On-orbit energy management methods and systems for damaged satellites

Abstract

This application provides an on-orbit energy management method and system, network equipment, computer-readable storage medium, and computer program product for damaged satellites. The method includes autonomously deciding whether to execute SADA (Satellite Adaptive Damage Control) rotation angle inference based on the satellite's current state data, and selecting an inference model under current energy constraints; simulating the regression relationship between the SADA rotation angle and the predicted solar panel current based on the selected inference model and the satellite's current state data; searching for the SADA target rotation angle that maximizes the predicted solar panel current from the regression relationship using a Bayesian optimization strategy, and converting the SADA target rotation angle into SADA control commands. This application constructs a damaged satellite energy management framework, enabling real-time and accurate prediction of the optimal solar angle for the peak current of the solar panels under the current damaged state, avoiding energy fluctuations caused by changes in satellite attitude and solar vector, and providing an important basis for autonomously maintaining energy supply onboard.

Description

Technical Field

[0001] This application mainly relates to the field of satellite attitude control technology, and in particular to an on-orbit energy management method and system, network equipment, computer-readable storage medium, and computer program product for damaged satellites. Background Technology

[0002] In spacecraft energy systems, the Solar Array Drive Assembly (SADA) is a core power generation component, and its proper functioning is crucial for maintaining the overall energy security of the spacecraft. Traditional SADA control relies on a dynamic model based on the rigid body assumption and uses pre-programmed procedures or ground commands for solar-oriented tracking to ensure maximum power generation efficiency. This technology is mature and effective provided the spacecraft structure remains intact.

[0003] However, when solar panels suffer structural damage due to abnormal events, the aforementioned traditional technical system becomes ineffective. For example, in a certain launch mission, an anomaly in the launch vehicle severely damaged the mechanical connection between the satellite's solar panels and the satellite body, creating a "weak connection" state. This state completely undermines the rigid body dynamics assumption upon which traditional SADA control relies, making any autonomous rotation control based on this model susceptible to sudden drops in energy output or even secondary structural damage due to unpredictable vibrations and deformations.

[0004] To address the aforementioned damage and ensure satellite energy supply and proper solar panel alignment, the current approach involves ground-based manual adjustment and command injection. SADA (Satellite Adaptive Dynamics and Data Acquisition) commands are sent to drive the solar panels to rotate near the potentiometer angle where peak currents might occur. However, this manual command injection method has several drawbacks: it operates in a "blind zone" outside the telemetry and control arc, during which ground monitoring of the satellite's status is impossible and telemetry information cannot be transmitted in real time, preventing manual intervention; and it requires manual assessment, approval, and command injection, resulting in a slow response time and an inability to make immediate adjustments based on satellite attitude and solar vector.

[0005] In summary, current technologies lack an on-orbit energy management method capable of autonomously, rapidly, and intelligently responding to damaged satellites. How to dynamically find and track the optimal power generation angle under damaged conditions in real time, without relying on ground-based manual intervention or rigid body models, has become a core technical challenge that urgently needs to be solved to salvage such on-orbit failure missions and ensure the long-term survival of satellites. Summary of the Invention

[0006] The technical problem this application aims to solve is the lack of an existing method for on-orbit energy management that can autonomously, quickly, and intelligently respond to damaged satellites.

[0007] To address the aforementioned technical problems, in a first aspect, this application provides an on-orbit energy recovery method for a damaged satellite, wherein the mechanical connection between the solar panels and the satellite body is damaged. The method includes: autonomously deciding whether to perform SADA rotation angle inference based on the current satellite status data, and selecting an inference model under current energy constraints; simulating a regression relationship between the SADA rotation angle and the predicted current of the solar panels based on the selected inference model and the current satellite status data; searching for the SADA target rotation angle that maximizes the predicted current of the solar panels from the regression relationship based on a Bayesian optimization strategy, and converting the SADA target rotation angle into an SADA control command.

[0008] Secondly, this application provides an on-orbit energy rescue system for a damaged satellite, wherein the mechanical connection between the solar panels and the satellite body is damaged. The system includes: a rescue decision module, used to autonomously decide whether to perform SADA rotation angle inference based on the current satellite status data, and to select an inference model under the current energy constraints; a solar panel current prediction module, used to simulate the regression relationship between the SADA rotation angle and the predicted solar panel current based on the selected inference model and the current satellite status data; and a rotation angle inference module, used to search for the SADA target rotation angle that maximizes the predicted solar panel current from the regression relationship based on a Bayesian optimization strategy, and to convert the SADA target rotation angle into SADA control commands.

[0009] Thirdly, this application provides a network device including a processor and a memory; the memory stores computer-executable instructions; the processor is used to execute the computer-executable instructions stored in the memory so that the network device performs the on-orbit energy rescue method described in this application.

[0010] Fourthly, this application provides a computer-readable storage medium including computer program instructions, which, when executed by a processor, enable the processor to perform the on-orbit energy rescue method described in this application.

[0011] Fifthly, this application provides a computer program product containing instructions that, when executed by a processor, cause the processor to perform the on-orbit energy rescue method described in this application.

[0012] Compared with the prior art, this application has the following advantages:

[0013] This application discloses an on-orbit energy management method and system, network equipment, computer-readable storage medium, and computer program product for damaged satellites. It constructs an energy management framework for damaged satellites, enabling real-time and accurate prediction of the optimal solar angle for peak current in the current damaged state of the solar panels. This avoids energy fluctuations caused by changes in satellite attitude and solar vector, providing crucial information for autonomously maintaining energy supply onboard. Simultaneously, it enhances autonomous management capabilities in overseas telemetry and control blind spots, significantly reducing the frequency of ground intervention and maintenance costs, and providing a solid technical foundation for subsequent deep space exploration missions. This technology is expected to be extended to the energy management of more types of spacecraft, ensuring that satellites can maintain basic functions even when some sensors or actuators fail, providing strong support for the long-term stable operation of space missions. Attached Figure Description

[0014] The accompanying drawings are included to provide a further understanding of this application. They are incorporated into and constitute a part of this application. The drawings illustrate embodiments of this application and, together with this specification, serve to explain the principles of this application.

