A method for realizing satellite platform and integrated electric intelligent cooperation based on heterogeneous computing power

Abstract

The application discloses a method for realizing satellite platform and integrated electric intelligent cooperation based on heterogeneous computing power, and relates to the technical field of satellite platform artificial intelligence, which comprises the following steps: deploying a platform AI subsystem and an integrated electric AI subsystem interconnected through an in-orbit bus in a single satellite; constructing a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine and a mode scheduler in the platform AI subsystem; constructing a student network model, a task planning module, a health perception model and an energy management execution module in the integrated electric AI subsystem; and cooperating with the model cooperation optimization mode, the task planning cooperation verification mode and the energy management cooperation optimization mode between the platform AI subsystem and the integrated electric AI subsystem according to preset priority rules through the mode scheduler, so that the AI model autonomous optimization, task planning simulation verification and energy orbit-level management are realized in the satellite.

Description

Technical Field

[0001] This invention belongs to the field of satellite platform artificial intelligence technology, specifically relating to a method for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power. Background Technology

[0002] Existing satellite platform AI (Artificial Intelligence) systems have gradually developed along with the improvement of onboard computing capabilities. In recent years, the development of onboard computing technology has enabled satellites to have on-orbit data processing capabilities. With the advancement of large-scale constellation construction, the scale of data acquired by satellites is growing exponentially. The traditional mode of relying on ground processing can no longer meet the timeliness requirements, and onboard artificial intelligence technology has emerged. Onboard intelligent technology refers to the technology of directly deploying and running artificial intelligence models and computing architectures with perception, cognition, reasoning and decision-making capabilities on the satellite's onboard computing platform to achieve autonomous information processing and closed-loop control on the satellite, thereby enabling the satellite to have autonomous perception capabilities and reducing dependence on ground control centers.

[0003] Despite the significant progress made in spaceborne AI technology in recent years, existing satellite platform AI systems still face several technical bottlenecks that urgently need to be addressed. First, the platform AI and the integrated power AI operate in isolation, lacking a system-level two-way collaborative mechanism. This prevents the computing advantages of the high-performance platform from effectively empowering the real-time decision-making of the integrated power AI. Second, existing onboard AI models rely on ground stations for parameter updates or model replacements after deployment. They cannot utilize the platform's surplus computing power to perform model distillation, compression, and optimization on-orbit, resulting in model update cycles that can last for days or even weeks, making it difficult to adapt to dynamic changes in the on-orbit environment. Third, existing satellite mission planning schemes are generated and executed directly by a single AI system, lacking a simulation pre-verification step by a high-performance platform, leading to execution risks such as energy exceeding limits and attitude unreachability. Fourth, existing AI energy management schemes are mainly geared towards ground data center scenarios, failing to consider constraints unique to satellites, such as orbital illumination cycles, solar array power generation characteristics, and deep battery discharge in the Earth's shadow. Fifth, model experience across AI subsystems cannot be transferred. The training data and optimization experience accumulated by the platform AI during long-term on-orbit operation cannot be effectively transferred to the integrated power AI for model iteration, causing the performance of the integrated power AI model to gradually degrade with environmental changes.

[0004] Therefore, how to provide a method for achieving intelligent collaboration between satellite platforms and integrated electrical systems based on heterogeneous computing power, which enables autonomous optimization of AI models, mission planning simulation verification, and energy orbital-level management within satellites, has become an important issue. Summary of the Invention

[0005] To address the aforementioned problems in the existing technology, this invention provides a method for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power, the method comprising: Deploy platform AI subsystems and integrated electrical AI subsystems interconnected via an intra-satellite bus within a single satellite; The platform's AI subsystem constructs a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine, and a pattern scheduler. The teacher network model is used to transfer model knowledge to the integrated power AI subsystem. The energy prediction deep reinforcement learning model is used to generate power allocation schemes based on satellite energy data. The multi-dimensional simulation engine is used to perform multi-dimensional feasibility simulation evaluations of mission planning schemes. The integrated power AI subsystem comprises a student network model, a task planning module, a health perception model, and an energy management execution module. The student network model is used to perform on-orbit real-time inference. The task planning module is used to generate the task planning scheme. The energy management execution module is used to execute the power allocation scheme. The health perception model is used to monitor hardware health. The mode scheduler operates in a collaborative optimization mode, a task planning collaborative verification mode, and an energy management collaborative optimization mode between the platform AI subsystem and the integrated power AI subsystem according to preset priority rules.

[0007] Secondly, the present invention provides a device for realizing intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power, the device comprising: The deployment module is used to deploy platform AI subsystems and integrated electrical AI subsystems interconnected via an intra-satellite bus within a single satellite; The first construction module is used to build a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine, and a mode scheduler in the platform's AI subsystem; the teacher network model is used to transfer model knowledge to the integrated power AI subsystem; the energy prediction deep reinforcement learning model is used to generate power allocation schemes based on satellite energy data; and the multi-dimensional simulation engine is used to perform multi-dimensional feasibility simulation evaluations of mission planning schemes. The second construction module is used to construct a student network model, a task planning module, a health perception model, and an energy management execution module in the integrated electric AI subsystem; the student network model is used to perform on-orbit real-time inference; the task planning module is used to generate the task planning scheme; the health perception model is used to monitor hardware health; and the energy management execution module is used to execute the power allocation scheme. The collaboration module is used to coordinate the operation of model collaboration optimization mode, task planning collaboration verification mode and energy management collaboration optimization mode between the platform AI subsystem and the integrated power AI subsystem according to preset priority rules through the mode scheduler.

[0008] Thirdly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When the processor executes a computer program stored in memory, it implements the steps described in any of the above methods for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power.

[0009] Fourthly, the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps described in any of the above-mentioned methods for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power.

[0010] This invention provides a method for achieving intelligent collaboration between a satellite platform and an integrated electrical system based on heterogeneous computing power. By deploying a high-computing-power platform AI subsystem and a low-power integrated electrical AI subsystem within a single satellite and constructing an intra-satellite bus interconnection, the mode scheduler uniformly arbitrates the allocation of computing power. This allows the high-computing-power side to undertake computationally intensive tasks such as distillation, simulation, and prediction, while the low-computing-power side undertakes tasks such as real-time inference, task generation, and instruction execution. This achieves the collaborative utilization of heterogeneous computing power within the satellite and improves the overall computing efficiency of the satellite platform.

[0011] By constructing a teacher network model deployed in the platform's AI subsystem and a student network model deployed in the integrated power AI subsystem, and establishing a model collaborative optimization mode, model knowledge can be transferred between the two AI subsystems within the satellite. The student network model can be optimized without relying on a ground station, shortening the model update cycle. By constructing a multi-dimensional simulation engine and a mission planning module, and establishing a mission planning collaborative verification mode, the mission planning schemes generated by the integrated power AI subsystem can be simulated and evaluated by the platform's AI subsystem before execution, avoiding potential execution risks such as energy exceeding limits and attitude unreachability from the source. By constructing a deep reinforcement learning model for energy prediction and an energy management execution module, and establishing an energy management collaborative optimization mode, energy prediction and power allocation are completed in the platform's AI subsystem, while power execution is completed in the integrated power AI subsystem, achieving cross-AI subsystem division of labor and collaboration in energy management tasks.

[0012] In addition, the platform's AI computing power is uniformly allocated by the mode scheduler according to preset priority rules, so that the three types of needs—energy security, task security, and model evolution—can still be met in an orderly manner under resource competition conditions. The three modes form a collaborative whole through a unified scheduling mechanism.

[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating a method for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power, provided by an embodiment of the present invention. Figure 2 This is a schematic diagram of the execution flow of the preset priority rules provided in the embodiments of the present invention; Figure 3 This is a schematic diagram of the structure of a device for realizing intelligent collaboration between a satellite platform and integrated electrical system based on heterogeneous computing power, provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0015] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0016] To address the limitations of existing technologies in achieving autonomous AI model optimization, mission planning simulation verification, and orbital-level energy management within satellites, this invention provides a method for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power. (See [link to previous document]). Figure 1 , Figure 1 This is a flowchart illustrating a method for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power, provided by an embodiment of the present invention. The method specifically includes the following steps: Step S101: Deploy a platform AI subsystem and an integrated electrical AI subsystem interconnected via an intra-satellite bus within a single satellite.

