On-orbit satellite load maintenance system and method based on edge perception and DRL

Abstract

The invention relates to an in-orbit satellite load maintenance system and method based on edge perception and DRL, and the system comprises an optical imaging module which is used for carrying out the imaging of the earth surface, and obtaining an image flow; the satellite edge AI processor is used for acquiring the image stream and telemetry data of a satellite and analyzing image quality based on the in-orbit image quality evaluation model on which the image stream passes to obtain image quality information; the focusing compensation module is used for generating a focusing compensation instruction and a thermal control power adjustment instruction by utilizing a deep reinforcement learning method according to the image quality information and telemetry data of a satellite; and the execution mechanism driver is used for driving the optical imaging module according to the focusing compensation instruction and driving a thermal control device on the satellite according to the thermal control power adjustment instruction. Maintenance is completed at the satellite-borne edge end in real time, the method has the advantages of being high in response speed, good in real-time performance and good in maintenance effect, the satellite imaging quality is improved, and the on-orbit service life of a satellite is prolonged.

Description

Technical Field

[0001] This disclosure relates to the field of remote sensing satellite technology, and in particular to an on-orbit satellite payload maintenance system and method based on edge sensing and DRL. Background Technology

[0002] As the on-orbit operation time of high-resolution remote sensing satellites increases, optical payload imaging systems face severe challenges from the space environment, including high-energy particle radiation causing thermal pixels in image sensors (CMOS / CCD), defocusing caused by thermal expansion and contraction of optical lens groups due to drastic temperature differences, and mechanical wear of the focusing mechanism.

[0003] Currently, satellite payload maintenance primarily relies on a "ground station monitoring mode." The satellite periodically transmits telemetry data (such as temperature, current, and historical imaging quality parameters) of the payload to the ground station via downlink. Ground engineers analyze this data using offline models to determine the payload's health status and issue downlink control commands (such as adjusting focus steps or modifying thermal control thresholds) during the next transit window.

[0004] In existing solutions, sensor signal processing typically follows the steps of signal amplification → analog-to-digital conversion (ADC) → data packetization and transmission. This approach is similar to the sensor parameter acquisition process, relying on complex analog circuits and ADC conversion.

[0005] The existing technology has the following disadvantages: (1) Poor real-time performance: Star-to-ground communication suffers from severe window period limitations and transmission delays, making it impossible to respond immediately to sudden payload performance degradation (such as instantaneous imaging anomalies caused by radiation); (2) The original high-definition image data is large and cannot be fully transmitted back to the ground for quality assessment, resulting in insufficient samples for maintenance decisions and affecting the maintenance effect. Summary of the Invention

[0006] Therefore, it is necessary to provide an on-orbit satellite payload maintenance system and method based on edge perception and DRL to address the problems of poor real-time performance and the need to improve maintenance effectiveness.

[0007] To solve the above problems, the present disclosure adopts the following technical solution: In a first aspect, this disclosure provides an on-orbit satellite payload maintenance system based on edge perception and DRL, comprising: an optical imaging module, a satellite edge AI processor, and an actuator driver; the optical imaging module is used to image the Earth's surface to obtain an image stream; the satellite edge AI processor carries an on-orbit image quality evaluation model, and is used to acquire the image stream and satellite telemetry data, to analyze image quality based on the image stream using the on-orbit image quality evaluation model to obtain image quality information, and to generate focus compensation commands and thermal control power adjustment commands using deep reinforcement learning methods based on the image quality information and satellite telemetry data; the actuator driver is used to drive the optical imaging module according to the focus compensation command and to drive the thermal control devices on the satellite according to the thermal control power adjustment command.

[0008] In a preferred embodiment, the image quality information includes sharpness and / or imaging signal-to-noise ratio, and the image quality information also includes noise distribution and / or spatial distribution of abnormal pixels in the thermal image; the telemetry data includes temperature data, radiation data, and current data.

[0009] In a preferred embodiment, the satellite edge AI processor is used to acquire weight update information from the ground station and update the on-orbit image quality assessment model accordingly.

[0010] In a preferred embodiment, the operation of generating focus compensation commands and thermal control power adjustment commands using deep reinforcement learning methods is performed when the satellite is in the observation area.

[0011] In a preferred embodiment, the state of the deep reinforcement learning method includes temperature, imaging contrast, noise variance, and cumulative radiation, and the action of the deep reinforcement learning method is a focus compensation command and a thermal control power adjustment command.

