Satellite-borne interference source doppler observation strategy optimization method

Abstract

The application discloses a kind of satellite-borne interference source Doppler observation strategy optimization methods, comprising: obtaining original radio frequency signal and satellite orbit data, predicting the mutation characteristics in Doppler data, generating adaptive observation scheme;According to adaptive observation scheme, energy-aware resource scheduling is carried out and observation is executed, and enhanced Doppler observation data are obtained;Based on enhanced Doppler observation data, positioning solution and performance evaluation are carried out, and the parameters for iterative optimization observation strategy are generated.The application realizes the intelligent matching of observation resource and signal information content, improves resource utilization efficiency, key event capture capability and final positioning accuracy.

Description

TECHNICAL FIELD

[0001] The application relates to satellite communication technology, in particular to a method for optimizing a Doppler observation strategy of a satellite-borne interference source. BACKGROUND

[0002] In the field of satellite-borne wireless communication and non-cooperative signal positioning, low earth orbit (LEO) satellite platforms have a significant advantage in performing detection and observation tasks of ground interference sources due to their flexible orbital coverage. Among them, passive positioning of non-cooperative interference sources based on Doppler shift information is a key technical means in the application of wireless communication interference monitoring and positioning. In practical applications, the reception quality of the interference source signal and the frequency observation accuracy are influenced by a variety of factors. In order to improve the spectrum utilization efficiency and the observation accuracy of the interference source, optimizing the observation strategy of the satellite-borne platform has become the core and premise of improving the performance of the entire positioning system.

[0003] In the field of general wireless communication resource optimization, especially in the emerging communication network assisted by intelligent reflecting surface (IRS), in order to cope with complex signal environments, researchers have proposed a variety of advanced joint optimization methods. For example, some works use traditional optimization theories such as block coordinate descent (BCD) and Lagrange dual transformation to iteratively solve the base station beamforming and IRS phase shift. In order to cope with higher-dimensional state and action spaces, deep reinforcement learning (DRL) methods are introduced. For example, in a multi-user MISO system integrating unmanned aerial vehicle IRS, some researchers use the deep deterministic policy gradient (DDPG) algorithm to achieve joint optimization of IRS trajectory and reflection coefficient; another group of scholars designed a segmented DRL framework with segmented action space to improve the joint beamforming effect in order to avoid the DRL algorithm falling into local optimum. These methods have shown great potential in dealing with high dynamic and multi-variable resource allocation problems, and have provided a feasible technical path for solving complex wireless communication system optimization problems.

[0004] However, although there are advanced optimization frameworks in related fields, when these methods are directly applied to the specific scenario of satellite-borne interference source Doppler observation, there are still profound and unsolved technical bottlenecks. These bottlenecks mainly arise from the contradiction between the unique event-driven information structure of the Doppler observation task and the absolute scarcity of on-board resources. Specifically, the existing technology has serious deficiencies in the foresight and semantic depth of state perception of the observation strategy, resulting in blindness of resource allocation and inefficiency of key information capture. SUMMARY

[0005] The application aims to provide a method for optimizing the Doppler observation strategy of a satellite-borne interference source to solve the problems of blind observation resource allocation and inaccurate mutation feature prediction in the prior art.

[0006] The technical scheme is a satellite-borne interference source Doppler observation strategy optimization method, comprising:

[0007] Obtaining original radio frequency signals of the satellite-borne interference source and satellite precise orbit data, predicting mutation characteristics in Doppler data, and generating an adaptive observation scheme;

[0008] According to the adaptive observation scheme, energy-aware resource scheduling is performed and observation is executed to obtain enhanced Doppler observation data;

[0009] Based on the enhanced Doppler observation data, positioning solution and performance evaluation are performed to generate parameters for iterative optimization of the observation strategy.

[0010] According to an aspect of the present application, obtaining enhanced Doppler observation data comprises:

[0011] Identifying an energy accumulation window defined by a low elevation angle period and a high-value observation period defined by Doppler mutation characteristics in the adaptive observation scheme;

[0012] In the energy accumulation window, the energy saving value is accumulated by adjusting the working mode of the receiver, and is charged into an energy reserve unit;

[0013] When entering the high-value observation period, the matching degree of the balance of the energy reserve unit and the power consumption demand of the period is evaluated;

[0014] Under the condition that the balance of the energy reserve unit meets the power consumption demand, high-density sampling is authorized to be performed to obtain enhanced Doppler observation data.

[0015] According to an aspect of the present application, evaluating the matching degree of the balance of the energy reserve unit and the power consumption demand of the period comprises:

[0016] An energy-accuracy benefit function representing the relationship between observation accuracy and energy consumption is constructed;

[0017] Based on the energy-accuracy benefit function, a segmented benefit evaluation matrix is generated for different power consumption modes in the high-value observation period;

[0018] A multi-constraint optimization model is established with the goal of maximizing total benefit and the balance of the energy reserve unit and the mode switching cost as constraint conditions;

[0019] The multi-constraint optimization model is solved to obtain an optimal power consumption allocation scheme, and high-density sampling is performed accordingly.

[0020] According to an aspect of the present application, it further comprises:

[0021] When it is determined through evaluation that the balance of the energy reserve unit cannot meet the power consumption demand, a degradation strategy is started;

[0022] The degradation strategy calculates a ratio of the current remaining energy to the expected demand of the subsequent task, and compares the ratio with at least one preset degradation threshold value;

[0023] If the ratio is lower than the degradation threshold value, a degradation sampling plan aiming to reduce the total energy consumption is generated, the degradation sampling plan is implemented by increasing the sampling interval of non-critical observation areas or shrinking the spatial range of high-density sampling;

[0024] Observation is performed according to the degradation sampling plan.

[0025] According to one aspect of the present application, when the original radio frequency signal is lower than the preset threshold value, the process of obtaining enhanced Doppler observation data is implemented by using a step-by-step Doppler frequency cascade fine estimation method, specifically:

[0026] Preliminary spectral peak positioning is performed according to the frequency grid adaptively adjusted according to the spectral energy distribution, and a Doppler coarse estimate value is obtained;

[0027] Based on the Doppler coarse estimate value, a reverse Doppler sequence is injected, and a solution that minimizes the flatness of the compensated signal spectrum is searched, to obtain a Doppler medium estimate value;

[0028] Based on the Doppler medium estimate value, a virtual mirror signal is constructed and iteratively adjusted, and the time variation rate of the beat component generated after the cross-correlation of the virtual mirror signal and the original signal is minimized, to obtain a Doppler fine estimate value, and the enhanced Doppler observation data is obtained by sampling according to the Doppler fine estimate value.

[0029] According to one aspect of the present application, the Doppler fine estimate value is obtained by constructing a virtual mirror signal and iteratively adjusting, including:

[0030] The conjugate of the original radio frequency signal is used to construct a virtual mirror signal based on the Doppler medium estimate value;

[0031] The cross-correlation function between the original radio frequency signal and the virtual mirror signal is calculated, and the beat component is separated therefrom;

[0032] The frequency used to construct the virtual mirror signal is iteratively adjusted until the time variation rate of the beat component converges to a preset minimum value, and the converged frequency is determined as the Doppler fine estimate value.

[0033] According to one aspect of the present application, when non-uniform sampling is used when obtaining enhanced Doppler observation data, a data reconstruction process is further included, specifically:

[0034] The Doppler fine estimate value is obtained, and the non-uniformly sampled observation data is pre-compensated for Doppler using the Doppler fine estimate value, to obtain a compensated stationary signal;

[0035] Solving the pre-constructed compressed sensing convex optimization model for the compensated stationary signal to obtain a reconstructed stationary signal component;

[0036] Re-adding the frequency trend represented by the Doppler fine estimate to the reconstructed stationary signal component to generate enhanced Doppler observation data.

[0037] According to one aspect of the present application, the mutation characteristics in the Doppler data are predicted, including:

[0038] Performing spectral analysis on the original radio frequency signal, and combining satellite precise orbit data, extracting time series Doppler frequency shift data through a frequency tracking loop;

[0039] Performing time-frequency transformation on the time series Doppler frequency shift data to generate a time-frequency spectrum, and processing it into a gray-scale image;

[0040] Extracting local binary patterns and gray-level co-occurrence matrices from the gray-scale image as texture features to construct a multi-dimensional texture feature vector;

[0041] Feeding the multi-dimensional texture feature vector into a pre-trained convolutional neural network to analyze and output the mutation type and mutation time in the Doppler data.

[0042] According to one aspect of the present application, predicting the mutation characteristics in the Doppler data can also be:

[0043] Extracting time series Doppler frequency shift data from the original radio frequency signal and calculating the corresponding Doppler second derivative sequence;

[0044] Calling a historical transit mutation point knowledge base and combining the current satellite precise orbit data to generate an initial predicted mutation position;

[0045] Real-time monitoring of the Doppler second derivative sequence, when the measured characteristics of the sequence deviate from the expected characteristics of the initial predicted mutation position and exceed the preset threshold, triggering a correction mechanism;

[0046] The correction mechanism fuses the initial prediction and the measured characteristics to output a corrected mutation window.

[0047] According to one aspect of the present application, the correction mechanism includes:

[0048] Using a Kalman filter to fuse the initial predicted mutation position and the measured characteristics extracted from the Doppler second derivative sequence;

[0049] The Kalman filter outputs an optimized corrected mutation position and a confidence weight coefficient representing the credibility of the corrected position;

[0050] According to the corrected mutation position and the confidence weight coefficient, the boundaries of the corrected mutation window are adaptively set.

[0051] The application introduces a forward-looking mutation prediction mechanism, solves the problem of resource allocation blindness caused by the inability to predict key events, can accurately predict future high-value observation windows, and drives the energy pool and non-uniform sampling strategy to perform optimal planning and targeted deployment of energy and sampling resources across time and future events. At the same time, it solves the problem of unclear optimization goal caused by the superficial state awareness of the existing method, and ensures the accuracy and efficiency of the optimization process. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 is a flowchart of a satellite-borne interference source Doppler observation strategy optimization method provided by an embodiment of the application.

[0053] Figure 2 is a flowchart of obtaining enhanced Doppler observation data provided by an embodiment of the application.

[0054] Figure 3 is a flowchart of evaluating the matching degree of the balance of the energy reserve unit and the power consumption demand of the period provided by an embodiment of the application.

[0055] Figure 4 is a flowchart of obtaining a Doppler fine estimate by constructing a virtual mirror signal and iteratively adjusting provided by an embodiment of the application.

[0056] Figure 5 is a flowchart of predicting mutation characteristics in Doppler data provided by an embodiment of the application. DETAILED DESCRIPTION

[0057] In order to enable persons skilled in the art to better understand the present application scheme, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by persons skilled in the art without creative labor should be within the scope of protection of the present application.

