Model training method and device of satellite system, equipment and medium

Abstract

The invention relates to the technical field of satellite system data transmission, in particular to a satellite system model training method and device, equipment and a medium. The method comprises the following steps: acquiring a distillation data set sent by each distillation satellite; performing centralized training on each distillation data set to obtain a first average gradient; gradient data sent by all the federated satellites are obtained, and a second average gradient of all the federated satellites is generated according to all the gradient data; determining the global gradient of each running satellite; updating the global model parameters according to the global gradient to obtain the global model parameters of the current round, and sending the global model parameters of the current round to the running satellite; and if the global model parameter of the current round does not meet the preset convergence condition, repeating the above steps until the target global model parameter is output to meet the preset convergence condition. According to the method and the device, the problems of overlarge communication load, relatively large privacy security problem and relatively low satellite data analysis capability in the prior art can be solved.

Description

Background Art

[0002] At present, the traditional satellite system model mainly transmits the data collected by the satellite to the ground base station, and then the ground base station transmits it to the cloud computing center and centrally trains the machine learning model. However, this method is increasingly unsuitable for current satellite scenarios. The main problems are: First, the satellite is in a free state in space, and the number and time of communication with the ground station are limited every day, so the satellite cannot analyze the collected data in real time. Second, the transmission of the satellite's raw image data requires a lot of bandwidth resources, which will lead to a huge communication load, and the original data set involves privacy and security issues.

[0003] Third, satellites launched in different eras have significant differences in energy supply status and computing processing capabilities. These differences may lead to inconsistencies in the time required for machine learning model training, especially when energy is limited, the model training process will be interrupted.

[0004] Fourth, the accuracy of existing satellites in analyzing the collected data is not high, which leads to insufficient satellite data analysis capabilities, and in turn causes large errors in the results predicted by the satellites when performing weather forecasts, climate analysis and other analysis based on the collected data. Summary of the invention

[0005] The present application provides a model training method, device, equipment and medium for a satellite system to solve one or more technical problems existing in the prior art and at least provide a beneficial choice or create conditions.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by the practice of the present application.

[0007] According to one aspect of an embodiment of the present application, a model training method for a satellite system is provided, wherein the satellite system includes a ground base station and an operating satellite; the operating satellite includes a distillation satellite and a federated satellite, wherein the distillation satellite and the federated satellite are satellites with relatively low functional configuration and satellites with good functional configuration, respectively, and the method is executed on the ground base station, and the method includes: S110, obtaining a distillation data set sent by each of the distillation satellites, wherein the distillation data set is obtained by extracting high-order features from the original data set collected by the distillation satellite; S120, performing centralized training on each of the distilled data sets to obtain a first average gradient of each of the distilled satellites; S130, acquiring gradient data sent by each of the federal satellites, and generating a second average gradient of each of the federal satellites according to each of the gradient data; S140, determining a global gradient of each of the operating satellites according to a preset first weight coefficient, a preset second weight coefficient, the first average gradient, and the second average gradient; S150, updating the global model parameters according to the global gradient to obtain the global model parameters of the current round, and sending the global model parameters of the current round to the operating satellites, so as to perform model training on each of the operating satellites according to the global model parameters of the current round; S160, if the global model parameters of the current round do not meet the preset convergence conditions, repeat steps S110-S150 until the iterative function outputs the target global model parameters to meet the preset convergence conditions; The preset convergence condition is that the model prediction accuracy of each of the operating satellites during data analysis after model training according to the target global model parameters is not less than a preset threshold.

[0008] In one embodiment of the present application, based on the above-mentioned solution, the centralized training of each of the distilled data sets to obtain the first average gradient of each of the distilled satellites includes: Obtaining a local loss function corresponding to each of the distilled data sets one by one, and performing data gradient training on each of the distilled data sets based on each of the local loss functions and preset model parameters to obtain a gradient value corresponding to each of the distilled data sets one by one; A first average gradient of each of the distillation satellites is calculated according to each of the gradient values.

[0009] In one embodiment of the present application, based on the above scheme, the distillation data set can be obtained by the following method: The distillation satellite performs iterative operations on the collected original data set according to the kernel ridge regression algorithm to obtain an initial distillation set; The initial distillation set is optimized by using a gradient clipping method and a stochastic gradient descent method to obtain the distillation data set.