[0012] In a preferred embodiment, the focus compensation command is the number of focus motor compensation pulses, and the thermal control power adjustment command is the current adjustment value of the thermoelectric cooler.

[0013] In a preferred embodiment, the deep learning algorithm employs a near-end strategy optimization algorithm.

[0014] In a preferred embodiment, the step of generating focus compensation instructions and thermal control power adjustment instructions using deep reinforcement learning methods based on the image quality information and satellite telemetry data includes the step of fusing the acquired image quality information and satellite telemetry data to form a reinforcement learning state space.

[0015] In a preferred embodiment, the system further includes a conversion module for converting the dark current drift of the photosensitive chip caused by temperature changes into a change in bias current; the input of the satellite edge AI processor includes the change in bias current caused by the increase in temperature of the space environment in which the satellite is located, and the satellite edge AI processor is specifically used to obtain image quality information by analyzing the image quality based on the change in bias current, the image stream, and the on-orbit image quality evaluation model thereon.

[0016] Secondly, this disclosure provides an on-orbit satellite payload maintenance method based on edge perception and DRL. The satellite is equipped with a satellite edge AI processor, which carries an on-orbit image quality assessment model. The method includes: The optical imaging module images the Earth's surface to obtain an image stream; The satellite edge AI processor acquires the image stream and satellite telemetry data, and analyzes the image quality based on the image stream using the on-orbit image quality evaluation model to obtain image quality information; The satellite edge AI processor generates focus compensation commands and thermal control power adjustment commands based on the image quality information and satellite telemetry data using deep reinforcement learning methods. The actuator driver drives the optical imaging module according to the focus compensation command; The actuator driver drives the thermal control devices on the satellite according to the thermal control power adjustment command.

[0017] The aforementioned on-orbit satellite payload maintenance system and method based on edge perception and DRL analyzes image quality using a satellite edge AI processor. Based on the image quality information and satellite telemetry data, it generates focus compensation commands and thermal control power adjustment commands using deep reinforcement learning. The actuator driver then drives the optical imaging module and thermal control devices to respond according to these commands, completing the process in real-time at the satellite edge. This provides advantages such as fast response speed and good real-time performance, effectively addressing sudden environmental interference and providing immediate response to sudden payload performance degradation. Furthermore, this disclosure eliminates the need to transmit images back to the ground; images are directly input into the satellite edge AI processor. This ensures sufficient samples for maintenance decisions, improving the maintenance effectiveness of on-orbit satellite payload maintenance, enhancing satellite imaging quality, and extending the satellite's on-orbit lifespan. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the system structure in one embodiment of the present disclosure; Figure 2 This is a flowchart illustrating a method in one embodiment of the present disclosure. Detailed Implementation

[0019] The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and preferred embodiments.

[0020] See Figure 1 This disclosure provides an on-orbit satellite payload maintenance system based on edge perception and DRL, comprising: an optical imaging module, a satellite edge AI processor, and an actuator driver; the optical imaging module is used to image the Earth's surface to obtain an image stream; the satellite edge AI processor carries an on-orbit image quality evaluation model, and is used to acquire the image stream and satellite telemetry data, analyze the image quality based on the image stream using the on-orbit image quality evaluation model to obtain image quality information, and generate focus compensation commands and thermal control power adjustment commands using deep reinforcement learning methods based on the image quality information and satellite telemetry data; the actuator driver is used to drive the optical imaging module according to the focus compensation command and to drive the thermal control devices on the satellite according to the thermal control power adjustment command.

[0021] Understandably, the optical imaging module, satellite edge AI processor, and actuator driver are all located on the satellite. Specifically, the satellite edge AI processor belongs to the satellite edge computing platform.

[0022] The optical imaging module on a satellite is used to perform high-precision, wide-area imaging of the Earth's surface from space, acquiring image information of surface targets. The optical imaging module includes an optical lens, a focusing mechanism, and an image sensor, arranged sequentially along the optical axis. The optical lens focuses incident light to form a clear optical image on the imaging plane. The image sensor is mounted at the focal plane of the optical system, directly opposite the end of the optical lens's optical path, and receives the focused light signal, converting it into an electrical signal. The focusing mechanism is used to fine-tune the position of the optical lens. The optical lens, focusing mechanism, and image sensor on the satellite work together to achieve high-precision imaging. The optical lens is responsible for "collecting light," the focusing mechanism for "focusing," and the image sensor, located on the imaging plane of the optical system, is responsible for converting the focused light signal into an electrical signal, which is then processed to form a digital image. Understandably, the optical imaging module images the Earth's surface multiple times to obtain an image stream.

