Space-based self-labeling two-level distributed model training system and method

Abstract

The present disclosure relates to a space-based self-labeling two-level distributed model training system and method, the system comprising: a first-level processing module deployed on each low-orbit satellite, configured to extract embedding features of batches of local data respectively and assign identification information to each embedding feature, and send each embedding feature and the assigned identification information to a corresponding medium-high-orbit satellite; a second-level processing module deployed on each medium-high-orbit satellite, configured to process each embedding feature sent by the low-orbit satellite into a corresponding task label based on a task demand sent by a control module deployed on a ground control center; and the first-level processing module deployed on each low-orbit satellite is further configured to construct a training data set based on the batches of local data, the task label corresponding to the embedding feature issued by the medium-high-orbit satellite, and the corresponding identification information, and train an initial model using the training data set to obtain a trained task model. Thus, efficient autonomous on-board collaborative and adaptive continuous model training can be achieved.

Description

Technical Field

[0001] This disclosure relates to the field of on-board intelligent processing, and in particular to a space-based self-labeled two-level distributed model training system and method. Background Technology

[0002] With the continuous improvement of low-Earth orbit (LEO) satellite data acquisition capabilities (such as the remote sensing data acquisition capabilities of LEO remote sensing satellites) and the increasing number of missions, the amount of local data generated on satellites daily is growing exponentially. Improving the efficiency and application value of on-board local data utilization is crucial for obtaining adaptive and high-precision mission models to perform corresponding tasks (such as classification and object detection) by training and continuously fine-tuning models based on on-board local data. However, due to limitations in the computing power, storage resources, and satellite-to-ground communication conditions of LEO satellites, existing methods for performing data processing and model training on LEO satellites still face many challenges.

[0003] Currently, the common approach is to acquire data on satellite and train it centrally on the ground. Raw data collected by low-Earth orbit (LEO) satellites is downloaded to ground stations, where ground control centers train the models. The trained models are then uploaded back to LEO satellites for deployment to perform tasks. However, this method requires high-speed satellite-to-ground data transmission, which is hampered by factors such as satellite overpass time, downlink capacity, the number of ground stations, and their deployment locations. Furthermore, during centralized ground training, data takes a long time to be transmitted in batches, resulting in poor model training timeliness. To reduce the pressure of data downlink, some studies have attempted to deploy pre-trained lightweight task models directly on LEO satellites for on-board inference, reducing the amount of data transmitted between satellite and ground. However, limited by on-board computing power and model size, the model's inference capability is limited, and it struggles to adapt to data drift during long-term operation. It also cannot adaptively update the model to suit different satellite imaging environments and mission requirements, resulting in poor model accuracy. Other studies have proposed distributed training or federated learning on-board or between satellite and ground, but the lack of high-quality data annotations on satellites also leads to poor training results. Therefore, how to achieve high-quality and adaptive continuous training of on-board models using local data collected by low-Earth orbit satellites without relying on large-scale on-board data downlink remains a technical problem that urgently needs to be solved. Summary of the Invention

[0004] In view of this, this disclosure proposes a space-based self-labeled two-level distributed model training system and method, which can make full use of the computing power, resources and communication capabilities of satellites in different orbits, as well as local data collected by low-orbit satellites, to achieve efficient and adaptive continuous training of on-board models without relying on large-scale on-board data downlink.

[0005] According to one aspect of this disclosure, a space-based self-annotated two-level distributed model training system is provided, comprising: a primary processing module deployed on each of multiple low-Earth orbit (LEO) satellites, a secondary processing module deployed on each of multiple medium-to-high Earth orbit (MEO) satellites, and a control module deployed at a ground control center; wherein, the primary processing module deployed on each LEO satellite is configured to: in response to receiving a model training instruction sent by a MEO satellite, extract the embedding features of each batch of local data and assign identification information to each embedding feature, and send each embedding feature and the assigned identification information to the corresponding MEO satellite, wherein the local data is data collected by the LEO satellite, and the identification information is used to distinguish the embedding features of different local data; the secondary processing module deployed on each MEO satellite is configured to: The configuration is as follows: based on the task requirements sent by the control module deployed at the ground control center, each embedded feature sent by the low-Earth orbit satellite is processed into a corresponding task label, and the task label and corresponding identification information corresponding to each embedded feature are sent to the corresponding low-Earth orbit satellite. The task requirements represent the tasks that the task model on the low-Earth orbit satellite needs to perform, and the task labels are used as labels for local data to train the task model. The primary processing module deployed on each low-Earth orbit satellite is also configured to: construct a training dataset based on the batch of local data, the task labels corresponding to the embedded features sent by the medium and high orbit satellites, and the corresponding identification information, and use the training dataset to train an initial model to obtain a trained task model. The task model is used to execute the tasks indicated by the task requirements.

[0006] In some embodiments, processing the embedded features sent by the low-Earth orbit satellite into corresponding task labels based on the task requirements sent by the control module deployed at the ground control center includes: extracting the depth features of each embedded feature using a depth feature extractor; generating task labels and label confidence scores for each embedded feature based on the task requirements using a corresponding label generator according to the depth features of each embedded feature, wherein the label confidence score is used to characterize the reliability of the task label; wherein, sending the task labels and corresponding identification information corresponding to each embedded feature to the corresponding low-Earth orbit satellite includes: sending high-confidence task labels with label confidence scores higher than a specified threshold and the corresponding identification information of the embedded features to the corresponding low-Earth orbit satellite, so as to train an initial model in the low-Earth orbit satellite using the high-confidence task labels.

[0007] In some embodiments, the secondary processing module deployed on each medium-high orbit satellite is further configured to: send low-confidence task labels with a label confidence level lower than the specified threshold and their corresponding embedding features to the ground control center; the control module deployed on the ground control center is configured to: correct the low-confidence task labels corresponding to the embedding features sent by the medium-high orbit satellite to obtain corrected task labels corresponding to the embedding features, and upload the corrected task labels corresponding to the embedding features to the corresponding medium-high orbit satellite; wherein, sending the task labels corresponding to each embedding feature and their corresponding identification information to the corresponding low-orbit satellite further includes: sending the corrected task labels and their corresponding identification information sent by the ground control center to the corresponding low-orbit satellite, so as to train the initial model on the low-orbit satellite using the corrected task labels.

[0008] In some embodiments, the primary processing module deployed on each low-Earth orbit satellite is further configured to: send the model parameters of the trained mission model to the corresponding medium-high orbit satellite; the secondary processing module deployed on each medium-high orbit satellite is further configured to: aggregate the model parameters sent by at least two low-Earth orbit satellites respectively to obtain the global model parameters of the current version, and distribute the global model parameters of the current version to the corresponding at least two low-Earth orbit satellites; the primary processing module deployed on each low-Earth orbit satellite is further configured to: in response to receiving the global model parameters of the current version sent by the medium-high orbit satellite, replace the model parameters of the trained mission model with the global model parameters of the current version to obtain the current version mission model, so as to use the current version mission model to execute the mission indicated by the mission requirements.

[0009] In some embodiments, the secondary processing module deployed on each medium-high orbit satellite is further configured to: send the current version of the global model parameters to the ground control center; the control module deployed on the ground control center is further configured to: perform an inference accuracy test on the current version of the global model parameters sent by the medium-high orbit satellite, and if the inference accuracy of the current version of the global model parameters does not reach the specified mission requirement accuracy, send a continue training instruction to the corresponding medium-high orbit satellite; the secondary processing module deployed on each medium-high orbit satellite is further configured to: forward the continue training instruction sent by the ground control center to at least two corresponding low-Earth orbit satellites; the primary processing module deployed on each low-Earth orbit satellite is further configured to: in response to receiving the continue training instruction, use the current version of the mission model as the initial model to re-execute the training of the initial model using the training dataset to obtain the trained mission model, until the inference accuracy of the global model parameters obtained by aggregating the model parameters of the trained mission model reaches the mission requirement accuracy.

[0010] In some embodiments, the control module deployed by the ground control center is further configured to: send a stop training command to the corresponding medium-high orbit satellite when the inference accuracy of the global model parameters in the current version has reached the accuracy required by the task; the secondary processing module deployed by each medium-high orbit satellite is further configured to: forward the stop training command sent by the ground control center to at least two corresponding low-orbit satellites; the primary processing module deployed by each low-orbit satellite is further configured to: stop model training and use the current version of the task model to execute the task indicated by the task requirements in response to receiving the stop training command.

[0011] In some embodiments, sending a continued training instruction to the corresponding medium-to-high orbit satellite when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy includes: sending a continued training instruction to the corresponding medium-to-high orbit satellite when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy and the inference accuracy of the global model parameters in the current version is the same as or improved compared to the global model parameters in the historical version; the global model parameters in the historical version are the global model parameters generated before the global model referenced in the current version; the control module deployed by the ground control center is further configured to: when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy and the inference accuracy of the global model parameters in the current version is the same as or improved compared to the global model parameters in the historical version, send a continued training instruction to the corresponding medium-to-high orbit satellite ... send a continued training instruction to the corresponding medium-to-high orbit satellite when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy and the global model parameters in the historical version are the global model parameters generated before the global model referenced in the historical version; the global model parameters in the historical version are the global model parameters generated before the global model referenced in the historical version; the global model parameters in the historical version are the global model parameters generated before the global model referenced in the historical version; the global model parameters in the historical version are the global model parameters generated before the global model referenced in the historical version; the global model parameters in the historical If the inference accuracy of the global model parameters decreases compared to the historical version, a rollback command is sent to the corresponding medium-high orbit satellites. The secondary processing module deployed on each medium-high orbit satellite is further configured to forward the rollback command sent by the ground control center to at least two corresponding low-Earth orbit satellites. The primary processing module deployed on each low-Earth orbit satellite is further configured to, in response to receiving the rollback command, restore the current version of the task model to the previous version and retrain it using the previous version as the initial model, or stop model training and execute the task indicated by the task requirements using the previous version of the task model. The previous version of the task model is the task model using the global model parameters before the inference accuracy decreases.