[0010] In one embodiment of the present application, based on the above scheme, the feature optimization of the initial distillation set is performed using the gradient clipping method and the stochastic gradient descent method to obtain the distillation data set, including: Extracting each sample data in the initial distillation set; Clipping the gradients corresponding to the sample data according to preset clipping parameters, and adding noise data conforming to the normal distribution to the clipped gradients to obtain target sample data; The stochastic gradient descent method is used to perform feature optimization on each of the target sample data to obtain the distilled data set.

[0011] In one embodiment of the present application, based on the aforementioned solution, determining the global gradient of each of the operating satellites according to the preset first weight coefficient, the preset second weight coefficient, the first average gradient, and the second average gradient includes: Multiplying the first weight coefficient by the first average gradient to obtain a first gradient value; Multiplying the second weight coefficient by the second average gradient to obtain a second gradient value; Taking the sum of the first gradient value and the second gradient value as the global gradient; The first weight coefficient is the quotient of the number of the distillation satellites divided by the number of the operating satellites, and the second weight coefficient is the quotient of the number of the federal satellites divided by the number of the operating satellites.

[0012] In one embodiment of the present application, based on the above solution, updating the global model parameters according to the global gradient to obtain the global model parameters of the current round includes: Obtain the preset target learning rate, and obtain the global model parameters of the previous round from the global model parameters obtained from multiple rounds of historical model training; The global model parameters of the current round can be obtained by the following formula:

[0013] in, are the global model parameters of the current round, are the global model parameters of the previous round, is the target learning rate, the is the global gradient.

[0014] In one embodiment of the present application, based on the above scheme, the gradient data of the federated satellite is obtained by the following method: After receiving the global model parameters of the previous round sent by the ground base station, training the collected local data set according to the global model of the previous round to obtain the gradient data of the federal satellite; Wherein, the steps S110-S150 are the model training of the current round, and the gradient data of the federated satellite is used for the model training of the current round.

[0015] According to one aspect of an embodiment of the present application, a model training device for a satellite system is provided, wherein the satellite system includes a ground base station and an operating satellite; the operating satellite includes a distillation satellite and a federated satellite, wherein the distillation satellite and the federated satellite are satellites with relatively low functional configuration and satellites with good functional configuration, respectively, and the device is applied to the ground base station, and the device includes: A first acquisition unit is used to acquire a distillation data set sent by each of the distillation satellites, where the distillation data set is obtained by extracting high-order features from the original data set collected by the distillation satellite; A training unit, configured to perform centralized training on each of the distilled data sets to obtain a first average gradient of each of the distilled satellites; A second acquisition unit, configured to acquire the gradient data sent by each of the federal satellites, and generate a second average gradient of each of the federal satellites according to each of the gradient data; a determining unit, configured to determine a global gradient of each of the operating satellites according to a preset first weight coefficient, a preset second weight coefficient, the first average gradient, and the second average gradient; An output unit updates the global model parameters according to the global gradient to obtain the global model parameters of the current round, and sends the global model parameters of the current round to the operating satellites, so as to perform model training on each of the operating satellites according to the global model parameters of the current round; An iteration unit, configured to repeat the above steps if the global model parameters of the current round do not meet the preset convergence condition, until the iterative function outputs the target global model parameters to meet the preset convergence condition; The preset convergence condition is that the model prediction accuracy of each of the operating satellites during data analysis after model training according to the target global model parameters is not less than a preset threshold.

[0016] According to one aspect of an embodiment of the present application, a computer-readable storage medium is provided, on which a computer program is stored. The computer program includes executable instructions. When the executable instructions are executed by a processor, the model training method described in the above embodiment is implemented.

[0017] According to one aspect of an embodiment of the present application, an electronic device is provided, comprising: one or more processors; and a memory for storing executable instructions of the processors, wherein when the executable instructions are executed by the one or more processors, the one or more processors implement the model training method as described in the above embodiment.

[0018] The beneficial effects of the present application are as follows: from the perspective of the overall logical architecture, the present application achieves efficient model training effects with reduced communication load by classifying satellites with low functional configurations and satellites with good functional configurations, and then performing different training methods according to different categories of satellites.

[0019] First, the distillation satellites with lower functional configuration are trained centrally with the distillation data set of the distillation satellites to obtain the first average gradient of each distillation satellite. The gradient data sent by the federated satellites with good functional configuration are processed to obtain the second average gradient, and then the first average gradient and the second average gradient are weightedly fused to obtain the global gradient, and then the global model parameters of the historical training are updated according to the global gradient, and then the global model parameters of the current round are obtained.