[0058] It should be noted that the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device that includes a series of steps or units does not have to be limited to only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0059] In the research, it is found that the existing optimization methods generally lack the ability of forward-looking prediction and planning for future key events. The Doppler observation of LEO satellites is a typical event-driven process, and most of its positioning information value is highly concentrated in the short mutation window when the Doppler frequency change rate reaches the peak. However, the existing DRL methods, such as DDPG, are essentially state-driven, which reactively make decisions based on the current channel or signal state. This mechanism cannot foresee the mutation window that will occur in the future, so it does not know to enter the energy-saving mode in the low-observation-efficiency stable period (such as low-elevation zone) to accumulate energy, nor can it plan in advance and prepare for high-density sampling before the mutation window arrives. This after-response rather than pre-planning mode results in the inability to optimally allocate the most valuable on-board resources of energy and sampling in the time dimension for future high-value events.

[0060] In addition, the static and blind allocation of observation resources has two aspects. First, the waste and mismatch of sampling resources. The information content of the Doppler signal is extremely uneven throughout the transit arc, and most of its positioning information is contained in the short mutation window where the frequency changes dramatically. In the vast stable period, the signal changes slowly and the information density is extremely low. Using a fixed and uniform sampling strategy, a large amount of redundant observation will inevitably be caused in the stable period, wasting valuable on-board storage and computing resources; while in the most critical mutation window, the fixed sampling rate may not be able to meet the Nyquist requirement of capturing signal transients, causing information loss, which is a typical mismatch between resources and information needs.

[0061] Secondly, the energy management is extensive and inefficient. Static power budget cannot be dynamically adjusted according to the real-time progress of the task and the actual situation of the signal. It lacks an intelligent energy pool mechanism and does not know how to actively degrade operation in the low-elevation period with poor signal quality and low observation benefit to accumulate energy, and then invest these valuable energy in the high-elevation period with high signal-to-noise ratio and good observation benefit, so as to realize the optimal time-domain allocation of energy resources in the whole task period. The existing method lacks the semantic depth of state awareness, that is, it lacks deep understanding and extraction of the inherent physical characteristics of Doppler signals. In IRS-assisted communication, the state input of DRL optimization is usually channel state information (CSI) or signal-to-noise ratio (SINR) and other macro indicators. However, in the Doppler positioning task, it is not the average strength of the signal that determines the observation value, but the inherent high-order dynamic characteristics that represent kinematics, such as the peak value of the second derivative of the Doppler, the texture fingerprint of the signal time-frequency spectrum, or the strength of the signal singularity. The existing optimization framework lacks a special semantic layer and cannot extract these deep-level mutation features with clear physical meaning from the original observation data as the basis for decision-making. This shallow understanding of state awareness makes the optimization algorithm unable to identify truly valuable observation periods, and the optimization process also loses the precise target, making it difficult to fundamentally improve the efficiency of capturing key information and the final success rate of the task. In the real on-orbit environment, especially when dealing with weak interference signals, simple derivative or frequency domain analysis methods are easily disturbed by noise. Random noise peaks may be mistaken for mutations (false alarms), while real weak mutation features may be submerged in noise and cannot be detected (missed alarms).

[0062] As shown in Figure 1 , a satellite-borne interference source Doppler observation strategy optimization method is proposed, which includes:

[0063] Obtain the original radio frequency signal of the satellite-borne interference source and the satellite precise orbit data, predict the mutation characteristics in the Doppler data, and generate an adaptive observation scheme.

[0064] For example, the intermediate frequency sampling data of the spaceborne receiver is acquired, and a fast Fourier transform (FFT) is performed to identify the spectral peak of the interference signal. Time synchronization and coordinate system conversion are performed in combination with GPS timing information and satellite attitude measurement data, and then a frequency tracking loop is used to extract the instantaneous Doppler frequency shift value sequence. At the same time, the power spectral density of the signal is calculated, which is used as an evaluation index of the subsequent data quality. On this basis, the calculation and analysis of the Doppler rate of change feature are carried out. Specifically, based on the instantaneous Doppler frequency shift value sequence extracted in the previous step, the five-point difference method can be used to calculate the first derivative, and then the Savitzky-Golay filter is used to calculate the Doppler second derivative sequence. By setting a threshold, the time period whose second derivative absolute value exceeds the threshold is identified and marked as a potential mutation interval, and its corresponding satellite elevation angle range is recorded. A prediction-correction double-loop mechanism is used to locate the mutation point. The mechanism calls a historical transit mutation point knowledge base, which stores statistical feature vectors (including mean, variance, etc.) of mutation points under similar conditions in the past 30 days. Based on the current satellite orbital elements and geographical position, the initial mutation point position of the current transit is matched and predicted from the knowledge base. During the observation process, the Doppler second derivative sequence is monitored in real time, and when the deviation between the measured value and the predicted value exceeds the preset threshold (for example, 15%), the correction mechanism is triggered, and the corrected mutation window is output by fusing the prediction and measurement information. According to the corrected mutation window, a non-uniform sampling strategy is generated. The strategy aims to densely sample in the mutation interval and sparsely sample in the smooth interval, so as to save resources while ensuring data quality. At the same time, according to the sampling strategy, combined with the power consumption of different working modes of the receiver, the expected energy consumption distribution of the entire observation task can be estimated.

[0065] According to the adaptive observation scheme, energy-aware resource scheduling and observation are performed to obtain enhanced Doppler observation data.

[0066] Specifically, the implementation of this step can include energy-aware dynamic resource scheduling and multipath effect identification and collaborative information enhancement. In terms of resource scheduling, the system reads the non-uniform sampling strategy and power consumption demand estimation value, and starts the energy pool mechanism. In the period when the satellite elevation angle is low (for example, less than 20°), that is, in the energy accumulation window, the saved energy is accumulated by reducing the working mode of the receiver (for example, switching from a high-precision mode of 15W to a power-saving mode of 3W). When the satellite enters the high-value observation period defined by the Doppler mutation feature, the system evaluates whether the energy reserve meets the power consumption demand of this period, and if it does, it authorizes the execution of high-density sampling observation. In terms of obtaining observation data, the system executes the final sampling plan. During the observation process, information enhancement can be performed using the multipath effect. Specifically, by analyzing the time-varying characteristics of the signal power spectral density, it is identified whether there is a multipath effect. If a multipath is identified, a subspace estimation algorithm such as the MUSIC algorithm can be used to separate the direct path signal component and one or more reflected path signal components from the received signal. By estimating the Doppler shift difference and relative time delay of each path, and using the geometric constraint relationship added by the reflected path to construct an over-determined equation system, solving the equation system can obtain a Doppler measurement value with enhanced parameter estimation accuracy, which is part of the enhanced Doppler observation data.

[0067] Based on the enhanced Doppler observation data, positioning solution and performance evaluation are performed to generate parameters for iterative optimization of the observation strategy.

[0068] For example, a state vector containing the position, velocity, and acceleration of the interference source is established, and based on the enhanced high-precision Doppler measurement value and satellite precise orbit data, a nonlinear observation equation is constructed. An extended Kalman filter (EKF) is used to recursively update the state vector, thereby estimating the motion state of the interference source. The pure Doppler data after motion compensation is input into the iterative positioning algorithm, and the geographical coordinates of the interference source are solved in combination with the observation constraints at multiple times. The positioning error ellipse and geometric dilution of precision (GDOP) and other indicators are calculated to evaluate the positioning performance of this transit, and a quantitative positioning quality score Q is generated. In this embodiment, in order to realize adaptive iterative optimization of the observation strategy, the positioning quality score Q is combined with the final balance of the energy reserve unit for this task and the accuracy of the mutation point prediction to form a multi-dimensional task performance indicator. Based on this multi-dimensional indicator, the reward function value R evaluating the comprehensive performance of this observation task is calculated. This reward function value R will serve as a feedback signal to update the internal parameters of the deep reinforcement learning network through gradient backpropagation algorithms (such as Actor and Critic parameter updates applied to TD3 networks). The updated network internal parameters constitute the optimization strategy parameter set for the iterative optimization of the observation strategy for the next task.

[0069] As Figure 2As shown, according to one aspect of the present application, enhanced Doppler observation data is obtained, including: identifying an energy accumulation window defined by a low elevation angle period in an adaptive observation scheme, and a high-value observation period defined by a Doppler mutation characteristic; within the energy accumulation window, accumulating saved energy values ​​by lowering the working mode of the receiver, and counting them into an energy reserve unit; when entering a high-value observation period, evaluating the matching degree between the balance of the energy reserve unit and the power consumption requirement of the period; under the condition that the balance of the energy reserve unit meets the power consumption requirement, authorizing the execution of high-density sampling to obtain enhanced Doppler observation data.

[0070] The energy accumulation window refers to the period when the satellite's elevation angle relative to the ground station is below a preset threshold (e.g., 20°). During this period, signal propagation loss is typically high and multipath effects are severely impacted, resulting in relatively low observation data quality. Therefore, it serves as a suitable window for implementing energy-saving strategies to accumulate energy. High-value observation periods are typically defined by the modified mutation window, when Doppler frequency is predicted to undergo significant changes.

[0071] Furthermore, if Figure 3 As shown in the figure, the matching degree between the balance of the energy reserve unit and the power consumption demand in the period is evaluated, specifically including: constructing an energy-precision benefit function that characterizes the relationship between observation accuracy and energy consumption; based on the energy-precision benefit function, generating a segmented benefit evaluation matrix for different power consumption modes in the high-value observation period; establishing a multi-constraint optimization model with the goal of maximizing the total benefit and integrating the balance of the energy reserve unit (energy reserve account) and the mode switching cost as constraints; solving the multi-constraint optimization model to obtain the optimal power consumption allocation plan, and performing high-density sampling based on the optimal power consumption allocation plan.

[0072] For example, the specific implementation of this process is as follows: In order to quantify the relationship between energy input and the return of observation data accuracy, the benefit function U(E) = α·log a (1+β(E / E ref ) ); where U(E) represents the accuracy benefit value obtained by investing energy E; E represents the energy allocated to the receiver during a specific observation period; α is the accuracy weighting factor, which can be set to α=0.8, for example; β is the energy sensitivity coefficient, which can be set to β=2.5, for example; E ref is the standard transit energy consumption benchmark value, which serves as a reference quantity. a can be 2, e, or 10. This logarithmic function reflects the principle of diminishing marginal benefits, that is, the initial energy input can bring significant accuracy improvement, but as the energy input increases, the benefit growth will tend to be flat. Based on this benefit function, a segmented benefit evaluation matrix M can be generated for different power consumption modes during high-value observation periods. benefit , each element M in the matrix ijrepresents the benefit value that can be brought by using the jth power consumption mode in the ith observation segment. A multi-constrained optimization model is established and solved. The objective function of the model is: max∑(M ij ·x ij ), where x ij is the selection variable of using the jth power consumption mode in the ith observation segment; and aims to maximize the total benefit during the entire observation mission. The constraint conditions include: energy constraint ∑(P seg [i]·t seg [i]·x ij )≤E avail , which ensures that the total energy consumption does not exceed the reserve, where P seg [i] is the power consumption of the selected power consumption mode in the ith observation segment, t seg [i] is the duration of the ith observation segment, E avail is the upper limit of the energy reserve of the system; continuity constraint |mode i -mode i+1 |≤1, which avoids drastic changes in the working mode of the receiver, where mode i is the working mode number of the receiver in the ith segment; and mode switching cost, for example, an additional consumption of 0.1 J per switching. In this embodiment, a dynamic programming algorithm is used to solve the above model to obtain the optimal power consumption allocation scheme X opt and the corresponding working mode sequence mode[i].