[0012] In some embodiments, the secondary processing module deployed on each medium-high orbit satellite is further configured to: after aggregating the model parameters sent by at least two low-orbit satellites to obtain the current version of the global model parameters, synchronize the current version of the global model parameters to the remaining medium-high orbit satellites, so that in the event of a failure of any medium-high orbit satellite, the non-failed medium-high orbit satellites can dock with at least two low-orbit satellites corresponding to the failed medium-high orbit satellite to continue to perform the functions required by the failed medium-high orbit satellite.

[0013] In some embodiments, the step of extracting the embedding features of each local data in a batch includes: preprocessing any local data in the batch to obtain preprocessed local data; and using an embedding feature extractor to generate embedding features of the preprocessed local data based on the preprocessed local data.

[0014] According to another aspect of this disclosure, a space-based self-annotated two-level distributed model training method is provided, applicable to multiple low-Earth orbit (LEO) satellites, multiple medium-Earth orbit (MEO) satellites, and a ground control center. The method includes: each LEO satellite, in response to receiving a model training instruction from a MEO satellite, extracts embedding features from a batch of local data and assigns identification information to each embedding feature; and sends each embedding feature and the assigned identification information to the corresponding MEO satellite. The local data is data collected by the LEO satellite, and the identification information is used to distinguish the embedding features of different local data. Each MEO satellite, based on the embedding features and assigned identification information sent by the LEO satellite, performs a two-level distributed model training based on the... The task requirements sent by the ground control center process the embedded features sent by the low-Earth orbit satellites into corresponding task labels, and then send the task labels and corresponding identification information of each embedded feature to the corresponding low-Earth orbit satellites. The task requirements represent the tasks that the task model on the low-Earth orbit satellites needs to perform. The task labels are used as labels for local data to train the task model. Each low-Earth orbit satellite constructs a training dataset based on the batch of local data, the task labels sent by the corresponding medium- and high-Earth orbit satellites, and the corresponding identification information. The initial model is then trained using this training dataset to obtain the trained task model. The task model is used to execute the tasks indicated by the task requirements.

[0015] According to various aspects of this disclosure, by utilizing the primary processing module deployed on low-Earth orbit satellites with limited computing power and resources to extract the embedded features of local data and send them to medium- and high-Earth orbit satellites, and by utilizing the secondary processing module deployed on medium- and high-Earth orbit satellites with strong computing power, wide communication coverage, and stable inter-satellite links to generate corresponding task labels based on task requirements, it is possible to achieve on-board self-labeling without relying on large-scale local data downstorage to the ground. Moreover, the amount of data with embedded features is smaller than the original local data, making inter-satellite data transmission more efficient. Furthermore, local data can be used for adaptive model training based on task requirements sent by the ground control center to adapt to different task requirements and data environments. In summary, it can fully utilize the computing power, resources, and communication capabilities of satellites in different orbits, decouple data label generation from lightweight model training, and achieve efficient, autonomous, on-board collaborative, and adaptive continuous model training.

[0016] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0017] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0018] Figure 1 The diagram illustrates an application scenario of a space-based self-annotated two-level distributed model training system according to an embodiment of the present disclosure.

[0019] Figure 2 This diagram illustrates a space-based self-labeled two-level distributed model training system according to an embodiment of the present disclosure.

[0020] Figure 3 This diagram illustrates a space-based self-labeled two-level distributed model training system according to an embodiment of the present disclosure.

[0021] Figure 4 This diagram illustrates the training timing of a space-based self-labeled two-level distributed model according to an embodiment of the present disclosure.

[0022] Figure 5 This diagram illustrates a training process for a space-based self-labeled two-level distributed model according to an embodiment of the present disclosure. Detailed Implementation

[0023] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0024] As used herein, the terms “comprising,” “including,” “having,” or variations thereof are open-ended and include one or more of the stated features, integrals, elements, steps, components, or functions, but do not exclude the presence or addition of one or more other features, integrals, elements, steps, components, functions, or groups thereof.

[0025] When an element is referred to as “connected,” “coupled,” “responding,” or a variation thereof relative to another element, it may be directly connected, coupled, or responding to another element, or there may be an intermediate element present.

[0026] Although the terms first, second, third, etc., may be used herein to describe various elements / operations, these elements / operations should not be limited by these terms. These terms are only used to distinguish one element / operation from another. Therefore, without departing from the teachings of the inventive concept, a first element / operation in some embodiments may be referred to as a second element / operation in other embodiments.

[0027] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0028] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0029] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant regions.

[0030] For example, Figure 1 This diagram illustrates an application scenario of a space-based self-labeled two-level distributed model training system according to an embodiment of the present disclosure. Figure 1 As shown, the application scenario includes: multiple low-Earth orbit satellites (L1, L2, L3, L4 in the figure), multiple medium-high orbit satellites (M1, M2 in the figure), and a ground control center G1. One medium-high orbit satellite can correspond to at least two low-Earth orbit satellites (for example, M1 corresponds to L1 and L2, and M2 corresponds to L3 and L4), or it can correspond to at least one low-Earth orbit satellite. Each medium-high orbit satellite has a communication link with the ground control center, each medium-high orbit satellite has a communication link with at least two corresponding low-Earth orbit satellites, and there are also communication links between the various medium-high orbit satellites.

[0031] It should be understood that Figure 1 The application scenario shown is only an exemplary application scenario. This disclosure does not limit the number of medium-low orbit satellites and medium-high orbit satellites used in the system in this disclosure embodiment, nor the number of low orbit satellites corresponding to each medium-high orbit satellite. The medium-high orbit satellites can be medium orbit satellites and / or high orbit satellites, and this disclosure does not limit them.

[0032] Based on the above Figure 1 The application scenarios shown Figure 2 This diagram illustrates a space-based self-labeling two-level distributed model training system according to an embodiment of the present disclosure, as shown below. Figure 2As shown, the training system includes: a primary processing module deployed on each of the multiple low-Earth orbit satellites (L1, L2, L3, L4 in the figure), a secondary processing module deployed on each of the multiple medium-high orbit satellites (M1, M2 in the figure), and a control module deployed at the ground control center G1.

[0033] The primary processing module deployed on each low-Earth orbit satellite is configured to: in response to receiving model training instructions sent by medium- and high-Earth orbit satellites, extract the embedding features of each batch of local data and assign identification information to each embedding feature, and send each embedding feature and the assigned identification information to the corresponding medium- and high-Earth orbit satellites. The local data is the data collected by the low-Earth orbit satellites, and the identification information is used to distinguish the embedding features of different local data.

[0034] The secondary processing module deployed on each medium-high orbit satellite is configured to: based on the task requirements sent by the control module deployed at the ground control center, process the embedded features sent by the low orbit satellite into corresponding task labels, and send the task labels and corresponding identification information corresponding to each embedded feature to the corresponding low orbit satellite. The task requirements represent the tasks that the task model on the low orbit satellite needs to perform, and the task labels are used as labels for local data to train the task model.

[0035] The primary processing module deployed on each low-Earth orbit satellite is also configured to: construct a training dataset based on batches of local data, task labels corresponding to the embedded features issued by medium- and high-Earth orbit satellites, and corresponding identification information; and use this training dataset to train an initial model to obtain a trained task model, which is used to execute the tasks indicated by the task requirements.

[0036] In practical applications, when a ground control center requires various low-Earth orbit (LEO) satellites to perform a certain task (such as classification or target detection), it can send task requirements to various medium- and high-Earth orbit (MEO) satellites through the control module. The MEO satellites can then send model training instructions to the corresponding LEO satellites. These instructions may include the task requirements, instructing the LEO satellites to train a task model corresponding to those requirements. It should be understood that this disclosure does not limit the type of task indicated by the task requirements.

[0037] After receiving model training instructions, the primary processing module in a low-Earth orbit satellite (such as L1, L2, L3, L4) can continuously collect batches of local data during the Earth observation cycle according to mission requirements for subsequent model training. Alternatively, it can use previously collected local data for subsequent model training; this disclosure does not limit this approach. It should be understood that various sensor devices can be installed on low-Earth orbit satellites, and different types of local data can be obtained through different sensor devices. The local data collected by the corresponding sensor devices can be acquired according to mission requirements. For example, the local data collected by low-Earth orbit satellites can be remote sensing data, including visible light, infrared, or synthetic aperture radar images. It can also be other data types, such as meteorological data. This disclosure does not limit the type or acquisition method of local data.

[0038] In some embodiments, a pre-trained embedding feature extractor can be pre-deployed on the low-Earth orbit (LEO) satellite. The primary processing module deployed on the LEO satellite can call this embedding feature extractor to extract the embedding features of each batch of local data. This embedding feature extractor is equivalent to an encoder performing preliminary encoding on the local data to obtain the embedding features. The embedding feature extractor can employ any neural network known in the art for extracting embedding features, such as a convolutional network. This disclosure does not limit the specific methods used. Since embedding features are typically represented as embedding vectors, the embedding feature extractor can map local data into fixed-dimensional embedding vectors. The embedding vectors can preferably be stored or transmitted using a low-precision numerical format, thereby further reducing the amount of data transmitted between satellites compared to directly transmitting local data.

[0039] Considering that the raw local data acquired by low-orbit satellites may contain noise, have low quality, and be large in size (e.g., a single remote sensing image is typically large), in order to facilitate efficient extraction of subsequent embedding features and improve the accuracy of the generated task labels, in some embodiments, the extraction of embedding features from batches of local data may include:

[0040] For any local data in the batch of local data, preprocess the local data to obtain preprocessed local data; use the embedding feature extractor to generate the embedding features of the preprocessed local data based on the preprocessed local data.