[0020] By sending the global model parameters of the current round to the operating satellites, model training is performed on each of the operating satellites according to the global model parameters of the current round. In this way, the operating satellites can analyze the collected image data by themselves after model training, without the need to transmit the collected image outputs to the ground base station for analysis, which greatly improves the analysis efficiency and reduces the communication load between the satellite and the ground station, and does not involve privacy and security issues.

[0021] Furthermore, if the global model parameters of the current round do not meet the preset convergence conditions, then it is necessary to iterate the global model parameters of the current round for multiple rounds, that is, repeat the above steps, obtain the distilled data set and gradient data after the current round of model training, and further update the global model of the next round until the target global model parameters are output to meet the preset convergence conditions. The preset convergence conditions are that the model prediction accuracy of each of the operating satellites after model training according to the target global model parameters during data analysis is not less than a preset threshold, that is, to maximize the accuracy of the operating satellite when performing image analysis.

[0022] Therefore, this application not only solves the problem of inefficiency caused by the limitation of the number of communications between satellites and ground stations, but also solves the problems of communication load and privacy security. It can also train satellites in a targeted manner according to different computing capabilities and functional configurations. The output target global model parameters can greatly improve the accuracy of satellites in image analysis, and optimize the data analysis capabilities of all operating satellites in the satellite system as a whole. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following is a brief description of the drawings required for use in the description of the embodiments. Obviously, the drawings described are only part of the embodiments of the present application, not all of the embodiments, and those skilled in the art can also obtain other design solutions and drawings based on these drawings without creative work.

[0024] Figure 1 This is a flow chart of a model training method for a satellite system according to an embodiment of the present application; Figure 2 The overall logic diagram of a single round of model training according to an embodiment of the present application is shown; Figure 3 It is a block diagram of a model training device for a satellite system according to an embodiment of the present application; Figure 4 It is a schematic diagram of the system structure of an electronic device according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more comprehensive and complete and fully convey the concept of the example embodiments to those skilled in the art.

[0026] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, known methods, devices, realizations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0027] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or micro-control node devices.

[0028] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0029] It should be noted that the "multiple" mentioned in this article refers to two or more. "And / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0030] The following is a detailed introduction to the technical background of the embodiments of the present application: In the models for satellite systems, the main problems are: 1. The bandwidth resources of satellites are limited. Transmitting a large number of original images to ground stations will cause a huge communication burden, and the original images involve privacy and security issues; 2. Satellites launched in different eras have significant differences in energy supply status and computing processing capabilities. These differences may lead to inconsistencies in the time required for machine learning model training, especially when energy is limited, the model training process will be interrupted.

[0031] Therefore, in order to solve the above problems and make the satellite system model (satellite federated learning model) more efficient and usable, this application proposes a model training method for a satellite system, namely, a semi-federated learning framework for cross-satellite data distillation and model collaboration for satellite constellations.

[0032] Specifically, the framework is divided into centralized training and federated training. This application classifies satellites with insufficient energy status and computing power (satellites with lower functional configuration) into centralized training. Such satellites are called distillation satellites. They only need to use the original data locally to generate distilled data and transmit it to the ground base station, and then perform centralized model training at the ground base station. Since the distilled data is very small and has strong privacy, it can well solve the problems of bandwidth limitation and privacy security.

[0033] Satellites with sufficient energy status and computing power (satellites with good functional configuration) belong to federated training. Such satellites are called federated satellites. They can perform more complex computing tasks and only need to transmit model updates (i.e. gradient data) to the ground station for global aggregation. Global aggregation is to weightedly fuse the first average gradient obtained from centralized training with the second average gradient obtained from federated training to obtain a new global gradient, and generate global model parameters based on the global gradient. Finally, the ground base station aggregates the centralized training model with the federated training model to generate the final global model parameters.

[0034] The implementation details of the technical solution of the embodiment of the present application are described in detail below: According to one aspect of an embodiment of the present application, a model training method for a satellite system is provided. Figure 1 This is a flow chart of a model training method for a satellite system according to an embodiment of the present application, wherein the satellite system includes a ground base station and an operating satellite; the operating satellite includes a distillation satellite and a federated satellite, wherein the distillation satellite and the federated satellite are satellites with a relatively low functional configuration and a satellite with a good functional configuration, respectively, and the method is executed on the ground base station; the method includes at least steps S110 to S160, which are described in detail as follows: In step S110, a distillation data set sent by each of the distillation satellites is obtained, where the distillation data set is obtained by extracting high-order features from the original data set collected by the distillation satellite.