[0023] In this embodiment, the satellite edge AI processor is also used to acquire the weight update information (of the on-orbit image quality assessment model) from the ground station and update the on-orbit image quality assessment model accordingly. Both the edge AI processor and the ground station have on-orbit image quality assessment models. The architecture of the on-orbit image quality assessment model carried by the satellite edge AI processor is the same as that of the on-orbit image quality assessment model of the ground station. Specifically, before updating according to the weight update information, the weight information and other parameters of the two models may differ, but their architecture and function are the same; after updating according to the weight update information, the on-orbit image quality assessment model carried by the satellite edge AI processor is completely identical to that of the on-orbit image quality assessment model of the ground station. In a specific embodiment, the weight update information of the on-orbit image quality assessment model of the ground station is transmitted to the onboard computer via a satellite-to-ground link. The onboard computer sends the weight update information to the satellite edge AI processor. The ground station is responsible for retraining the on-orbit image quality assessment model, while the satellite edge AI processor at the satellite edge is responsible for real-time data acquisition, status assessment, and maintenance decisions.

[0024] In this embodiment, the telemetry data acquired by the satellite (i.e., the satellite's telemetry data) refers to the key physical quantities of the spacecraft's internal or external environment collected in real time by onboard sensors. The telemetry data includes temperature data, radiation data, and current data. Temperature data represents the operating temperature of onboard electronic equipment, batteries, optical payloads, etc., monitored by the satellite, and measures the high-energy particle radiation level in the space where the satellite is located to prevent overheating or cryogenic failure. Current data represents the relevant current of the satellite's onboard power system, used to reflect the output and load of the power system. Radiation dose represents the high-energy particle radiation level in the space where the satellite is located, used to assess the damage caused by high-energy particles in space to chips and materials, and to provide early warning of single-event upset and other fault risks. Temperature data is acquired using temperature sensors, current data is acquired using current transformers, and radiation dose data is acquired using radiation detectors.

[0025] In one embodiment, the image quality information includes sharpness (using modulation transfer function MTF as a sharpness metric) and / or imaging signal-to-noise ratio, and the image quality information also includes noise distribution and / or the spatial distribution of anomalous pixels in the thermal image.

[0026] In this embodiment, after the optical imaging module captures an image, the satellite edge AI processor extracts image quality information through a convolutional neural network, such as the modulation transfer function (MTF) and noise level of the current imaging.

[0027] In one embodiment, the edge AI processor is activated (i.e., turned on) during the imaging period, and / or the edge AI processor is activated during a preset period.

[0028] In one embodiment, a "switching ring oscillator" control logic is introduced to reduce power consumption. The satellite edge AI processor is controlled by task scheduling pulses. The operation of generating focus compensation commands and thermal power adjustment commands using deep reinforcement learning methods by the satellite edge AI processor is executed when the satellite is in the observation zone. That is, when the satellite is in the observation zone, the satellite edge AI processor generates focus compensation commands and thermal power adjustment commands based on the image quality information and satellite telemetry data, using deep reinforcement learning methods. When the satellite enters the observation zone, the control voltage is set high, and the satellite edge AI processor starts running the deep reinforcement learning (DRL) strategy; when the satellite enters the non-observation zone, specifically, when the satellite enters the shadow zone and / or the non-task zone, the control voltage is set low, the satellite edge AI processor enters sleep mode, and the satellite only maintains basic hardware health monitoring. It is understood that in some embodiments, the shadow zone can be defined as the non-observation zone.

[0029] This design allows computing resources to be used only when necessary, reducing the drain on the satellite's limited energy supply from the satellite's edge AI processor.

[0030] The deep reinforcement learning method includes: The acquired telemetry data from the satellite is fused with the image quality information (image features) to form a state space for reinforcement learning.

[0031] State S: (One-to-one correspondence of temperature, imaging contrast, noise variance, and cumulative radiation).

[0032] Action A: Outputs the number of compensation pulses for the focusing motor and the current adjustment value for the TEC (thermoelectric cooler).