[0041] Preprocessing may include, for example, data segmentation and data filtering. Data segmentation can divide the original local data into multiple data blocks. For instance, when the local data is a remote sensing image, the acquired remote sensing image can be segmented into blocks of reasonable sizes such as 256×256, 512×512, or 1024×1024 pixels, while setting a reasonable overlap ratio to ensure the integrity of boundary targets without introducing excessive redundancy. Data filtering can be used to remove local data or data blocks with high noise or low quality. It should be understood that those skilled in the art can use any known noise detection or quality detection technology to perform noise detection or quality detection on local data, and then perform data filtering based on the detection results. This disclosure does not limit this approach. In some embodiments, when the local data is a remote sensing image, preprocessing may also include radiometric correction, geometric correction, scale transformation, cloud and haze removal, etc., to improve data quality without changing the image semantics, thereby improving the accuracy of subsequent task labeling. It should be understood that those skilled in the art can set corresponding preprocessing methods for different types of local data, and this disclosure does not limit such methods.

[0042] It should be understood that when preprocessing includes data segmentation, generating embedding features of the preprocessed local data using an embedding feature extractor may include: inputting each data block into which the local data is segmented to the embedding feature extractor to obtain the embedding features of each data block of the local data; and then assigning identification information to the embedding features of each data block of the local data, so that the embedding features of any local data include the embedding features of each data block of the local data. When preprocessing does not include data segmentation, generating embedding features of the preprocessed local data using an embedding feature extractor may include: inputting the entire local data set to the embedding feature extractor to obtain the embedding features of each local data set; and then assigning identification information to the embedding features of each local data set.

[0043] In some embodiments, the identification information may include, but is not limited to, data indexes, timestamps, geographic locations, imaging parameters, quality indicators, etc., as long as they can identify the embedding features of different local data or different data blocks. This disclosure does not impose any limitations on this aspect. Utilizing the identification information facilitates subsequent label generation and the construction of the training dataset.

[0044] It should be understood that each low-Earth orbit (LEO) satellite can send the embedded features and assigned identification information extracted from its local data to the corresponding medium- and high-Earth orbit (MEO) satellites. Each MEO satellite can receive the embedded features and identification information sent by at least two corresponding LEO satellites. For example, MEO satellite M1 receives embedded feature information (including embedded features and identification information) sent by LEO satellites L1 and L2, and MEO satellite M2 receives embedded feature information sent by LEO satellites L3 and L4.

[0045] In some embodiments, a deep feature extractor and a label generator can be pre-deployed in each medium-to-high orbit satellite to fulfill the task requirements sent by the control module deployed at the ground control center. This involves processing the embedded features sent by the low-orbit satellites into corresponding task labels. Specifically, in some embodiments, the process of processing the embedded features sent by the low-orbit satellites into corresponding task labels, based on the task requirements sent by the control module deployed at the ground control center, includes:

[0046] The deep feature extractor is used to extract the deep features of each embedded feature; based on the task requirements, the corresponding label generator is used to generate the task label and label confidence score for each embedded feature according to the deep features of each embedded feature. The label confidence score is used to characterize the reliability of the task label.

[0047] The above-mentioned task labels and corresponding identification information corresponding to each embedded feature are sent to the corresponding low-Earth orbit satellites, including: sending high-confidence task labels with a label confidence level higher than a specified threshold and the corresponding identification information of the embedded features to the corresponding low-Earth orbit satellites, so as to train the initial model in the low-Earth orbit satellites using the high-confidence task labels.

[0048] The deep feature extractor can employ any known neural network for extracting deep features, such as a backbone network. The backbone network is a base model pre-trained on the ground or trained offline, and does not participate in training during runtime. The secondary processing module deployed on each medium-to-high orbit satellite can call the deep feature extractor to extract the deep features of each embedded feature, which is equivalent to performing deep parsing encoding on each embedded feature to obtain the deep features.

[0049] The label generator acts as a decoder, used to decode depth features to obtain corresponding task labels and label confidence scores. It should be understood that different types of tasks can use different types of label generators; that is, multiple label generators can be deployed in each medium-to-high orbit satellite to generate different types of task labels and label confidence scores. In practical applications, the secondary processing module deployed in each medium-to-high orbit satellite can call the corresponding label generator to generate corresponding task labels and label confidence scores based on the task type indicated by the task requirements. In other words, depth features are input into the corresponding label generator to obtain task labels and label confidence scores, thereby outputting different task labels according to different tasks. For example, if the task is classification, the task labels include, but are not limited to, Top-k categories; if the task is object detection or segmentation, the output labels include, but are not limited to, object bounding box location and category. It should be understood that the task labels corresponding to different tasks are different. In addition to the above-mentioned categories and locations, task labels may also include, but are not limited to, attributes, states, etc., which are not limited in this embodiment.

[0050] It should be understood that the above-mentioned use of deep feature extractors and label generators to obtain task labels for each embedded feature can actually be understood as using a high-precision model with a large number of parameters to perform the corresponding task (such as performing classification, object detection, or segmentation). The results obtained by the high-precision model with a large number of parameters performing the corresponding task are used as task labels for subsequent training of the task model on low-Earth orbit satellites. This is because the computing power and resources of low-Earth orbit satellites are limited, and the task models deployed on low-Earth orbit satellites are usually lightweight models with a small number of parameters. In this way, using the results generated by the high-precision model with a large number of parameters as task labels to train the lightweight model with a small number of parameters can make full use of the computing power and resource capabilities of satellites in different orbits, realize the self-labeling of on-board training data, and improve the training efficiency of models on low-Earth orbit satellites.

[0051] It is known that decoders in this field typically output the confidence level of the decoded result along with the predicted confidence level. The confidence level characterizes the reliability (or trustworthiness) of the decoded result; a higher confidence level indicates a more reliable decoding result. Therefore, the confidence level of the task labels generated by the task label generator can be used to select high-confidence task labels for subsequent task model training, thereby improving the training accuracy of the task model. Specifically, the secondary processing module deployed on medium- and high-orbit satellites can perform quality control on task labels using the label confidence level. When the label confidence level is higher than a specified threshold, the corresponding task label is marked as a high-confidence task label, and the high-confidence task labels with higher confidence levels and their corresponding embedded feature identifiers are sent to the corresponding low-orbit satellites. This allows the initial model to be trained on the low-orbit satellites using the high-confidence task labels, i.e., supervised training of the initial model. For task labels with confidence levels lower than the specified threshold, these task labels can be marked as low-confidence task labels and can be excluded from subsequent supervised training of the initial model.

[0052] In some embodiments, to fully utilize local data collected by low-Earth orbit (LEO) satellites, medium- and high-Earth orbit (MEO) satellites can also send the low-confidence mission labels and their corresponding embedding features to a ground control center. The ground control center then corrects these low-confidence mission labels and uploads the corrected labels to the MEO satellites. The MEO satellites integrate the high-confidence mission labels with the ground-corrected labels and can then also send the corrected mission labels and their corresponding identification information back to the LEO satellites for model training. Therefore, specifically, the secondary processing module deployed on each MEO satellite can also be configured to: send low-confidence mission labels with a confidence level below a specified threshold and their corresponding embedding features to the ground control center.

[0053] The control module deployed in the ground control center can be configured to: correct the low-confidence task labels corresponding to the embedded features issued by medium- and high-orbit satellites, obtain the corrected task labels corresponding to the embedded features, and upload the corrected task labels corresponding to the embedded features to the corresponding medium- and high-orbit satellites;

[0054] The process of sending the task labels and corresponding identification information corresponding to each embedded feature to the corresponding low-Earth orbit satellite also includes sending the correction task labels and corresponding identification information sent by the ground control center to the corresponding low-Earth orbit satellite, so as to train the initial model on the low-Earth orbit satellite using the correction task labels.

[0055] In practical applications, the control module of the ground control center can, for example, use manual methods or high-precision large models to correct the low-confidence task labels corresponding to the embedded features (that is, to confirm or correct them, which is also equivalent to regenerating high-confidence task labels based on the embedded features). This embodiment of the present disclosure does not limit this. Then, the corrected task labels are uploaded to the corresponding medium-high orbit satellite. For example, for the low-confidence task labels issued by medium-high orbit satellite M1, the ground control center G1 corrects them and then transmits the corrected task labels back to medium-high orbit satellite M1.

[0056] It should be understood that the identification information can be used to distinguish embedded features, and naturally it can also be used to distinguish the task labels of embedded features. When medium- and high-orbit satellites transmit embedded features and corresponding low-confidence task labels, they can also transmit the identification information of the embedded features at the same time. When the ground control center transmits the calibration task label, it can also transmit the identification information of the corresponding embedded features at the same time to identify the calibration task label. After receiving the calibration task label, the medium- and high-orbit satellites can determine the embedded features corresponding to the calibration task label based on the identification information, and can transmit the calibration task label and the identification information of the corresponding embedded features to the corresponding low-orbit satellite. For example, for the low-confidence task label in the embedded features transmitted by low-orbit satellite L1, after receiving the calibration task label transmitted by the ground control center, medium- and high-orbit satellite M1 transmits the calibration task label and the corresponding identification information to low-orbit satellite L1.

[0057] After the low-Earth orbit satellites (such as L1, L2, L3, and L4) receive the mission tags (including high-confidence mission tags and / or correction mission tags) and corresponding identification information from the medium- and high-Earth orbit satellites M1 and M2, the primary processing module deployed on the low-Earth orbit satellites can match the mission tags with the batch of local data stored locally on the low-Earth orbit satellites based on the identification information corresponding to the mission tags, in order to construct an independent training dataset for the low-Earth orbit satellites. The training dataset may include at least a portion of the local data in the batch of local data and the mission tag of each local data in the at least a portion of the local data. Then, the primary processing modules deployed on each low-Earth orbit satellite can adopt any known supervised training method in the field to perform local iterative supervised training on the initial model using their own constructed training datasets. The number of local training rounds can be customized. For example, in the nth round of training, the local data in the training dataset can be input into the (n-1)th generation model (when n=1, the 0th generation model is the initial model) to obtain the output result of the (n-1)th generation model. The loss is calculated using the difference between the output result of the (n-1)th generation model and the task label corresponding to the local data. Based on this loss, the model parameters of the (n-1)th generation model are optimized to obtain the nth generation model. This process continues until the preset total number of training rounds is reached.