[0073] Alternatively, in some embodiments, the algorithm used to solve the multi-constrained optimization model can also be other optimization algorithms, such as genetic algorithm, simulated annealing algorithm, or greedy algorithm with low computational complexity when fast real-time adjustment is required. In addition, the form of the energy-precision benefit function U(E) can also be adjusted according to the specific satellite mission requirements and hardware characteristics, for example, a piecewise linear function or an exponential function can be used to model. Finally, the observation is performed according to the optimal power consumption allocation scheme. This embodiment solves the contradiction between the limited resources (energy) in orbit and the high performance requirements of the observation mission, and realizes the fine and intelligent management of resources.

[0074] Further, in order to cope with the situation that the actual energy consumption may not match the expectation during the in-orbit execution, the energy intelligent allocation process also includes a real-time adjustment mechanism. The mechanism is evaluated once every fixed time period (for example, every 10 seconds), and the deviation ΔE between the actual consumed energy E actual and the planned consumed energy E planned is calculated. When the absolute value |ΔE| of the deviation exceeds the preset adjustment threshold (for example, 0.5 J), it indicates that there is a non-negligible unplanned energy consumption change, and the system will immediately recalculate the benefit evaluation matrix Mbenefit_new But in order to ensure real-time, the system no longer calls the complex dynamic programming algorithm, but uses a fast greedy algorithm with lower computational complexity (for example, O(n)) to update the working mode of the subsequent period and generate a real-time mode switching instruction sequence cmd[i]. Through this fast feedback and adjustment cycle, it can be ensured that the energy use efficiency of the satellite remains near the optimal state throughout the transit period.

[0075] According to one aspect of the present application, it further comprises: when the evaluation determines that the balance of the energy reserve unit cannot meet the power consumption demand, starting a degradation strategy; the degradation strategy calculates the ratio of the current remaining energy to the expected demand of the subsequent task, and compares the ratio with at least one preset degradation threshold; if the ratio is lower than the degradation threshold, a degradation sampling plan aimed at reducing the total energy consumption is generated, and the degradation sampling plan is realized by increasing the sampling interval of non-critical observation areas or shrinking the spatial range of high-density sampling; and the observation is performed according to the degradation sampling plan.

[0076] Among them, the degradation strategy is a robust mechanism to ensure the completion of core tasks, and its motivation is to avoid the complete failure of the entire observation task (especially the observation of high-value mutation periods) due to energy depletion, and to realize the graceful degradation of the system. Specifically, the implementation of the degradation strategy is as follows: the system will continuously monitor the remaining energy E remain and the expected demand of the subsequent task E needed , and calculate the energy ratio r energy =E remain / E needed . E needed can be dynamically calculated according to the subsequent observation plan that has not been executed and the power consumption estimation model. According to the different intervals of the energy ratio r energy , the degradation sampling plan is generated in stages. The system internally presets at least one, preferably multiple, degradation thresholds to realize hierarchical degradation management. For example, two degradation thresholds can be set: a first-level degradation threshold (for example, 0.7) and a second-level degradation threshold (for example, 0.4). Trigger and execute the first-level degradation: when it is monitored that 0.4≤r energy <0.7, the first-level degradation is triggered. At this time, the system will update the original sampling plan to generate a first-level degradation sampling plan, the core idea of which is to sacrifice the secondary and protect the primary. The specific measure can be to increase the sampling interval of all non-mutation windows (i.e. smooth intervals) by a fixed percentage, for example, by 50%, based on the original plan. This can reduce the overall energy consumption without affecting the data density of the core mutation area. Trigger and execute the second-level degradation: when it is monitored that r energyWhen the energy condition is very critical, i.e., when the energy level is less than 0.4, a secondary degradation strategy is triggered. The secondary degradation strategy is more stringent and aims to protect the core. The specific measures can be that the system generates a sampling plan in an emergency mode, only retains a high-density intensive sampling plan within a small range (e.g., ±3°) before and after the determined and modified mutation window, and forces all other regions outside the range to switch to the power-saving mode and the basic sampling rate regardless of whether the original plan is intensive or sparse. This ensures that even in extremely unfavorable energy conditions, the observation task of the Doppler mutation, which is the highest value scientific event, can be completed with a high probability (e.g., ensuring that at least 80% of the high-value observation tasks are completed), thereby improving the task success rate and reliability of the entire observation system.

[0077] Optionally, in some embodiments, the adjustment manner of the degradation strategy can be more smooth. For example, the increase amount of the sampling interval can be a function continuously related to r energy , instead of a fixed 50%. In addition, the boundary of the shrinking high-density sampling space range can also be dynamic, instead of a fixed ±3°, and the specific size can be determined by a function considering r energy and the mutation intensity prediction value.

[0078] The embodiment solves the problem of waste of on-board energy resources due to extensive and inefficient energy management. Specifically, the introduction of the energy pool mechanism enables the observation strategy to have energy planning capabilities across time scales. By actively adjusting and reducing the receiver operating mode in low-elevation energy accumulation windows with low observation benefits, the saved energy is stored in a virtual account, and then in the predicted high-value observation period, the optimal withdrawal is made according to the energy-precision benefit function. Combined with a multi-constraint optimization model that aims to maximize the total benefit and incorporates the mode switching cost, it is ensured that every bit of energy is used to the best effect. The traditional static and large-margin power budget is transformed into a fine and real-time dynamic adjustment of optimal energy flow management according to the task, improving the utilization efficiency of on-board energy, which is the most scarce resource.

[0079] According to an aspect of the present application, the mutation feature in the predicted Doppler data can be: extracting time series Doppler frequency shift data from the original radio frequency signal and calculating the corresponding Doppler second derivative sequence; calling a historical transit mutation point knowledge base and combining the current satellite precise orbit data to generate an initial predicted mutation position; real-time monitoring of the Doppler second derivative sequence, when the measured characteristics of the sequence deviate from the expected characteristics of the initial predicted mutation position and exceed the preset threshold, triggering a correction mechanism; the correction mechanism combines the initial prediction and the measured characteristics to output a modified mutation window for generating an adaptive observation scheme.

[0080] Further, the correction mechanism includes: adopting a Kalman filter to fuse the initial predicted mutation position and the measured features extracted from the Doppler second derivative sequence; the Kalman filter outputs an optimized corrected mutation position and a confidence weight coefficient representing the credibility of the corrected position; and the corrected mutation window boundary is adaptively set according to the corrected mutation position and the confidence weight coefficient, wherein a higher confidence weight corresponds to a narrower window, and a lower confidence weight corresponds to a wider window.

[0081] In the embodiment, the system first reads all historical observation records of past transits within a certain period of time (for example, 30 days), which contain the satellite elevation angle at the time of each successful observation of the mutation point, the atmospheric conditions at that time, and various related factors such as season. Using clustering analysis and other algorithms, mutation point data with similar environmental conditions (such as similar orbit type, season, and geographic region) are grouped together. The statistical feature vector F hist is calculated for each group of data, which contains statistical quantities such as the mean, variance, and skewness of the mutation elevation angle under that condition, to quantify its distribution characteristics. A classification index table is established for these feature vectors to facilitate fast retrieval and matching based on current conditions in subsequent tasks. Before a new transit task begins, the system obtains the current orbit elements and satellite geographic position, and uses this information to match the classification index table established in the first step to find the historical condition category that is most similar to the current condition. The system then extracts the statistical feature vector F hist corresponding to this category. To improve the accuracy of the prediction, the model also combines the real-time atmospheric refractive index and ionospheric total electron content (TEC) values obtained through other sensors, and through the framework of Bayesian inference, fuses the historical statistical prior and real-time environmental information to finally calculate the initial predicted elevation angle θ pred and its confidence interval [θ low , θ high ] for this transit. During the satellite transit, the system continuously monitors the real-time calculated Doppler second derivative sequence. When the absolute values of the second derivatives of three consecutive sampling points all exceed the preset trigger threshold (for example, 0.5 Hz / s 2 ), the system determines that the real mutation signal has been captured, triggering the correction program immediately. After triggering, the system extracts the measured mutation features (such as the elevation angle at the time of the mutation), and calculates the deviation Δθ between the initial predicted elevation angle θ pred and the measured mutation elevation angle. At this time, the Kalman filter is called, which takes θ pred as the prior prediction value and the measured mutation elevation angle as the observation value, and through the state update process of the filter, generates the optimal estimation value that fuses the two kinds of information, i.e., the corrected elevation angle θ corr , while outputting the confidence weight coefficient w corrThe system is based on the modified elevation angle θ corr and the confidence weight coefficient w corr , an adaptive window algorithm is used to dynamically determine the final mutation window boundary [θ start ,θ end ]. The rule of this algorithm is that when w corr When the value is high (such as w corr >0.8), indicating that the system is very confident in the judgment of the mutation location, and a narrower tight window will be set at this time, such as [θ corr -2°, θ corr +2°] to concentrate resources. On the contrary, when w corr When the value is low (such as w corr <0.5), indicating that there is a large uncertainty, the system will set a wider conservative window, such as [θ corr -5°, θ corr +5°] to ensure that the target is not missed. This dynamically adjusted window will be directly used to generate subsequent non-uniform sampling strategies, thereby achieving precise targeted delivery of observation resources.

[0082] This embodiment addresses the failure to capture critical information due to prediction lag and low robustness by introducing two forward-looking mutation feature prediction methods: a prediction-correction dual-loop mechanism and texture fingerprint recognition. Specifically, the prediction-correction mechanism optimally integrates statistical priors from historical data (via a knowledge base) with current evidence from real-time observations (via second-order derivative monitoring) within a Kalman filter framework. This eliminates the need for the system to confirm mutations after the fact, instead enabling the formation of a high-probability prediction window before a mutation occurs, thus achieving forward-looking decision-making. The texture fingerprint recognition method, on the other hand, transforms one-dimensional time series analysis into a two-dimensional image recognition problem, leveraging a convolutional neural network (CNN) to extract physically meaningful patterns from the deep textures of the time-spectrogram. Its robustness far exceeds that of traditional single-threshold decision methods, particularly in low signal-to-noise ratio environments, improving its ability to capture subtle mutation features. This ensures that observation strategies can accurately target the highest-value observation periods in the future, providing a reliable foundation for the targeted allocation of all subsequent resources.