[0035] Specifically, the distilled data set is obtained by extracting high-order features from the original data set collected by the distillation satellite. For example, after the distillation satellite collects the original image data, due to the large amount of data in the original image data, transmitting the original image data to the ground base station will cause excessive communication load and occupy more bandwidth resources. At the same time, the original data involves privacy and security issues. Therefore, the distilled data set obtained by extracting high-order features from the original data set occupies a small bandwidth, and after the high-order feature extraction, conventional privacy parameter encryption will be performed, which solves the problem of privacy and security.

[0036] In one embodiment of the present application, the distillation dataset can be obtained by the following method: The distillation satellite performs iterative operations on the collected original data set according to the kernel ridge regression algorithm to obtain an initial distillation set; The initial distillation set is optimized by using a gradient clipping method and a stochastic gradient descent method to obtain the distillation data set.

[0037] Specifically, this application proposes a data distillation method based on differential privacy kernel induction points, which generates differential privacy distillation samples by clipping the gradient of the distilled dataset and adding calibration noise during training. In the data distillation process, the overall goal is to find a small dataset , which is a large raw dataset drawn from a data distribution P with respect to a learning algorithm A and a loss function l The approximation of ξ can be expressed as follows: ; In the nuclear induction point, the loss The classification accuracy is expressed based on the distance between the true label and the predicted label. The learning algorithm A is kernel ridge regression (KRR). In this process, the distillation dataset is first randomly initialized, and then iteratively refined by minimizing the kernel ridge regression (KRR) loss. : ; Here, is the regularization parameter and >0, is the kernel matrix. During the training phase, a gradient-based optimization method (such as stochastic gradient descent) is used to update the distilled dataset until certain preset convergence conditions are met. Labels are extracted from a uniform distribution of the number of classes. ,The original KIP algorithm has an option to optimize the labels by giving a ,optimized distilled image.

[0038] In one embodiment of the present application, the feature optimization of the initial distillation set by using the gradient clipping method and the stochastic gradient descent method to obtain the distillation data set includes: Extracting each sample data in the initial distillation set; Clipping the gradients corresponding to the sample data according to preset clipping parameters, and adding noise data conforming to the normal distribution to the clipped gradients to obtain target sample data; The stochastic gradient descent method is used to perform feature optimization on each of the target sample data to obtain the distilled data set.

[0039] Specifically, in the specific distillation process, the distillation data set is first initialized, and each sample data (input data ) is drawn from a normal distribution, the privacy parameter (the scale parameter used to add noise) is calculated according to the set privacy level, and the gradient is clipped according to the preset clipping parameter to ensure that the norm of the gradient does not exceed the threshold C. Specifically, the gradient Clipped to , that is, if the gradient The norm exceeds , then scale its size. Add noise to the clipped gradient to get the target sample data. The noise conforms to the normal distribution, and accumulate all gradients. Use stochastic gradient descent (SGD) to update the distilled data set. , and finally generate a support dataset with privacy protection .

[0040] In step S120, centralized training is performed on each of the distillation data sets to obtain a first average gradient of each of the distillation satellites.

[0041] In one embodiment of the present application, the centralized training of each of the distilled data sets to obtain the first average gradient of each of the distilled satellites includes: Obtaining a local loss function corresponding to each of the distilled data sets one by one, and performing data gradient training on each of the distilled data sets based on each of the local loss functions and preset model parameters to obtain a gradient value corresponding to each of the distilled data sets one by one; A first average gradient of each of the distillation satellites is calculated according to each of the gradient values.

[0042] Specifically, in centralized training, assuming that satellites, and each extracts its own distilled dataset, which together form the data set This training method transmits the distilled data to the ground station through satellite-to-ground communication for centralized training. Based on this data set, the first average gradient , can be expressed by the following formula: ; Among them, w is the model parameter (i.e. the global model parameter of the previous round), is the gradient value obtained by each centralized training satellite, and this value can be expressed as , represents the total number of distilled samples (which can be analyzed from the distilled dataset), and From the centralized training Satellite distilled data samples (which can be analyzed from the distilled dataset), is the local loss function (which can be calculated based on the global model parameters of the previous round).