[0058] It should be understood that since the calibration task label is generated by the ground control center, the time required for the calibration task label to be generated by the ground control center and then transmitted to the medium-high orbit satellite and then sent to the low orbit satellite is relatively long. Therefore, the medium-high orbit satellite can first send the high confidence task label to the low orbit satellite, allowing the low orbit satellite to start model training based on the training dataset constructed by the high confidence task label. After receiving the calibration task label, it can then combine the calibration task label to construct the training dataset and continue training. Alternatively, after the medium-high orbit satellite receives the calibration task label, it can integrate the calibration task label and the high confidence task label and send them to the low orbit satellite together. This disclosure does not limit this.

[0059] In practical applications, the initial model can be a lightweight model pre-stored on a low-Earth orbit (LEO) satellite, possessing certain capabilities to perform corresponding tasks. LEO satellites can pre-store multiple lightweight models with different task capabilities. Upon receiving a model training instruction, the primary processing module deployed on the LEO satellite can retrieve the corresponding lightweight model as the initial module for training, based on the task type indicated in the task requirements of the model training instruction. Alternatively, when the ground control center sends a task requirement to a medium- or high-Earth orbit (MEO) satellite, it can simultaneously send a pre-trained lightweight model capable of performing the corresponding task. MEO satellites can send this pre-trained lightweight model to the corresponding LEO satellite when sending the model training instruction. This lightweight model serves as the initial model, and then supervised training can be performed using the constructed training dataset to obtain the trained task model. This improves the training efficiency of the task model. Furthermore, since the training dataset is constructed from local data of the LEO satellite, the trained task model is more adaptable to the data acquisition environment of each LEO satellite, thus improving the adaptability of the trained task model for performing tasks on LEO satellites.

[0060] To further improve the adaptability and generalization ability of mission models on low-Earth orbit satellites in the face of complex environments and break down data silos between low-Earth orbit satellites, multi-satellite collaborative modeling can be carried out by referring to the existing federated learning implementation method to improve model performance. Therefore, in some embodiments, the primary processing module deployed on each low-Earth orbit satellite can also be configured to send the model parameters of the trained mission model to the corresponding medium- and high-Earth orbit satellites.

[0061] The secondary processing module deployed on each medium-high orbit satellite can also be configured to: aggregate the model parameters sent by at least two low orbit satellites respectively to obtain the global model parameters of the current version, and send the global model parameters of the current version to the corresponding at least two low orbit satellites;

[0062] Each low-Earth orbit satellite's primary processing module can also be configured to: in response to receiving the current version of global model parameters sent by medium- and high-Earth orbit satellites, replace the model parameters of the trained mission model with the current version of global model parameters to obtain the current version of the mission model, so as to use the current version of the mission model to execute the mission indicated by the mission requirements.

[0063] It should be understood that after any low-Earth orbit (LEO) satellite completes its local iterative training of the initial model, the LEO satellite can upload the updated model parameters (i.e., the model parameters of the trained mission model) to the corresponding medium-high orbit (MEO) satellite. The original local data does not need to be transmitted, which helps reduce inter-satellite data transmission costs. Upon receiving the model parameters from the corresponding LEO satellite, the secondary processing module deployed on any MEO satellite can perform parameter aggregation. For example, the secondary processing module deployed on MEO satellite M1 receives model parameters from LEO satellites L1 and L2, aggregates these parameters to obtain the current version of the global model parameters, and sends these current version global model parameters back to LEO satellites L1 and L2. Similarly, the secondary processing module deployed on MEO satellite M2 receives model parameters from LEO satellites L3 and L4, aggregates these parameters to obtain the current version of the global model parameters, and sends these current version global model parameters back to LEO satellites L3 and L4. In this process, any parameter aggregation method known in the art can be used, such as FedAvg (Federated Average Algorithm), weighted average, etc., and this disclosure does not limit the specific methods used.

[0064] After the medium- and high-orbit satellites distribute the global model parameters to the corresponding low-orbit satellites, the primary processing module deployed on the low-orbit satellites can receive the global model parameters and replace the model parameters of the trained task model with the global model parameters. In other words, the global model parameters are used to complete a local task model parameter update to obtain the current version of the task model. Then, the task indicated by the task requirements can be executed using the current version of the task model.

[0065] Considering that medium- and high-orbit satellites may malfunction or malfunction and be unable to continue performing corresponding operations, in some embodiments, the secondary processing module deployed on each medium- and high-orbit satellite can also be configured to: aggregate the model parameters sent by at least two low-orbit satellites to obtain the current version of the global model parameters, and then synchronize the current version of the global model parameters to the remaining medium- and high-orbit satellites. In the event of a failure of any medium- and high-orbit satellite, the non-failed medium- and high-orbit satellites can connect with at least two low-orbit satellites corresponding to the failed medium- and high-orbit satellite to continue performing the functions required by the failed medium- and high-orbit satellite. That is, the non-failed medium- and high-orbit satellites take over the functions of the failed medium- and high-orbit satellites, which helps to improve the robustness of the entire system.

[0066] Among them, the remaining medium-high orbit (MEO) satellites are those other than the one in the MEO orbit itself. For example, MEO satellites M1 and M2 can synchronize and aggregate their global model parameters after generating them. That is, M1 synchronizes its locally generated global model parameters to M2, and M2 synchronizes its locally generated global model parameters to M1. This is equivalent to backing up the global model parameters generated by any MEO satellite on other MEO satellites. In this way, if MEO satellite M1 fails, MEO satellite M2 can take over from low-Earth orbit (LEO) satellites L1 and L2 to continue performing subsequent operations. Simultaneously, while MEO satellites M1 and M2 are synchronizing and aggregating their global model parameters, they can also distribute the global model parameters to their respective LEO satellites L1, L2, L3, and L4. LEO satellites L1, L2, L3, and L4 receive the global model parameters and replace them, completing one model parameter update.

[0067] Considering that the global model parameters obtained through local iterative training on low-Earth orbit satellites and aggregation of model parameters from medium- and high-Earth orbit satellites may not necessarily meet the specified mission accuracy requirements, the training effect of the on-board model can be monitored by the control module deployed at the ground control center to ensure that the mission model performing the mission on the low-Earth orbit satellites meets the mission accuracy requirements. Therefore, in some embodiments, the secondary processing module deployed on each medium- and high-Earth orbit satellite can also be configured to: send the current version of the global model parameters to the ground control center.

[0068] The control module deployed in the ground control center can also be configured to: perform inference accuracy tests on the current version of the global model parameters issued by the medium- and high-orbit satellites; if the inference accuracy of the current version of the global model parameters does not meet the specified mission requirements accuracy, send a training instruction to the corresponding medium- and high-orbit satellites.

[0069] The secondary processing module deployed on each medium-high orbit satellite can also be configured to forward the continued training instructions sent by the ground control center to at least two corresponding low orbit satellites;

[0070] Each low-Earth orbit satellite's primary processing module can also be configured to: in response to receiving a continued training instruction, use the current version of the mission model as the initial model to retrain the initial model using the training dataset, thereby obtaining the trained mission model, until the inference accuracy of the global model parameters obtained by aggregating the model parameters of the trained mission model reaches the mission-required accuracy.

[0071] For example, while the medium-high orbit satellite M1 distributes the global model parameters to the low-Earth orbit satellites L1 and L2, it can also distribute the global model parameters to the ground control center G1. The control module of the ground control center G1 then performs inference accuracy testing on the global model parameters distributed by the medium-high orbit satellite M1. In practical applications, the control module deployed at the ground control center can use existing local datasets to test the inference accuracy of the global model parameters, that is, to test the inference accuracy of the global model with global model parameters (which has the same model structure as the initial model and the mission model mentioned above). It should be understood that those skilled in the art can use any known model testing technology in the art to test the model inference accuracy, and this disclosure does not limit this.

[0072] Furthermore, when the inference accuracy of the current version of the global model parameters does not meet the specified mission accuracy requirements, the control module deployed at the ground control center can send a continued training instruction to the corresponding medium-high orbit satellites. The medium-high orbit satellites then forward the continued training instruction to at least two corresponding low-Earth orbit satellites. Upon receiving the continued training instruction, the primary processing module deployed on each low-Earth orbit satellite can use the current version of the mission model as the initial model to re-train the initial model using the training dataset, obtaining the trained mission model. That is, it uses the training dataset to perform local iterative training on the current version of the mission model to obtain the re-trained mission model. Then, it re-sends the model parameters of the trained mission model to the medium-high orbit satellites. The secondary processing module deployed on the medium-high orbit satellites aggregates the model parameters to obtain the global model parameters, which are simultaneously distributed to the low-Earth orbit satellites and the ground control center. The control module deployed at the ground control center then tests the inference accuracy of the global model parameters again. This process is repeated until the inference accuracy of the global model parameters obtained by aggregating the model parameters of the trained mission model reaches the mission accuracy requirements. At this point, model training stops, and the mission model with global model parameters that have reached the mission accuracy requirements is used to execute the mission as instructed by the mission requirements.

[0073] That is, the control module deployed in the ground control center can also be configured to send a stop training command to the corresponding medium and high orbit satellites when the inference accuracy of the global model parameters in the current version has reached the accuracy required for the mission.

[0074] The secondary processing module deployed on each medium-high orbit satellite can also be configured to forward the stop training command sent by the ground control center to at least two corresponding low orbit satellites;

[0075] The primary processing module deployed on each low-Earth orbit satellite can also be configured to: in response to receiving a stop training command, stop model training and use the current version of the mission model to perform the tasks indicated by the mission requirements.

[0076] Considering that during the process of iterative local training of low-Earth orbit satellites, aggregation of model parameters for medium- and high-Earth orbit satellites, and testing of model accuracy at the ground control center, the inference accuracy of the global model parameters may remain unchanged or increase, or it may decrease. Therefore, the above-mentioned sending a continued training instruction to the corresponding medium- and high-Earth orbit satellite when the inference accuracy of the current version of the global model parameters does not meet the specified mission requirement accuracy can include: sending a continued training instruction to the corresponding medium- and high-Earth orbit satellite when the inference accuracy of the current version of the global model parameters does not meet the specified mission requirement accuracy and the inference accuracy of the current version of the global model parameters is the same as or improved compared to the inference accuracy of the historical version of the global model parameters; the historical version of the global model parameters are the global model parameters generated before the current version of the global model.