[0083] Alternatively, as Figure 5As shown, the mutation feature in the predicted Doppler data can also be: performing spectral analysis on the original radio frequency signal, and combining satellite precise orbit data, extracting time sequence Doppler frequency shift data through a frequency tracking loop; performing time-frequency transformation on the time sequence Doppler frequency shift data to generate a time-frequency spectrum, and processing the time-frequency spectrum into a gray scale image; extracting a local binary pattern and a gray level co-occurrence matrix from the gray scale image as texture features to construct a multi-dimensional texture feature vector; feeding the multi-dimensional texture feature vector into a pre-trained convolutional neural network to analyze and output the mutation type and mutation time in the Doppler data, which is used to generate an adaptive observation scheme.

[0084] In the embodiment, the Doppler frequency shift time sequence f D (t) obtained after preprocessing is subjected to time-frequency transformation. Specifically, a short-time Fourier transform (STFT) can be used. For example, a sliding time window (the window length can be set to 5 seconds, and the sliding step length can be set to 0.5 seconds) is constructed, and Fourier transform is performed on the data segment in each window to generate a time-frequency spectrum S(t, f). The time-frequency spectrum is a two-dimensional representation, the x-axis represents time t, the y-axis represents frequency f, and the brightness or color of each point (t, f) in the time-frequency spectrum represents the energy intensity of the signal at that time and that frequency. The generated time-frequency spectrum S(t, f) is subjected to normalization processing, the energy intensity value is mapped to the integer interval [0, 255], and is converted into a gray scale image I(x, y). In this way, the change pattern of the Doppler signal over time is directly converted into a texture pattern in the image. Features that can represent the local and global texture characteristics of the gray scale image I(x, y) are extracted. In the embodiment, at least the following features are included: Local Binary Patterns (LBP): LBP compares the gray scale values of a pixel and other pixels in its neighborhood, encodes the comparison results into binary numbers, and thus effectively captures the microscopic structure information of the image, such as edges, corners and spots, etc. Gray-Level Co-occurrence Matrix (GLCM): GLCM calculates the gray scale value appearance probability of two pixel pairs separated by a certain distance and direction in the image to reflect the comprehensive information of the image gray scale about direction, adjacent interval, change amplitude, etc. A plurality of statistical features, such as contrast, correlation, energy and homogeneity, etc., can be further calculated from GLCM. Gabor filter response: Gabor filter performs excellently in extracting local spatial frequency features of the image, and can capture texture information from different scales and directions. The plurality of features extracted from the above are combined to construct a high-dimensional mutation texture feature vector V texture . For example, a 128-dimensional feature vector can be constructed, which comprehensively describes the fingerprint of the Doppler signal in the time-frequency domain. Based on the CNN-based mutation classification and positioning, the constructed texture feature vector V textureinto a pre-trained convolutional neural network (CNN). The convolutional neural network is a deep learning model whose network structure (e.g., containing 3 convolutional layers and 2 fully connected layers) is particularly suitable for processing image and high-dimensional feature vector classification tasks. Through training on a large number of labeled (containing various mutation types) historical data, the network has learned the complex mapping relationship from a specific texture feature vector to a specific mutation pattern. The network receives V texture , and outputs the classification result of the current data window, including the mutation type (e.g., step type, pulse type, oscillation type, or gradual type) and the accurate estimation of the mutation time t mut .

[0085] Optionally, in some low signal-to-noise ratio and weak signal scenarios, in order to enhance the saliency of texture features, the multiple consecutive time-frequency spectrograms can be accumulated and averaged before generating the gray image I(x, y), that is, I enh =Σ(I i ) / N, where N is the number of accumulated frames. Random noise can be effectively suppressed, and the real texture structure related to the mutation event that persists can be highlighted, thereby improving the success rate and accuracy of mutation detection in a weak signal environment.

[0086] Regarding the steps of constructing a multi-dimensional texture feature vector, further to: the multi-dimensional texture feature vector, for example, a 128-dimensional feature vector, can be constructed by splicing multiple different types of texture features to achieve comprehensive and non-redundant description of the time-frequency spectrogram information. An exemplary construction method is: extracting the local binary pattern (LBP) statistical histogram feature from the gray image I(x, y); calculating the gray level co-occurrence matrix (GLCM) of the image and extracting multiple statistics such as contrast, correlation, energy, and homogeneity from it; constructing a set of Gabor filter banks covering different scales and directions, and filtering the image with them, taking their response energy as Gabor texture features. Splice the three groups of features (LBP features, GLCM features, and Gabor features) in a predetermined order, and you can form a final 128-dimensional feature vector V texture that can fully represent the Doppler mutation texture fingerprint.

[0087] According to one aspect of the present application, the step of generating the adaptive observation scheme comprises a process of generating a non-uniform sampling strategy, which comprises: calculating the ratio of the maximum value to the average value of the second derivative of Doppler within the identified Doppler mutation window; adaptively determining the proportional coefficient in the inverse proportional sampling function according to the ratio, wherein a larger ratio corresponds to a smaller proportional coefficient; applying the inverse proportional sampling function containing the adaptive proportional coefficient to generate dense sampling intervals for the mutation interval and sparse sampling intervals for the stationary interval; constructing a transition zone at the boundary between the dense sampling interval and the sparse sampling interval, and using a S-shaped function to smoothly connect the dense and sparse sampling intervals to ensure continuous and non-impulsive change of the sampling rate.

[0088] Specifically, after determining the mutation window, the system first calculates the maximum value d 2 f max and the average value d 2 f avg of the absolute value of the second derivative of Doppler within the window. The ratio r = d 2 f max / d 2 f avg of the two is calculated, which is used to evaluate the severity of the mutation. The system uses a piecewise function to adaptively determine the inverse proportional coefficient k according to the value of r: when r > 10, set k = 0.1; when 5 < r ≤ 10, set k = 0.2; when r ≤ 5, set k = 0.5. This makes the sampling strategy not only respond to instantaneous changes, but also respond to the macroscopic characteristics of the overall event. A sampling function Δt = k / |d 2 f / dt 2 | containing the adaptive coefficient k is generated. In order to avoid the mutation of the sampling rate at the boundary of different intervals, the system sets a transition zone at the boundary of the mutation window (e.g. [θ start -1°, θ start ] and [θ end , θ end +1°]). In the transition zone, a S-shaped function is used to smoothly connect the dense sampling interval Δt dense and the sparse sampling interval Δt sparse . The smoothing function can be specifically expressed as: Δt trans = Δt sparse + (Δt dense - Δt sparse ) / (1 + exp (-10 (θ - θ boundary). The rate of change of time interval between adjacent sampling points can be ensured not to exceed a preset smooth threshold, for example, 50%, thereby ensuring the smoothness of hardware execution. After generating the sampling interval function At(0), an energy consumption estimation model is constructed. The model maps different sampling intervals At to instantaneous power P(0) based on the power consumption characteristics of the receiver (for example, high-precision mode 15W, energy-saving mode 3W). For example, when At < 0.5s, corresponding to high-precision mode, P(0) = 15W; when At > 2s, corresponding to energy-saving mode, P(0) = 3W; in the intermediate region, the power value can be linearly interpolated. By integrating the instantaneous power P(0) of the entire transit arc, the expected total energy consumption E of this task can be calculated total , and a segmented power consumption demand sequence P seg [i] is generated. From the initial time t0 of the task, according to the continuous sampling interval function At(0), the 0 i of each time is obtained through orbit prediction, and each sampling time t i+1 is recursively calculated as t i + At(0 i ), generating a complete non-uniform sampling time sequence {t i}. Before outputting the final executable sampling plan, the system will also verify it to ensure that it meets the basic quality requirements, for example, verifying that the number of sampling points in the mutation window is not less than 20, and the sampling point interval in the smooth region is not more than 5 seconds.

[0089] The embodiment solves the problems of waste and mismatch of sampling resources caused by fixed uniform sampling by designing an adaptive inverse proportional sampling function and combining an S-shaped function for smooth transition. Specifically, the introduction of the inverse proportional sampling function makes the sampling rate and the information content of the signal (represented by the second derivative of the Doppler) accurately negatively correlated, that is, the most intensive sampling is performed at the mutation point with the most information, and sparse sampling is performed in the smooth area with sparse information, thereby realizing the on-demand allocation of sampling resources. Further, the proportionality coefficient k is adaptively adjusted according to the overall severity of the mutation event (through the ratio r), realizing dual adaptation of macro and micro. The S-shaped function is used for smooth connection at the interval boundary, ensuring continuous and non-impulsive change of the sampling rate, not only improving the stability of subsequent signal processing, but also protecting the hardware life of the on-board execution mechanism. The limited sampling resources are optimally utilized.

[0090] According to one aspect of the present application, when the original radio frequency signal is below a threshold value, the process of obtaining enhanced Doppler observation data is implemented using a step-by-step Doppler frequency cascade fine estimation method, specifically: a preliminary spectral peak positioning is performed according to a frequency grid that is adaptively adjusted based on spectral energy distribution, to obtain a Doppler coarse estimate; based on the Doppler coarse estimate, a reverse Doppler sequence is injected and a solution that minimizes the flatness of the compensated signal spectrum is searched, to obtain a Doppler medium estimate; based on the Doppler medium estimate, a virtual mirror signal is constructed and iteratively adjusted to minimize the time rate of change of the beat component generated after cross-correlation of the virtual mirror signal and the original signal, to obtain a final Doppler fine estimate. Based on the Doppler fine estimate, enhanced Doppler observation data is sampled.

[0091] The present embodiment balances between computational efficiency and estimation accuracy through a coarse-to-fine strategy. It first uses a method with lower computational cost to quickly narrow down the frequency uncertainty range, and then uses a method with higher computational cost but more accurate to perform fine search, thereby efficiently obtaining a high-precision estimation result. The specific implementation includes: coarse estimation based on adaptive frequency grid evolution D_coarse The goal of this stage is to quickly locate the approximate position of the Doppler frequency from a larger frequency uncertainty range (e.g. ±500Hz). Based on the preliminary Doppler shift estimate f i i = f D_coarse + i x Δf, i ∈ [-100, 100], where the initial grid resolution Δf can be set to 10Hz. Calculate the spectral energy E i of the signal at each grid point f i . The grid density is adaptively adjusted based on the energy distribution. Specifically, in the area with higher energy (e.g. E i > 0.7 x max(E)), the grid is densified (e.g. the resolution becomes Δf / 4); while in the area with lower energy (e.g. E i < 0.3 x max(E)), the grid is sparsified (e.g. the resolution becomes 4Δf). After several (e.g. 5) iterations of evolution, the grid points will naturally gather around the true spectral peak. At this time, in the densest grid area, the spectral peak position is fitted by methods such as parabolic interpolation, thereby outputting a Doppler coarse estimate f D_grid with high precision, which can reach about ±50Hz.