[0017] In this embodiment of the invention, a platform AI subsystem and an integrated electrical AI subsystem are deployed within a single satellite via an intra-satellite bus, and the platform AI subsystem and the integrated electrical AI subsystem are physically isolated from each other.

[0018] The platform's AI subsystem is deployed on the satellite platform's high-performance processing unit, equipped with a high-performance AI acceleration chip with a computing power of no less than 248 TOPS and a large-capacity high-speed storage of no less than 32GB. This AI subsystem is capable of running large-scale deep learning models. It has an independent power supply module and does not share a power path with the integrated power AI subsystem.

[0019] The integrated electronic system's AI subsystem is deployed in the low-power AI processing unit of the satellite's integrated electronic system. It is equipped with an embedded AI acceleration chip with a computing power of no more than 2 TOPS and limited storage capacity of no more than 4GB. The integrated electronic system's AI subsystem has the ability to run lightweight inference models and execute real-time control commands. The integrated electronic system's AI subsystem also has an independent power supply module.

[0020] The two AI subsystems are interconnected bidirectionally via an onboard gigabit Ethernet bus, supporting model file transfer (typical size 20~50MB, transmission time less than 120 seconds), planning scheme summary transmission (typical size 10~50KB), and real-time exchange of status monitoring data.

[0021] Step S102: Construct a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine, and a mode scheduler in the platform's AI subsystem. The teacher network model is used to transfer model knowledge to the integrated power AI subsystem. The energy prediction deep reinforcement learning model is used to generate power allocation schemes based on satellite energy data. The multi-dimensional simulation engine is used to conduct multi-dimensional feasibility simulation evaluations of mission planning schemes.

[0022] In this embodiment of the invention, the following models and software modules have been pre-deployed in the platform's AI subsystem before the satellite is launched into orbit.

[0023] The teacher network model employs a deep residual network architecture (on the scale of ResNet-50), with approximately 25 million parameters. It is deployed in the large-capacity storage of the platform's AI subsystem, serving as the teacher-side component for knowledge distillation. This teacher network model was pre-trained using a large-scale ground-based remote sensing dataset before the satellite was launched into orbit.

[0024] Through the aforementioned pre-training, the teacher network model possesses strong feature extraction and pattern recognition capabilities before the satellite is launched into orbit. The function of the teacher network model is to transfer model knowledge to the integrated electronic AI subsystem. Specifically, when the student network model of the integrated electronic AI subsystem experiences performance degradation due to changes in the on-orbit environment, such as changes in lighting conditions or data distribution drift, the platform AI subsystem can utilize the teacher network model as a knowledge source, transferring the recognition capabilities inherent in the teacher network model to the student network model through knowledge distillation technology. Here, pattern recognition capability refers to the ability to recognize the satellite's operational state in space orbit. The satellite's operational states in space orbit include safe mode, orbit change mode, solar cruise mode, Earth photography mode, illumination data transmission mode, and shadow data transmission mode.

[0025] The energy prediction DRL (Deep Reinforcement Learning) model employs a deep reinforcement learning network based on the PPO (Proximal Policy Optimization) algorithm. It includes a policy network and a value function network, each consisting of three fully connected layers with hidden layer dimensions of 256, 128, and 64, respectively. The energy prediction deep reinforcement learning model undergoes initial training using historical satellite energy data before satellite launch and orbit insertion. The trained energy prediction deep reinforcement learning model can then be used to generate power allocation schemes based on satellite energy data.

[0026] In this embodiment of the invention, the multi-dimensional simulation engine is a comprehensive simulation platform integrating a satellite attitude dynamics simulation model, an orbital illumination condition simulation model, an energy charging and discharging simulation model, and a ground station communication window calculation model. It is used to perform multi-dimensional feasibility simulation evaluations of mission planning schemes submitted by the integrated power AI subsystem. The multi-dimensional feasibility simulation evaluation can include illumination condition feasibility assessment, attitude reachability analysis, energy balance verification, and communication window checking.

[0027] The Mode Scheduler is a scheduler running on the platform's AI subsystem. It is responsible for coordinating runtime resource allocation and priority arbitration for three collaborative modes: Model Co-optimization (MCO), Mission Planning Co-validation (MPCV), and Energy Management Co-optimization (EMCO).

[0028] Step S103: Construct a student network model, a task planning module, a health perception model, and an energy management execution module in the integrated power AI subsystem; the student network model is used to perform on-orbit real-time inference; the task planning module is used to generate task planning schemes; the energy management execution module is used to execute power allocation schemes; and the health perception model is used to monitor hardware health.

[0029] In this embodiment of the invention, the following models and software modules have been pre-deployed in the integrated electronic AI subsystem before the satellite is launched into orbit.

[0030] In this embodiment of the invention, the student network model employs a lightweight convolutional architecture (on the scale of MobileNetV3), with approximately 3M parameters, and is deployed in the limited storage of the integrated electronic AI subsystem. This student network model is a lightweight version of the teacher network model after knowledge distillation and compression. During on-orbit operation, this student network model is used to perform on-orbit real-time inference based on real-time data input from the satellite payload.

[0031] The mission planning module adopts a mission planning engine based on a constraint satisfaction problem solver, which can integrate input conditions such as mission priority, orbit position, and energy constraints to generate a mission planning scheme that includes observation timing, attitude maneuver sequence, data transmission window arrangement, and energy budget allocation.

[0032] The health sensing model employs a lightweight LSTM (Long Short Term Memory) autoencoder, which contains two LSTM units with hidden layer dimensions of 64 and 32, respectively. It is used to monitor the current and temperature timing data of each subsystem in real time and monitor the hardware health status by calculating the reconstruction error of the autoencoder.

[0033] Specifically, during on-orbit operation, the integrated electrical AI subsystem continuously collects multi-channel telemetry data from sensors of various satellite subsystems at a fixed sampling period (typically 1 second). This data includes, but is not limited to: real-time current values ​​reflecting load status, voltage values ​​reflecting power supply stability, temperature values ​​reflecting thermal control status, and drive currents and feedback signals of key actuators. This multi-channel time-series data is fed into an autoencoder in real time. The encoder compresses the high-dimensional multi-channel time-series data into a low-dimensional latent space feature vector, and the decoder reconstructs the original-dimensional time-series data from the low-dimensional latent space feature vector. The health perception model calculates the reconstruction error between the original input multi-channel time-series data and the reconstructed output time-series data, using the average reconstruction error within a sliding window as the basis for health scoring. The sliding window size can be N=100 sampling points.

[0034] The energy management execution module receives the power allocation scheme issued by the platform's AI subsystem, issues power adjustment commands to each subsystem via the satellite bus, and monitors the battery's SOC (State of Charge) and the load status of each subsystem in real time.

[0035] In one implementation, the integrated electronic AI subsystem also includes a model version manager for managing the version information of the student network model. The model version manager has the functions of hot-updating the model and rolling back the version in case of an anomaly, wherein hot-updating does not interrupt the execution of the current on-orbit inference task.

[0036] Step S104: The mode scheduler operates the model collaborative optimization mode, task planning collaborative verification mode, and energy management collaborative optimization mode between the platform AI subsystem and the integrated power AI subsystem according to the preset priority rules.

[0037] This invention defines three collaborative optimization modes: model collaborative optimization mode, task planning collaborative verification mode, and energy management collaborative optimization mode. These three collaborative optimization modes share the computing resources of the platform's AI subsystem, therefore requiring explicit priority rules and collaborative logic. The static configuration is as follows: The energy management collaborative optimization mode enjoys the highest priority. Energy security is a prerequisite for satellite survival; therefore, the energy management collaborative optimization mode operates continuously at a fixed cycle, such as once per orbital cycle, and can preempt the computing power of other collaborative optimization modes' platform AI subsystems at any time. The energy prediction data produced by the energy management collaborative optimization mode is written to the shared data area for the mission planning collaborative verification mode to read during energy balance verification.

[0038] The task planning collaborative verification mode is set to medium priority. The security verification of task planning directly affects the success or failure of task execution; therefore, the task planning collaborative verification mode is triggered upon receiving a planning verification request submitted by the integrated power AI subsystem. When performing the energy balance verification dimension, the task planning collaborative verification mode reads the latest energy forecast data produced by the energy management collaborative optimization mode from the shared data area as input.