[0077] The control module deployed in the ground control center can also be configured to send a rollback command to the corresponding medium- and high-orbit satellites when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirements and the inference accuracy of the global model parameters in the current version is lower than that of the global model parameters in the historical version.

[0078] The secondary processing module deployed on each medium-high orbit satellite can also be configured to forward the rollback command sent by the ground control center to at least two corresponding low orbit satellites;

[0079] The primary processing module deployed on each low-Earth orbit satellite can also be configured to: in response to receiving a rollback command, restore the current version of the mission model to the previous version of the mission model and retrain it using the previous version of the mission model as the initial model; or, stop model training and use the previous version of the mission model to execute the mission as indicated by the mission requirements; the previous version of the mission model is the mission model using the global model parameters before the inference accuracy degradation.

[0080] The above process can be understood as follows: at the ground control center, the inference accuracy of the current version of the global model parameters is tested using a local dataset. If the inference accuracy of the current version of the global model parameters remains stable or improves compared to historical versions, a continued training instruction is sent to the corresponding medium-high orbit satellites to repeat the process of local iterative training on low-Earth orbit satellites, aggregation of model parameters on medium-high orbit satellites, and testing of inference accuracy at the ground control center, until the inference accuracy reaches the mission requirements. If the inference accuracy of the current version of the global model parameters continues to decline compared to historical versions, the ground control center sends a rollback instruction, instructing the medium-high orbit satellites to instruct the primary processing modules deployed on the corresponding low-Earth orbit satellites to restore the current version of the mission model to the global model parameters of the previous version, i.e., restore... The task model uses global model parameters before the inference accuracy decreases. Then, by adjusting hyperparameters such as the learning rate, the previous version of the task model can be retrained (equivalent to using the previous version as the initial model and performing local iterative training using the training dataset). The subsequent process of aggregating model parameters for medium and high-orbit satellites and testing inference accuracy at the ground control center is then executed again until the inference accuracy reaches the specified task requirement or shows an improvement. Alternatively, the primary processing module deployed on low-orbit satellites can directly stop training, allowing the previous version of the task model to execute the tasks specified in the task requirements. Although in this case, the previous version of the task model may not have reached the required accuracy, it may have already achieved the highest possible accuracy. It should be understood that after the ground control center sends a rollback command, the low-orbit satellite can be customized to either restore the previous version of the global model parameters for retraining or stop training. Alternatively, the rollback command may include instructions to modify hyperparameters for retraining or to stop training, allowing for corresponding processing according to the instructions in the rollback command. This disclosure does not limit this aspect.

[0081] It should be understood that the aforementioned process of local iterative training for low-Earth orbit (LEO) satellites, aggregation of model parameters for medium- and high-Earth orbit (MEO) satellites, and testing of inference accuracy by the ground control center can be repeated cyclically until the model accuracy of the global module parameters reaches the mission requirements or a stop-training command is sent from the ground, resulting in the final trained mission model. The final trained mission model can then be used to perform the tasks specified in the mission requirements for the next observation cycle. After the ground control center sends new mission requirements, the aforementioned system can be used to perform processes such as local data acquisition, tag generation for MEO satellites, local iterative training for LEO satellites, aggregation of model parameters for MEO satellites, and testing of inference accuracy by the ground control center.

[0082] According to the training system of this disclosure, by utilizing the primary processing module deployed on a low-Earth orbit satellite with limited computing power and resources to extract the embedding features of local data and send them to a medium- and high-Earth orbit satellite, and by utilizing the secondary processing module deployed on a medium- and high-Earth orbit satellite with strong computing power, wide communication coverage, and stable inter-satellite links to generate corresponding task labels based on task requirements, it is possible to achieve on-board self-labeling without relying on large-scale local data downlink to the ground. Moreover, the amount of data with embedded features is smaller than the original local data, making inter-satellite data transmission more efficient. Furthermore, it can use local data for adaptive model training based on task requirements sent by the ground control center to adapt to different task requirements and data environments. In summary, it can fully utilize the computing power, resources, and communication capabilities of satellites in different orbits, decouple data label generation from lightweight model training, and achieve efficient, autonomous, on-board collaborative, and adaptive continuous model training.

[0083] In the training system of this embodiment, due to the limited computing power, power consumption, and storage resources of low-Earth orbit (LEO) satellites, they are mainly responsible for Earth observation, generating local data, and performing data preprocessing. They also deploy lightweight embedding feature extractors to generate embedding features and perform model training. Medium- and high-Earth orbit (MEO) satellites, with relatively abundant computing resources, undertake the deployment of deep feature extractors with a large number of parameters, parsing the embedding features of LEO satellites, generating task labels, and completing model aggregation. The ground control center is responsible for label correction and model evaluation. This eliminates the need to download large amounts of raw local data to the ground for model training and updates. LEO satellites only need to train their local lightweight models, while MEO satellites perform label generation and parameter aggregation, achieving on-board model training and updates while meeting the actual on-board computing power requirements.

[0084] Firstly, the training system of this embodiment deploys a deep feature extractor with a large parameter and computational scale on medium- and high-orbit satellites with strong computing resources, wide communication coverage, and stable inter-satellite links. This extractor is used to perform high-level semantic parsing on the embedding vectors uploaded by low-orbit satellites and generate labels. On low-orbit satellites with limited computing power and power consumption, only a smaller embedding feature extractor with a smaller parameter and computational scale and a lightweight trainable model for the task are deployed. The model is trained locally using the task labels generated by the medium- and high-orbit satellites. The trained model parameters are then aggregated centrally or in a distributed manner, thereby enabling collaborative adaptation of models from multiple low-orbit satellites in different observation areas, under different observation conditions, and with different observation objects.

[0085] Based on the training system of the above-described embodiments of this disclosure, exemplarily, Figure 3 This diagram illustrates a space-based self-labeling two-level distributed model training system according to an embodiment of the present disclosure, as shown below. Figure 3As shown, the training system includes: a first-level local training layer (i.e., a first-level processing module), a second-level space-based annotation and parameter aggregation layer (i.e., a second-level processing module), and a ground-based training and annotation supervision layer (i.e., a control module).

[0086] The first-level local training layer is deployed on multiple low-Earth orbit satellites to preprocess local data and generate embedded features. It constructs a local training dataset based on the mission labels issued by medium- and high-Earth orbit satellites and trains the initial model locally.

[0087] The secondary space-based annotation and parameter aggregation layer is deployed on multiple medium- and high-orbit satellites. It is used to perform deep feature extraction and task label generation on the embedded features sent by low-orbit satellites, receive and aggregate model parameters from multiple low-orbit satellites, and synchronize the aggregated global model parameters among multiple medium- and high-orbit satellites and send them down to low-orbit satellites.

[0088] The ground training and labeling supervision layer, deployed at the ground control center, is used to correct low-confidence task labels, optimize the quality of task label generation, evaluate and version control the global model parameters generated by the aggregation of medium and high orbit satellites, and does not participate in the two-level distributed deployment and training of the model.

[0089] Specifically, the first-level local training layer deployed in each low-Earth orbit satellite includes the following sub-modules:

[0090] The preprocessing module is used to preprocess the local data acquired by low-orbit satellites to obtain data that can be input into the feature generation module.

[0091] The embedded feature generation module is used to input preprocessed local data into the embedded feature extractor to generate embedded features, and to attach identification information to the embedded features. The model parameters of the embedded feature extractor are not updated during system training.

[0092] The training module is used to build a local training dataset based on the mission labels and local data issued by medium and high orbit satellites, and to perform local iterative training on the mission model.

[0093] Specifically, the second-level space-based annotation and parameter aggregation layer deployed for each medium-to-high orbit satellite includes the following sub-modules:

[0094] The deep feature extraction module is used to input the embedded features sent by low-orbit satellites into the deep feature extractor (such as the backbone network) to generate corresponding deep features. The backbone network is not updated during system training.

[0095] The label generation module is used to input deep features into the label generator to generate task labels and label confidence scores, including but not limited to category, location, attribute, state, etc.

[0096] The parameter aggregation module is used to receive model parameters from multiple low-Earth orbit satellites and aggregate them to generate global model parameters.

[0097] Specifically, the training and annotation supervision layer deployed on the ground in the ground control center includes the following sub-modules:

[0098] The model monitoring and rollback module is used to test the inference accuracy of the global model parameters issued by medium and high orbit satellites using ground datasets, and to send instructions to continue training, stop training, or rollback based on the test results.

[0099] The label correction module is used to quickly correct low-confidence samples (i.e., low-confidence task labels and corresponding embedded features) sent by medium- and high-orbit satellites using ground-based manual methods or high-precision large models, and then send the corrected task labels back to the medium- and high-orbit satellites.

[0100] In summary, the primary local training layer is deployed on multiple low-Earth orbit (LEO) satellites. Each LEO satellite's primary local training layer includes at least three sub-modules: a preprocessing module, an embedding feature generation module, and a training module. In some embodiments, each LEO satellite may also include a data acquisition module. The data acquisition module collects local data, and the preprocessing module performs preprocessing operations such as segmentation and filtering on the raw local data to obtain local data that can be processed by the embedding feature generation module. The embedding feature generation module receives the local data and extracts embedding features; the training module uses the task labels transmitted from medium- and high-Earth orbit (MEO) satellites to construct a local dataset and performs local iterative training on the model to obtain local model parameters. By deploying a lightweight embedding feature generation module and a lightweight model corresponding to the task on LEO satellites, adaptation of datasets from multiple satellites can be achieved without sharing the original local data.