[0015] Figure 1 This is a structural schematic diagram of one embodiment of the satellite.

[0016] Figure 2 This is a flowchart of an on-orbit energy management method for a damaged satellite according to an embodiment of this application.

[0017] Figure 3 yes Figure 2 A flowchart of an embodiment of step S2.

[0018] Figure 4 yes Figure 2 A flowchart of an embodiment of step S3.

[0019] Figure 5 This is a system block diagram of an on-orbit energy recovery system for a damaged satellite according to an embodiment of this application.

[0020] Figure 6 This is a system block diagram of a network device according to an embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0022] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, these terms have no special meaning and therefore should not be construed as limiting the scope of protection of this application. In addition, although the terminology used in this application is selected from commonly known and used terms, some terms mentioned in this application's specification may have been chosen by the applicant according to his or her judgment, and their detailed meanings are explained in the relevant sections of this description. Moreover, this application should be understood not only through the actual terms used, but also through the meaning implied by each term.

[0023] Flowcharts are used in this application to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed in exact order. Instead, various steps can be processed in reverse order or simultaneously. Furthermore, other operations may be added to these processes, or one or more steps may be removed from these processes.

[0024] To make the objectives, technical solutions, and advantages of this application clearer, a further detailed description of this application will be provided below with reference to the accompanying drawings. The specific operating methods and functional descriptions in the method embodiments can also be applied to the device embodiments or system embodiments.

[0025] Figure 1 This is a structural schematic diagram of one embodiment of the satellite. (See diagram below.) Figure 1 As shown, satellite 1 includes a satellite body 11, a solar panel drive unit (SADA) 13, and solar panels 13. Under normal design conditions, the SADA 12 is rigidly connected to both the satellite body 11 and the solar panels 13. The solar panels 13 rotate under the drive of the SADA 12, achieving one-dimensional solar-oriented tracking. Given a dynamic model, the target rotation angle for solar panel tracking can be calculated using the known satellite attitude and solar vector.

[0026] When a satellite suffers structural damage due to an abnormal event, the original dynamic model based on rigid connection is not applicable because the solar panels are only weakly rigidly connected to the satellite body and there is a risk of cable snagging. Considering the overall energy security of the satellite, it is not recommended that SADA autonomously drive the solar panels to rotate in orbit for an extended period of time.

[0027] To address the aforementioned challenges, this application proposes an on-orbit energy management method for damaged satellites. This method enables real-time and accurate prediction of the optimal solar angle (i.e., the SADA target rotation angle) for peak current in the current damaged state of the solar panels, avoiding energy fluctuations caused by attitude and solar vector changes, and providing crucial information for autonomous energy supply maintenance onboard. Simultaneously, it enhances the autonomous management capabilities in overseas telemetry and control blind spots, significantly reducing the frequency of ground intervention and maintenance costs, and providing a solid technical foundation for subsequent deep space exploration missions. This method is expected to be extended to energy management of more types of spacecraft, ensuring that satellites can maintain basic functions even when some sensors or actuators fail, providing strong support for the long-term stable operation of space missions.

[0028] Figure 2 This is a flowchart of an on-orbit energy management method for a damaged satellite according to an embodiment of this application. Figure 2 As shown, the on-orbit energy management method 200 for damaged satellites includes:

[0029] Step S1: Based on the current satellite status data, autonomously decide whether to perform the SADA rotation angle inference and select the inference model under the current energy constraints;

[0030] Step S2: Simulate the regression relationship between the SADA rotation angle and the predicted current of the solar panels based on the selected inference model and the current satellite status data;

[0031] Step S3: Based on the Bayesian optimization strategy, search the regression relationship for the SADA target rotation angle that maximizes the predicted current of the solar panel, and convert the SADA target rotation angle into SADA control commands.

[0032] The following provides a detailed explanation of steps S1 to S3:

[0033] In step S1, the satellite's current status data includes at least the solar panel power generation at the current moment, the satellite's total power consumption, the remaining battery power at the current moment, and the estimated remaining mission time.

[0034] The following state variables represent the various data points in the current satellite state data:

[0035] E available (t): The remaining charge of the battery at time t.

[0036] P total_consumption (t): Total power consumption of the satellite at time t. In some embodiments, in addition to the energy subsystem, the satellite also includes several other subsystems that consume energy:

[0037]

[0038] in, For the power consumption of the energy subsystem, Let be the power consumption of the i-th task.

[0039] P solar_gen (t): Solar power generation at time t.

[0040] T rem (t): The estimated remaining time for the mission at time t. This time can be estimated based on different predetermined orbital strategies, such as a preset duration for the initial orbit insertion segment, a preset duration for the transfer orbit segment, and the remaining time after entering orbit can be calculated based on mission requirements.

[0041] In meeting the energy constraint: E available (t)≥E min_threshold Under the premise that the current remaining battery power is greater than or equal to the minimum energy threshold limit, decide whether to perform rotation angle inference Dangle∈{0,1}.

[0042] First, calculate the current net power consumption of the system:

[0043]

[0044] if A negative value indicates that current power generation exceeds consumption, and there is sufficient surplus energy. Therefore, the reasoning based on the rotation angle will not be considered, i.e., D. angle =0; otherwise, it enters an energy-limited state.

[0045] Calculate how long the current remaining energy can sustain the system:

[0046]

[0047] Set a safety margin time and the battery's critical charge ,if or If the remaining energy is insufficient to sustain the current task or the remaining power is below the critical level, then at time t, the rotation angle reasoning is executed immediately, i.e., D. angle =1, and select the appropriate inference model based on the current state.