[0039] The model co-optimization mode has the lowest priority. Model performance optimization falls under the category of capability enhancement and does not affect satellite safety. Therefore, the model co-optimization mode is only initiated when the integrated power AI subsystem reports a model performance degradation flag and the platform AI subsystem's current computing power is idle, i.e., neither the energy management co-optimization mode nor the mission planning co-verification mode is occupied. During the execution of the model co-optimization mode, if it is preempted by the energy management co-optimization mode or the mission planning co-verification mode, the platform AI subsystem saves the intermediate state of the current distillation and compression process; after the higher priority mode has completed execution, the model co-optimization mode resumes execution from the saved breakpoint.

[0040] In this embodiment of the invention, the platform AI subsystem and the integrated electrical AI subsystem share the following shared data area via the intra-satellite bus, i.e., the intra-satellite gigabit Ethernet bus, as the basis for data exchange during operation in the three collaborative modes. The shared data area specifically includes: Energy State Sharing Area: The current real-time battery SOC value is continuously written by the Integrated Electricity AI subsystem; it is also written by the Energy Management Collaborative Optimization Mode, storing predicted energy availability for multiple future orbital cycles, power allocation schemes for each subsystem, and the current battery state of charge prediction curve. The Task Planning Collaborative Verification Mode reads the latest energy prediction data from the Energy State Sharing Area for energy balance verification.

[0041] Model Performance Monitoring Area: Continuously written to by the Integrated Electronics AI Subsystem, this area stores performance metrics such as real-time inference accuracy, false positive rate, and response latency of the student network model. The model scheduler reads these performance metrics from the model performance monitoring area to determine whether to trigger the model co-optimization process.

[0042] Task planning interaction area: The integrated power AI subsystem writes the preliminary planning scheme summary, and the platform AI subsystem writes the multi-dimensional evaluation results and improvement suggestion parameter package, realizing two-way data exchange in the task planning collaborative verification mode.

[0043] Model update transmission area: The platform's AI subsystem writes lightweight model files that have undergone distillation and compression, and the integrated electric AI subsystem reads and completes the hot update deployment.

[0044] In this embodiment of the invention, a mode scheduler, according to preset priority rules, collaboratively operates a model collaborative optimization mode, a task planning collaborative verification mode, and an energy management collaborative optimization mode between the platform AI subsystem and the integrated power AI subsystem, including: The system reads the real-time battery SOC value through the mode scheduler and determines whether the real-time battery SOC value is lower than the preset emergency threshold; if it is lower than the preset emergency threshold, the energy management collaborative optimization mode is activated. If the threshold is not lower than the preset emergency threshold, check if there is a planning verification request flag in the task planning interaction area; if there is a planning verification request flag and the energy management collaborative optimization mode is not currently being executed, then start the task planning collaborative verification mode. If there is no planning verification request flag, check if there is a model co-optimization trigger flag; if there is a model co-optimization trigger flag and the platform's AI subsystem is currently idle, then start the model co-optimization mode.

[0045] In one implementation, if the model co-optimization trigger flag exists but the platform's AI subsystem is not currently idle, the model co-optimization trigger flag remains valid and awaits re-checking in the next scheduling cycle.

[0046] If none of the three collaborative optimization modes mentioned above are required at the current moment, the system enters an idle waiting state until the next scheduling cycle arrives. Then, the system will re-execute the steps of reading the real-time battery SOC value through the mode scheduler and determining whether the real-time battery SOC value is lower than the preset emergency threshold.

[0047] See Figure 2 , Figure 2 This is a schematic diagram of the execution flow of the preset priority rules provided in this embodiment of the invention. The mode scheduler runs in the platform's AI subsystem and periodically executes the following scheduling logic at fixed time intervals (typically 1 second): In step S201, the mode scheduler first reads the real-time battery SOC value written to the shared data area by the integrated power AI subsystem and determines whether the real-time battery SOC value is lower than a preset emergency threshold (typically 20%). If it is lower than the preset emergency threshold, the mode scheduler immediately interrupts the currently executing MCO mode or MPCV mode, allocates all computing power of the platform AI subsystem to the EMCO mode, and starts the EMCO mode execution process, where the EMCO mode performs emergency power reallocation. This step ensures that energy security always receives the highest response priority.

[0048] In step S202, if the real-time battery SOC value is not lower than the preset emergency threshold, the mode scheduler checks whether there is a new planning verification request flag in the task planning interaction area. If it exists, and the platform AI subsystem is not currently performing emergency calculations in EMCO mode, the mode scheduler allocates the necessary simulation computing resources for MPCV mode and starts the MPCV mode execution process. If it does not exist, proceed to step S203.

[0049] Step S203: If not present, the mode scheduler reads the performance metrics from the model performance monitoring area to determine if an MCO trigger flag exists. This MCO trigger flag is written by the integrated AI subsystem when it detects a decrease in model performance. If an MCO trigger flag exists, and the platform's AI subsystem currently has idle computing power (i.e., neither the EMCO mode nor the MPCV mode is consuming computing resources), the mode scheduler allocates the computing resources required for distillation and compression for the MCO mode and initiates the MCO mode execution process. If the computing power is not idle, the MCO trigger flag remains valid, awaiting re-checking in the next scheduling cycle.

[0050] In step S204, if none of the three collaborative optimization modes need to be executed, the mode scheduler enters an idle waiting state until the next scheduling cycle arrives and step S201 is executed again.

[0051] The four steps described above are executed sequentially in the order of S201, S202, S203, and S204 within each scheduling cycle. The result of each step determines whether subsequent steps are executed: if step S201 triggers EMCO mode, steps S202 and S203 are skipped; if step S202 activates MPCV mode, step S203 is skipped; subsequent steps are executed only if none of the preceding steps have triggered a task. This sequential priority check ensures that high-priority modes always receive computing resources first.

[0052] In this embodiment of the invention, the collaborative operation model and collaborative optimization mode between the platform AI subsystem and the integrated power AI subsystem include: The integrated electrical AI subsystem extracts false positive and false negative samples from the sliding window and generates feature data after forward propagation through the student network model. This data is then reported to the platform AI subsystem as a feedback data packet. After receiving the feedback data packet, the platform's AI subsystem performs cluster analysis on the feedback data packet and makes targeted fine-tuning to the teacher network model based on the analysis results. Using the fine-tuned teacher network model as the knowledge source, it performs knowledge distillation on the student network model, compresses the student network model obtained by distillation, obtains the compressed student network model file, and performs accuracy verification. After successful verification, the platform's AI subsystem will push the compressed student network model file to the integrated electrical AI subsystem via the satellite bus. During idle periods when the student network model is not performing on-orbit real-time inference, the model version manager of the integrated AI subsystem replaces the current student network model with a compressed student network model file to complete the hot update deployment of the model.

[0053] In this embodiment of the invention, knowledge distillation occurs between two AI subsystems within a single satellite, rather than through a traditional ground-satellite link. The triggering mechanism is automatic, triggered after the integrated electronic AI subsystem autonomously detects a performance degradation in the model, requiring no ground intervention. The entire process of distillation, compression, verification, and deployment is completed in an on-orbit closed loop. The specific execution flow of the MCO mode is as follows: While performing on-orbit inference tasks, the student network model of the integrated electronic AI subsystem continuously calculates the performance indicators of recent inference results. The on-orbit inference task refers to the process during the satellite's on-orbit operation, where the integrated electronic AI subsystem continuously receives real-time image data output by the satellite payload system, inputs each frame of image into the student network model to perform forward propagation calculations, and outputs target detection and recognition results.

[0054] The integrated AI subsystem maintains a sliding window (1000 frames in size) and calculates the inference accuracy, false positive rate, and average response latency of the on-orbit inference task based on the inference results within the sliding window. When the accuracy decreases relative to a preset baseline by more than a preset threshold (typically a decrease of no less than 20%), or the false positive rate increases relative to a preset baseline by more than a preset threshold (typically an increase of no less than 3 percentage points), the integrated AI subsystem performs the following operations: First, extract the image metadata of false positives and / or false negatives from the sliding window. This image metadata does not contain the original image, but only the feature vector and label information. Pack the image metadata, i.e. the feature data, into a feedback data packet of 50~200MB in size. Then, write the MCO trigger flag in the model performance monitoring area. Finally, send the feedback data packet to the model update transmission area of ​​the platform's AI subsystem via the intra-satellite bus.

