to the fusion module.

4. The robust satellite time estimation method based on multi-source fusion according to claim 1, characterized in that, The process of predicting the nonlinear drift of the satellite clock and generating the predicted deviation value for the next time step includes: The preprocessed standardized time series samples and environmental feature vectors are input into the trained deep learning model; The model uses a recurrent neural network structure to generate initial predictions and uses physical regularization terms as loss constraints in the inference process. Output the predicted satellite clock deviation value for the next moment and record the prediction confidence index.

5. The robust satellite time estimation method based on multi-source fusion according to claim 1, characterized in that, The process of extracting signal features and establishing a credibility model includes: The historical continuity, packet loss rate, and time delay jitter of each signal source are statistically analyzed as reliability characteristics, and the current ionospheric disturbance index, blockage indication, and multipath index are collected as environmental characteristics. Historical reliability features and environmental features are input into the credibility modeler, and a Bayesian network structure is used to calculate the posterior reliability probability of each signal source. Output the dynamic credibility index of each signal source at the current moment for use by the credibility assessment module.

6. The robust satellite time estimation method based on multi-source fusion according to claim 1, characterized in that, The credibility assessment and spoofing detection of multi-source signals includes: Extract feature vectors of amplitude, phase, code correlation, instantaneous signal-to-noise ratio, and temporal consistency from multi-source observations; The feature vector is input into the deception detector and the credibility model. Anomaly detection algorithm is used to identify abnormal signals and the prior probability of the corresponding signal is adjusted from the credibility model. Output the confidence values ​​of each signal source after detection and adjustment, and mark the signal sources that are judged to be abnormal.

Images