[0033] temperature Temperature data from satellite telemetry data, and image quality information including imaging contrast. and noise distribution, noise variance The cumulative radiation amount is determined based on the noise distribution. The data is obtained from the radiation data in the satellite's telemetry data. The focus compensation command is the number of compensation pulses of the focus motor, and the thermal control power adjustment command is the current adjustment value of the TEC (thermoelectric cooler), which serves as the thermal control device on the satellite.

[0034] In some embodiments, the action output also includes an image sensor bias voltage adjustment command. This allows for direct, microsecond-level dynamic adjustment of the sensor bias voltage and the focus motor drive pulses, avoiding the cumbersome ground intervention procedures of traditional solutions.

[0035] The deep reinforcement learning algorithm can employ the PPO (Proximal Policy Optimization) algorithm, which uses payload imaging sharpness (modulation transfer function MTF) and power consumption balance as reward objectives, and automatically outputs focus compensation commands and thermal control power adjustment commands.

[0036] In other embodiments, the PPO algorithm can be replaced by a lightweight Transformer model to enhance the ability to process long sequence environmental data.

[0037] Preferably, the deep reinforcement learning algorithm has self-evolution capability and can automatically correct the control strategy according to the nonlinear wear generated by the load over time, without the need for frequent ground patching.

[0038] In one specific embodiment, the satellite edge AI processor includes an image quality assessment module and a DRL (Deep Reinforcement Learning) module. The image quality assessment module carries an on-orbit image quality assessment model. The function of the image quality assessment module includes analyzing image quality based on the image stream using the on-orbit image quality assessment model to obtain image quality information. The DRL module includes a policy network unit (edge-end neural functional network unit), a state space unit, and a reward function unit. The edge-end neural functional network unit includes an input layer, a hidden layer, an Actor head (also known as a policy head, acting as a decision-maker), and a Critic head (also known as a value head, acting as an evaluator). The input layer connects to the hidden layer, and the hidden layer connects to the Actor head and the Critic head. State features can be extracted through the input layer and hidden layer (fully connected / ResNet residual network), divided into two branches. The Actor head is responsible for outputting specific hardware control parameters (such as cooling power and focus compensation), i.e., generating specific action policies and outputting actions, including thermal power adjustment commands and focus compensation commands. The state space unit is an intermediate layer that defines three key state variables of the environment: imaging quality (modulation transfer function MTF), ambient temperature, and so on. Cumulative radiation These three states form the basis for the agent's decision-making. The reward function unit uses image quality satisfaction as a positive incentive and energy consumption and radiation consumption as negative penalties. Based on the positive incentive and negative penalty, it obtains the total reward R, which is then fed back to the Critic head after a comprehensive balance. The Critic head is responsible for evaluating the value of the current state and balancing image quality and hardware lifespan through the reward function.

[0039] During satellite operation, it experiences frequent transitions into and out of Earth's shadow, resulting in severe temperature fluctuations (up to ±150°C or higher). This thermal radiation shock causes significant drift in the dark current of CMOS or CCD image sensors, introducing image noise, reducing the signal-to-noise ratio, and in severe cases, even causing payload imaging failure. Traditional ground calibration methods cannot adapt to real-time thermal environment changes in orbit. To address this issue, this embodiment integrates a temperature-adaptive dark current compensation circuit (also known as a "temperature and radiation drift cancellation circuit") at the front end of the satellite payload's imaging unit. The core function of this circuit is to sense the dark current disturbance in the image sensor caused by temperature and radiation changes in real time, and dynamically generate a compensation current of equal amplitude but opposite polarity, injecting it into the signal path to cancel out the dark current drift at its source and maintain the stability of the output signal.

[0040] The change in bias current is calculated using the formula: in, This represents the change in bias current, i.e., the output current. Indicates carrier mobility. Indicates the gate oxide capacitance. Indicates the width of the channel. Indicates the length of the channel. This represents the voltage between the gate and the source. Indicates the gate-source threshold voltage; The change in bias current caused by the increased temperature of the space environment in which the satellite is located. The values ​​are fed back to the input of the satellite edge AI processor as environmental baseline correction values, thereby offsetting the effect of temperature on the imaging signal-to-noise ratio.

[0041] The on-orbit satellite payload maintenance system also includes a conversion module for converting the dark current drift of the photosensitive chip caused by temperature changes into a change in bias current. The satellite edge AI processor receives input from the change in bias current. The satellite edge AI processor analyzes the image quality based on the image stream, using the on-orbit image quality evaluation model, based on the change in bias current caused by temperature increase, to obtain image quality information. Specifically, it analyzes the imaging signal-to-noise ratio using the on-orbit image quality evaluation model.