[0055] False positives refer to samples where the detection results output by the student network model during the on-orbit inference task are inconsistent with the ground truth labels of the corresponding input images. Specifically, the integrated electronic AI subsystem maintains a sliding window that stores the detection results output by the student network model after inference on several recent frames of input images. This window includes the target category label, location bounding box, and the ground truth label for each frame. The integrated electronic AI subsystem compares the detection results of each frame in the sliding window with the corresponding ground truth labels one by one. If the detection result is inconsistent with the ground truth label, the frame is determined to be a false positive.

[0056] A missed detection is defined as a frame in which the ground truth label contains a target, but the student network model fails to detect that target. The ground truth label can be viewed in the satellite telemetry data.

[0057] The feature vectors corresponding to false positives or false negatives are extracted from the feature extraction layer of the student network model. The feature vectors are packaged together with the real labels to generate feature data, which is used as a feedback data package.

[0058] In this embodiment of the invention, after the mode scheduler of the platform AI subsystem detects the MCO trigger flag, it allocates computing resources for the MCO mode. The platform AI subsystem first reads feedback data packets from the model update transmission area, and performs feature clustering analysis on the false positive and false negative samples using the K-means clustering algorithm (a widely used unsupervised machine learning method) to identify the specific scenario types where the student network model performs poorly. Based on the identified specific scenario types, the corresponding weak scenario samples are selected from the feedback data packets. Subsequently, the platform AI subsystem uses the weak scenario samples to perform targeted fine-tuning on the pre-deployed teacher network model, where the fine-tuning learning rate is set to... The magnitude and number of fine-tuning rounds are adaptively determined based on the sample size, such as 5 to 20 rounds, to enhance the teacher network model's ability to identify weak scenes while maintaining its general capabilities. Specific scene types that exhibit weak performance include overexposed scenes at specific lighting angles or confusing scenes involving specific types of ground features.

[0059] After fine-tuning, the platform's AI subsystem performs a knowledge distillation process, using the fine-tuned teacher network model as the knowledge source and the architecture of the currently running student network model in the integrated electronics AI subsystem as the target, employing the KL divergence (Kullback-Leibler Divergence, relative entropy) loss function. Knowledge transfer is being carried out. Distillation temperature parameters. Set between 4 and 10, such as the typical value. =6. A higher distillation temperature parameter can make the soft label distribution of the teacher network model output smoother, which is beneficial for the student network model to learn the inter-class similarity relationship.

[0060] In this embodiment of the invention, the distillation loss function Defined as: ; in, This is the distillation weighting coefficient, with a typical value of 0.7; and The teacher network model and the student network model are respectively based on the distillation temperature parameter. The soft label outputs below specifically represent the class probability distributions of the teacher network model and the student network model after softening by the distillation temperature parameter; For standard cross-entropy loss, These are the real labels, i.e., the real categories; For student network models.

[0061] The number of training rounds for distillation is dynamically determined based on the loss convergence, with a typical value of 30 to 100 rounds.

[0062] In this embodiment of the invention, the platform AI subsystem receives the parameters of the distilled student network model and performs the following compression operations in sequence: The first step is to perform structured pruning: Calculate the importance score of each convolutional channel in the student network model. The scoring criterion is the L1-norm (L1 norm) criterion. The sum of the absolute values ​​of all convolutional channels is used as the denominator, and the absolute value of each convolutional channel is used as the numerator. The larger the value, the higher the score. Sort the channels by score from low to high, and remove the lowest-scoring channels, with a removal ratio of 30% to 50%, thereby reducing the number of model parameters and computational cost.

[0063] The second step is to perform INT8 quantization to obtain the compressed student network model: the 32-bit floating-point weight parameters are mapped to 8-bit integer representations, and the quantization formula is as follows: ; in, Represents an 8-bit integer; This represents the rounding function; This represents the original floating-point weights, i.e., 32-bit floating-point weight parameters; Indicates the zero-point offset of the mapping; Indicates the scaling factor. , This represents the maximum floating-point value representing all weights. The smallest floating-point value representing all weights.

[0064] The number of parameters in the student network model is compressed to about one-quarter of its original size, increasing inference speed by 2 to 4 times. After compression, the student network model file size is typically controlled within the range of 20 to 50 MB, ensuring that transmission can be completed within 120 seconds via the intra-satellite bus.

[0065] The platform's AI subsystem uses a historically accumulated test dataset to verify the accuracy of the compressed student network model within its built-in simulation environment. The test dataset contains labeled samples for various scenario types.

[0066] The validation metrics include accuracy loss, false positive rate increase, and inference latency change. Accuracy loss is calculated based on the accuracy of the teacher's network model after fine-tuning. Passing the validation requires an accuracy loss of no more than 5 percentage points and a false positive rate increase of no more than 1 percentage point. If validation passes, the platform's AI subsystem writes the compressed student network model file to the model update transmission area and includes a model update ready flag. If validation fails, the platform's AI subsystem returns to adjust the distillation temperature parameters or pruning ratio and re-executes distillation and compression. The maximum number of retries is 3. If the validation still fails, the current MCO process is abandoned, and the system waits for the next trigger.

[0067] In this embodiment of the invention, the model version manager of the integrated electrical AI subsystem continuously monitors the model update readiness flag in the model update transmission area. Upon detecting this flag, the model version manager reads the new compressed student network model file and loads it into the backup model storage area. After loading, the model version manager performs a model pointer switching operation during the next inference frame interval, i.e., the idle time window between two inference tasks (approximately 50-100 milliseconds): switching the model reference pointer of the inference engine from the current student network model to the newly loaded compressed student network model file. The pointer switching operation is completed in milliseconds, without interrupting the currently executing on-orbit inference task. After the switch, the old model version information is saved to the version history to support version rollback in case of anomalies; the old model version information includes the version number, accuracy metrics, and description of applicable scenarios. After the new model is put into operation, the integrated electrical AI subsystem resets the performance monitoring sliding window and restarts performance metric statistics, forming a closed loop. After this step is completed, the current round of MCO mode execution ends.

[0068] In this embodiment of the invention, the collaborative operation mode for task planning and verification between the platform AI subsystem and the integrated power AI subsystem includes: The integrated electrical AI subsystem uses the task planning module to generate a preliminary task planning scheme; The platform's AI subsystem utilizes a multi-dimensional simulation engine to perform multi-dimensional feasibility simulation evaluations on the initial task planning scheme, and obtains multi-dimensional evaluation results. The platform's AI subsystem assesses the preliminary task planning scheme based on the multi-dimensional evaluation results. When the scheme is deemed risky or infeasible, it generates an improvement suggestion parameter package and returns it to the integrated power AI subsystem along with the assessment results. The integrated AI subsystem inputs the improved suggestion parameter package as additional constraints into the task planning module, regenerates the revised task planning scheme, and uses a multi-dimensional simulation engine to verify the revised task planning scheme again until the platform AI subsystem determines that the revised task planning scheme is feasible or the number of iterations reaches the preset limit, thus obtaining the final task planning scheme.

[0069] In this embodiment of the invention, after receiving task requirement input, such as task instructions injected by the ground station or task requests triggered autonomously by the satellite, the task planning module of the integrated electric AI subsystem generates a preliminary task planning scheme based on comprehensive constraints. The constraints include task priority and time window requirements, current orbital position and predicted ephemeris, real-time battery SOC value and expected charge-discharge curve, and operating conditions and resource requirements of each payload.

[0070] The output of the mission planning scheme is a structured data packet containing four sequences: observation time series, attitude maneuver sequence, data transmission window arrangement sequence, and energy budget allocation sequence. The observation time series contains the start and end times and observation parameters of each observation target; the attitude maneuver sequence contains the target attitude angles and angular velocities at each time point; the data transmission window arrangement sequence contains the start and end times of each data transmission task and the target ground station; and the energy budget allocation sequence contains the subsystem power allocation plan for each time period. This structured data packet is written into the mission planning interaction area, and simultaneously, the integrated power AI subsystem writes a planning verification request flag into the mission planning interaction area.

[0071] After detecting the planning verification request flag, the platform's AI subsystem's mode scheduler allocates simulation computing resources for the MPCV mode. The platform's AI subsystem reads the structured data packet of the preliminary task planning scheme from the task planning interaction area and performs a multi-dimensional feasibility simulation evaluation in the multi-dimensional simulation engine. Specifically, it performs simulation evaluations in the following dimensions sequentially: (1) The first dimension is the feasibility assessment of illumination conditions: Based on the observation time sequence and orbital ephemeris in the preliminary mission planning scheme, calculate the solar altitude angle and the predicted power generation curve of the solar cell array for each observation period, and determine whether the power generation meets the energy demand for that period.