[0043] In step S130, the gradient data sent by each of the federal satellites is obtained, and a second average gradient of each of the federal satellites is generated according to each of the gradient data.

[0044] Specifically, in federated training (FL), assuming that Satellites, and all of them generate corresponding gradient updates, and obtain the gradient data of each federated satellite, which together form the gradient set These gradient data will be uploaded to the ground base station for federated averaging, and finally generate the second average gradient , can be expressed by the following formula: ; In one embodiment of the present application, the gradient data of the federated satellite is obtained by the following method: After receiving the global model parameters of the previous round sent by the ground base station, training the collected local data set according to the global model of the previous round to obtain the gradient data of the federal satellite; Wherein, the steps S110-S150 are the model training of the current round, and the gradient data of the federated satellite is used for the model training of the current round.

[0045] Specifically, the gradient data of the federal satellite used in the current round of model training is obtained after the global model parameter training in the previous round. Therefore, the present application can continuously perform multiple rounds of iterative updates on the global model parameters, so that the output of the final target global model parameters can meet the preset convergence conditions.

[0046] In step S140, the global gradient of each of the operating satellites is determined according to a preset first weight coefficient, a preset second weight coefficient, the first average gradient, and the second average gradient.

[0047] The determining of the global gradient of each of the operating satellites according to the preset first weight coefficient, the preset second weight coefficient, the first average gradient, and the second average gradient comprises: Multiplying the first weight coefficient by the first average gradient to obtain a first gradient value; Multiplying the second weight coefficient by the second average gradient to obtain a second gradient value; Taking the sum of the first gradient value and the second gradient value as the global gradient; The first weight coefficient is the quotient of the number of the distillation satellites divided by the number of the operating satellites, and the second weight coefficient is the quotient of the number of the federal satellites divided by the number of the operating satellites.

[0048] Specifically, the ground base station will use the first average gradient obtained by centralized training in the current round of model training And the second average gradient obtained by federated training The weights used to generate the global gradient G are used to update the global model parameters, as shown in the following formula.

[0049] ; in, is the preset first weight coefficient, is the preset second weight coefficient.

[0050] In step S150, the global model parameters are updated according to the global gradient to obtain the global model parameters of the current round, and the global model parameters of the current round are sent to the operating satellites so as to perform model training on each of the operating satellites according to the global model parameters of the current round.

[0051] In one embodiment of the present application, updating the global model parameters according to the global gradient to obtain the global model parameters of the current round includes: Obtain the preset target learning rate, and obtain the global model parameters of the previous round from the global model parameters obtained from multiple rounds of historical model training; The global model parameters of the current round can be obtained by the following preset iterative function formula:

[0052] in, are the global model parameters of the current round, are the global model parameters of the previous round, is the target learning rate, the The global gradient of the current round of model training.

[0053] In step S160, if the global model parameters of the current round do not meet the preset convergence conditions, steps S110-S150 are repeated until the iterative function outputs the target global model parameters to meet the preset convergence conditions; The preset convergence condition is that the model prediction accuracy of each of the operating satellites during data analysis after model training according to the target global model parameters is not less than a preset threshold.

[0054] If the global model parameters of the current round do not meet the preset convergence conditions, steps S110-S150 will be repeated, that is, iterative training will be performed. In a new round of training, that is, the next round of model training, the operating satellite will first send the global model parameters of the current round when establishing a connection with the ground base station to provide local model training for each operating satellite. After the operating satellite has performed local model training, the ground base station will continue to perform iterative rounds of centralized training and federated training on the distilled data set or gradient data transmitted by the operating satellite until the training is stopped when the target global model parameters are output.

[0055] Therefore, from the perspective of the overall satellite system model, it is assumed that the satellite system is a constellation consisting of K satellites and G ground stations. Satellite k∈ ={1,2,...,K} collect and store data sets , and then the satellites collaborate to learn the global model w∈Rd by minimizing the global objective function. The process can be expressed as follows:

[0056] in is the size of the data set, Represents the size of the sum of all satellite data sets, satellite local objective function Is the data set The overall goal is to minimize the global model loss function. The loss on all client datasets, and the loss function can be expressed as:

[0057] in It is a data point and the training loss of the model parameters. If the ground station can download the complete data set from the satellite, the formula can be solved In other words, through formula (1), the global model parameters that can minimize the overall loss rate can be found in the process of continuous iterative training. The overall loss rate is the sum of the loss rates of all satellites when performing data analysis, thereby improving the data analysis capabilities of all operating satellites.