[0092] Medium estimation based on reverse Doppler injection This stage uses signal processing techniques to convert the frequency estimation problem into an optimization problem. Specifically, based on the obtained f D_grid , a smaller range (e.g. [f D_grid -50Hz, f​D_grid +50Hz]), a series of tentative inverse Doppler sequences exp(-j2πf probe ×t). Multiply these sequences one by one with the original signal s(t) to obtain the compensated signal s comp (t)=s(t)×exp(-j2πf probe ×t). When the test frequency f probe When the frequency is exactly equal to the true Doppler frequency, the Doppler effect is compensated, s comp The spectrum of (t) will become the flattest. Calculate each s comp The spectrum flatness index F=σ(|S comp (f)|) / μ(|S comp (f)|), where σ and μ are the standard deviation and mean of the spectrum amplitude respectively. Search for the f that minimizes the spectrum flatness index F. probe , which is the Doppler mean value f D_inject To accelerate convergence, optimization algorithms such as the golden section search can be used. This can improve the estimation accuracy to approximately ±0.1 Hz.

[0093] like Figure 4 As shown, a Doppler precise estimate is obtained by constructing a virtual mirror signal and iteratively adjusting it, including: constructing a virtual mirror signal using the conjugate of the original RF signal and taking the Doppler mean estimate as the frequency basis; calculating the cross-correlation function between the original RF signal and the virtual mirror signal, and separating the beat frequency component from the cross-correlation function; iteratively adjusting the frequency used to construct the virtual mirror signal until the time rate of change of the beat frequency component converges to a preset minimum value, and determining the converged frequency as the Doppler precise estimate.

[0094] Specifically, based on the median estimate f D_inject , construct a virtual mirror signal s mirror (t)=s*(t)×exp(j4πf D_inject ×t), where s*(t) is the conjugate of the original signal. Calculate the original signal s(t) and the mirror signal s mirror (t) is the cross-correlation function R(τ). When there is a slight frequency difference between the two, the cross-correlation result will contain a beat frequency component with a very low frequency. The beat frequency f beat It can be extracted from the phase difference of adjacent sampling points of the cross-correlation function. Iteratively fine-tune the frequency f used to construct the image signal D_inject , and observe the time rate of change of the beat frequency df beat / dt. When f D_inject When it is consistent with the true Doppler frequency, df beat / dt will approach zero theoretically. beatWhen / dt converges to the preset minimum value, the iteration ends, and the frequency used to generate the image signal is the final Doppler precision estimate f D_mirror , its accuracy can reach ±0.01Hz or even higher. Optionally, a three-level estimation error model can be constructed, and the errors at each level can be back-propagated and compensated through the Kalman filter, further improving the final accuracy to, for example, below ±0.001Hz.

[0095] According to one aspect of the present application, the step of obtaining enhanced Doppler observation data, when sampling is performed in a non-uniform manner, also includes a data reconstruction process, specifically: obtaining a Doppler fine estimate, and using the Doppler fine estimate to pre-Doppler compensate the non-uniformly sampled observation data to obtain a compensated stationary signal; solving a pre-constructed compressed sensing convex optimization model for the compensated stationary signal to obtain a reconstructed stationary signal component; and re-assigning the frequency trend represented by the Doppler fine estimate to the reconstructed stationary signal component to generate enhanced Doppler observation data.

[0096] Standard compressed sensing algorithms are not very effective in processing non-stationary signals with strong time-varying characteristics (such as Doppler signals). This embodiment first uses high-precision Doppler estimation to remove the non-stationary trend, thereby transforming the difficult non-stationary signal reconstruction problem into a relatively easy stationary signal reconstruction problem. The specific implementation method is: the non-uniformly sampled observation data y is combined with the reverse Doppler sequence exp(-j2πf D_final ×t) to obtain an approximately stationary signal s comp Where f D_final is the final Doppler precision estimate obtained by the above cascade precision estimation. comp , solve the pre-built compressed sensing convex optimization model: min||α||1subject to||s comp -ΦD stationary ×α|| 2 ≤ε. Where α is the signal in a certain transformation domain (defined by the dictionary D stationary Definition, for example, the sparse coefficients under the discrete cosine transform (DCT) basis, Φ is the measurement matrix that characterizes the non-uniform sampling process, and ε is the noise tolerance. This optimization problem can be efficiently solved using algorithms such as the alternating direction multiplier method (ADMM). The sparse coefficient α is obtained by solving final After that, through the inverse transformation D stationary ×α final The stationary signal component can be reconstructed. By re-adding the previously stripped Doppler frequency trend, a complete, high-resolution reconstructed signal can be obtained.

[0097] The description of the cascade refinement method and the data reconstruction process can be further refined as follows: in the coarse estimation based on adaptive frequency grid evolution, adaptive adjustment specifically refers to: in the high-energy region (for example, E i > 70% of the maximum energy), the grid resolution is encrypted to one quarter of the original (Δf / 4); and in the low-energy region (for example, E i < 30% of the maximum energy), the grid resolution is sparse to four times the original (4Δf). After the entire cascade refinement process, a cascade error backpropagation compensation step can be added to further improve the estimation accuracy. Specifically, a three-level estimation error model is constructed to calculate the coarse estimation error ε1, the intermediate estimation error ε2 and the fine estimation error ε3. After estimating the statistical characteristics of each level of error by a Kalman filter, a backpropagation compensator is designed, for example, the following formula is used to compensate the estimation results of the first two levels: f D_grid_new =f D_grid -α1·ε3 and f D_inject_new =f D_inject -α2·ε3; where the compensation coefficients can be set as α1=0.3 and α2=0.7. After compensation, the intermediate estimation and fine estimation steps can be re-executed, and the process is iterated until the final fine estimation error |ε3| is less than a small threshold value, for example, 0.001 Hz, thereby obtaining a final Doppler fine estimation value f D_final with high confidence.

[0098] In the data reconstruction process, the Doppler-assisted reconstruction method cleverly converts the difficult non-stationary signal reconstruction problem into a relatively easy stationary signal reconstruction problem, and when using the alternating direction multiplier method (ADMM) to solve it, the convergence speed and peak signal-to-noise ratio (PSNR) are improved.

[0099] According to one aspect of the present application, the step of obtaining enhanced Doppler observation data further includes a four-fold feature cross-validation process for confirming a Doppler mutation event, which includes: extracting time series Doppler frequency shift data from the observation data; performing at least four heterogeneous feature extraction analyses on the time series Doppler frequency shift data in parallel; cross- verifying the consistency of the features output by the four analyses in time and amplitude; when the consistency meets a predetermined standard, determining it as a reliable mutation event, and generating a fused mutation feature parameter set containing the mutation time, duration and intensity by weighting and fusing the four groups of features.

[0100] Specifically, the four heterogeneous feature extraction analyses include:

[0101] Wavelet multi-scale decomposition-based feature recognition. This analysis uses a Daubechies-4 wavelet basis to decompose the Doppler frequency shift time series f D (t) into 5 layers. By calculating the detail coefficients dj energy distribution and set an energy jump detection threshold (e.g. energy ratio in a sliding window exceeds 3 times) to identify the scale and time of the jump occurrence. The specific start and end points of the jump t start_w and t end_w are located on the detail coefficients d3 and d4, and the duration of the jump Δt w and the amplitude of the jump A w are calculated, and finally the wavelet domain jump feature vector F wavelet is generated.

[0102] Instantaneous feature extraction based on time-frequency joint analysis. This analysis performs a short-time Fourier transform (STFT) on f D (t), where the window length can be set to 0.5 seconds and the overlap rate can be set to 75%, generating a time-frequency spectrogram S(t, f). The instantaneous frequency trajectory f inst (t) is extracted by a ridge tracking algorithm, and its rate of change df inst / dt is calculated. When |df inst / dt| exceeds a pre-set threshold (e.g. 5 Hz / s), it is identified as a frequency jump point, and features such as jump amplitude Δf jump and jump slope k jump are extracted, generating a time-frequency domain jump feature vector F timefreq .

[0103] Singular point quantification based on multi-resolution singularity detection. This analysis calculates the continuous wavelet transform modulus maxima line of the second derivative of the Doppler d 2 f D / dt 2 , and tracks its evolution trajectory at different scales. The singularity strength of the signal is quantified by calculating the Lipschitz exponent α = log|W(a, t)| / log(a), where W(a, t) is the wavelet coefficient at scale a and time t. Points with α < 0.5 are identified as strong singular points, which are the core positions t singular of the jump, and features such as singularity peak α min and singularity zone width w singular are extracted, generating a singularity feature vector F singular .

[0104] Spectrum flatness time-varying characteristic analysis based on reverse Doppler injection. This is a newly added verification dimension. This analysis generates a reverse Doppler sequence and injects it into the signal within the jump candidate period at intervals of 0.1 seconds. By observing the time-varying characteristics of the spectrum flatness F(t), it is found that F(t) will appear a step due to the dramatic change of signal statistical characteristics at the true jump point. By calculating the time derivative of the flatness |dF / dt|, when its value exceeds three times the standard deviation of the background fluctuations (i.e. |dF / dt| > 3σ F ), it is confirmed that the jump exists.

[0105] After obtaining the four sets of features, the system performs cross-validation and fusion. Specifically, the time consistency C t and the amplitude consistency C A between the features are calculated. t When C A > 0.8 and C pred > 0.7, it is determined to be a reliable mutation. For the confirmed mutations, the estimation results of the four sets of features are fused by weighted average based on the signal-to-noise ratio of each method, and finally a high-confidence fused mutation feature parameter set is output, which reduces the false positive rate.

[0106] According to one aspect of the present application, the step of generating parameters for iterative optimization of observation strategy includes: obtaining a positioning quality score after positioning solution and performance evaluation, and combining the final balance of the energy reserve unit of the current task and the accuracy of the mutation point prediction to generate a multi-dimensional task performance indicator; fusing the multi-dimensional task performance indicators such as positioning quality score, energy use efficiency and prediction accuracy to calculate a reward function value evaluating the comprehensive performance of the current observation task; applying the reward function value as a feedback signal to update the internal parameters of the deep reinforcement learning network through gradient backpropagation algorithm; the updated internal parameters of the deep reinforcement learning network constitute the optimization strategy parameter set for iterative optimization of observation strategy.