[0072] (2) The second dimension is attitude reachability analysis: The attitude maneuver sequence in the preliminary mission planning scheme is input into the satellite attitude dynamics simulation model to verify whether each attitude maneuver meets the physical capability constraints of the attitude control actuator (including flywheel and magnetic torque device). The constraints include maximum angular velocity constraints and maximum angular acceleration constraints, and the execution time margin of each attitude maneuver is calculated.

[0073] (3) The third dimension is energy balance verification: combining the power generation prediction of the integrated illumination dimension and the power consumption plan of each subsystem, the battery SOC change trajectory is simulated during the entire preliminary task planning cycle to confirm that the minimum SOC point is not lower than the safety threshold, which is typically 15%. This dimension requires reading the latest energy prediction data generated by the EMCO mode from the energy state sharing area as an auxiliary input to improve prediction accuracy.

[0074] (4) The fourth dimension is the communication window check: Based on the orbital ephemeris and the location of the ground station, calculate the visible time window of each ground station and verify whether the data transmission period in the preliminary mission planning scheme matches the visible window.

[0075] The simulation results from the four dimensions are compiled into an evaluation report as a multidimensional evaluation result.

[0076] Based on the multi-dimensional evaluation results, the platform's AI subsystem classifies the preliminary mission planning scheme into one of the following three categories: Category 1 is feasible, indicating that all four dimensions have passed verification and there are no risky items; Category 2 is risky, indicating that some dimensions are in a critical state or slightly exceed limits, such as the lowest SOC point being 14%, slightly below the 15% threshold; Category 3 is infeasible, indicating that at least one key dimension is severely exceeded. Key dimensions refer to dimensions that play a crucial role in satellite operation, including battery power, attitude angular velocity, etc., such as attitude angular velocity exceeding the actuator's limit value.

[0077] For the first category, the platform's AI subsystem writes a verification pass flag in the task planning interaction area. The integrated power AI subsystem reads this flag and directly executes the preliminary task planning scheme, ending the MPCV process. For the second or third category, the platform's AI subsystem generates a specific improvement suggestion parameter package based on simulation data. This simulation data is obtained by the platform's AI subsystem using computing power to simulate the preliminary task planning scheme provided by the integrated power AI subsystem. For example, if energy is insufficient, it is recommended to shorten the observation time or postpone high-power task periods; if attitude maneuvers exceed limits, it is recommended to reduce the angular velocity and correspondingly extend the maneuver time; if the data transmission window does not match, it is recommended to postpone the data transmission period to the next visible window. The improvement suggestion parameter package is written to the task planning interaction area, and a verification failure flag is simultaneously written: "Please correct."

[0078] The task planning module of the integrated power system's AI subsystem detected that the verification in the task planning interaction area had failed. After correcting the flag, it reads the improved suggestion parameter package and inputs it as an additional constraint into the task planning module to regenerate the corrected planning scheme. The corrected task planning scheme is rewritten into the task planning interaction area, and the planning verification request flag is set again, triggering a new round of multi-dimensional feasibility simulation evaluation. There is an upper limit to the number of iterations, typically 5. If the verification passes within the upper limit, the integrated power system's AI subsystem executes the final task planning scheme; if it still fails after reaching the upper limit, the integrated power system's AI subsystem sends a request for assistance to the ground station, requesting manual intervention for decision-making.

[0079] In this embodiment of the invention, the collaborative operation of an energy management optimization mode between the platform AI subsystem and the integrated power AI subsystem includes: The platform's AI subsystem collects the satellite's state vector sequence; it inputs the state vector sequence into the energy prediction deep reinforcement learning model to output the power adjustment factor sequence of each subsystem in multiple future orbital cycles, which serves as a power allocation scheme. The platform's AI subsystem verifies the safety of the power allocation scheme by simulating the battery state of charge change curves for multiple future orbital cycles after the power allocation scheme is executed. If the lowest state of charge value in the battery state of charge change curve is lower than the preset safety threshold, the energy prediction deep reinforcement learning model is called again to generate a new power allocation scheme. The process continues until the verification is passed or the retry limit is reached. Then, the platform's AI subsystem sends the final power allocation scheme to the integrated power AI subsystem. The energy management execution module issues power adjustment instructions to each subsystem according to the final power allocation scheme, and after execution, packages the actual state of charge change curve and the actual task completion status into execution feedback data, which is then sent to the platform AI subsystem via the intra-satellite bus. The platform's AI subsystem updates the energy prediction deep reinforcement learning model based on execution feedback data and employs a policy gradient update algorithm.

[0080] In this embodiment of the invention, the platform's AI subsystem continuously collects the following four types of real-time data from sensors in various satellite subsystems at a fixed sampling period, such as 10 seconds: (a) Real-time power generation of solar panels is obtained by reading the current output voltage and current of the photovoltaic array from the output power sensor of the solar cell array and calculating the real-time power generation. (b) Battery pack status, including current battery SOC, battery temperature, terminal voltage and internal resistance read from the battery management system; (c) Instantaneous power consumption of each subsystem is obtained by reading the real-time current values ​​of each subsystem, such as the attitude control system, communication system, data transmission system, load system, and integrated electric AI subsystem, from the power distribution unit and calculating their respective power consumption. (d) Orbital ephemeris data: by reading current orbital parameters from the satellite computer, including orbital altitude, inclination, right ascension of the ascending node, etc., the current solar altitude angle and future shadow period forecast are calculated.

[0081] Based on the collected data, the current battery SOC value, predicted solar altitude angle, predicted mission load power, and satellite orbit phase are extracted to construct a state vector, which is then stored in the time-series database of the platform's AI subsystem. Then, the state vectors from multiple orbital periods are used as a state vector sequence.

[0082] The collected state vector sequence is used as input and fed into a pre-deployed deep reinforcement learning model for energy prediction for inference.

[0083] The state space of the deep reinforcement learning model for energy prediction is defined as follows: ; in, express The state vector at any given time; express The state of charge of the battery at any given time; express Predicted solar altitude angle at any given time; express Predicted task load power at any given time; express Satellite orbit phase at any given moment; Action space is defined as: ; in, express Action vector at any given moment; This represents the power regulation factor of the attitude control subsystem; This represents the power regulation factor of the communication subsystem; This represents the power regulation factor of the data transmission subsystem; This represents the power adjustment factor of the load subsystem; the power adjustment factor of each subsystem is a real number between 0 and 1, representing the scaling ratio of the rated power of each subsystem.

[0084] The reward function is defined as: ; in, express Momentary rewards; Indicating the first step in the task planning scheme The weight of each task; This indicates a task completion flag: 0 indicates incomplete, and 1 indicates complete. This represents the safety margin factor, with a typical value of 0.3; Represents a function that finds the maximum value; This indicates the lower limit of battery safety, with a typical value of 15%.

[0085] The deep reinforcement learning model for energy prediction, based on the input state vector sequence, uses the PPO algorithm for inference and outputs the optimal power adjustment factor sequence for each control slot over multiple future orbital cycles, serving as a power allocation optimization scheme. The state space and reward function of the PPO algorithm are specifically designed for the satellite orbital environment, and the inference process of the PPO algorithm itself adopts existing standard procedures.

[0086] In this embodiment of the invention, the platform AI subsystem performs safety verification on the power allocation scheme. The verification method is as follows: The power values ​​of each subsystem in the power allocation scheme are substituted into the energy simulation model. Combined with the collected current battery state and orbital illumination prediction, the battery state of charge (SOC) change curves for multiple future orbital cycles after implementing the power allocation scheme are simulated, and the minimum SOC value is extracted. The energy simulation model is an existing model pre-deployed within the platform AI subsystem. If the minimum SOC value is not lower than a preset safety threshold, such as 15%, the verification passes. If the minimum SOC value is lower than the preset safety threshold, the platform AI subsystem automatically tightens the safety constraints, i.e., ... The power allocation is temporarily increased by 2 percentage points, and the PPO inference is re-executed to generate a more conservative power allocation scheme. This process is repeated a maximum of 3 times. If the safety requirements are still not met, the platform's AI subsystem generates a minimum power consumption safety net scheme, maintaining only the minimum operating power of the spaceborne computer and communication systems, and uses this as the final power allocation scheme. Then, the platform's AI subsystem distributes the final power allocation scheme to the integrated power AI subsystem.