[0058] In one embodiment of the present application, the present application also proposes a communication model: In the geocentric inertial coordinate system, the present application assumes that the satellite k and the ground station g have trajectories and , where t is the continuous wall-clock time. If the link between the satellite and the ground station is to be connected, the satellite k must be at the minimum elevation angle The visible range is called the visible window in the present invention. The angle of the window is calculated as follows:

[0059] in, represents the difference between the ground station trajectory and the satellite trajectory, That is, the angle between the satellite and the ground station. When the satellite is within the visible angle range of the ground station, the satellite and the ground station can establish a communication connection.

[0060] Connection strategy: In this application, the first orbital edge computing simulator and runtime service are used to build a satellite constellation scenario from low-Earth orbit satellites to ground base stations. It includes 100 operational satellites and 10 ground base stations. The ground part consists of two circular arrays, each of which is evenly distributed with 5 ground stations. The two rings are located at 87° north latitude and 87° south latitude. The 100 satellites in the space part operate in a polar orbit with an inclination of 97.3°.

[0061] The overall logic of this application can be referred to Figure 2 As shown, this application combines the advantages of federated learning and centralized learning, combining data set distillation technology with centralized learning, thereby reducing the communication pressure between satellite and ground communications, and differential privacy improves the privacy security of data, and fully utilizes the computing power of ground base stations by transmitting distilled data to ground base stations. For satellites with sufficient computing power and energy status, local models (gradient data) are updated and transmitted to ground base stations through federated learning, fully utilizing the computing resources of federated satellites, while also effectively protecting the security of original data and achieving efficient model training.

[0062] Figure 3This is a block diagram of a model training device 300 for a satellite system according to an embodiment of the present application. According to a model training device 300 for a satellite system according to an embodiment of the present application, the satellite system includes a ground base station and an operating satellite; the operating satellite includes a distillation satellite and a federal satellite, the distillation satellite and the federal satellite are satellites with lower functional configuration and satellites with better functional configuration, respectively, and the device is applied to the ground base station; the device 300 includes: a first acquisition unit 301, a training unit 302, a second acquisition unit 303, a determination unit 304 and an output unit 305.

[0063] A first acquisition unit 301 is used to acquire a distillation data set sent by each of the distillation satellites, where the distillation data set is obtained by extracting high-order features from the original data set collected by the distillation satellite; A training unit 302 is used to perform centralized training on each of the distilled data sets to obtain a first average gradient of each of the distilled satellites; A second acquisition unit 303 is used to acquire the gradient data sent by each of the federal satellites, and generate a second average gradient of each of the federal satellites according to each of the gradient data; A determination unit 304 is used to determine the global gradient of each of the operating satellites according to a preset first weight coefficient, a preset second weight coefficient, the first average gradient, and the second average gradient; An output unit 305 is used to update the global model parameters according to the global gradient, obtain the global model parameters of the current round, and send the global model parameters of the current round to the operating satellites, so as to perform model training on each of the operating satellites according to the global model parameters of the current round; An iteration unit, configured to repeat the above steps if the global model parameters of the current round do not meet the preset convergence condition, until the iterative function outputs the target global model parameters to meet the preset convergence condition; The preset convergence condition is that the model prediction accuracy of each of the operating satellites during data analysis after model training according to the target global model parameters is not less than a preset threshold.

[0064] As another aspect, the present application further provides a computer-readable storage medium on which a program product capable of implementing the method provided above in this specification is stored. In some possible implementations, various aspects of the present application may also be implemented in the form of a program product, which includes a program code, and when the program product is run on a terminal device, the program code is used to cause the terminal device to execute the steps according to various exemplary implementations of the present application described in the above "Embodiment Method" section of this specification.

[0065] According to the program product for implementing the above method in the embodiment of the present application, it can adopt a portable compact disk read-only memory (CD-ROM) and include program code, and can be run on a terminal device, such as a personal computer. However, the program product of the present application is not limited thereto. In this document, a readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, an apparatus or a device.

[0066] The program product may be any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0067] Computer readable signal media may include data signals propagated in baseband or as part of a carrier wave, in which readable program code is carried. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Readable signal media may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0068] The program code embodied on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the foregoing.

[0069] Program code for performing the operations of the present application may be written in any combination of one or more programming languages, including object-oriented programming languages, such as Java, C++, etc., and conventional procedural programming languages, such as "C" or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, as a separate software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0070] As another aspect, the present application also provides an electronic device capable of implementing the above method.