[0107] Specifically, the generation of multi-dimensional task performance indicators is performed after the completion of a transit observation task. The system first collects performance indicators that can comprehensively reflect the execution effect of the current task from various sub-modules, including at least: positioning quality score Q: the score is output by the positioning solution and accuracy evaluation module, and is a direct quantitative evaluation of the core goal of the current task: interference source positioning. The Q value can be calculated by comprehensively calculating parameters such as the size of the positioning error ellipse and the GDOP value. The higher the Q value, the more accurate the positioning. Energy use efficiency indicator: this indicator is directly represented by the final balance of the energy reserve unit in the energy pool mechanism. A higher final balance means that less energy is consumed to complete the given goal, reflecting higher energy use efficiency. Mutation point prediction accuracy indicator: this indicator is calculated by comparing the initial predicted elevation θ pred output by the mutation point prediction module with the final mutation position θ corr obtained after real-time correction. The smaller the deviation, the more accurate the prediction. The calculation of the comprehensive reward function value is to fuse the above multi-dimensional performance indicators into a single scalar value - reward function value R. The reward function is a core concept in deep reinforcement learning, which provides a quantitative evaluation of the good or bad of each complete behavior of the agent (in this case, the observation strategy decision network). A reasonable reward function design is the key to the success of deep reinforcement learning. In this embodiment, R can be designed in the form of a weighted sum: R=wq • Q + w e • E final - w p • A error ; where E final is the final balance of the energy reserve unit; A error is the absolute value of the error of the mutation point prediction; w q , w e , w p are the weight coefficients of positioning quality, energy efficiency, and prediction accuracy, respectively. These weight coefficients can be set according to the overall priority of the task. For example, in a task where the highest accuracy is the primary goal, the weight of w q may be set relatively high.

[0108] The calculated reward value R will serve as the core feedback signal for updating the deep reinforcement learning network pre-installed in the system. In this embodiment, the network adopts the TD3 (Twin Delayed Deep Deterministic Policy Gradient) architecture, which is a deep reinforcement learning algorithm suitable for continuous action space. The TD3 network contains an Actor network and two Critic networks. The Actor network is responsible for outputting specific policies (for example, determining the key parameters of the sampling policy), while the Critic network is responsible for evaluating the quality of the Actor network output policy. The reward value R is mainly used to update the Critic network so that it can more accurately evaluate the value of state-action pairs. Through the gradient backpropagation algorithm, the Actor network will adjust its parameters in the direction that allows the Critic network to give a higher evaluation value. The essence of this process is that the actor will revise its performance according to the feedback from the critic in order to obtain a higher reward (reward value R) next time. After the above updating process, the weights and biases and other parameters inside the TD3 network have changed. These updated network parameters themselves constitute the optimized policy parameter set for iterative optimization of the observation policy. When the next new observation task begins, the system will use this evolved TD3 network to make decisions, such as adjusting the weights of the mutation point prediction model or optimizing the generation parameters of the sampling policy.

[0109] Optionally, in some embodiments, other advanced deep reinforcement learning algorithms can be used instead of TD3, such as SAC (Soft Actor-Critic) or PPO (Proximal Policy Optimization). The composition of the reward function R can also be more complex, such as introducing a nonlinear term or including consideration of other dimensions such as task completion time. The present embodiment can achieve continuous self-evolution of the policy, thereby maintaining high efficiency and intelligence in the long term.

[0110] The embodiment gives the entire observation system the ability of self-evolution by constructing a multi-dimensional reward function containing positioning quality, energy efficiency and prediction accuracy, and applying a deep reinforcement learning algorithm for closed-loop iteration, and solves the fundamental problem that the existing optimization framework cannot learn and evolve strategies according to the final comprehensive effect of the task. Specifically, the final result of a complete transit observation task, including the completion degree of the core target (positioning quality Q), the economy of resource consumption (energy final balance), and the accuracy of process prediction (prediction error), is fused into a single, quantitative reward value. The reward value, as the final judgment of the entire strategy behavior, drives the internal parameters of the deep reinforcement learning network to update through gradient backpropagation. This enables the system to learn from each successful experience and failure lesson, autonomously and continuously optimize its internal decision model, and thus gradually adapt to the changing on-orbit environment without human intervention, and maintain high efficiency and intelligence in the long term.

[0111] According to one aspect of the present application, the step of obtaining the enhanced Doppler observation data further comprises an information enhancement process using multipath effects, which includes: using a subspace estimation algorithm to separate the direct path signal component and at least one reflected path signal component from the received signal; estimating the Doppler frequency shift and relative time delay of the direct path and at least one reflected path respectively to obtain a Doppler difference sequence and a relative time delay sequence; jointly processing the Doppler difference sequence and the relative time delay sequence, and using the geometric constraint relationship attached to the reflected path to construct an over-determined equation set; and obtaining the Doppler measurement value with enhanced parameter estimation accuracy as a component of the enhanced Doppler observation data by solving the over-determined equation set.

[0112] Specifically, the existence of multipath effects is detected by analyzing the power spectral density time series of the received signal. When the power spectrum presents multiple peaks, and the interval and intensity relationship of these peaks conforms to the typical multipath propagation geometric model, it can be determined that there is a multipath effect. Once the existence of multipath is confirmed, the system will use a high-resolution subspace estimation algorithm, preferably an improved MUSIC (Multiple Signal Classification) algorithm, to separate the signal components. This algorithm can construct a signal subspace projection matrix to project the mixed received signal onto different subspaces, thereby separating the signal components from the direct path and one or more reflected paths (e.g., reflected by the earth's surface). After successfully separating the signal components, the system independently estimates the parameters of each component to obtain: the Doppler frequency shift f0 of the direct path; the Doppler frequency shift f i ; the relative time delay τ iBased on these estimates, the Doppler difference sequence Δf i = f i - f0and the relative time delay sequence τ i The reflected paths are not random, they carry geometric information about the propagation environment. For example, for a path that is reflected once by the Earth's surface, its path length, angle of arrival, are all strictly geometrically constrained by the positions of the satellite, the interferer, and the reflection point on the Earth's surface. This embodiment mathematically formulates these known geometric constraints into a series of constraint equations. These geometric constraint equations are then combined with the observation equations (i.e. the equations relating Δf i and τ i ) obtained from the parameter estimation. Since the number of reflected paths (and thus the number of observation equations) is usually larger than the number of unknown parameters (e.g. the corrections to the interferer's position), this naturally forms an overdetermined system of equations.

[0113] The overdetermined system of equations refers to a system of equations where the number of equations is larger than the number of unknowns, which usually does not have an exact solution. For the constructed overdetermined system of equations, a weighted least squares method is used to solve it. The goal of the solution is to find a set of estimates for the unknown parameters such that the sum of the squared residuals of all equations is minimized. The weighting refers to the fact that different equations can be given different weights according to the signal-to-noise ratio or strength of the corresponding reflected path, with paths having higher signal-to-noise ratios having more influence on the solution. By solving this overdetermined system of equations, the Doppler measurements with enhanced accuracy are obtained. This enhanced data combines information from the direct path and all utilized reflected paths, and its accuracy and reliability, especially in low elevation angle regions where multipath effects are significant, are superior to results obtained using only direct path signals.

[0114] Optionally, in some embodiments, other types of subspace algorithms, such as the ESPRIT algorithm, can be used. The geometric constraint model can also be extended as needed, for example, by considering the Earth's curvature or introducing a more refined ground reflection model. This embodiment expands the information acquisition capability of the system under poor observation conditions.

[0115] The embodiment improves the lower limit of system performance and data quality under poor observation conditions. On the one hand, for weak signal scenarios, a three-stage cascade estimation strategy consisting of adaptive grid coarse estimation, reverse injection intermediate estimation and virtual beat frequency fine estimation solves the problem of easy failure of traditional single FFT or phase-locked loop under low signal-to-noise ratio by gradually reducing the uncertainty range from coarse to fine, ensuring high final accuracy while considering the calculation efficiency. On the other hand, for the multipath effect generally existing in the low elevation period, instead of regarding it as pure interference, it is separated by subspace algorithm, and the additional geometric constraint information is used to construct overdetermined equations for solving. The additional propagation path is converted into an additional information source, which improves the data accuracy of the period with the worst observation quality in the traditional method and enhances the robustness and full-arc information acquisition capability of the whole system.

[0116] According to one aspect of the present application, the step of obtaining enhanced Doppler observation data further includes an intelligent data splicing and reconstruction process, which includes: based on the identified mutation characteristics, dividing the observation data into at least one smooth data segment and one mutation transition data segment; evaluating the statistical characteristics of each data segment including smoothness and change rate, and adaptively assigning different interpolation strategies to data segments with different characteristics; at the junction of the smooth data segment and the mutation transition data segment, a five-order polynomial transition function is constructed to ensure the continuity of the second derivative of Doppler; the transition function is applied to the data segments with different interpolation strategies to perform smooth splicing and reconstruct a continuous and smooth complete data sequence.

[0117] In the embodiment, the entire transit observation data sequence is segmented according to the fused mutation feature parameter set. Specifically, the confirmed mutation center time t mut is taken as the boundary, the data is divided into three main parts: the smooth segment before mutation [t start , t mut -Δt mut ], the mutation transition segment [t mut -Δt mut , t mut +Δt mut ], and the smooth segment after mutation [t mut +Δt mut , t end ], where t start is the starting time point of the observation data sequence, Δt mut is the half-width of the mutation window, and t end is the termination time point of the observation data sequence. For each data segment, the system will independently extract its statistical characteristics, such as mean μ i , variance σ i 2 , and autocorrelation function ACF i(τ), thus constructing the piecewise feature description matrix F seg Based on the piecewise feature description matrix F seg , the characteristics of each data segment are evaluated and the most appropriate interpolation strategy is tailored for it. The interpolation strategy refers to the method of constructing new data points between discrete data points. The evaluation indicators can include: smoothness indicator: S i =σ i 2 / μ i 2 ; rate of change indicator: R i =max|df / dt| The rules of adaptive assignment can be: for smooth data segments (S i and R i are small, for example S i <0.1 and R i <1): cubic spline interpolation is selected. This method can generate a globally smooth curve, which is very suitable for describing smooth and slowly changing signals. For monotonically changing segments (S i is moderate, for example 0.1≤S i <0.5): piecewise cubic Hermite interpolation is selected. The advantage of this method is that it can guarantee the monotonicity of the interpolated curve, avoiding unnecessary oscillation (i.e. the Runge phenomenon) in the monotonic data. For abrupt transition data segments (S i is large or R i is large): radial basis function (RBF) interpolation is selected. RBF interpolation has strong flexibility and fitting ability for handling complex and irregularly distributed data points.

[0118] Boundary smoothing splicing based on a fifth-order polynomial In order to avoid the occurrence of creases or sharp points at the junction of data segments using different interpolation strategies, the embodiment constructs a transition function that can ensure high continuity. Specifically, a very short transition zone (e.g. t boundary ±0.5s) is set before and after each junction point t boundary , and a fifth-order polynomial transition function p(t)=a0+a1t+a2t 2 +a3t 3 +a4t 4 +a5t 5 is used for connection in this area. The motivation for choosing a fifth-order polynomial is that it has six undetermined coefficients (a0 to a5), which allows it to satisfy six boundary conditions simultaneously: i.e. at the start and end of the transition zone, the function value (p), the first derivative value (p', representing the velocity) and the second derivative value (p'', representing the acceleration) are equal to those of the interpolation functions of the left and right segments. By solving these six simultaneous boundary conditions, the coefficients of the polynomial can be determined, thus constructing a transition function that is not only C 1 continuous (tangent continuous) but also C2 Smooth stitching with continuous curvature. Meaning that not only there is no kink at the stitching point, but also the smoothness of the kink is continuous.