[0087] The energy management execution module of the integrated power AI subsystem receives the final power allocation scheme via the intra-satellite bus. This power allocation scheme includes the power adjustment factors for each subsystem in future control time slots. Based on the power adjustment factors, the integrated power AI subsystem issues power adjustment commands to the power management units of each subsystem via the intra-satellite bus, dynamically adjusting the operating cycle and power budget of each subsystem. During execution, the integrated power AI subsystem continuously reads actual SOC change data from the battery management system and records the actual completion status of each task, such as whether it was completed on time or interrupted due to insufficient power. After completing a full orbital cycle, the integrated power AI subsystem packages the actual state of charge change curve and task completion status into execution feedback data and sends it to the platform AI subsystem via the intra-satellite bus. Upon receiving the execution feedback data, the platform AI subsystem uses it as new experience samples and employs the standard PPO strategy gradient update algorithm to update the parameters of the energy prediction deep reinforcement learning model, enabling the energy management strategy to be continuously optimized based on actual execution results.

[0088] In this embodiment of the invention, bidirectional health monitoring and cross-verification of key instructions are implemented between the integrated power AI subsystem and the platform AI subsystem to ensure the overall reliability of the system under conditions of asymmetric computing power, wherein the computing power gap between the platform AI subsystem and the integrated power AI subsystem exceeds 100 times.

[0089] During the operation of the above three collaborative optimization modes, the following high-reliability guarantee mechanisms continue to operate in parallel: The platform AI subsystem and the integrated electrical AI subsystem each send heartbeat signals to each other at fixed intervals (typically 1 second). These heartbeat signals contain a summary of their respective operational statuses, including CPU utilization, memory usage, and critical process status. If one subsystem does not receive a heartbeat signal from the other within a preset timeout window (typically 5 seconds), it determines the other subsystem is abnormal and takes over its critical tasks. When the integrated electrical AI subsystem takes over, it only executes the most critical space management functions. The platform AI subsystem, however, can execute all of the other subsystem's functions, but is limited by its interface compatibility.

[0090] In one implementation, the method for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power provided by the embodiments of the present invention further includes: When the integrated power AI subsystem issues safety-critical control commands to various subsystems, the integrated power AI subsystem will simultaneously send the command calculation request to itself and the platform AI subsystem. The integrated electrical AI subsystem and the platform AI subsystem each calculate the instruction parameters independently. After the calculation is completed, the two systems exchange the calculation results and compare them. If the comparison results match, the safety-critical control instructions will be executed; if the comparison results do not match, the safety-critical control instructions will be rejected.

[0091] For safety-critical control commands, such as attitude control commands, energy protection commands, and load emergency shutdown commands, a mechanism of independent calculation followed by cross-validation by dual AIs is adopted. The specific process is as follows: the command initiator, typically the integrated electrical AI subsystem, sends the command calculation request to both its own and the platform's AI subsystems simultaneously; each AI independently calculates the command parameters; after calculation, the two systems exchange and compare the results; if the results are consistent or the difference is within a preset tolerance range, the safety-critical control command is confirmed for execution; if the results are inconsistent, i.e., the difference exceeds the preset tolerance range, the safety-critical control command is rejected, and the system enters a safety mode awaiting ground intervention. The preset tolerance range can be set by technical personnel based on experience.

[0092] In one implementation, the power modules of the platform AI subsystem and the integrated power AI subsystem are physically isolated.

[0093] The power modules of the platform AI subsystem and the integrated power AI subsystem are physically isolated, so that a power supply failure in one subsystem will not affect the normal operation of the other subsystem. Under extreme energy conditions, the power allocation scheme of the energy management collaborative optimization mode prioritizes the minimum operating power of the integrated power AI subsystem.

[0094] In this embodiment of the invention, by deploying a high-performance AI platform subsystem and a low-power integrated electrical AI subsystem within a single satellite and constructing an intra-satellite bus interconnection, the mode scheduler uniformly arbitrates the allocation of computing power, enabling the high-performance side to undertake computationally intensive tasks such as distillation, simulation, and prediction, while the low-performance side undertakes tasks such as real-time inference, task generation, and instruction execution, thereby achieving the collaborative utilization of heterogeneous computing power within the satellite and improving the overall computing efficiency of the satellite platform.

[0095] By constructing a teacher network model deployed in the platform's AI subsystem and a student network model deployed in the integrated power AI subsystem, and establishing a model collaborative optimization mode, model knowledge can be transferred between the two AI subsystems within the satellite. The student network model can be optimized without relying on a ground station, shortening the model update cycle. By constructing a multi-dimensional simulation engine and a mission planning module, and establishing a mission planning collaborative verification mode, the mission planning schemes generated by the integrated power AI subsystem can be simulated and evaluated by the platform's AI subsystem before execution, avoiding potential execution risks such as energy exceeding limits and attitude unreachability from the source. By constructing a deep reinforcement learning model for energy prediction and an energy management execution module, and establishing an energy management collaborative optimization mode, energy prediction and power allocation are completed in the platform's AI subsystem, while power execution is completed in the integrated power AI subsystem, achieving cross-AI subsystem division of labor and collaboration in energy management tasks.

[0096] In addition, the platform's AI computing power is uniformly allocated by the mode scheduler according to preset priority rules, so that the three types of needs—energy security, task security, and model evolution—can still be met in an orderly manner under resource competition conditions. The three modes form a collaborative whole through a unified scheduling mechanism.

[0097] The following will use several specific examples to further illustrate the method for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power, as provided in the embodiments of the present invention.

[0098] Example 1 is a typical scenario of MCO mode: The scenario background is that a remote sensing satellite is in orbit for 180 days. Due to the high solar altitude angle period in summer, the target detection performed by the integrated electronic AI subsystem has increased the false detection rate from the baseline of 2.1% to 8.7%.

[0099] The execution process includes: The integrated AI subsystem detects that the false detection rate exceeds the threshold (5%) for three consecutive orbital cycles. It extracts 500 frames of feature data containing false detection samples (approximately 80MB in total) from the sliding window, packages them, and sends them to the platform AI subsystem. The platform AI subsystem's mode scheduler detects the MCO trigger flag. After confirming that neither the EMCO mode nor the MPCV mode is consuming computing power, it initiates the MCO mode execution process. The platform AI subsystem performs cluster analysis on the feedback data packets, identifying that overexposed samples under high solar altitude angle conditions account for 78% of the total false detections. The platform AI subsystem performs targeted fine-tuning of the teacher network model, with a learning rate of 5e-6 and 10 rounds of fine-tuning. Then, it performs knowledge distillation for 50 rounds at T=6 until the loss converges to 0.023. Subsequently, it performs structured pruning and INT8 quantization to remove 30% of redundant channels, compressing the model from 8.7MB to 2.1MB and improving the INT8 inference speed by 3.2 times. Accuracy verification showed that the false detection rate of the compressed student network model decreased to 2.4%, meeting the requirement of not exceeding 3%, and the verification passed. The model version manager of the integrated electrical AI subsystem completed the hot update of the model during the next inference frame interval, with the update taking 60 seconds and not interrupting the current Earth observation mission. The new student network model version is recorded as v2.3, and the old version is retained for rollback.

[0100] Example 2 is a typical scenario of MPCV mode: The scenario background is that a satellite receives an emergency fire monitoring mission from a ground station, requiring it to complete attitude maneuvering to the target area, continuous imaging, and data downlink within 90 minutes.

[0101] The execution process includes: the task planning module of the integrated power AI subsystem, considering both the mission urgency and current orbital conditions, generates a preliminary task planning scheme. This scheme includes completing attitude maneuvers within 15 minutes of the ground shadow emerging, followed by 12 minutes of continuous imaging, and finally 15 minutes of data transmission. A scheme summary (approximately 20KB) is written to the task planning interaction area. The platform AI subsystem's mode scheduler detects the verification request flag and initiates MPCV simulation evaluation. The simulation results reveal two issues: the battery SOC will drop to 12% during the imaging period, below the 15% safety threshold; and there is an 8-minute deviation between the data transmission period and the next ground station's visible window. The platform AI subsystem generates an improved parameter package, including compressing the imaging duration to 6 minutes, covering only the core area of ​​the fire, and delaying the data transmission period by 12 minutes to align with the ground station window. The integrated power AI subsystem receives the revised task planning scheme and resubmits it for verification. The platform AI subsystem passes the second simulation verification. The integrated power AI subsystem executes the revised task planning scheme, and the mission is successfully completed.