[0048] Specifically, if In high-power tasks (such as K-to-ground data transmission or laser link establishment), given the significant constraints on energy and computing resources, the system will prioritize the inference model with the lowest inference latency. For example, the LightGBMXT_BAG_L1_FULL inference model is selected from the 12 inference models in Table 1. Although this model sacrifices a slight amount of accuracy (RMSE of -0.1457), its inference latency is only 6.06 milliseconds, greatly reducing the computational load and ensuring the system's real-time response capability and stability under extreme resource-constrained conditions.

[0049] Additionally, if This indicates the entry into a low-power mission phase (such as initial orbit insertion or solar panel deployment), where system energy is most abundant, and the primary goal is to achieve the highest possible prediction accuracy. At this stage, the solution will employ the high-precision inference model WeightedEnsemble_L2_FULL. With its outstanding RMSE performance of -0.1448, this model ensures the most reliable current predictions at critical moments, even with an inference latency of 45.6 milliseconds, significantly improving the accuracy of onboard health assessments.

[0050] In addition, to cope with special operating conditions, such as transient high power consumption moments (caused by sudden startup of propulsion components or sensors), the system can quickly identify sudden energy surges, immediately start the aforementioned inference model with the least inference latency to shorten the reaction time, and automatically restore the global model strategy after the disturbance ends.

[0051] Table 1. Alternative Inference Models

[0052]

[0053] Figure 3 yes Figure 2 A flowchart of an embodiment of step S2. (See attached flowchart.) Figure 3 As shown, the regression relationship between the SADA rotation angle and the predicted current of the solar panels, simulated based on the selected inference model and the current satellite status data, includes:

[0054] Step S21: Obtain the basic feature values ​​that are independent of angle from the current satellite status data;

[0055] Step S22: Initialize a Bayesian optimizer. The Bayesian optimizer uses a Gaussian process regression model to model the objective function, which is the regression relationship between the SADA rotation angle and the predicted current of the solar panel given the basic feature values.

[0056] Step S23: Randomly sample N discrete initial angle points within the angle range of SADA rotation angle. For each sampling point, combine its sine and cosine values ​​with the basic feature values ​​to form a complete feature vector and input it into the inference model to obtain the solar panel predicted current. Use the initial angle points and their corresponding solar panel predicted currents to train the Gaussian process regression model.

[0057] The following is a framework based on Bayesian optimization, but enhanced for the specific problem of "angle" (which is periodic, i.e., 360° equals 0°).

[0058] The inputs to the Gaussian process regression model are the basic eigenvalues ​​x_b and the SADA rotation angle. The output is the predicted current of the solar panel. The fundamental eigenvalue x_b is a d-dimensional fundamental eigenvector that contains factors affecting the current output other than the angle, such as: satellite attitude, solar vector, operating period, load current, charging current, discharging current, bus voltage, battery pack voltage, etc.

[0059] The modeling process is as follows:

[0060] 1) Periodic transformation:

[0061]

[0062] The purpose of this step is to address the issue that 359° and 1° differ significantly in numerical value but are nearly identical in physical sense. By mapping them to a sine-cosine space, the two points become very close. This is a preprocessing technique used in this application for periodic variables.

[0063] 2) Initial sampling

[0064] Latin hypercube sampling (LHS) was used to select samples within a range of 1° to 360°. An initial angle Latin hypercube sampling ensures that samples are sufficiently and uniformly distributed in space, making it more efficient than random sampling.

[0065] 3) Gaussian process modeling

[0066] For each initial angle The selected inference model f is used to predict the current that the solar panel can generate at that angle. This model f establishes "basic features x_b + angle". "and "output current" The relationship between "".

[0067] Construct the initial dataset D0 as {(transformed angles)} Predicted current value The set of )}.

[0068] A Gaussian process regression (GPR) model is used as a surrogate model to simulate the unknown objective function, which is the regression relationship between the SADA rotation angle and the predicted current of the solar panel given the basic eigenvalues.

[0069] In some embodiments, a Gaussian process regression model with the Matérn kernel function is chosen to model the objective function. The Matérn kernel function is better able to handle potentially non-smooth or "coarse" function shapes than the commonly used RBF kernel, which is more in line with the actual situation of damaged solar panels.

[0070] After obtaining the regression relationship between the SADA rotation angle and the predicted current of the solar panel, the traditional method to find the optimal SADA target rotation angle is through exhaustive search (i.e., traversing 1° to 360°) to find the angle corresponding to the peak current. However, this method has obvious inefficiency. This paper proposes an intelligent optimization scheme based on Bayesian optimization. The core of this scheme lies in integrating the Bayesian optimization strategy. By designing an inference model bias correction mechanism and the Bayesian optimization optimization strategy, the target rotation angle of the SADA can be identified efficiently and accurately.

[0071] The inference model bias correction mechanism includes: calculating a global scaling factor to correct the deviation between the solar panel predicted current and the actual solar panel current; performing a standardized inverse transform on the solar panel predicted current, and multiplying the inversely transformed solar panel predicted current by the global scaling factor to obtain the solar panel corrected current; and training a Gaussian process regression model using the initial angle point and its corresponding solar panel corrected current.

[0072] To address the potential systematic discrepancy between model predictions and actual current values, this application introduces a wrapper called ScaledPredictorWrapper. The design of this wrapper stems from observations in previous experiments: while the model can accurately fit the trend of current variation with angle, it often exhibits a fixed proportional deviation in terms of numerical magnitude. Therefore, the core function of this wrapper is to calculate a global scaling factor to correct for this systematic bias.