[0119] Integrity verification and quality assessment After the interpolation and stitching of all data segments are completed, the complete data sequence after reconstruction is obtained. The system also performs several final tests on it: continuity test: calculate the difference between adjacent data points to ensure that there is no jump. Smoothness test: calculate the local curvature to ensure that there is no sharp kink. Physical plausibility test: ensure that the reconstructed Doppler shift value is within its physically possible maximum range. Calculate the comprehensive reconstruction quality index Q, and generate a data quality report to provide reliability reference for subsequent applications using the data.

[0120] Through the above steps, the embodiment can intelligently reconstruct the incomplete and discontinuous original data generated due to various reasons (such as non-uniform sampling, signal interruption) into high-quality, complete and smooth final data products.

[0121] Regarding the intelligent data stitching and reconstruction process, it can be further refined as follows: After the multi-scale data segmentation step based on the mutation feature, the segmentation feature description matrix F seg and the inter-segment relationship matrix R seg will be directly input for the subsequent adaptive interpolation strategy selection. In the integrity verification and quality assessment step, the comprehensive reconstruction quality index Q can be designed in the form of a weighted sum, for example: Q = w1·C cont + w2·C smooth + w3·C physics . Where C cont , C smooth and C physics are the quantitative scores of continuity, smoothness and physical plausibility test respectively, and the weights w1, w2, w3 can be configured according to task requirements, for example, set to 0.4, 0.3 and 0.3 respectively. Finally, the system will output the high-quality reconstructed data sequence, accompanied by a data quality report, which will mark the Q value and the time period found in the test that may have problems for subsequent analysis reference.

[0122] In a specific embodiment, it is assumed that a high-confidence Doppler mutation window has been identified, and the following input parameters are available: Mutation window: determined, the mutation window of this transit occurs in the interval of satellite elevation angles [θ start , θ end ]= [30°, 45°]. Second derivative feature: calculated, the maximum value of the second derivative of Doppler in this window is d 2 f max = 1.2 Hz / s 2 , and the average value is d2 f avg =0.1Hz / s 2 Boundary sampling interval: According to the task requirements, the target sampling interval Δt in the stable sparse area sparse =2s; at the boundary of the mutation window, the target sampling interval Δt in the dense area dense =0.2s. Calculate the ratio of mutation severity r: r=d 2 f max / d 2 f avg =1.2 / 0.1=12; Determine the coefficient k based on the r value: According to the piecewise function rule, since r=12>10, it indicates that the mutation is very drastic, so the smallest proportional coefficient k=0.1 is selected. At the point where the mutation is most drastic, |d 2 f / dt 2 |=d 2 f max =1.2Hz / s 2 The calculated sampling interval is Δt=k / |d 2 f / dt 2 |=0.1 / 1.2≈0.083s. In actual implementation, this value may be limited by the system's minimum sampling interval capability (e.g., 0.1s), so the actual sampling interval can be set to 0.1s. At a point in the mutation window where the change is relatively slow, assuming |d 2 f / dt 2 |=0.2Hz / s 2 The calculated sampling interval is Δt=0.1 / 0.2=0.5s. The boundary from the stable region to the mutation region, i.e. θ boundary =30°. Apply the S-type function for smooth connection, the formula is: Δt trans =Δt sparse +(Δt dense -Δt sparse ) / (1+exp(-a(θ-θ boundary ))); Substitute the parameters (assuming the steepness coefficient a=10): Δt trans =2+(0.2-2) / (1+exp(-10(θ-30))); calculate the sampling interval at a point just entering the transition zone, for example, when θ=30.1°: exp(-10(30.1-30))=exp(-1)≈0.368Δt trans= 2 + (-1.8) / (1 + 0.368) ~ 2 - 1.316 = 0.684 s; it can be seen that when the elevation angle just crosses the 30° boundary by 0.1°, the sampling interval has been rapidly but smoothly reduced from 2 s in the plateau region to about 0.684 s, and will continue to smoothly reduce to the target interval 0.2 s in the dense region as the elevation angle further increases. The step change from 2 s to 0.2 s at 30° is avoided. By iterating the above calculation for the entire transit arc in terms of elevation angle or time, the complete, stepless, non-uniform sampling time sequence {t i} is generated. The system also verifies the sequence, for example, checks and ensures that the total number of sampling points in the [30°, 45°] window is not less than 20, while the sampling interval in the plateau region is not more than 5 seconds, to ensure the effectiveness and completeness of the final plan. This embodiment clearly shows the transformation of abstract mathematical models and rules into specific, executable, intelligent observation sampling plans.

[0123] According to one aspect of the present application, a step of real-time energy monitoring and adaptive adjustment is also included. The specific implementation of this step is that the system continuously and frequently monitors the voltage of the main power bus and the output current of the battery pack on the spacecraft through dedicated sensors. A safety threshold for the battery voltage is set inside the system. During normal operation, if it is detected that the bus voltage drops rapidly within a short time and the drop amplitude exceeds the preset safety threshold, the system determines that there may be an abnormal high-power consumption or battery failure, and immediately triggers an emergency energy management mode. In this emergency mode, the decision logic of the system is no longer to maximize efficiency, but to ensure survival and core tasks. The system will immediately execute the pre-set power reduction plan, for example, temporarily shut down or downgrade non-critical scientific payloads, and prioritize power supply for the current observation task with the highest value (for example, the dense sampling within the mutation window). At the same time, the system will dynamically update an executable sampling schedule based on the current real-time power consumption and remaining power, which may reduce some of the subsequent observation points, but the core goal is to ensure that the current high-value observation task can at least complete 80% of its core part. This mechanism provides a final safety guarantee for the entire observation task.

[0124] According to one aspect of the present application, the original GNSS observations contain multiple error sources (such as ionospheric delay, multipath effect, cycle slip, gross error, etc.), which must be eliminated or suppressed through a systematic processing flow. Therefore, before extracting the Doppler frequency shift data, a multi-source heterogeneous filtering fusion step based on dual-frequency GNSS observations is included, which is specifically: reading the dual-frequency pseudorange observations (ρ L1 , ρ L2 ), dual-frequency carrier phase observations (Φ L1, Φ L2 ) and the signal-to-noise ratio sequence SNR[i] of each channel. Based on these raw data, the quality indicators of each observation are calculated. For example, by combining the pseudorange and carrier phase, the pseudorange multipath indicator MP = |p L1 -p L2 -2.546(Φ L1 -Φ L2 )| is calculated, the magnitude of which can reflect the degree of influence of multipath error. Based on the signal-to-noise ratio and the multipath indicator, the initial weight of each observation is calculated, such as the pseudorange weight w ρ = SNR 2 / (SNR 2 + MP 2 ), the carrier phase weight w Φ = exp(-σ Φ 2 / σ ref 2 ), where σ Φ is the standard deviation of the carrier phase noise corresponding to the current observation segment or a certain observation point, and σ ref is the ideal carrier phase noise level used as a standard reference. All these weights are finally combined into the observation weight matrix W obs , which is used in the subsequent filtering process.

[0125] The state vector X containing state quantities such as satellite position (x, y, z), velocity (vx, vy, vz), receiver clock error, and clock drift is constructed. The state transition matrix is established based on the satellite's orbit dynamics model. In the update step of Kalman filtering, the W obs matrix is read to dynamically adjust the observation noise covariance matrix R, so as to give different trust degrees to observations of different qualities. Further, the system uses a fading memory factor λ to dynamically adjust the process noise Q, so that the filter can adaptively compensate for satellite maneuvers or model errors, achieving adaptive Kalman filtering state estimation.

[0126] In order to eliminate the ionospheric delay, which is the largest error source, the system uses dual-frequency pseudorange and carrier phase to construct ionosphere-free combined observations P IF and L IF . At the same time, in order to detect the possible cycle slip (i.e., sudden jump of integer ambiguity) in the carrier phase observation, a geometry-independent combined observation GF = λ1Φ L1 - λ2Φ L2 is also constructed, where λ1 is the carrier wavelength of L1 band and λ2 is the carrier wavelength of L2 band. By performing sliding window detection on the GF sequence, the time of cycle slip occurrence can be effectively identified, and a cycle slip identification sequence is generated for subsequent processing.

[0127] After Kalman filtering, in order to further eliminate the possible gross errors in the data (i.e., outliers far away from the normal data group), the system adopts the robust estimation method. Specifically, the standardized residual v of each observation is calculated. std . Using the IGG-III robustness weight function, according to v std The size of |v recalculate the weight for each observation. For example, when |v std When | is very small (such as less than 1.5), the weight is 1; when |v std When | is very large (such as greater than 2.5), the weight is directly set to 0; in the middle area, the weight is based on |v std The size of | is smoothly transitioned. The residual calculation and weight update are performed iteratively until the weight converges. Finally, those observations with a weight of 0 are regarded as gross errors and eliminated from the solution, thus outputting a robust positioning solution X with high accuracy and reliability, eliminating various errors. robust And the corresponding positioning precision dilution DOP.

[0128] According to one aspect of the present application, the step of obtaining enhanced Doppler observation data may further include a joint reconstruction process based on multi-resolution analysis. The specific implementation of this process is: obtaining downsampled or non-uniformly sampled observation data, including global data y with low sampling rate L and possible high-sampling rate local data segments y H . Although the signal itself is very complex, it is sparse in a certain transform domain, that is, it can be represented by very few non-zero coefficients. In order to find the most suitable transform domain, an over-complete dictionary D is constructed. The over-complete dictionary is a matrix containing a variety of different basis functions. For example, it also contains the discrete cosine transform (DCT) basis for representing the stationary component, the Gabor atoms for representing the oscillating component, and a variety of wavelet bases for representing transients and singular points. The greedy algorithm of orthogonal matching pursuit (OMP) is used to solve the sparse coefficient vector α. Its goal is to find the sparsest α (that is, the one with the least non-zero items) so that the signal y can be approximately represented as Dα.

[0129] Optimization of the compressive sensing measurement matrix. The measurement matrix Φ mathematically describes the process from the original high-resolution signal to the actually obtained low-resolution or non-uniformly sampled data. To ensure that the original signal can be successfully reconstructed from the under-sampled data, the measurement matrix Φ must satisfy the key condition of the restricted isometry property (RIP). To achieve this, the system can first initialize Φ with a random Gaussian matrix, and then iteratively optimize it using a gradient descent method through a mutual correlation minimization algorithm. The goal of the optimization is to make Φ'Φ (where Φ' is the transpose of Φ) as close to the identity matrix as possible, thereby ensuring that the measurement matrix Φ has good RIP properties, for example, with an RIP constant less than 0.3.