[0071] Those skilled in the art will appreciate that various aspects of the present application may be implemented as a system, method or program product. Therefore, various aspects of the present application may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to as "circuit", "module" or "system" herein.

[0072] Refer to the following Figure 4 The electronic device 400 according to this embodiment of the present application is described. Figure 4 The electronic device 400 shown is merely an example and should not bring any limitation to the functions and scope of use of the embodiments of the present application.

[0073] like Figure 4 As shown, the electronic device 400 is in the form of a general computing device. The components of the electronic device 400 may include but are not limited to: at least one processing unit 410, at least one storage unit 420, and a bus 430 connecting different system components (including the storage unit 420 and the processing unit 410).

[0074] The storage unit stores program codes, which can be executed by the processing unit 410, so that the processing unit 410 executes the steps described in the above “Example Method” section of this specification according to various exemplary implementations of the present application.

[0075] The storage unit 420 may include a readable medium in the form of a volatile storage unit, such as a random access memory unit (RAM) 421 and / or a cache memory unit 422 , and may further include a read-only memory unit (ROM) 423 .

[0076] The storage unit 420 may also include a program / utility 424 having a set (at least one) of program modules 425, such program modules 425 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include an implementation of a network environment.

[0077] Bus 430 may represent one or more of several types of bus structures, including a memory unit bus or memory unit controller node, a peripheral bus, an accelerated graphics port, a processing unit, or a local bus using any of a variety of bus architectures.

[0078] The electronic device 400 may also communicate with one or more external devices 1200 (e.g., keyboards, pointing devices, Bluetooth devices, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 400, and / or communicate with any device that enables the electronic device 400 to communicate with one or more other computing devices (e.g., routers, modems, etc.). This communication may be performed through an input / output (I / O) interface 450. In addition, the electronic device 400 may also communicate with one or more networks (e.g., local area networks (LANs), wide area networks (WANs), and / or public networks, such as the Internet) through a network adapter 460. As shown, the network adapter 460 communicates with other modules of the electronic device 400 through a bus 430. It should be understood that, although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 400, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0079] Through the description of the above implementation methods, it is easy for those skilled in the art to understand that the example implementation methods described here can be implemented by software, or by combining software with necessary hardware. Therefore, the technical solution according to the implementation method of the present application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the implementation method of the present application.

[0080] In addition, the above-mentioned figures are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present application, and are not intended to be limiting. It is easy to understand that the processes shown in the above-mentioned figures do not indicate or limit the time sequence of these processes. In addition, it is also easy to understand that these processes can be performed synchronously or asynchronously, for example, in multiple modules.

[0081] It should be understood that the present application is not limited to the precise structures that have been described above and shown in the drawings, and that various modifications and changes may be performed without departing from the scope thereof. The scope of the present application is limited only by the appended claims.

Claims

1. A model training method for a satellite system, characterized in that: The satellite system includes a ground base station and an operating satellite; the operating satellite includes a distillation satellite and a federation satellite, wherein the distillation satellite and the federation satellite are satellites with relatively low functional configuration and satellites with good functional configuration respectively; The ground base station stores global model parameters obtained from multiple rounds of historical model training. The method is executed on the ground base station, and the method includes: S110, obtaining a distillation data set sent by each of the distillation satellites, wherein the distillation data set is obtained by extracting high-order features from the original data set collected by the distillation satellite; S120, performing centralized training on each of the distilled data sets to obtain a first average gradient of each of the distilled satellites; S130, acquiring gradient data sent by each of the federal satellites, and generating a second average gradient of each of the federal satellites according to each of the gradient data; S140, determining a global gradient of each of the operating satellites according to a preset first weight coefficient, a preset second weight coefficient, the first average gradient, and the second average gradient; S150, updating the global model parameters according to the global gradient to obtain the global model parameters of the current round, and sending the global model parameters of the current round to the operating satellites, so as to perform model training on each of the operating satellites according to the global model parameters of the current round; S160, if the global model parameters of the current round do not meet the preset convergence conditions, repeat steps S110-S150 until the iterative function outputs the target global model parameters to meet the preset convergence conditions; The preset convergence condition is that the model prediction accuracy of each of the operating satellites during data analysis after model training according to the target global model parameters is not less than a preset threshold.

