Deployment method and device of large model at satellite end and storage medium

Abstract

The invention discloses a method and device for deploying a large model at a satellite end and a storage medium, relates to the technical field of information, and can reduce the deployment difficulty of the large model on a satellite and improve the deployment efficiency. Comprising the steps of obtaining configuration attribute information of a target satellite terminal and task attribute information of a predetermined task in response to a deployment signal of a large model of the target satellite terminal, and obtaining an initial large model; obtaining a ground end sample data set, and training the initial large model by using the ground end sample data set to obtain a task large model; determining model attribute information of the task large model, and determining a target compression mode suitable for the task large model based on the model attribute information and the task type of the predetermined task; compressing the task large model by using a target compression mode to obtain an initial satellite-borne large model; and based on the configuration attribute information, performing target satellite terminal adaptation adjustment on the initial satellite-borne large model to obtain a satellite-borne large model, and deploying the satellite-borne large model to the target satellite terminal.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method, device and storage medium for deploying a large model on a satellite terminal. Background Art

[0002] With the continuous development of society, large models have entered a period of explosive growth, showing great potential and value in various application fields. The remote sensing field has also entered the era of large models. my country's digital economy development level is in a leading position in the world, and the research and development and application of domestic remote sensing large model technology have also made breakthrough progress. It has played an important role in resource monitoring, emergency rescue, agricultural finance, urban management and other fields, showing significant social and economic benefits. my country's remote sensing large models have developed rapidly. Since 2022, many scientific research institutes and technology companies have successively released self-developed remote sensing large models, and continuously iterated algorithms in practical applications, and the model performance has been continuously improved. However, the current large models are only used for the interpretation of satellite remote sensing data, and are not widely used on satellite terminals. Based on this, in order to make satellite terminals smarter, large models need to be deployed on satellite terminals.

[0003] At present, large ground models are usually deployed directly to satellites. However, this large model deployment method increases the difficulty of deploying large models on satellites due to the complex structure of large models and limited resources on the satellite side, which in turn leads to reduced performance of large models after deployment on satellites or the inability to operate the models normally. At the same time, large models with complex structures consume a lot of bandwidth and time during transmission, which reduces the efficiency of deployment on satellites. Summary of the invention

[0004] The present invention provides a method, device and storage medium for deploying a large model on a satellite, which are mainly capable of reducing the difficulty of deploying the large model on a satellite and improving the deployment efficiency, enhancing the performance of the large model after being deployed on the satellite, and ensuring the stable operation of the large model after being deployed on the satellite.

[0005] According to a first aspect of the present invention, a method for deploying a large model on a satellite terminal is provided, comprising:

[0006] In response to a deployment signal of a target satellite terminal large model, obtaining configuration attribute information of the target satellite terminal and task attribute information of a predetermined task, and obtaining an initial large model;

[0007] Acquire a ground-side sample data set corresponding to the task attribute information, and use the ground-side sample data set to train the initial large model to obtain a task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data;

[0008] Determine model attribute information of the task large model, and determine a target compression method applicable to the task large model based on the model attribute information and the task type of the predetermined task;

[0009] Using the target compression method to compress the mission large model to obtain an initial onboard large model;

[0010] Based on the configuration attribute information, the initial onboard large model is adapted and adjusted for the target satellite terminal to obtain the onboard large model, and the onboard large model is deployed to the target satellite terminal.

[0011] Optionally, if the target compression method is a knowledge transfer compression method, the target compression method is used to compress the task large model to obtain an initial onboard large model, including:

[0012] Acquire a basic single model and a first sample compressed data set corresponding to the task attribute information, and determine a temperature parameter based on the number of model parameters of the basic single model, wherein the number of model parameters of the basic single model is less than a preset number threshold, the structural complexity of the model structure of the basic single model is less than a preset complexity threshold, and the first sample compressed data set includes first sample input data and actual annotation information corresponding to the first sample input data;

[0013] Inputting the first sample input data into the task big model for information prediction to obtain task big model prediction information, and inputting the first sample input data into the basic single model for information prediction to obtain single model prediction information;

[0014] Based on the temperature parameter, the task large model prediction information is adjusted to obtain adjusted task large model prediction information;

[0015] Determine the cross entropy loss function of the basic single model based on the actual annotation information corresponding to the first sample input data and the single model prediction information, and determine the information gain loss function of the basic single model based on the adjusted task large model prediction information and the single model prediction information;

[0016] Determine the weight coefficients corresponding to the cross entropy loss function and the information gain loss function respectively, and based on the weight coefficients, add the cross entropy loss function and the information gain loss function to obtain a comprehensive loss function;

[0017] Based on the comprehensive loss function, the basic single model is iteratively trained to obtain the initial satellite-borne large model.