[0042] In this embodiment, the actuator driver directly controls the thermal control via pulse width modulation (PWM), and can also directly control the focus via PWM, thereby improving the accuracy and linearity of the adjustment.

[0043] In one specific embodiment, the satellite has a physical hardware layer and an edge control layer. The edge control layer includes a health monitoring module and a DRL (Deep Reinforcement Learning) maintenance controller. The physical hardware layer includes an optical lens, a focusing mechanism, an image sensor, and a thermoelectric cooler. The optical lens generates temperature drift / vibration, the focusing mechanism focuses, and the image sensor achieves imaging. The thermoelectric cooler can handle the thermal noise of the image sensor. The health monitoring module can monitor relevant data of the optical lens and the image sensor and detect degradation. Based on the detected degradation results, the DRL maintenance controller controls the thermoelectric cooler through PWM and controls the focusing mechanism through step pulses to correct the parameters of the image sensor. That is, the DRL maintenance controller dynamically adjusts the cooling power of the TEC according to the thermal noise feedback of the image sensor CMOS and predictively compensates the focusing mechanism according to the trend of image sharpness degradation, thereby avoiding imaging failure caused by drastic environmental changes.

[0044] This disclosure provides an on-orbit satellite payload maintenance method based on edge perception and DRL. The satellite is equipped with a satellite edge AI processor, which carries an on-orbit image quality assessment model. See [link to relevant documentation]. Figure 2 The method includes: The optical imaging module images the Earth's surface to obtain an image stream; The satellite edge AI processor acquires the image stream and satellite telemetry data, and analyzes the image quality based on the image stream using the on-orbit image quality evaluation model to obtain image quality information; The satellite edge AI processor generates focus compensation commands and thermal control power adjustment commands based on the image quality information and satellite telemetry data using deep reinforcement learning methods. The actuator driver drives the optical imaging module according to the focus compensation command; The actuator driver drives the thermal control devices on the satellite according to the thermal control power adjustment command.

[0045] It should be understood that the above method does not imply or suggest that all steps must be performed in this order.

[0046] In one specific embodiment, the method includes: the satellite edge AI processor acquiring the weight update information of the on-orbit image quality assessment model of the ground station and updating the on-orbit image quality assessment model thereon accordingly.

[0047] In one specific embodiment, the method includes: The satellite acquires image streams and satellite telemetry data in real time; The satellite edge AI processor acquires the image stream and satellite telemetry data; The satellite edge AI processor analyzes the image quality based on the image stream using the on-orbit image quality evaluation model to obtain sharpness and / or imaging signal-to-noise ratio; The satellite edge AI processor analyzes the image quality based on the image stream and obtains the noise distribution and / or the spatial distribution of abnormal pixels in the thermal image through the on-orbit image quality evaluation model on it; Based on the image quality information and satellite telemetry data, a deep reinforcement learning method is used. The deep reinforcement learning includes the construction of the state space, neural network inference (using the PPO algorithm), and the generation of focus compensation commands, thermal control power adjustment commands, and image sensor bias voltage control commands. The actuator driver drives the thermal control devices on the satellite and the image sensor bias voltage control commands according to the focus compensation command and thermal control power adjustment command; Determine whether the satellite's performance (e.g., imaging performance) has recovered; If the fault is not recovered, the system switches to a backup component or logs the failure. If the fault is recovered, the satellite edge AI processor continues to operate.

[0048] In specific implementation, the on-orbit satellite payload maintenance method based on edge perception and DRL can refer to the implementation of the on-orbit satellite payload maintenance method system based on edge perception and DRL in any of the above embodiments, and will not be described in detail here.

[0049] The simulation results of the system and method disclosed herein are as follows: Through joint debugging with STK (Satellite Toolbox) and digital twin platform, the results show that under simulated reaction flywheel vibration and extreme temperature environment, the system can keep the MTF value of payload imaging stable above 0.15. Compared with the traditional ground maintenance scheme, the effective imaging duration is increased by 22% and the total energy consumption of the payload is reduced by 15%.