[0101] The secondary space-based annotation and parameter aggregation layer is deployed on multiple medium-to-high orbit satellites. Each medium-to-high orbit satellite's secondary space-based annotation and parameter aggregation layer includes at least three sub-modules: a depth feature extraction module, a label generation module, and a parameter aggregation module. The depth feature extraction module receives embedded features uploaded from low-orbit satellites and generates depth features. The label generation module receives depth features from local data and outputs the corresponding task labels for the local data. The parameter aggregation module aggregates the local model parameters trained on the low-orbit satellites. After each round of parameter aggregation, multiple medium-to-high orbit satellites will synchronize global model parameters and distribute the global model parameters to the low-orbit satellites and the ground control center.

[0102] The ground training and annotation supervision layer is deployed at the ground control center. This layer comprises at least two sub-modules: a model monitoring and rollback module and a label correction module. The ground control center primarily oversees and manages the process, not participating in the specific details of the two-level distributed model training between low-Earth orbit (LEO) and medium-to-high Earth orbit (MEO) satellites. The model monitoring and rollback module uses existing local datasets on the ground to test the global model parameters transmitted from MEO satellites. If the model accuracy remains stable or increases during multiple training rounds, a training continuation command is sent; otherwise, training is stopped, the model is rolled back to the previous version, or hyperparameters are modified and retrained. The label correction module ensures the accuracy of task labels. Task labels with confidence levels below a specified threshold, along with their corresponding embedded data, are re-labeled manually or using a high-precision large model to generate corrected task labels, which are then transmitted back to the MEO satellites for training.

[0103] In the aforementioned training system, by deploying a primary local training layer on low-Earth orbit satellites and a secondary space-based annotation and parameter aggregation layer on medium- and high-Earth orbit satellites, the ground control center can manage and control the on-board model training and self-annotation process. This allows full utilization of the computing power, resources, and communication capabilities of satellites in different orbits, decoupling task label generation from lightweight model training, and enabling on-board continuous model training for space-based autonomous data annotation and distributed multi-scenario adaptive tasks.

[0104] In the aforementioned training system, the label generator in the label generation module is a functional module that can connect to the output of the backbone network. It is used to map deep features to task labels and label confidence scores, adaptable to specific downstream tasks such as classification, detection, or segmentation. Labels are supervisory information generated by medium- and high-orbit satellites or ground control centers. Task labels include label content (such as category, location, etc.), and label confidence scores characterize the credibility (reliability) of the task labels and determine whether ground correction is triggered. Taking image classification as an example, the output label is the image category and the confidence score for that category; taking object detection as an example, the output label includes the object bounding box location, the object bounding box category, and the confidence score for that object bounding box category. Embedded features refer to low-dimensional feature vectors extracted from local data by the embedded feature extractor in the embedded feature generation module. They are used to represent the semantic information of local data without transmitting the original local data. The deep feature extractor in the deep feature extraction module is a pre-trained backbone network used to provide general deep feature parsing capabilities and does not participate in the model training process. The system comprises multiple medium-to-high orbit (MEO) satellites connected by inter-satellite links, forming a logically centralized server cluster. Label generation and model aggregation tasks are synchronously replicated and migrated within this network, ensuring that if any MEO satellite node fails, other satellites can take over the corresponding functions. A ground-based training and annotation supervision layer is used to improve label accuracy and monitor model training performance. This two-tiered distributed deployment and training framework based on satellite annotation avoids the limitations of traditional centralized ground-based training due to limited satellite-to-ground links and solves the problem of scarce on-satellite training data. Furthermore, by aggregating model parameters from multiple low-Earth orbit (LEO) satellites, the system enhances the model's generalization and adaptability in complex scenarios.

[0105] Based on the above training system, taking local data as remote sensing images as an example... Figure 4 This diagram illustrates a training timeline of a space-based self-labeled two-level distributed model according to an embodiment of the present disclosure, as shown below. Figure 4As shown, the training sequence includes: Low-Earth orbit (LEO) satellites perform image acquisition and switching preprocessing; image patches are input to the embedding feature generation module to obtain embedding vectors and image information (i.e., image patch identification information) and transmitted to the medium-Earth orbit (MEO) satellite; the MEO satellite inputs the embedding vectors to the depth feature extraction module to obtain depth features; the depth features are input to the label generation model to obtain task labels and label confidence levels; high-confidence labels corresponding to image patches are sent to the LEO satellite, while low-confidence labels and their corresponding embedding vectors are sent to the ground control center; the LEO satellite uses the labels to train the local model, obtaining local model parameters, which are then sent to the MEO satellite; the ground control center confirms or corrects the low-confidence labels and their corresponding embedding vectors. The generated high-confidence labels (i.e., ground calibration labels) are transmitted back to the medium- and high-orbit satellites. The medium-orbit satellites forward the ground calibration labels to the low-orbit satellites to participate in model training. The medium-orbit satellites receive the local model parameters sent by the low-orbit satellites and perform model parameter aggregation and synchronization. The aggregated global model parameters are simultaneously sent to the low-orbit satellites and the ground control center. The low-orbit satellites perform local model updates and retraining. After obtaining the local model parameters, they are sent back to the medium-orbit satellites for model parameter aggregation and synchronization. The ground control center performs model monitoring (i.e., inference accuracy testing) and rollback based on the global model parameters discovered by the medium-orbit satellites. Based on the test results, it sends instructions to the medium- and high-orbit satellites to continue training, stop training, or modify hyperparameters and retrain.

[0106] Based on the above training system, taking local data as remote sensing images as an example... Figure 5 This diagram illustrates a training process for a space-based self-labeled two-level distributed model according to an embodiment of the present disclosure, as shown below. Figure 5 As shown, the training process includes:

[0107] Low-Earth orbit satellites L1, L2, L3, and L4 acquire remote sensing images, which are then segmented and preprocessed.

[0108] Low-Earth orbit satellites L1, L2, L3, and L4 input preprocessed image patches into the embedding feature generation module to obtain the embedding vectors corresponding to the image patches;

[0109] Medium- and high-orbit satellites M1 and M2 receive the embedding vectors from low-orbit satellites L1 and L2 and L3 and L4, respectively. The embedding vectors are then input into the depth feature extraction module for feature parsing to obtain the corresponding depth features.

[0110] Medium and high orbit satellites M1 and M2 input depth features into the label generation module to obtain high-confidence labels, which are then transmitted to low orbit satellites L1 and L2 and L3 and L4, respectively. Low-confidence labels are transmitted to the ground control center G1.

[0111] Ground control center G1 receives images with low-confidence tags, annotates them using manual / high-precision large models, and sends them back to medium-high orbit satellites M1 and M2. Medium-high orbit satellites M1 and M2 save the tag correction results and resend the tags to the corresponding low-orbit satellites.

[0112] Low-Earth orbit satellites L1, L2, L3, and L4 receive mission labels and match them with image patches stored on the satellites to construct a training dataset corresponding to the low-Earth orbit satellites.

[0113] Without starting training, the low-Earth orbit satellites L1, L2, L3, and L4 continue to acquire remote sensing images and perform image segmentation and preprocessing.

[0114] When training begins, the low-Earth orbit satellites L1, L2, L3, and L4 select the corresponding initial models according to the mission requirements and use the dataset constructed with labels for local training until the local training is completed.

[0115] After local training is completed, medium- and high-orbit satellites M1 and M2 receive model parameters from low-orbit satellites L1, L2, L3, and L4, respectively, and perform parameter aggregation.

[0116] Once the parameters have been aggregated, the medium- and high-orbit satellites M1 and M2 synchronize the global model parameters with each other and send the global model parameters to the low-orbit satellites L1, L2, L3, and L4 and the ground control center G1. The ground control center G1 performs model testing (i.e., model accuracy testing) and issues instructions for model training / stopping / modifying hyperparameters based on the test results of the global model parameters.

[0117] If the ground control center G1 indicates that training should not be terminated, continue to use the data constructed by the labels for local training, then perform parameter aggregation on medium and high orbit satellites after local training, and finally conduct model testing at the ground control center.

[0118] With the ground control center G1 instructing the end of training, the model training is complete. Low-Earth orbit satellites L1, L2, L3, and L4 continue to collect data and use the trained mission model to execute the mission.

[0119] Based on the above training system, this disclosure also provides a method for deploying and training a space-based self-labeled two-level distributed model, including:

[0120] Step 1: The low-orbit satellite acquires the local raw data, and then segments and preprocesses the local raw data to obtain the corresponding processable data blocks;

[0121] Step 2: The low-Earth orbit satellite inputs the data block into the embedding feature generation module to generate the corresponding embedding vector, and at the same time attaches data information and uploads it to the medium- and high-Earth orbit satellite;

[0122] Step 3: The medium- and high-orbit satellites input the embedded vectors into the deep feature extraction module for feature parsing, and obtain the labels corresponding to the embedded vectors through the label generation module; for labels with low confidence and corresponding data, the medium- and high-orbit satellites send them to the ground control center, and regenerate the labels using the label correction module; the remaining high-confidence labels are directly sent to the low-orbit satellites.

[0123] Step 4: The low-Earth orbit satellite constructs a training dataset based on local data blocks and labels, performs local training on the mission model, and obtains the model parameters.

[0124] Step 5: Medium- and high-orbit satellites receive model parameters from multiple low-orbit satellites, aggregate the model parameters, and generate a global model that is synchronized among multiple medium- and high-orbit satellites and then distributed to low-orbit satellites.

[0125] Step Six: The ground control center periodically acquires the global model after aggregating medium and high orbit satellites, tests it against the existing ground validation set, and performs version control on the model based on the test results;

[0126] Step 7: Repeat steps 4, 5 and 6, using low-Earth orbit satellites for local training and medium- and high-Earth orbit satellite model parameter aggregation, until the model reaches the accuracy required for the mission or the ground sends a command to stop training; if the model training is complete, proceed to the next cycle and continue executing steps 1 to 7.