[0102] Example 3 is a typical scenario of EMCO mode: The scenario background is that a satellite is about to enter a long shadow area that lasts for 45 minutes, and at the same time needs to perform a high-priority Earth observation task.

[0103] The execution process included: the platform's AI subsystem collecting the current status, including a battery SOC of 85%, solar array power generation of 0W (already in the Earth's shadow), attitude control power consumption of 5W, communication and data transmission power consumption of 50W, payload power consumption of 30W, and an orbital ephemeris showing a shadow duration of 45 minutes. The energy prediction deep reinforcement learning model predicted that if the observation tasks were executed as originally planned, the SOC would drop to 8% after 45 minutes, far below the 15% safety threshold. The safety verification failed. The platform's AI subsystem tightened safety constraints and re-optimized, generating a new power allocation scheme, including a payload power of 10W for the first 15 minutes in the shadow (partial observations), 6W for the middle 20 minutes (lowest power mode), and 8W for the last 10 minutes (resuming observations), with an expected SOC of 18% upon exiting the shadow. The safety verification passed, and the new power allocation scheme was issued to the integrated power AI subsystem for execution. After actual execution, the SOC upon exiting the shadow was 19% (slightly better than predicted, with an error within 5%), and the key observation tasks were completed within 5 minutes of exiting the shadow. The platform's AI subsystem updates the parameters of the energy prediction deep reinforcement learning model based on actual SOC data, improving the strategy security score by 3.2%.

[0104] In this embodiment of the invention, by deploying a high-performance AI platform subsystem and a low-power integrated electrical AI subsystem within a single satellite and constructing an intra-satellite bus to interconnect the two, the high-performance AI subsystem on the platform side undertakes computationally intensive tasks such as knowledge distillation, simulation verification, and DRL prediction, while the low-performance AI subsystem on the integrated electrical side undertakes low-latency tasks such as real-time inference, task execution, and instruction issuance. The mode scheduler enables on-demand allocation and priority arbitration of computing resources, significantly improving the overall computing power utilization efficiency of the satellite platform.

[0105] Through the model collaborative optimization mode, the integrated electrical AI subsystem can autonomously trigger a report when the performance of the student network model deteriorates. The platform AI subsystem completes knowledge distillation, model compression, and accuracy verification in a closed loop within the satellite, and pushes the optimized student network model back to the integrated electrical AI to complete hot update deployment. The entire process does not require ground station intervention, shortening the model update cycle from several days that traditional solutions rely on ground station uploading to several minutes that are completed autonomously within the satellite, greatly improving the satellite's ability to adapt to dynamic changes in the on-orbit environment.

[0106] By adopting a mission planning collaborative verification mode, the mission planning scheme generated by the integrated power AI subsystem is submitted to the multi-dimensional simulation engine of the platform AI subsystem before execution for feasibility verification in four dimensions. If the verification fails, the platform AI subsystem generates a specific improvement suggestion parameter package for the integrated power AI subsystem to iteratively correct, thus avoiding execution risks such as energy over-limit and attitude unreachability from the source and improving the success rate and safety of satellite mission execution.

[0107] Through a collaborative optimization model for energy management, a deep reinforcement learning model specifically designed for satellite orbital environments is adopted. Its state space includes satellite-specific parameters such as orbital phase and solar altitude angle, and its reward function includes battery safety margin constraints. This enables continuous optimization management of orbital-level energy under extreme conditions such as deep discharge in the Earth's shadow region, effectively extending satellite battery life and ensuring energy security.

[0108] The energy forecast data produced by the EMCO mode is used by the MPCV mode for energy dimension verification, and the task execution results of the MPCV mode are used by the MCO mode for model performance monitoring. The three modes form an organic whole through shared data structure and priority scheduling, rather than a simple superposition of three independent functions, realizing unified and coordinated scheduling of the three types of needs of satellite platform energy management, mission planning and model evolution.

[0109] Through a triple protection mechanism of independent power supply for dual AI subsystems, two-way health monitoring, and cross-verification of key commands, the other automatically takes over critical tasks when either AI subsystem malfunctions. Safety-critical control commands can only be executed after being independently calculated and cross-verified by both AI subsystems. This ensures that the satellite's critical functions can continue to operate under single-point failure conditions, providing system-level security for the stable operation of the above three collaborative modes.

[0110] Based on the same inventive concept, embodiments of the present invention also provide a device for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power, see [link to related document]. Figure 3 , Figure 3 This is a schematic diagram of a device for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power, provided by an embodiment of the present invention. The device includes: Deployment module 301 is used to deploy a platform AI subsystem and an integrated electrical AI subsystem interconnected via an intra-satellite bus within a single satellite; The first building module 302 is used to build a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine, and a mode scheduler in the platform's AI subsystem. The teacher network model is used to transfer model knowledge to the integrated power AI subsystem. The energy prediction deep reinforcement learning model is used to generate power allocation schemes based on satellite energy data. The multi-dimensional simulation engine is used to perform multi-dimensional feasibility simulation evaluations of mission planning schemes. The second building module 303 is used to build a student network model, a task planning module, a health perception model, and an energy management execution module in the integrated electric AI subsystem; the student network model is used to perform on-orbit real-time inference; the task planning module is used to generate task planning schemes; the health perception model is used to monitor hardware health; and the energy management execution module is used to execute power allocation schemes. The collaboration module 304 is used to coordinate the operation of model collaboration optimization mode, task planning collaboration verification mode and energy management collaboration optimization mode between the platform AI subsystem and the integrated power AI subsystem according to preset priority rules through the mode scheduler.

[0111] In this embodiment of the invention, by deploying a high-performance AI platform subsystem and a low-power integrated electrical AI subsystem within a single satellite and constructing an intra-satellite bus interconnection, the mode scheduler uniformly arbitrates the allocation of computing power, enabling the high-performance side to undertake computationally intensive tasks such as distillation, simulation, and prediction, while the low-performance side undertakes tasks such as real-time inference, task generation, and instruction execution, thereby achieving the collaborative utilization of heterogeneous computing power within the satellite and improving the overall computing efficiency of the satellite platform.

[0112] By constructing a teacher network model deployed in the platform's AI subsystem and a student network model deployed in the integrated power AI subsystem, and establishing a model collaborative optimization mode, model knowledge can be transferred between the two AI subsystems within the satellite. The student network model can be optimized without relying on a ground station, shortening the model update cycle. By constructing a multi-dimensional simulation engine and a mission planning module, and establishing a mission planning collaborative verification mode, the mission planning schemes generated by the integrated power AI subsystem can be simulated and evaluated by the platform's AI subsystem before execution, avoiding potential execution risks such as energy exceeding limits and attitude unreachability from the source. By constructing a deep reinforcement learning model for energy prediction and an energy management execution module, and establishing an energy management collaborative optimization mode, energy prediction and power allocation are completed in the platform's AI subsystem, while power execution is completed in the integrated power AI subsystem, achieving cross-AI subsystem division of labor and collaboration in energy management tasks.

[0113] In addition, the platform's AI computing power is uniformly allocated by the mode scheduler according to preset priority rules, so that the three types of needs—energy security, task security, and model evolution—can still be met in an orderly manner under resource competition conditions. The three modes form a collaborative whole through a unified scheduling mechanism.

[0114] This invention also provides an electronic device, such as... Figure 4 As shown, it includes a processor 401, a communication interface 402, a memory 403, and a communication bus 404, wherein the processor 401, the communication interface 402, and the memory 403 communicate with each other through the communication bus 404. Memory 403 is used to store computer programs; When the processor 401 executes the program stored in the memory 403, it implements the method steps of any of the above-mentioned methods for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power.

[0115] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the diagram, but this does not indicate that there is only one bus or one type of bus.

[0116] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0117] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0118] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0119] The present invention also provides a computer-readable storage medium. A computer program is stored in the computer-readable storage medium, and when executed by a processor, the computer program implements the method steps of any of the above-described methods for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power.

[0120] Optionally, the computer-readable storage medium may be non-volatile memory (NVM), such as at least one disk storage device.

[0121] Optionally, the aforementioned computer-readable storage medium may also be at least one storage device located remotely from the aforementioned processor.

[0122] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute the steps described in any of the above-described methods for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power.