2. The model training method according to claim 1, characterized in that: The centralized training of each of the distilled data sets to obtain a first average gradient of each of the distilled satellites includes: Obtaining a local loss function corresponding to each of the distilled data sets one by one, and performing data gradient training on each of the distilled data sets based on each of the local loss functions and preset model parameters to obtain a gradient value corresponding to each of the distilled data sets one by one; A first average gradient of each of the distillation satellites is calculated according to each of the gradient values.

3. The model training method according to claim 2, characterized in that: The distillation dataset can be obtained by the following method: The distillation satellite performs iterative operations on the collected original data set according to the kernel ridge regression algorithm to obtain an initial distillation set; The initial distillation set is optimized by using a gradient clipping method and a stochastic gradient descent method to obtain the distillation data set.

4. The model training method according to claim 3, characterized in that: The initial distillation set is optimized by using a gradient clipping method and a stochastic gradient descent method to obtain the distillation data set, including: Extracting each sample data in the initial distillation set; Clipping the gradients corresponding to the sample data according to preset clipping parameters, and adding noise data conforming to the normal distribution to the clipped gradients to obtain target sample data; The stochastic gradient descent method is used to perform feature optimization on each of the target sample data to obtain the distilled data set.

5. The model training method according to claim 4, characterized in that: The determining of the global gradient of each of the operating satellites according to the preset first weight coefficient, the preset second weight coefficient, the first average gradient, and the second average gradient comprises: Multiplying the first weight coefficient by the first average gradient to obtain a first gradient value; Multiplying the second weight coefficient by the second average gradient to obtain a second gradient value; Taking the sum of the first gradient value and the second gradient value as the global gradient; The first weight coefficient is the quotient of the number of the distillation satellites divided by the number of the operating satellites, and the second weight coefficient is the quotient of the number of the federal satellites divided by the number of the operating satellites.

6. The model training method according to claim 5, characterized in that: The updating of the global model parameters according to the global gradient to obtain the global model parameters of the current round includes: Obtain the preset target learning rate, and obtain the global model parameters of the previous round from the global model parameters obtained from multiple rounds of historical model training; The global model parameters of the current round can be obtained by the following formula: in, are the global model parameters of the current round, are the global model parameters of the previous round, is the target learning rate, the is the global gradient.

7. The model training method according to claim 6, characterized in that: The gradient data of the federal satellite is obtained by the following method: After receiving the global model parameters of the previous round sent by the ground base station, training the collected local data set according to the global model of the previous round to obtain the gradient data of the federal satellite; Wherein, the steps S110-S150 are the model training of the current round, and the gradient data of the federated satellite is used for the model training of the current round.

8. A model training device for a satellite system, characterized in that: The satellite system includes a ground base station and an operating satellite; the operating satellite includes a distillation satellite and a federation satellite, the distillation satellite and the federation satellite are satellites with low functional configuration and satellites with good functional configuration, respectively, the device is applied to the ground base station, and the device includes: A first acquisition unit is used to acquire a distillation data set sent by each of the distillation satellites, where the distillation data set is obtained by extracting high-order features from the original data set collected by the distillation satellite; A training unit, configured to perform centralized training on each of the distilled data sets to obtain a first average gradient of each of the distilled satellites; A second acquisition unit, configured to acquire the gradient data sent by each of the federal satellites, and generate a second average gradient of each of the federal satellites according to each of the gradient data; a determining unit, configured to determine a global gradient of each of the operating satellites according to a preset first weight coefficient, a preset second weight coefficient, the first average gradient, and the second average gradient; An output unit updates the global model parameters according to the global gradient to obtain the global model parameters of the current round, and sends the global model parameters of the current round to the operating satellites, so as to perform model training on each of the operating satellites according to the global model parameters of the current round; An iteration unit, configured to repeat the above steps if the global model parameters of the current round do not meet the preset convergence condition, until the iterative function outputs the target global model parameters to meet the preset convergence condition; The preset convergence condition is that the model prediction accuracy of each of the operating satellites during data analysis after model training according to the target global model parameters is not less than a preset threshold.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores at least one program code, and the at least one program code is loaded and executed by a processor to implement the operations performed by the method according to any one of claims 1 to 7.

10. An electronic device, characterized in that: The electronic device includes one or more processors and one or more memories, wherein the one or more memories store at least one program code, and the at least one program code is loaded and executed by the one or more processors to implement the operations performed by the method according to any one of claims 1 to 7.

Images