[0073] The scaling factor calculation process is as follows: The wrapper automatically reads the training dataset and selects the first 100 valid samples (i.e., excluding samples with current values ​​close to zero to avoid numerical instability). For each valid sample, the system constructs a standardized set of basic feature vectors, including basic features such as satellite attitude, solar vector, operating period, charging and discharging current, and battery voltage, and fixes the angle feature. Subsequently, the original model predicts these constructed features. The prediction results are first subjected to a standardized inverse transform (i.e., the predicted value is multiplied by the standard deviation of the training data current and the mean is added) to restore it to an order of magnitude close to the original current. Next, the scaling factor of a single sample is calculated by dividing the actual current value of the sample by the predicted value after the inverse transform. After collecting the scaling factors of all valid samples, the system uses the median as the initial estimate and further filters out outliers that deviate too far from the median. Finally, the average of all filtered scaling factors is taken as the global scaling factor.

[0074] ScaledPredictorWrapper not only addresses the systematic magnitude bias in model predictions but also plays a crucial role in the actual prediction process. During Bayesian optimization, at each prediction stage, the wrapper first calls the selected inference model to obtain the raw prediction value, then applies the inverse normalization transform of the training data current. Next, this inversely transformed prediction value is multiplied by a pre-calculated global scaling factor for final correction, ensuring the output is non-negative. This mechanism effectively compensates for the potential loss of normalized parameters during model saving and loading, ensuring the accuracy and stability of the prediction results.

[0075] The Bayesian optimization strategy consists of two phases. The first phase is the initial exploration phase, where the system randomly samples five discrete angle points. For each sampled point, a complete feature vector is constructed (containing basic features and the corresponding sine and cosine values ​​of the angle), and the packaged inference model is used for current prediction. These prediction results will be used to initially train the Gaussian process model.

[0076] Then, the Bayesian optimization iterative phase begins. In each iteration, the Gaussian process model calculates the acquisition function based on the existing evaluation history. By maximizing this acquisition function, the optimizer can intelligently select the next angle point that is most likely to find a better result.

[0077] Figure 4 yes Figure 2 A flowchart of an embodiment of step S3. (See attached flowchart.) Figure 4 As shown, step S3 includes:

[0078] Step 31: Calculate the acquisition function based on the Gaussian process regression model, and select the next angle point to be evaluated by maximizing the acquisition function;

[0079] Step 32: Combine the sine and cosine values ​​of the next angle point to be evaluated with the basic feature values ​​to form a complete feature vector and input it into the inference model to obtain the solar panel prediction current. Use the next angle point to be evaluated and its corresponding solar panel prediction current to train the Gaussian process regression model.

[0080] Step 33: Repeat steps 31 to 32 until the predetermined termination condition is met. Finally, the angle that maximizes the predicted current of the solar panel found during the iteration process is determined as the SADA target rotation angle.

[0081] The acquisition function balances "exploration" (trying new areas with high uncertainty) and "exploitation" (further searching areas that are known to perform well). Instead of using a single function, the algorithm innovatively employs a dynamically weighted combination.

[0082] In some embodiments, the acquisition function is a dynamically weighted hybrid acquisition function, which is a weighted sum of the expected improvement (EI), the upper confidence boundary (UCB), and the probability of improvement (POI), wherein the weights are dynamically adjusted according to the progress of the iteration.

[0083] Specifically, the Expected Improvement Index (EI) tends to select points that represent a significant improvement over the current best value. The Upper Confidence Boundary (UCB) tends to select points where the model's predicted value is high and the uncertainty is also high. The Probability of Improvement (POI) tends to select points where there is a high probability of exceeding the current best value.

[0084] In one embodiment, the weight values ​​are adjusted as follows during the iteration process, where w_EI is the weight for the desired improvement, w_UCB is the weight for the upper confidence boundary, and w_POI is the weight for the improvement probability:

[0085] First half (t ≤ T / 2): w_EI = 0.5, w_UCB = 0.3, w_POI = 0.2. At this stage, exploration is emphasized (high weight for EI), aiming to gain a broad understanding of the performance across the entire angle space.

[0086] In the second half (t > T / 2): w_UCB = 0.5, w_EI = 0.3, w_POI = 0.2. At this point, the emphasis is placed on utilizing (higher UCB weight) to focus on a refined search within the identified promising regions.

[0087] The weight of POI is always fixed at 0.2, which plays a stabilizing role.

[0088] By maximizing the acquisition function obtained in the previous step, the next most worthwhile angle to evaluate is found within the angle boundary. The inference model f is used again to "evaluate" this new angle θ_{t+1}, obtaining its predicted current y_{t+1}. This new data point (t(θ_{t+1}), y_{t+1}) is added to the dataset D_{t+1}, and the Gaussian process regression is updated to make its understanding of the objective function more accurate.

[0089] Find the angle θ_best that has the largest predicted current value from all evaluated points in the final dataset D_T.

[0090] In some embodiments, the method further includes: performing a local search within a preset range of the target rotation angle of the SADA to find the SADA rotation angles that make the predicted current of the solar panel among the top K values ​​within the preset range, and using the SADA rotation angles corresponding to the top K values ​​as candidate values ​​for the target rotation angle of the SADA.

[0091] For example, a fine local search is performed within ±3° of θ_best to find the exact optimal point θ* within this small interval. This is because an error of ±3° is acceptable in engineering, so there is no need to pursue absolute mathematical optimality; finding a good solution within this small range is sufficient. The final step of local refinement fully considers the fault tolerance requirements and computational efficiency of practical engineering, pursuing practical optimality rather than mathematical perfection.

[0092] In summary, the SADA angle reasoning process of this application consists of two steps: solar panel current prediction and SADA target rotation angle reasoning. First, a regression problem is constructed to predict the current of the solar panel under different SADA angles under different operating conditions. Then, based on the predicted current value, a Bayesian optimization algorithm is used to find the Top-3 potentiometer angles derived from the optimal current value. The Top-1 angle value is used as the preferred recommendation, and the other angle values ​​are used as alternatives, serving as the decision-making basis for the subsequent fault handling module.