[0018] Optionally, if the target compression method is a parameter removal compression method, the target compression method is used to compress the mission large model to obtain an initial onboard large model, including:

[0019] Acquire a second sample compressed data set corresponding to the task attribute information, wherein the second sample compressed data set includes second sample input data and actual annotation information corresponding to the second sample input data, input the second sample input data into the task large model for forward propagation, and determine the input data of each network layer and the output data of each network layer in the task large model during the forward propagation process;

[0020] Determine the loss function of the task large model based on the actual annotation information corresponding to the last network layer output data in the task large model and the second sample input data;

[0021] Taking any network layer except the last network layer in each network layer in the task large model as a current network layer, determining the loss partial derivative of the loss function with respect to the output data of the subsequent network layer corresponding to the current network layer, and determining the activation derivative of the activation function of the subsequent network layer at the input data of the subsequent network layer;

[0022] Multiplying the loss partial derivative of the output data of the subsequent network layer by the activation derivative at the input data of the subsequent network layer to obtain the loss partial derivative of the loss function with respect to the input data of the subsequent network layer;

[0023] Determine the activation derivative of the activation function of the current network layer at the input data of the current network layer, multiply the activation derivative at the input data of the current network layer by the loss partial derivative of the input data of the subsequent network layer, and multiply the product by the transpose of the weight of the subsequent network layer to obtain the loss partial derivative of the loss function with respect to the input data of the current network layer;

[0024] Multiplying the transpose of the current network layer output data by the loss partial derivative of the subsequent network layer input data to obtain the sensitivity of the weight parameter in the subsequent network layer to the loss function, and determining the loss partial derivative of the subsequent network layer input data as the sensitivity of the bias parameter in the subsequent network layer to the loss function;

[0025] Multiplying the transpose of the second sample input data by the loss partial derivative of the current network layer input data to obtain the sensitivity of the weight parameter in the current network layer to the loss function, and determining the loss partial derivative of the current network layer input data as the sensitivity of the bias parameter in the current network layer to the loss function;

[0026] A target sensitivity less than a preset sensitivity threshold is determined among the sensitivity of the weight parameters of each network layer in the mission large model to the loss function and the sensitivity of the bias parameters to the loss function, the weight parameters and bias parameters corresponding to the target sensitivity are determined as redundant parameters, and the redundant parameters in the mission large model are removed to obtain the initial satellite-borne large model.

[0027] Optionally, if the target compression method is a matrix conversion compression method, compressing the mission large model using the target compression method to obtain an initial onboard large model includes:

[0028] Taking any network layer in the task model as a target network layer, determining a weight matrix of the target network layer, and determining a transposed matrix of the weight matrix;

[0029] Multiplying the weight matrix by the transposed matrix to obtain a first matrix, and multiplying the transposed matrix by the weight matrix to obtain a second matrix;

[0030] Determine a first eigenvalue and a first eigenvector of the first matrix, and determine a second eigenvalue and a second eigenvector of the second matrix;

[0031] Based on the first eigenvector, construct a first orthogonal matrix, and based on the second eigenvector, construct a second orthogonal matrix;

[0032] Determine a square root of the first eigenvalue or the second eigenvalue, and construct a diagonal matrix based on the square root;

[0033] The first orthogonal matrix, the second orthogonal matrix and the diagonal matrix are multiplied to obtain a low-rank matrix, and the low-rank matrix is ​​used to replace the weight matrix of the target network layer, and the model constituted by the target network layer with the low-rank matrix is ​​determined as the initial onboard large model.