[0050] This disclosure discloses an on-orbit satellite payload maintenance system and method based on edge perception and DRL. It utilizes a satellite edge AI processor for image quality analysis, and based on the image quality information and satellite telemetry data, generates focus compensation commands and thermal control power adjustment commands using deep reinforcement learning. The actuator driver then drives the optical imaging module and thermal control devices to respond according to these commands. This system offers advantages such as fast response speed and good real-time performance. Compared to existing ground station monitoring methods, this disclosure completes the process in real-time at the satellite edge, reducing response time from "hours" to "milliseconds." It effectively addresses sudden environmental interference and provides immediate responses to sudden payload performance degradation (such as instantaneous imaging anomalies caused by radiation). Unlike ground station monitoring methods, this disclosure eliminates the need to transmit images back to the ground. Images are directly used as input to the satellite edge AI processor, ensuring sufficient samples for maintenance decisions and improving maintenance effectiveness. This disclosure achieves on-orbit satellite payload maintenance based on edge computing and deep reinforcement learning. The satellite payload maintenance has good real-time performance and good maintenance effect, realizes autonomous predictive maintenance, improves satellite imaging quality and extends satellite on-orbit life.

[0051] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0052] The embodiments described above are merely illustrative of several implementations of this disclosure, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this disclosure, and these all fall within the protection scope of this disclosure. Therefore, the protection scope of this patent should be determined by the appended claims.

Claims

1. An on-orbit satellite payload maintenance system based on edge sensing and DRL, characterized in that, include: Optical imaging module, satellite edge AI processor, actuator driver; The optical imaging module is used to image the Earth's surface to obtain an image stream; the satellite edge AI processor carries an on-orbit image quality evaluation model, and the satellite edge AI processor is used to acquire the image stream and satellite telemetry data, to analyze the image quality based on the image stream through the on-orbit image quality evaluation model to obtain image quality information, and to generate focus compensation commands and thermal control power adjustment commands based on the image quality information and satellite telemetry data using deep reinforcement learning methods; The actuator driver is used to drive the optical imaging module according to the focus compensation command and to drive the thermal control device on the satellite according to the thermal control power adjustment command.

2. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The image quality information includes sharpness and / or imaging signal-to-noise ratio, and also includes noise distribution and / or spatial distribution of abnormal pixels in the thermal image; the telemetry data includes temperature data, radiation data, and current data.

3. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The satellite edge AI processor is used to obtain weight update information from the ground station and update the on-orbit image quality evaluation model accordingly.

4. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The operation of generating focus compensation commands and thermal control power adjustment commands using deep reinforcement learning methods is performed when the satellite is in the observation area.

5. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The states of the deep reinforcement learning method include temperature, imaging contrast, noise variance, and cumulative radiation. The actions of the deep reinforcement learning method are focus compensation commands and thermal control power adjustment commands.

6. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 5, characterized in that, The focus compensation command is the number of compensation pulses for the focus motor, and the thermal control power adjustment command is the current adjustment value of the thermoelectric cooler.

7. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The deep learning algorithm employs a near-end strategy optimization algorithm.

8. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The step of generating focus compensation commands and thermal control power adjustment commands using deep reinforcement learning methods based on the image quality information and satellite telemetry data includes the step of fusing the acquired image quality information and satellite telemetry data to form a reinforcement learning state space.

9. The on-orbit satellite payload maintenance system based on edge sensing and DRL according to claim 1, characterized in that, The system also includes a conversion module for converting the dark current drift of the photosensitive chip caused by temperature changes into a change in bias current. The input of the satellite edge AI processor includes the change in bias current caused by the increase in temperature of the space environment in which the satellite is located. Specifically, the satellite edge AI processor is used to analyze the image quality based on the change in bias current, the image stream, and the on-orbit image quality evaluation model to obtain image quality information.

10. An on-orbit satellite payload maintenance method based on edge sensing and DRL, characterized in that, The satellite is equipped with a satellite edge AI processor, which carries an on-orbit image quality assessment model. The method includes: The optical imaging module images the Earth's surface to obtain an image stream; The satellite edge AI processor acquires the image stream and satellite telemetry data, and analyzes the image quality based on the image stream using the on-orbit image quality evaluation model to obtain image quality information; The satellite edge AI processor generates focus compensation commands and thermal control power adjustment commands based on the image quality information and satellite telemetry data using deep reinforcement learning methods. The actuator driver drives the optical imaging module according to the focus compensation command; The actuator driver drives the thermal control devices on the satellite according to the thermal control power adjustment command.

Images