[0093] In some embodiments, the method further includes constructing multiple inference models and uploading these models to the satellite. Constructing multiple inference models includes: receiving multi-source telemetry data from the satellite; preprocessing and selecting features from the telemetry data to obtain feature vectors for predicting solar panel currents; using the feature vectors as input values ​​to the inference models and the predicted solar panel currents as output values, constructing multiple inference models based on different types of regression models.

[0094] Since there are many telemetry measurements related to satellite and solar panel energy, in order to ensure the accuracy of data inference, this paper first analyzes the correlation of data features and designs a feature selection integration strategy. Highly relevant features are selected through voting using multiple methods to ensure the accuracy of subsequent model construction.

[0095] Specifically, telemetry measurements related to solar panel energy include: SADA adapter unit 5V voltage, SADA adapter unit 3.3V voltage, SADA adapter unit 1.8V voltage, SADA reference voltage 1, SADA reference voltage 2, SADA +28V voltage, SADA -12V voltage, SADA +12V voltage, SADA -5V voltage, SADA current value, Y1 solar cell current, Y2 solar cell current, load current (DC), charging current (DC), discharging current (DC), bus voltage (DC), battery pack voltage (DC), positive Y potentiometer angle (i.e., positive Y-wing SADA rotation angle), and negative Y potentiometer angle (i.e., negative Y-wing SADA rotation angle). Furthermore, according to the above formulas, when the satellite SADA and solar panel are in a normal rigid body connection state, it is necessary to consider the solar vector and satellite attitude (both including X, Y, and Z axis characteristics) to calculate the solar panel's angle relative to the sun to ensure maximum energy supply. Therefore, this scheme also takes these two variables into account. Furthermore, since the SADA and the sail are currently connected in a weakly rigid body manner, different thruster jetting conditions may correspond to different mechanical connection relationships, implying different mapping relationships. Therefore, we consider the jetting conditions of different thrusters as "operating conditions" features to simulate the mapping relationships of different weakly rigid body connections through training a single model. Additionally, to normalize the feature data, non-angular feature values ​​are normalized to the range [-1, 1]. For potentiometer angle feature values, considering that the sine and cosine values ​​corresponding to the potentiometer angle also fall within the range [-1, 1], the normalization of the potentiometer angle values ​​involves directly converting them into the corresponding sine and cosine values ​​as additional input features. The final predicted potentiometer angle needs to be transformed by arcsine and cosine before it can be used as the output result.

[0096] Accordingly, this application summarizes a dataset containing 30 feature variables and 1 target variable (solar panel current). The dataset is loaded into a feature matrix X and a target vector y using MATLAB data files. During the data cleaning stage, missing values ​​were first detected and processed by replacing all missing values ​​(NaN) in the feature matrix X and target vector y with 1 to ensure data integrity. Outlier handling was then performed using Winsorization, calculating the 1% and 99th quantiles of each feature as cutoff boundaries to limit feature values ​​to the corresponding intervals, effectively reducing the interference of extreme values ​​on subsequent analysis. Analysis revealed that when the data variance was below 0.1, the feature data showed very small variations across different time periods, which could be considered constant features. These features were directly deleted during the preprocessing stage to avoid significant interference with subsequent correlation calculations. Finally, the processed data was Z-score standardized to eliminate dimensional differences, ensuring that each feature had zero mean and unit variance, laying the foundation for subsequent analysis.

[0097] In the feature selection stage, this application employs three methods for comparative analysis. First, the filtering method selects features based on their statistical correlation with the target variable. It calculates and ranks the correlation coefficients between each feature and the target current, selecting the top 10 most relevant features. Second, the wrapping method uses a recursive feature elimination (RFE) strategy. Using LightGBM as the base model, it progressively eliminates redundant features through 5-fold cross-validation and backward selection, determining the optimal feature subset based on the principle of minimizing cross-validation error. Third, the embedding method utilizes the feature selection mechanism built into Lasso regression. It determines the optimal regularization parameter through 10-fold cross-validation, retaining features corresponding to non-zero coefficients.

[0098] To integrate the advantages of the three methods, this application designs a feature selection integration strategy, which counts the frequency of each feature being selected, retains features that are selected by at least two methods, and forms the final feature subset.

[0099] This application utilizes the voting results of three feature selection methods on 26 features. The key features selected by all three methods include satellite attitude, solar vector, operating period, potentiometer angle (SADA rotation angle), and angle source. Subsequent modeling will also be based on these features.

[0100] The feature selection and integration strategy proposed in this application is interpretable. Through screening, it was found that features with weak correlation to solar panel current mainly fall into two categories: one is telemetry data with a slow refresh cycle (i.e., "slow telemetry"), such as solar panel temperature. These features cannot be used as auxiliary judgment criteria for real-time monitoring status during rotation, and their granularity is coarse, only reflecting the general trend of solar rotation. The other category is constant features with low variance, such as SADA voltage / current, load / charging / discharging current, etc. These features are health status indicators; when abnormal conditions occur, they deviate from standard values. Since the SADA adaptation unit is working normally under the current satellite scenario, the relevant voltage and current values ​​are within the normal range and do not change significantly over time, therefore, a strong correlation cannot be established with the solar panel current value. The remaining screened features are not only highly correlated with solar panel current but also belong to short-cycle, rapidly transmitted telemetry data (i.e., "fast telemetry"), which can effectively ensure the accuracy and speed of inference and improve the efficiency of onboard autonomous management.

[0101] To accurately predict the solar panel current value, the construction of the inference model can be regarded as a regression problem. Specifically, the solar vector, satellite attitude, operating time period, angle source (source 1 indicates a solar panel in the positive Y direction, source 2 indicates a solar panel in the negative Y direction), potentiometer angle, load current, charging current, discharging current, and battery voltage are used as input features, and the ±Y-wing solar panel current intensity is used as a label to construct the corresponding regression model. This application uses Autogluon to train and predict various regression models, and constructs 12 inference models as shown in Table 1.