[0034] Optionally, the configuration attribute information includes hardware attribute information, software environment information, and model framework information that can be carried of the target satellite terminal;

[0035] The step of adapting the initial onboard large model to the target satellite terminal based on the configuration attribute information to obtain the onboard large model includes:

[0036] Determine the target model file format adapted by the target satellite terminal based on the hardware attribute information, the software environment information, and the model framework information;

[0037] Determine whether the model file format of the initial onboard large model matches the target model file format; if so, determine the initial onboard large model as the onboard large model; otherwise, convert the model file format of the initial onboard large model based on the target model file format, and determine the initial onboard large model after the conversion as the onboard large model;

[0038] After the initial onboard large model is adapted and adjusted for the target satellite terminal based on the configuration attribute information to obtain the onboard large model, the method further includes:

[0039] Acquire a test data set corresponding to the task attribute information, wherein the test data set includes sample input test data and annotation information corresponding to the sample input test data;

[0040] Based on the test data set, the onboard large model is tested to obtain the prediction accuracy of the onboard large model;

[0041] The step of deploying the onboard large model to the target satellite terminal comprises:

[0042] Determine whether the prediction accuracy of the onboard large model is greater than a preset accuracy threshold; if so, directly deploy the onboard large model to the target satellite terminal; otherwise, retrain, compress and adapt the initial large model.

[0043] Optionally, after deploying the onboard large model to the target satellite terminal, the method further includes:

[0044] Responding in real time to a parameter update signal of the onboard large model on the target satellite terminal, acquiring computing resource information of the target satellite terminal;

[0045] Based on the hardware attribute information, software environment information, and computing resource information of the target satellite terminal, determining whether the target satellite terminal has the ability to update model parameters;

[0046] If available, a new satellite-side data set adapted to the mission attribute information is obtained, and the model parameters of the onboard large model of the target satellite terminal are updated by using the new satellite-side data set to obtain the onboard large model after the parameters are updated; otherwise, a new ground-side data set adapted to the mission attribute information is obtained, and the initial large model is trained by using the new ground-side data set to obtain an updated mission large model, wherein the new satellite-side data set includes new satellite-side input data and actual annotation information corresponding to the new satellite-side input data, and the new ground-side data set includes new ground-side training data and actual annotation information corresponding to the new ground-side training data;

[0047] Compressing the updated mission large model to obtain an updated initial onboard large model;

[0048] Based on the configuration attribute information, the updated initial onboard large model is adapted and adjusted for the target satellite terminal to obtain an updated onboard large model, and the updated onboard large model is deployed to the target satellite terminal.

[0049] Optionally, after deploying the onboard large model to the target satellite terminal, the method further includes:

[0050] Responding to a task execution instruction inputted by the ground terminal through a control interface, and transmitting the task execution instruction to the satellite service computing device of the target satellite terminal through a receiving device of the ground terminal;

[0051] Controlling the satellite service computing device to transmit the task execution instruction to the satellite-borne computing device of the target satellite terminal, and using the satellite-borne computing device to parse the task requirement information from the task execution instruction;

[0052] Based on the task requirement information, the onboard large model of the target satellite terminal is called to execute the task, and the task execution result of the onboard large model is obtained, and the task execution result is sent to the control interface of the ground end for display.

[0053] According to a second aspect of the present invention, there is provided a large model deployment device on a satellite end, comprising:

[0054] an acquisition unit, configured to acquire configuration attribute information of the target satellite terminal and task attribute information of a predetermined task in response to a deployment signal of the target satellite terminal large model, and to acquire an initial large model;

[0055] A training unit, used to obtain a ground-side sample data set corresponding to the task attribute information, and use the ground-side sample data set to train the initial large model to obtain a task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data;

[0056] A determination unit, used to determine model attribute information of the task large model, and determine a target compression method applicable to the task large model based on the model attribute information and the task type of the predetermined task;

[0057] A compression unit, used for compressing the mission large model using the target compression method to obtain an initial onboard large model;

[0058] A deployment unit is used to adapt the initial onboard large model to the target satellite terminal based on the configuration attribute information to obtain the onboard large model, and deploy the onboard large model to the target satellite terminal.

[0059] According to a third aspect of the present invention, there is provided a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above method for deploying a large model on a satellite end.

[0060] According to a fourth aspect of the present invention, there is provided a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above method for deploying the large model on a satellite terminal when executing the program.