[0130] After constructing the over-complete dictionary D and the optimized measurement matrix Φ, the signal reconstruction problem can be constructed as a convex optimization problem, mathematically expressed as: min ||α||1 subject to ||y - ΦDα||2 ≤ ε. 2 The meaning of this formula is: among all the sparse solutions α that satisfy the error between the reconstructed signal after measurement and the actual observed value y does not exceed the noise level ε, find the solution with the smallest L1 norm ||α||1. This problem can be solved using an efficient algorithm called the alternating direction method of multipliers (ADMM). ADMM decomposes the original problem into several sub-problems that are easier to solve by introducing auxiliary variables and augmented Lagrangian functions, and iteratively updates them alternately until the convergence criterion (e.g., the ratio of the residual to the norm of the solution is less than 10 -4 ) is satisfied. Finally, the optimal sparse coefficient α final is obtained, and the reconstructed signal is x rec =Dα final .

[0131] To further improve the quality of the reconstructed signal, especially when dealing with signals that contain both slowly varying trends and high-frequency details, the system performs multi-resolution fusion on the reconstructed signal x rec . Specifically, x rec is decomposed into a low-frequency component x low (frequency below 10 Hz) and a high-frequency component x high (frequency above 10 Hz). For the low-frequency component x low , a smoothing method such as bicubic interpolation can be used to further improve its resolution and visual effect. For the high-frequency component x high , its original high-resolution features are maintained. In the overlapping area of different resolution data processing, a weighted fusion method is used for smooth transition, for example, a cosine window function is used for weighting to ensure the phase continuity of the final fused signal. The final output high-resolution reconstructed signal x finalThe quality of the reconstructed image can be evaluated by multiple indexes such as mean square error (MSE), peak signal to noise ratio (PSNR) and structural similarity (SSIM).

[0132] According to an aspect of the present application, the raw observed Doppler shift is the result of superposition of two motion components: one is the motion of the satellite relative to the earth, and the other is the motion of the jammer itself. Conventional positioning algorithms usually assume that the jammer is stationary, therefore the motion of the jammer itself becomes a major error source. Therefore, before performing the positioning calculation, a step of predictive Doppler compensation can be included, which is based on the motion parameter estimates of the jammer outputted by the state estimation module (e.g. an extended Kalman filter based module). These parameters include the best estimates of the position, velocity and even acceleration of the jammer at the current time. Based on these motion state parameters, the motion trajectory of the jammer in a small future time window (e.g. 5 to 10 seconds) is predicted forward by kinematic equations. According to the predicted motion trajectory, the expected Doppler shift variation caused only by the motion of the jammer in this time window is accurately calculated. The calculated expected Doppler shift variation caused by the motion of the jammer is subtracted from the raw, mixed observation data. This process is the forward compensation. After the compensation, the data sequence obtained will ideally only contain the Doppler shift component caused by the motion of the satellite, i.e. the pure Doppler data after compensation. The pure Doppler data after the predictive compensation is inputted into the subsequent iterative positioning algorithm for calculation. Since the major error source in the input data has been eliminated in advance, the convergence speed of the positioning algorithm and the final positioning accuracy will be effectively improved. Optionally, in some embodiments, the length of the time window used for the prediction can be adaptively adjusted according to the severity of the motion of the jammer. The more severe the motion, the shorter the prediction window can be to ensure the accuracy of the prediction.

[0133] The application solves two core technical bottlenecks by introducing a mutation event prediction mechanism with foresight and a state perception ability with semantic depth. First, to address the problem of lack of foresight in observation strategy, leading to blind resource allocation, the system is endowed with the ability to predict future key events through two prediction methods. The first is a prediction-correction double-loop mechanism that can generate a high-probability expected window before the mutation window arrives by fusing the prior statistics of the historical knowledge base and the current evidence of real-time observation. The second is a texture fingerprint recognition method that uses a convolutional neural network to learn and identify deep patterns from the time-frequency spectrum of the signal that indicate that a mutation is about to occur. This is the key prerequisite for all subsequent resource optimization. On this basis, the energy pool mechanism can actively degrade operation during low-observation-benefit stable periods to accumulate energy, and then optimize and concentrate these energies into the upcoming high-value mutation window according to the energy-precision benefit function. At the same time, the non-uniform sampling strategy based on the adaptive inverse proportional function also allocates sampling resources accurately to the time period with the highest information density according to the prediction result. The traditional static and passive resource allocation mode is transformed into a dynamic, future key event-oriented, and foresight-based intelligent planning mode.

[0134] Second, to address the problem of lack of semantic depth in state perception, leading to unclear optimization goals, a deep state perception system that can understand the internal physical characteristics of Doppler signals is constructed. Instead of relying solely on macroscopic indicators such as signal strength, it delves into the dynamic nature of the signal, using high-order physical quantities such as Doppler second derivative, texture features of time-frequency spectrum, and signal singularity strength to directly represent kinematic characteristics as the core of state perception. For example, through cross-validation of four heterogeneous features, the system can quantify and confirm the physical properties of mutation events from different dimensions, providing a clear, robust, and explicitly physically meaningful target for optimization decisions. More importantly, the deep reinforcement learning closed-loop optimization framework used includes not only the final positioning result but also the accuracy of the prediction process in the reward function. This encourages the agent (observation strategy network) to continuously learn how to better understand these physical features, thereby achieving continuous self-evolution of semantic depth in state perception.

[0135] The above detailed the preferred embodiments of the application, but the application is not limited to the specific details in the above embodiments, and various equivalent transformations of the technical solutions of the application can be made within the technical concept of the application, and these equivalent transformations all belong to the protection scope of the application.

Claims

1. A Doppler observation strategy optimization method for satellite-borne interference sources, characterized in that: include: Obtain the original RF signal of the onboard interference source and the precise satellite orbit data, predict the mutation characteristics in the Doppler data, and generate an adaptive observation plan; According to the adaptive observation scheme, energy-aware resource scheduling is performed and observations are performed to obtain enhanced Doppler observation data; Perform positioning calculations and performance evaluation based on enhanced Doppler observation data, and generate parameters for iteratively optimizing observation strategies. Obtain enhanced Doppler observation data, including: Identify energy accumulation windows defined by low elevation angle periods in adaptive observing schemes, and high-value observing periods defined by Doppler mutation characteristics; In the energy accumulation window, by reducing the working mode of the receiver, the energy saved is accumulated and counted into the energy reserve unit; When entering a high-value observation period, evaluate the matching degree between the balance of the energy storage unit and the power consumption demand of that period; Under the condition that the balance of the energy reserve unit meets the power consumption requirement, high-density sampling is authorized to obtain enhanced Doppler observation data.

2. The method according to claim 1, characterized in that Evaluate the matching degree between the balance of the energy reserve unit and the power consumption demand during the period, including: Construct an energy-accuracy benefit function to characterize the relationship between observation accuracy and energy consumption; Based on the energy-accuracy benefit function, a segmented benefit evaluation matrix is ​​generated for different power consumption modes during high-value observation periods. Establish a multi-constraint optimization model with the goal of maximizing total benefits and integrating energy storage unit balance and mode switching cost as constraints; The multi-constraint optimization model is solved to obtain the optimal power allocation scheme, and high-density sampling is performed accordingly.

3. The method according to claim 1, characterized in that Also includes: When the assessment determines that the balance of the energy storage unit cannot meet the power consumption demand, the degradation strategy is initiated; The degradation strategy calculates the ratio of the current remaining energy to the expected demand of the subsequent task and compares it with at least one preset degradation threshold; If the ratio is lower than the degradation threshold, a downgraded sampling plan is generated to reduce the total energy consumption. The downgraded sampling plan is achieved by increasing the sampling interval of non-critical observation areas or shrinking the spatial range of high-density sampling. Observations are performed according to a downsampling plan.

4. The method according to claim 1, wherein When the original RF signal is lower than the preset threshold, the process of obtaining enhanced Doppler observation data is achieved by using a step-by-step Doppler frequency cascade estimation method, specifically: The frequency grid is adaptively adjusted based on the spectrum energy distribution to perform preliminary spectrum peak location and obtain a rough Doppler estimate. Based on the coarse Doppler estimate, the intermediate Doppler estimate is obtained by injecting the reverse Doppler sequence and searching for the solution that minimizes the spectral flatness of the compensated signal. Based on the Doppler mean estimation, a virtual mirror signal is constructed and iteratively adjusted to minimize the time rate of change of the beat frequency component generated by its cross-correlation with the original signal, and the Doppler precise estimation is obtained. Based on this, the enhanced Doppler observation data is sampled.

5. The method according to claim 4, characterized in that By constructing a virtual mirror signal and iteratively adjusting it, a precise Doppler estimation is obtained, including: Using the conjugate of the original RF signal and taking the Doppler mean value as the frequency basis, a virtual mirror signal is constructed; Calculate the cross-correlation function between the original RF signal and the virtual mirror signal, and separate the beat frequency component therefrom; The frequency used to construct the virtual mirror signal is iteratively adjusted until the time rate of change of the beat frequency component converges to a preset minimum value, and the converged frequency is determined as the Doppler precise estimate.

6. The method according to claim 4, characterized in that When acquiring enhanced Doppler observation data, when sampling is performed in a non-uniform manner, a data reconstruction process is also included, specifically: Obtaining a precise Doppler estimate and using it to pre-compensate the non-uniformly sampled observation data to obtain a compensated stationary signal; For the compensated stationary signal, solve the pre-built compressed sensing convex optimization model to obtain the reconstructed stationary signal component; The frequency trend represented by the Doppler precise estimate is re-assigned to the reconstructed stationary signal component to generate enhanced Doppler observation data.

7. The method according to claim 1, characterized in that Predicting mutation signatures in Doppler data, including: Perform spectrum analysis on the original RF signal and, combined with the satellite's precise orbit data, extract the time-series Doppler shift data through a frequency tracking loop; Perform time-frequency transformation on the time series Doppler frequency shift data to generate a time-frequency spectrum and process it into a grayscale image; Extract local binary patterns and gray-level co-occurrence matrices from grayscale images as texture features and construct multi-dimensional texture feature vectors; The multidimensional texture feature vector is fed into a pre-trained convolutional neural network to parse and output the mutation type and mutation time in the Doppler data.

8. The method according to claim 1, characterized in that Predicting mutation features in Doppler data can also be: Extract the time-series Doppler frequency shift data from the original RF signal and calculate the corresponding Doppler second-order derivative sequence; The historical transit mutation point knowledge base is called up and combined with the current satellite precise orbit data to generate the initial predicted mutation position; Real-time monitoring of the Doppler second-order derivative sequence. When the measured characteristics of the sequence deviate from the expected characteristics of the initially predicted mutation position and exceed a preset threshold, a correction mechanism is triggered. The correction mechanism fuses the initial prediction with the measured features and outputs a corrected mutation window.

9. The method according to claim 8, characterized in that Correction mechanisms include: A Kalman filter is used to fuse the initial predicted mutation position with the measured features extracted from the Doppler second-order derivative sequence; The Kalman filter outputs an optimized corrected mutation position and a confidence weight coefficient representing the credibility of the corrected position; The boundary of the modified mutation window is adaptively set according to the modified mutation position and the confidence weight coefficient.

Images