[0102] In some embodiments, autonomous updating of the inference model is also included. During satellite operation, telemetry data and physical environment characteristics change dynamically over time, leading to a decline in model performance (concept drift). Traditional methods rely on ground-based retraining and uploading of the model, which is limited by satellite-to-ground communication bandwidth and link stability, and has a slow response time. Limited onboard computing resources (CPU, memory, power consumption) also make complete model retraining impractical. Therefore, designing a lightweight and robust online update mechanism is crucial to ensure that the model always accurately reflects the current operational status of the satellite.

[0103] The data required for online updates comes from the telemetry data stream continuously collected during the satellite's on-orbit operation. This data includes real-time satellite attitude, solar vector, SADA potentiometer angles, and actual solar panel current values. The update triggering mechanism can employ the following strategies:

[0104] Periodic updates: The system automatically initiates a model fine-tuning every fixed time interval (such as 24 hours, 7 days or a specific orbital cycle).

[0105] Performance degradation trigger: Continuously monitor the model's prediction error (e.g., by comparing the model's predicted current with the actual current after SADA rotation). If the moving average of the prediction error continuously exceeds a preset threshold... This will trigger a model update.

[0106] Specific event triggering: After completing a major maneuver (such as track control or attitude adjustment) or detecting an abnormal physical event (such as an impact), an update is immediately triggered to adapt to the new operating conditions.

[0107] Online updates to the inference model employ incremental learning or model fine-tuning strategies, rather than resource-intensive full retraining. For lightweight tree models like LightGBM, their inherent support for incremental training is utilized. When an update is triggered, the latest on-orbit data is collected as a mini-batch dataset. .Model New models can be formed by adding new trees or fine-tuning the parameters of existing trees to adapt to new data. Here, a smaller learning rate will be used for fine-tuning to avoid catastrophic forgetting of historical knowledge.

[0108] For stacked ensemble models (such as WeightedEnsemble_L2_FULL), retraining all base models and ensemble layers is impractical due to their complexity. This solution employs online adjustment of ensemble layer weights. During online updates, the latest on-orbit data is utilized. The base model is redefined by minimizing the error on the validation set. The weights are determined by the data distribution. This method is computationally inexpensive and can quickly adapt to changes in data distribution without retraining complex sub-models.

[0109] In addition, to ensure the stability and security of online updates, rigorous quality checks and outlier filtering must be performed on the on-orbit data used for updates to prevent noisy data from contaminating the model. The system must retain the previous stable version of the current model. If the updated model exhibits performance degradation or abnormal behavior in the test set or during the initial operation phase, it can be immediately rolled back to the previous version. Within the ground tracking and control arc, the update process can be monitored or authorized by the ground, and the ground can directly intervene to stop or trigger the update if necessary.

[0110] Through the aforementioned comprehensive online update strategy, the satellite can continuously optimize its solar panel current prediction and SADA rotation angle optimization capabilities, effectively responding to new damage modes and environmental changes that may occur during long-term on-orbit operation, and significantly improving the autonomy, adaptability, and reliability of onboard energy health management.

[0111] Experimental evaluation

[0112] 1. Data Preparation: This experiment is based on real telemetry information transmitted from the satellite for evaluation. Information items include: satellite attitude, solar vector, solar panel current value, potentiometer angle, and operating time period. Currently, data from 8 time periods has been collected, with approximately 78,000 samples per time period, totaling approximately 620,000 samples.

[0113] 2. Experimental Results:

[0114] 1) Optimal angle prediction accuracy: 92%

[0115] 2) Optimal angle prediction time: 0.2466s / sample

[0116] The effectiveness of this method was verified based on actual on-orbit telemetry data from the satellite. Results show that the model can autonomously identify changes in satellite attitude and solar vector and implement millisecond-level rapid responses. The prediction accuracy of the SADA target rotation angle reaches 92%, effectively improving the reliability of on-orbit operation. This optimized control algorithm reduces energy loss during solar panel adjustment, achieving efficient and stable operation of the satellite energy system in complex space environments and improving energy utilization efficiency. Simultaneously, it enhances the autonomous management capability of overseas telemetry and control blind spots, significantly reducing the frequency of ground intervention and maintenance costs, providing a solid technical foundation for subsequent deep space exploration missions. This technology is expected to be extended to the energy management of more types of spacecraft, ensuring that satellites can maintain basic functions even in the event of partial sensor or actuator failure, providing strong support for the long-term stable operation of space missions.

[0117] This application also provides an on-orbit energy recovery system for damaged satellites.

[0118] Figure 5 This is a system block diagram of an on-orbit energy recovery system for a damaged satellite, according to an embodiment of this application. Figure 5 As shown, the on-orbit energy rescue system 500 for damaged satellites includes: a rescue decision module 51, a solar panel current prediction module 52, and a rotation angle reasoning module 53.

[0119] Among them, the rescue decision module 51 is used to autonomously decide whether to execute the SADA rotation angle reasoning based on the current satellite status data, and to select the reasoning model under the current energy limitation conditions;

[0120] The solar panel current prediction module 52 is used to simulate the regression relationship between the SADA rotation angle and the solar panel predicted current based on the selected inference model and the current state data of the satellite.

[0121] The rotation angle reasoning module 53 is used to search for the SADA target rotation angle that maximizes the predicted current of the solar panel from the regression relationship based on the Bayesian optimization strategy, and convert the SADA target rotation angle into SADA control commands.

[0122] This application also provides a network device 600. For example... Figure 6 As shown, network device 600 includes: bus 601, processor 602, memory 604, and communication interface 603. Processor 602, memory 604, and communication interface 603 communicate via bus 601. Network device 600 can be a server or a terminal device. It should be understood that this application does not limit the number of processors and memories in network device 600.