[0061] According to a method, device and storage medium for deploying a large model on a satellite side provided by the present invention, compared with the current method of directly deploying a ground large model on a satellite, the present invention obtains the configuration attribute information of the target satellite terminal and the task attribute information of a predetermined task, and obtains an initial large model by responding to a deployment signal of the target satellite terminal large model; and obtains a ground side sample data set corresponding to the task attribute information, and uses the ground side sample data set to train the initial large model to obtain a task large model, wherein the ground side sample data set includes sample input data and actual annotation information corresponding to the sample input data; at the same time, the model attribute information of the task large model is determined, and based on the model attribute information and the task type of the predetermined task, the target compression method applicable to the task large model is determined; then the task large model is compressed using the target compression method to obtain an initial satellite-borne large model; finally, based on the configuration attribute information, the initial satellite-borne large model is adapted and adjusted for the target satellite terminal to obtain a satellite-borne large model, and the satellite-borne large model is deployed to the target satellite terminal. Therefore, by training the large model on the ground side, since the computing resources on the ground side are relatively abundant, the abundant computing resources can support efficient training of the large model, thereby improving the training accuracy of the large model; at the same time, by comprehensively analyzing the model attribute information, task type, and configuration attribute information to select a suitable model compression method, the accuracy of the compression method can be improved, so that the compressed large model can maintain a high performance; before the trained large model is deployed on the satellite, by compressing the large model, the size of the model can be significantly reduced, and the demand for storage and computing resources can be reduced, so that the compressed model is easier to deploy and integrate on the satellite, reducing the difficulty and cost of deployment. At the same time, due to the reduction in the size of the model, the bandwidth and time of data transmission are also reduced, thereby improving the efficiency of deployment; before the large model is deployed on the satellite, by adapting and adjusting the large model, the stable and efficient operation of the large model on the hardware and operating system of the satellite can be ensured. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0063] Figure 1 A flow chart of a method for deploying a large model on a satellite terminal provided by an embodiment of the present invention is shown;

[0064] Figure 2 A flow chart of another method for deploying a large model on a satellite terminal provided by an embodiment of the present invention is shown;

[0065] Figure 3 A schematic diagram of the structure of a large model deployment device on a satellite end provided by an embodiment of the present invention is shown;

[0066] Figure 4 A schematic diagram of the structure of another large model deployment device on a satellite end provided by an embodiment of the present invention is shown;

[0067] Figure 5 A schematic diagram of the physical structure of a computer device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0068] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0069] At present, the method of directly deploying large ground models on satellites will increase the difficulty of deploying large models on satellites due to the complex structure of large models. At the same time, large models with complex structures require a lot of bandwidth and time during the transmission process, which will reduce the efficiency of deployment on satellites.

[0070] In order to solve the above problems, an embodiment of the present invention provides a method for deploying a large model on a satellite end, such as Figure 1 As shown, the method includes:

[0071] 101. In response to a deployment signal of a target satellite terminal large model, obtain configuration attribute information of the target satellite terminal and task attribute information of a predetermined task, and obtain an initial large model.

[0072] Among them, the configuration attribute information includes: hardware attribute information of the target satellite terminal, software environment information, and model framework information that the target satellite terminal can carry. The hardware attribute information package refers to the computing power of the target satellite terminal (usually the computing power can be determined by the processor type, main frequency, core number and other parameters carried by the target satellite terminal), storage capacity (usually the storage capacity can be determined by the memory and hard disk and other parameters of the target satellite terminal), energy consumption (energy consumption can be reflected by the power supply, heat dissipation capacity and other information of the target satellite terminal), communication capability (communication capability can be determined by the communication frequency band, rate and stability and other parameters of the target satellite terminal); the software environment information refers to the operating system attribute information of the target satellite terminal (operating system attribute information such as the type, version and stability of the operating system carried by the target satellite terminal), programming language attribute information (programming language attribute information such as the type and version of the programming language supported by the target satellite terminal), software library and tool attribute information (software library and tool attribute information such as the deep learning framework and data processing tools installed on the target satellite), and other information; the model framework information refers to the model framework compatibility, model size limit, model performance requirements and other information of the target satellite terminal. The task attribute information of the scheduled task includes: task scenario of the scheduled task (task scenarios such as task planning scenario, task scheduling scenario, data interpretation scenario, human-computer interaction scenario, etc.), task characteristics, task type (task types such as multi-target observation planning task type, emergency planning task type, sudden failure planning task type, orbit adjustment planning task type, etc. in task planning scenario; observation task type, data transmission task type, data calculation task type, real-time task type, customized task type, etc. in task scheduling scenario; image segmentation task type, target detection task type, change detection task type in data interpretation scenario; instruction understanding task type, instruction segmentation task type, function call task type, multi-modal output task type, etc. in human-computer interaction scenario) and other information.