[0123] It should be noted that the terms "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention.

[0124] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Furthermore, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0125] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings and the disclosure in carrying out the claimed invention. In the description of the invention, the word "comprising" does not exclude other components or steps, "a" or "an" does not exclude a plurality, and "a plurality" means two or more, unless otherwise explicitly specified. Furthermore, while different embodiments may describe certain measures, this does not mean that these measures cannot be combined to produce good results.

[0126] The method provided in this invention can be applied to electronic devices. Specifically, the electronic device can be a desktop computer, a portable computer, a smart mobile terminal, a server, etc. No limitation is made herein; any electronic device that can implement this invention falls within the protection scope of this invention.

[0127] For the embodiments of the device / electronic device / storage medium, since they are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments.

[0128] It should be noted that the device, electronic device and storage medium in the embodiments of the present invention are respectively the device, electronic device and storage medium for applying the above-mentioned method for realizing intelligent collaboration between satellite platform and integrated electrical system based on heterogeneous computing power. Therefore, all embodiments of the above-mentioned method for realizing intelligent collaboration between satellite platform and integrated electrical system based on heterogeneous computing power are applicable to the device, electronic device and storage medium, and can achieve the same or similar beneficial effects.

[0129] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for achieving intelligent collaboration between a satellite platform and integrated electronic systems based on heterogeneous computing power, characterized in that, The method includes: Deploy platform AI subsystems and integrated electrical AI subsystems interconnected via an intra-satellite bus within a single satellite; The platform's AI subsystem constructs a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine, and a pattern scheduler. The teacher network model is used to transfer model knowledge to the integrated power AI subsystem. The energy prediction deep reinforcement learning model is used to generate power allocation schemes based on satellite energy data. The multi-dimensional simulation engine is used to perform multi-dimensional feasibility simulation evaluations of mission planning schemes. The integrated power AI subsystem comprises a student network model, a task planning module, a health perception model, and an energy management execution module. The student network model is used to perform on-orbit real-time inference. The task planning module is used to generate the task planning scheme. The energy management execution module is used to execute the power allocation scheme. The health perception model is used to monitor hardware health. The mode scheduler operates in a collaborative optimization mode, a task planning collaborative verification mode, and an energy management collaborative optimization mode between the platform AI subsystem and the integrated power AI subsystem according to preset priority rules.

2. The method according to claim 1, characterized in that, The mode scheduler, according to preset priority rules, coordinates the operation of three modes between the platform AI subsystem and the integrated power AI subsystem: collaborative optimization mode, task planning collaborative verification mode, and energy management collaborative optimization mode. These modes include: The mode scheduler reads the real-time battery SOC value and determines whether the real-time battery SOC value is lower than a preset emergency threshold; if it is lower than the preset emergency threshold, the energy management collaborative optimization mode is activated. If the threshold is not lower than the preset emergency threshold, check if there is a planning verification request flag in the task planning interaction area; if the planning verification request flag exists and the energy management collaborative optimization mode is not currently being executed, then start the task planning collaborative verification mode. If the planning verification request flag does not exist, check if the model co-optimization trigger flag exists; if the model co-optimization trigger flag exists and the platform AI subsystem is currently idle, then start the model co-optimization mode.

3. The method according to claim 1, characterized in that, The collaborative operation model and collaborative optimization mode between the platform AI subsystem and the integrated power AI subsystem include: The integrated electrical AI subsystem extracts false detection samples and missed detection samples from the sliding window and generates feature data after forward propagation through the student network model. This data is then reported to the platform AI subsystem as a feedback data packet. After receiving the feedback data packet, the platform's AI subsystem performs cluster analysis on the feedback data packet and fine-tunes the teacher network model based on the analysis results. Using the fine-tuned teacher network model as a knowledge source, it performs knowledge distillation on the student network model, compresses the distilled student network model, obtains a compressed student network model file, and verifies its accuracy. After successful verification, the platform AI subsystem pushes the compressed student network model file to the integrated electrical AI subsystem via the satellite bus. During idle periods when the student network model is not performing on-orbit real-time inference, the model version manager of the integrated electric AI subsystem replaces the current student network model with the compressed student network model file to complete the hot update deployment of the model.

4. The method according to claim 1, characterized in that, The collaborative operation mode for task planning and verification between the platform AI subsystem and the integrated power AI subsystem includes: The integrated electric AI subsystem uses the task planning module to generate a preliminary task planning scheme; The platform's AI subsystem uses the multi-dimensional simulation engine to perform a multi-dimensional feasibility simulation evaluation on the preliminary task planning scheme, and obtains multi-dimensional evaluation results. The platform AI subsystem determines the preliminary task planning scheme based on the multi-dimensional evaluation results; when it determines that there is a risk or that it is not feasible, it generates an improvement suggestion parameter package and returns it to the integrated power AI subsystem along with the determination result. The integrated electric vehicle AI subsystem inputs the improved suggestion parameter package as additional constraints into the task planning module, regenerates the revised task planning scheme, and uses the multi-dimensional simulation engine to verify the revised task planning scheme again until the platform AI subsystem determines that the revised task planning scheme is feasible or the number of iterations reaches a preset limit, thus obtaining the final task planning scheme.

5. The method according to claim 1, characterized in that, The energy management collaborative optimization mode operates collaboratively between the platform AI subsystem and the integrated power AI subsystem, including: The platform's AI subsystem collects the satellite's state vector sequence; the state vector sequence is input into the energy prediction deep reinforcement learning model to output the power adjustment factor sequence of each subsystem in multiple future orbital cycles, which serves as the power allocation scheme. The platform AI subsystem performs security verification on the power allocation scheme, simulates the battery state of charge change curve for multiple future orbit cycles after the power allocation scheme is executed, and if the lowest state of charge value in the battery state of charge change curve is lower than the preset safety threshold, the energy prediction deep reinforcement learning model is called again to generate a new power allocation scheme until the verification is passed or the retry limit is reached, the platform AI subsystem sends the final power allocation scheme to the integrated power AI subsystem. The energy management execution module issues power adjustment commands to each subsystem according to the final power allocation scheme, and after execution, packages the actual state of charge change curve and the actual task completion status into execution feedback data, which is then sent to the platform AI subsystem via the intra-satellite bus. The platform's AI subsystem updates the energy prediction deep reinforcement learning model based on the execution feedback data using a policy gradient update algorithm.

6. The method according to claim 1, characterized in that, The method further includes: When the integrated electric vehicle AI subsystem issues safety-critical control commands to each subsystem, the integrated electric vehicle AI subsystem simultaneously sends the command calculation request to itself and the platform AI subsystem. The integrated electric AI subsystem and the platform AI subsystem each independently calculate the instruction parameters. After the calculation is completed, the two systems exchange the calculation results and compare them. If the comparison results match, the execution of the safety-critical control command is confirmed; if the comparison results do not match, the execution of the safety-critical control command is rejected.

7. The method according to claim 1, characterized in that, The power modules of the platform AI subsystem and the integrated power AI subsystem are physically isolated.

8. A device for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power, characterized in that, The device includes: The deployment module is used to deploy platform AI subsystems and integrated electrical AI subsystems interconnected via an intra-satellite bus within a single satellite; The first construction module is used to build a teacher network model, an energy prediction deep reinforcement learning model, a multi-dimensional simulation engine, and a mode scheduler in the platform's AI subsystem; the teacher network model is used to transfer model knowledge to the integrated power AI subsystem; the energy prediction deep reinforcement learning model is used to generate power allocation schemes based on satellite energy data; and the multi-dimensional simulation engine is used to perform multi-dimensional feasibility simulation evaluations of mission planning schemes. The second construction module is used to construct a student network model, a task planning module, a health perception model, and an energy management execution module in the integrated electric AI subsystem; the student network model is used to perform on-orbit real-time inference; the task planning module is used to generate the task planning scheme; the health perception model is used to monitor hardware health; and the energy management execution module is used to execute the power allocation scheme. The collaboration module is used to coordinate the operation of model collaboration optimization mode, task planning collaboration verification mode and energy management collaboration optimization mode between the platform AI subsystem and the integrated power AI subsystem according to preset priority rules through the mode scheduler.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; When a processor executes a computer program stored in a memory, it implements the method for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the method for achieving intelligent collaboration between a satellite platform and integrated electrical systems based on heterogeneous computing power as described in any one of claims 1-7.

Images