[0123] Bus 601 can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be divided into address buses, data buses, control buses, etc. For ease of representation, Figure 6 The bus 601 may be represented by a single line, but this does not mean that there is only one bus or one type of bus. The bus 601 may include a path for transmitting information between various components of the network device 600 (e.g., memory 604, processor 602, communication interface 603).

[0124] Processor 602 may include any one or more processors such as a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP).

[0125] Memory 604 may include volatile memory, such as random access memory (RAM). Processor 602 may also include non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid state drive (SSD).

[0126] The memory 604 stores executable program code, which the processor 602 executes to implement the aforementioned on-orbit energy rescue methods. That is, the memory 604 stores instructions for executing the on-orbit energy rescue methods.

[0127] The communication interface 603 uses transceiver modules such as, but not limited to, network interface cards and transceivers to enable communication between the network device 600 and other devices or communication networks.

[0128] This application also provides a computer program product containing instructions. The computer program product may be a software or program product containing instructions, capable of running on a network device or stored on any available medium. When the computer program product is run on at least one network device, it causes the at least one network device to perform an on-orbit energy rescue method.

[0129] This application also provides a computer-readable storage medium. The computer-readable storage medium can be any available medium that a network device can store, or a data storage device such as a data center containing one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive). The computer-readable storage medium includes instructions that instruct the network device to perform an on-orbit energy rescue method.

[0130] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the protection scope of the technical solutions of the embodiments of the present invention.

[0131] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0132] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for on-orbit energy recovery of a damaged satellite, wherein the mechanical connection between the solar panels and the satellite body of the damaged satellite is damaged, characterized in that, include: Based on the current satellite status data, the system autonomously decides whether to perform the SADA rotation angle inference and selects the inference model under the current energy constraints. Based on the selected inference model and the current satellite status data, the regression relationship between the SADA rotation angle and the predicted current of the solar panels was simulated. Based on the Bayesian optimization strategy, the SADA target rotation angle that maximizes the predicted current of the solar panel is searched from the regression relationship, and the SADA target rotation angle is converted into SADA control command. The reasoning process for autonomously deciding whether to execute the SADA rotation angle based on the satellite's current status data includes: obtaining the solar panel power generation and the satellite's total power consumption at the current moment from the satellite's current status data; if the solar panel power generation at the current moment is less than the satellite's total power consumption, obtaining the remaining battery power and the estimated remaining mission time from the satellite's current status data; calculating the satellite's net power consumption based on the solar panel power generation and the satellite's total power consumption at the current moment, calculating the current remaining energy sustainability time based on the satellite's net power consumption and the current remaining battery power, and determining whether the current remaining energy sustainability time is less than the sum of the estimated remaining mission time and the safety margin time; if so, executing the SADA rotation angle reasoning process. The selection of the inference model under the current energy constraint conditions includes: if the current remaining energy sustainability time is less than or equal to the estimated remaining time of the task, then the inference model with the smallest inference latency is selected; if the current remaining energy sustainability time is greater than the estimated remaining time of the task but less than the sum of the estimated remaining time of the task and the safety margin time, then the inference model with the highest inference accuracy is selected.

2. The on-orbit energy rescue method as described in claim 1, characterized in that, Also includes: Determine whether the power generation of the solar panels at the current moment is greater than or equal to the total power consumption of the satellite. If so, do not perform the SADA rotation angle inference.

3. The on-orbit energy rescue method as described in claim 1, characterized in that, Also includes: Determine whether the remaining battery power at the current moment is less than a preset critical power level. If so, perform the reasoning for the SADA rotation angle.

4. The on-orbit energy rescue method as described in claim 1, characterized in that, Based on the selected inference model and the current satellite status data, the regression relationship between the SADA rotation angle and the predicted current of the solar panels was simulated, including: Obtain angle-independent basic feature values ​​from the current satellite status data; Initialize a Bayesian optimizer, which uses a Gaussian process regression model to model the objective function, which is the regression relationship between the SADA rotation angle and the solar panel predicted current given the basic feature values; Within the rotation angle range of SADA, N discrete initial angle points are randomly sampled. For each sampling point, its sine and cosine values ​​are combined with the basic feature values ​​to form a complete feature vector, which is then input into the inference model to obtain the solar panel predicted current. The Gaussian process regression model is then trained using the initial angle points and their corresponding solar panel predicted currents.

5. The on-orbit energy rescue method as described in claim 4, characterized in that, The SADA target rotation angle that maximizes the predicted current of the solar panel, based on the Bayesian optimization strategy, is searched from the regression relationship, including: Step a: Calculate the acquisition function based on the Gaussian process regression model, and select the next angle point to be evaluated by maximizing the acquisition function; Step b: Combine the sine and cosine values ​​of the next angle point to be evaluated with the basic feature values ​​to form a complete feature vector and input it into the inference model to obtain the solar panel prediction current. Use the next angle point to be evaluated and its corresponding solar panel prediction current to train the Gaussian process regression model. Step c: Repeat steps a~b until the predetermined termination condition is met. Finally, the angle that maximizes the predicted current of the solar panel found during the iteration process is determined as the SADA target rotation angle.

6. The on-orbit energy rescue method as described in claim 5, characterized in that, The acquisition function is a dynamically weighted hybrid acquisition function, which is a weighted sum of the expected improvement, the upper confidence boundary, and the improvement probability, wherein the weights are dynamically adjusted according to the progress of the iteration.

7. The on-orbit energy rescue method as described in claim 6, characterized in that, In the first half of the iteration, the weights expected to improve are greater than the weights of the upper confidence boundary; In the latter half of the iteration, the weights expected to improve are less than the weights of the upper confidence boundary; During the iteration process, the weight of the improvement probability is a fixed value.

8. The on-orbit energy rescue method as described in claim 6, characterized in that, Using Latin hypercube sampling, N discrete initial angle points are randomly sampled within the angle range of the SADA rotation angle.

9. The on-orbit energy rescue method as described in claim 4, characterized in that, The SADA rotation angle range is [1°, 360°].