[0127] Based on the above training system, taking local data as remote sensing images as an example, this disclosure also provides a method for deploying and training a space-based self-annotated two-level distributed model, including:

[0128] Step one: Low-Earth orbit satellites L1, L2, L3, and L4 continuously generate remote sensing images during their Earth observation cycle, according to mission requirements. These images may include visible light, infrared, or synthetic aperture radar (SAR) images. Next, the acquired remote sensing images are segmented into blocks of reasonable sizes, such as 256×256, 512×512, or 1024×1024 pixels, with a suitable overlap ratio set to ensure the integrity of boundary targets without introducing excessive redundancy. Then, on-board preprocessing operations are performed on each image block. These preprocessing operations include, but are not limited to, radiometric correction, geometric correction, scale transformation, and cloud / haze removal, improving the accuracy of model training and annotation without altering the image semantics.

[0129] Step two: Low-Earth orbit satellites L1, L2, L3, and L4 input preprocessed image patches into the embedding feature generation module to generate corresponding embedding vectors. The embedding layer can be a convolutional module, used to map image patches into fixed-dimensional embedding vectors. These embedding vectors are preferentially stored or transmitted using low-precision numerical formats to further reduce the amount of data transmitted between satellites. Simultaneously, image data information, including but not limited to: image patch index, timestamp, geographical location, imaging parameters, and quality indicators, is added to the embedding vector of each image patch to facilitate subsequent label selection.

[0130] Step 3: Medium-high orbit satellite M1 receives embedding vector information from low-Earth orbit satellites L1 and L2, and medium-high orbit satellite M2 receives embedding vector information from low-Earth orbit satellites L3 and L4. These embedding vectors are then input into the deep feature extraction module for feature parsing to obtain the corresponding deep feature representations. The deep feature extraction module can utilize a frozen backbone network, which is a base model obtained through ground pre-training or offline training and does not participate in training during runtime.

[0131] Step four: Medium- and high-orbit satellites M1 and M2 input depth features into the label generation module to obtain labels. The labels output differ depending on the task. If the task is classification, the output labels include, but are not limited to, Top-k categories and corresponding confidence scores. If the task is detection / segmentation, the output labels include, but are not limited to, bounding box positions, categories, and confidence scores. Medium- and high-orbit satellites M1 and M2 perform quality control on the labels using confidence scores. When the confidence score is below a set threshold, the label is marked as low confidence and is not used as a supervised training label. Simultaneously, the image embedding features and corresponding labels are sent to the ground. When the confidence score is above the set threshold, the label is marked as high confidence and used for supervised training. The ground control center G1 uses manual verification or a high-precision large model to confirm or correct the labels, and uploads the corrected labels to medium- and high-orbit satellites M1 and M2. Medium- and high-orbit satellites M1 and M2 integrate the high-confidence labels with the ground-corrected labels and send the image patch index and corresponding labels to low-orbit satellites L1, L2, L3, and L4.

[0132] Step 5: Low-Earth orbit (LEO) satellites L1, L2, L3, and L4 receive tags from medium- and high-Earth orbit (MEO) satellites M1 and M2, match them with image patches stored on the satellites, and construct an independent training dataset for the LEO satellites. LEO satellites L1, L2, L3, and L4 use this training dataset for local iterative training; the number of local training rounds is adjustable. After local training is complete, LEO satellites L1, L2, L3, and L4 upload the updated model parameters to MEO satellites M1 and M2; the original image data is not transmitted.

[0133] Step Six: Medium-to-high Earth orbit (MEO) satellites M1 and M2 receive model parameters from low-Earth orbit (LEO) satellites L1, L2, L3, and L4, and aggregate these parameters to obtain a global model. The aggregation method can be FedAvg, weighted average, etc., and is not limited here. MEO satellites M1 and M2 synchronize the aggregated global model and distribute the model parameters to the LEO satellites L1, L2, L3, and L4 participating in the training. LEO satellites L1, L2, L3, and L4 receive the model parameters and replace them, completing one global model update. Simultaneously with the distribution of model parameters to LEO satellites L1, L2, L3, and L4, MEO satellites M1 and M2 distribute the global model parameters to the ground control center G1, where the model monitoring and rollback module is used to perform inference testing on the model. The model inference accuracy is tested on the existing local dataset at the ground control center. If the inference accuracy of different versions of the model remains stable or improves during training, a training continuation command is sent to the medium-high orbit satellites M1 and M2. If the inference accuracy continues to decline, a rollback command is sent from the ground to restore the model parameters of the previous version to the medium-high orbit satellites M1 and M2, and adjust hyperparameters such as the learning rate for retraining or stopping training.

[0134] Step 7: Repeat steps 5 and 6 to perform local training of low-Earth orbit (LEO) satellites and aggregation of medium- and high-Earth orbit (MEO) satellite models until the required model accuracy for the mission is achieved or a ground-based command is sent to stop training. If model training is complete, proceed to the next observation cycle and repeat steps 1 to 7 to continue the process of image acquisition, image annotation, local training of LEO satellites, and aggregation of MEO satellite models.

[0135] This disclosure also provides a space-based self-annotated two-level distributed model training method, which is applied to multiple low-Earth orbit satellites, multiple medium-high orbit satellites, and a ground control center. The training method includes:

[0136] Step S11: In response to receiving the model training instruction sent by the medium-high orbit satellite, each low-Earth orbit satellite extracts the embedding features of each batch of local data and assigns identification information to each embedding feature. The embedding features and the assigned identification information are then sent to the corresponding medium-high orbit satellite. The local data is the data collected by the low-Earth orbit satellite, and the identification information is used to distinguish the embedding features of different local data.

[0137] Step S12: For each medium-high orbit satellite, based on the task requirements sent by the ground control center, each embedded feature sent by the low orbit satellite is processed into a corresponding task label for each embedded feature and the corresponding identification information sent by the low orbit satellite. The task label and the corresponding identification information are then sent to the corresponding low orbit satellite. The task requirements represent the tasks that the task model on the low orbit satellite needs to perform. The task labels are used as labels for training the task model in local data.

[0138] Step S13: Each low-Earth orbit satellite constructs a training dataset based on the batch of local data, the corresponding mission tags issued by the medium- and high-Earth orbit satellites, and the corresponding identification information. The initial model is then trained using this training dataset to obtain the trained mission model. The mission model is used to execute the mission indicated by the mission requirements.

[0139] In some embodiments, processing the embedded features sent by the low-Earth orbit satellite into corresponding task labels based on the task requirements sent by the control module deployed at the ground control center includes: extracting the depth features of each embedded feature using a depth feature extractor; generating task labels and label confidence scores for each embedded feature based on the task requirements using a corresponding label generator according to the depth features of each embedded feature, wherein the label confidence score is used to characterize the reliability of the task label; wherein, sending the task labels and corresponding identification information corresponding to each embedded feature to the corresponding low-Earth orbit satellite includes: sending high-confidence task labels with label confidence scores higher than a specified threshold and the corresponding identification information of the embedded features to the corresponding low-Earth orbit satellite, so as to train an initial model in the low-Earth orbit satellite using the high-confidence task labels.

[0140] In some embodiments, the method further includes: each medium-high orbit satellite sending low-confidence task labels with a label confidence level lower than the specified threshold and corresponding embedding features to the ground control center; the ground control center deploys a system to correct the low-confidence task labels corresponding to the embedding features sent by the medium-high orbit satellites, obtaining corrected task labels corresponding to the embedding features, and uploading the corrected task labels corresponding to the embedding features to the corresponding medium-high orbit satellites; wherein, sending the task labels corresponding to each embedding feature and the corresponding identification information to the corresponding low-orbit satellites further includes: sending the corrected task labels and the corresponding identification information sent by the ground control center to the corresponding low-orbit satellites, so as to train an initial model on the low-orbit satellites using the corrected task labels.

[0141] In some embodiments, the method further includes: each low-Earth orbit (LEO) satellite sending the model parameters of the trained mission model to the corresponding medium-high orbit (MEO) satellite; each MEO satellite aggregating the model parameters sent by at least two LEO satellites to obtain the global model parameters of the current version, and distributing the global model parameters of the current version to the corresponding at least two LEO satellites; each LEO satellite, in response to receiving the global model parameters of the current version sent by the MEO satellite, replacing the model parameters of the trained mission model with the global model parameters of the current version to obtain the current version mission model, so as to execute the mission indicated by the mission requirements using the current version mission model.

[0142] In some embodiments, the method further includes: each medium-high orbit satellite sending the current version of the global model parameters to the ground control center; the ground control center deploying a test to infer accuracy of the current version of the global model parameters sent by the medium-high orbit satellite; if the inference accuracy of the current version of the global model parameters does not meet the specified mission requirement accuracy, sending a continued training instruction to the corresponding medium-high orbit satellite; each medium-high orbit satellite forwarding the continued training instruction sent by the ground control center to at least two corresponding low-Earth orbit satellites; each low-Earth orbit satellite, in response to receiving the continued training instruction, using the current version of the mission model as the initial model, to re-execute training of the initial model using the training dataset to obtain the trained mission model, until the inference accuracy of the global model parameters obtained by aggregating the model parameters of the trained mission model reaches the mission requirement accuracy.

[0143] In some embodiments, the method further includes: the ground control center sending a stop training command to the corresponding medium-high orbit satellites when the inference accuracy of the global model parameters in the current version has reached the accuracy required by the mission; each medium-high orbit satellite forwarding the stop training command sent by the ground control center to at least two corresponding low orbit satellites; each low orbit satellite, in response to receiving the stop training command, stopping model training and using the current version of the mission model to perform the mission indicated by the mission requirements.

[0144] In some embodiments, sending a continued training instruction to the corresponding medium-high orbit satellite when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy includes: sending a continued training instruction to the corresponding medium-high orbit satellite when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy and the inference accuracy of the global model parameters in the current version is the same as or improved compared to the global model parameters in the historical version; the global model parameters in the historical version are the global model parameters generated before the global model in the current version is referenced; the method further includes: when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirement accuracy and the inference accuracy of the global model parameters in the current version decreases compared to the global model parameters in the historical version, the ground control center sends a rollback instruction to the corresponding medium-high orbit satellite; each medium-high orbit satellite forwards the rollback instruction sent by the ground control center to at least two corresponding low-Earth orbit satellites; each low-Earth orbit satellite, in response to receiving the rollback instruction, restores the current version mission model to the previous version mission model and retrains the previous version mission model as the initial model, or stops model training and uses the previous version mission model to execute the task indicated by the mission requirement; the previous version mission model is the mission model using the global model parameters before the decrease in inference accuracy.