10. The on-orbit energy rescue method as described in claim 4, characterized in that, We choose a Gaussian process regression model with the Matérn kernel function to model the objective function.

11. The on-orbit energy rescue method as described in claim 4, characterized in that, Also includes: Calculate a global scaling factor to correct the discrepancy between the predicted current of the solar panel and the actual current of the solar panel; The predicted solar panel current is subjected to a standardized inverse transform, and the predicted solar panel current after inverse transform is multiplied by the global scaling factor to obtain the corrected solar panel current. The Gaussian process regression model is trained using the initial angle point and its corresponding solar panel correction current.

12. The on-orbit energy rescue method as described in claim 11, characterized in that, Calculating a global scaling factor includes: Multiple samples were acquired, each including the satellite's current status data, the current rotation angle of the SADA system, and the actual current of the solar panels. The current satellite status data and the current rotation angle of SADA are input into the inference model, and the inference model outputs the first solar panel predicted current corresponding to the current rotation angle of SADA. Perform a standardized inverse transform on the predicted current of the first solar panel, and divide the actual current of the solar panel by the inversely transformed predicted current of the first solar panel to obtain the scaling factor for a single sample. Collect the scaling factors of all samples, filter out scaling factors that deviate from the median of the scaling factors by a preset distance, and finally take the average of all filtered scaling factors as the global scaling factor.

13. The on-orbit energy rescue method as described in claim 5, characterized in that, Also includes: A local search is performed within a preset range of the target SADA rotation angle to find the SADA rotation angles that make the solar panel's predicted current among the top K values ​​in the preset range. The SADA rotation angles corresponding to the top K values ​​are then used as candidate values ​​for the target SADA rotation angle.

14. The on-orbit energy rescue method as described in claim 1, characterized in that, It also includes constructing multiple inference models and uploading these inference models to the satellite, wherein constructing multiple inference models includes: Receive multi-source telemetry data from satellites, preprocess and select features from the telemetry data to obtain a feature vector for predicting solar panel current; Using the feature vector as the input value of the inference model and the solar panel predicted current as the output value of the inference model, the multiple inference models are constructed based on different types of regression models.

15. The on-orbit energy rescue method as described in claim 14, characterized in that, Feature selection of the telemetry data includes: Calculate and sort the correlation coefficients between each feature and the predicted current of the solar panel, and select the top m features as the feature vector. A recursive feature elimination strategy is adopted, with LightGBM as the base model. Redundant features are gradually eliminated through multi-fold cross-validation and backward selection. The feature vector is determined according to the principle of minimizing cross-validation error. Lasso regression's built-in feature selection mechanism is used to determine the optimal regularization parameter through multi-fold cross-validation, and the features corresponding to non-zero coefficients are retained as feature vectors. The frequency of each feature being selected is counted, and features that are selected by at least two methods are retained to form the final feature vector.

16. The on-orbit energy rescue method as described in claim 14, characterized in that, Also includes: The inference model is updated based on a preset update strategy; Among them, the inference model is updated using incremental learning or model fine-tuning strategies. For tree-based inference models, when an update is triggered, the latest on-orbit data is collected as a small batch dataset. The model can adapt to the new data by adding new trees or fine-tuning the parameters of existing trees, thus forming a new inference model. For the inference model of the stacked ensemble model, when an update is triggered, the ensemble layer weights are adjusted online. The weights of the base model are re-determined by minimizing the error on the validation set using the latest on-orbit data.

17. The on-orbit energy rescue method as described in claim 16, characterized in that, The default update strategies include: The inference model is updated at fixed intervals. When the average deviation between the solar panel predicted current by the inference model and the actual current after SADA rotation is greater than a preset threshold, the inference model is triggered to update. The inference model is updated after the satellite experiences a specific event.

18. An on-orbit energy recovery system for a damaged satellite, wherein the mechanical connection between the solar panels and the satellite body is damaged, characterized in that, include: The rescue decision module is used to autonomously decide whether to execute the SADA rotation angle based on the current satellite status data, and to select the inference model under the current energy constraints. The solar panel current prediction module is used to simulate the regression relationship between the SADA rotation angle and the predicted solar panel current based on the selected inference model and the current state data of the satellite. The rotation angle reasoning module is used to search for the SADA target rotation angle that maximizes the predicted current of the solar panel from the regression relationship based on a Bayesian optimization strategy, and to convert the SADA target rotation angle into SADA control commands. The reasoning process for autonomously deciding whether to execute the SADA rotation angle based on the satellite's current status data includes: obtaining the solar panel power generation and the satellite's total power consumption at the current moment from the satellite's current status data; if the solar panel power generation at the current moment is less than the satellite's total power consumption, obtaining the remaining battery power and the estimated remaining mission time from the satellite's current status data; calculating the satellite's net power consumption based on the solar panel power generation and the satellite's total power consumption at the current moment, calculating the current remaining energy sustainability time based on the satellite's net power consumption and the current remaining battery power, and determining whether the current remaining energy sustainability time is less than the sum of the estimated remaining mission time and the safety margin time; if so, executing the SADA rotation angle reasoning process. The selection of the inference model under the current energy constraint conditions includes: if the current remaining energy sustainability time is less than or equal to the estimated remaining time of the task, then the inference model with the smallest inference latency is selected; if the current remaining energy sustainability time is greater than the estimated remaining time of the task but less than the sum of the estimated remaining time of the task and the safety margin time, then the inference model with the highest inference accuracy is selected.

19. A network device, characterized in that, The device includes a processor and a memory; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory to cause the network device to perform the method as described in any one of claims 1-17.

20. A computer-readable storage medium, characterized in that, Includes computer program instructions, which, when executed by a processor, cause the processor to perform the method as described in any one of claims 1-17.

21. A computer program product containing instructions, characterized in that, When the instruction is executed by the processor, the processor performs the method as described in any one of claims 1-17.

Images