[0145] In some embodiments, the method further includes: after each medium-high orbit satellite aggregates the model parameters sent by at least two low-orbit satellites to obtain the current version of the global model parameters, it synchronizes the current version of the global model parameters to the remaining medium-high orbit satellites, so that when any medium-high orbit satellite fails, the non-failed medium-high orbit satellites can dock with at least two low-orbit satellites corresponding to the failed medium-high orbit satellite to continue to perform the functions required by the failed medium-high orbit satellite.

[0146] In some embodiments, the step of extracting the embedding features of each local data in a batch includes: preprocessing any local data in the batch to obtain preprocessed local data; and using an embedding feature extractor to generate embedding features of the preprocessed local data based on the preprocessed local data.

[0147] The specific implementation of the steps in the training method provided in this embodiment can be referred to the description of the system embodiment above, and will not be repeated here for the sake of brevity.

[0148] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation that may be implemented according to various embodiments of the system and method of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction, which contains one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0149] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A space-based self-labeled two-level distributed model training system, characterized in that, include: The system consists of a primary processing module deployed on each of multiple low-Earth orbit satellites, a secondary processing module deployed on each of multiple medium- and high-Earth orbit satellites, and a control module deployed at the ground control center. The primary processing module deployed on each low-Earth orbit satellite is configured to: in response to receiving model training instructions sent by medium- and high-Earth orbit satellites, extract the embedding features of each batch of local data and assign identification information to each embedding feature, and send each embedding feature and the assigned identification information to the corresponding medium- and high-Earth orbit satellites. The local data is data collected by the low-Earth orbit satellites, and the identification information is used to distinguish the embedding features of different local data. The secondary processing module deployed on each medium-high orbit satellite is configured to: based on the task requirements sent by the control module deployed at the ground control center, process each embedded feature sent by the low orbit satellite into a corresponding task label, and send the task label and corresponding identification information corresponding to each embedded feature to the corresponding low orbit satellite. The task requirements represent the tasks that the task model on the low orbit satellite needs to perform, and the task labels are used as labels for training the task model with local data. The primary processing module deployed on each low-Earth orbit satellite is also configured to: construct a training dataset based on the batch of local data, the task labels corresponding to the embedded features issued by the medium- and high-Earth orbit satellites, and the corresponding identification information; and use the training dataset to train an initial model to obtain a trained task model, which is used to execute the task indicated by the task requirements.

2. The system according to claim 1, characterized in that, The task requirements sent by the control module deployed at the ground control center process the embedded features sent by the low-orbit satellites into corresponding task tags, including: The deep features of each embedded feature are extracted using a deep feature extractor; Based on the task requirements, the corresponding label generator is used to generate task labels and label confidence scores for each embedded feature according to the deep features of each embedded feature. The label confidence scores are used to characterize the reliability of the task labels. The step of sending the task tags and corresponding identification information corresponding to each embedded feature to the corresponding low-Earth orbit satellite includes: High-confidence task labels with a confidence level higher than a specified threshold, along with the corresponding embedded feature identification information, are sent to the corresponding low-Earth orbit satellites to train the initial model using the high-confidence task labels in the low-Earth orbit satellites.

3. The system according to claim 2, characterized in that, The secondary processing module deployed on each medium-high orbit satellite is also configured to: send low-confidence task labels with a label confidence level lower than the specified threshold and their corresponding embedded features to the ground control center; The control module deployed in the ground control center is configured to: correct the low-confidence task labels corresponding to the embedded features issued by the medium- and high-orbit satellites, obtain the corrected task labels corresponding to the embedded features, and upload the corrected task labels corresponding to the embedded features to the corresponding medium- and high-orbit satellites. The step of sending the task tags and corresponding identification information corresponding to each embedded feature to the corresponding low-Earth orbit satellite also includes: The calibration task label and corresponding identification information sent by the ground control center are distributed to the corresponding low-Earth orbit satellites so that the low-Earth orbit satellites can use the calibration task label to train the initial model.

4. The system according to claim 1, characterized in that, The primary processing module deployed on each low-Earth orbit satellite is also configured to send the model parameters of the trained mission model to the corresponding medium- and high-Earth orbit satellites. The secondary processing module deployed on each medium-high orbit satellite is also configured to: aggregate the model parameters sent by at least two low orbit satellites respectively to obtain the global model parameters of the current version, and send the global model parameters of the current version to the corresponding at least two low orbit satellites; The primary processing module deployed on each low-Earth orbit satellite is also configured to: in response to receiving the current version of global model parameters sent by medium- and high-Earth orbit satellites, replace the model parameters of the trained mission model with the current version of global model parameters to obtain the current version of the mission model, so as to use the current version of the mission model to execute the mission indicated by the mission requirements.

5. The system according to claim 4, characterized in that, The secondary processing module deployed on each medium-to-high orbit satellite is also configured to: send the current version of the global model parameters to the ground control center; The control module deployed in the ground control center is also configured to: perform inference accuracy tests on the current version of the global model parameters issued by the medium- and high-orbit satellites; if the inference accuracy of the current version of the global model parameters does not meet the specified mission requirements accuracy, send a training instruction to the corresponding medium- and high-orbit satellites. The secondary processing module deployed on each medium-high orbit satellite is also configured to forward the continued training instructions sent by the ground control center to at least two corresponding low orbit satellites; The primary processing module deployed on each low-Earth orbit satellite is also configured to: in response to receiving a continued training instruction, use the current version of the mission model as the initial model to re-train the initial model using the training dataset to obtain the trained mission model, until the inference accuracy of the global model parameters obtained by aggregating the model parameters of the trained mission model reaches the accuracy required by the mission.

6. The system according to claim 5, characterized in that, The control module deployed in the ground control center is also configured to: send a stop training command to the corresponding medium- and high-orbit satellites when the inference accuracy of the global model parameters in the current version has reached the accuracy required by the mission. The secondary processing module deployed on each medium-high orbit satellite is also configured to forward the stop training command sent by the ground control center to at least two corresponding low orbit satellites; The primary processing module deployed on each low-Earth orbit satellite is also configured to: in response to receiving a stop training command, stop model training and execute the task indicated by the task requirement using the current version of the task model.

7. The system according to claim 5, characterized in that, If the inference accuracy of the global model parameters in the current version does not meet the specified mission accuracy requirements, a training continuation instruction will be sent to the corresponding medium- and high-orbit satellites, including: If the inference accuracy of the current version of the global model parameters does not meet the specified mission requirements, and the inference accuracy of the current version of the global model parameters is the same as or improved compared to the historical version of the global model parameters, a continued training instruction is sent to the corresponding medium-high orbit satellite; the historical version of the global model parameters refers to the global model parameters generated before the current version of the global model references them. The control module deployed in the ground control center is also configured to send a rollback command to the corresponding medium- and high-orbit satellites when the inference accuracy of the global model parameters in the current version does not meet the specified mission requirements and the inference accuracy of the global model parameters in the current version decreases compared to the global model parameters in the historical version. The secondary processing module deployed on each medium-high orbit satellite is also configured to forward the rollback command sent by the ground control center to at least two corresponding low orbit satellites; The primary processing module deployed on each low-Earth orbit satellite is also configured to: in response to receiving a rollback command, restore the current version of the mission model to the previous version of the mission model and retrain the previous version of the mission model as the initial model, or stop model training and use the previous version of the mission model to execute the task indicated by the mission requirement; the previous version of the mission model is the mission model using the global model parameters before the inference accuracy degradation.

8. The system according to claim 4, characterized in that, The secondary processing module deployed on each medium-high orbit satellite is also configured to: aggregate the model parameters sent by at least two low-orbit satellites to obtain the current version of the global model parameters, and then synchronize the current version of the global model parameters to the remaining medium-high orbit satellites. In the event of a failure of any medium-high orbit satellite, the undisturbed medium-high orbit satellites will connect with at least two low-orbit satellites corresponding to the failed medium-high orbit satellite to continue to perform the functions required by the failed medium-high orbit satellite.

9. The system according to claim 1, characterized in that, The extraction of embedding features from batches of local data includes: For any local data in the batch of local data, preprocess the local data to obtain preprocessed local data; The embedding feature extractor is used to generate embedding features of the preprocessed local data based on the preprocessed local data.

10. A space-based self-labeled two-level distributed model training method, characterized in that, The method, applicable to multiple low-Earth orbit satellites, multiple medium- and high-Earth orbit satellites, and a ground control center, includes: In response to receiving the model training instruction sent by the medium- and high-orbit satellites, each low-Earth orbit satellite extracts the embedding features of its batch of local data and assigns identification information to each embedding feature. It then sends each embedding feature and the assigned identification information to the corresponding medium- and high-orbit satellites. The local data is the data collected by the low-Earth orbit satellites, and the identification information is used to distinguish the embedding features of different local data. For each medium-high orbit satellite, based on the embedded features and assigned identification information sent by the low orbit satellite, and the task requirements sent by the ground control center, each embedded feature sent by the low orbit satellite is processed into a corresponding task label, and the task label and corresponding identification information corresponding to each embedded feature are sent to the corresponding low orbit satellite. The task requirements represent the tasks that the task model on the low orbit satellite needs to perform, and the task labels are used as labels for local data to train the task model. Each low-Earth orbit satellite constructs a training dataset based on the batch of local data, the corresponding mission tags issued by medium- and high-Earth orbit satellites, and the corresponding identification information. The initial model is then trained using this training dataset to obtain the trained mission model, which is used to execute the mission as indicated by the mission requirements.

Images