[0073] For the embodiment of the present invention, first, a suitable model architecture is selected according to the task attribute information of the predetermined task, and the hierarchical structure of the model is determined according to factors such as the task complexity of the predetermined task, the computing resource consumption of executing the predetermined task, and the required model performance. Finally, an initial large model is constructed according to information such as the model architecture and the hierarchical structure. In another embodiment of the present invention, a corresponding initial large model can also be directly obtained in the network according to information such as task attributes, and the parameters of the initial large model are initialized.

[0074] 102. Obtain a ground-side sample data set corresponding to the task attribute information, and use the ground-side sample data set to train the initial large model to obtain the task large model.

[0075] Among them, the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data.

[0076] For the embodiment of the present invention, the ground-side sample data set is adapted to the predetermined task. For example, if the predetermined task is to identify the target object of a ship at sea, the ground-side sample data set includes a sample sea area image and the annotation information of the target in the sample sea area image. The annotation information can be information such as the annotation box of the target and the annotation type of the target in the annotation box. At this time, the satellite-borne large model should be a satellite-borne large model for ship identification. For example, if the predetermined task is to identify the movement or change of an object in a spatial position, the ground-side sample data set includes a spatial position image and the annotation information of the object change in the spatial position. The annotation information can be the annotation box of the target and the position coordinate information corresponding to the target in the annotation box. For different predetermined tasks, the model architecture of the corresponding initial large model can be the same or different, depending on information such as task complexity, task type, and data characteristics. When training the initial large model, the sample data set can be divided into a training data set and a test data set. For example, the sea area image in the training data set is used as input data, and the ship annotation information in the training data set is used as output data. The input data and the output data are used to train the initial large model, and then the test data set is used to test the trained initial large model. Finally, the trained initial large model that meets the test conditions is determined as the task large model, wherein the test conditions can be that the prediction accuracy of the model meets specific requirements, or the number of training times reaches a preset number. Thus, by using the ground-side sample data set to train the initial large model, since the ground-side sample data set is usually richer and more diverse, covering various scenarios and conditions, the large model trained on the ground side can learn more features and patterns, and improve the generalization ability of the model. At the same time, the computing resources on the ground side are usually richer and more powerful than those on the satellite platform, including high-performance CPU, GPU and other hardware acceleration devices, which enables the training of the initial large model on the ground side to make fuller use of these resources, improve the training efficiency of the large model, and thus improve the deployment capability of the large model on the satellite side.

[0077] 103. Determine model attribute information of the task model, and determine a target compression method applicable to the task model based on the model attribute information and the task type of the predetermined task.

[0078] Among them, the model attribute information includes the model type, parameter quantity, number of layers, storage space and computing resources required when executing the task, the required accuracy of the task model for the predetermined task, etc.; task types include: image classification tasks, natural language processing tasks, speech recognition tasks, information recommendation tasks, etc.; the target compression method can be knowledge transfer compression method, parameter removal compression method, matrix transformation compression method, model branch reduction compression method, weight sharing compression method, structure search compression method, etc.

[0079] For the embodiment of the present invention, the model attribute information and the task type of the scheduled task are comprehensively analyzed, and the target compression method corresponding to the task large model is determined according to the analysis results. For example, by analyzing the model attribute information, the key features of the model such as the type, structure, and parameter quantity can be understood, so as to select a compression method that matches it. This precise matching can ensure that the compressed model can minimize the volume and computational complexity while maintaining high performance. The task type determines the performance requirements that the model needs to meet. For example, for tasks with high real-time requirements, it may be necessary to select a compression method that can significantly improve the reasoning speed; and for tasks with high precision requirements, it is necessary to select a compression method that has less impact on precision. Comprehensive analysis of the task type helps to select the most appropriate compression strategy. At the same time, when selecting the target compression method, the configuration attribute information of the satellite terminal can also be comprehensively considered. The configuration attribute information of the satellite terminal includes computing resources, memory resources, storage resources, and power consumption restrictions. By comprehensively analyzing this information, a compression model that can run efficiently under the resource restrictions of the satellite terminal can be selected, which helps to avoid resource waste and improve resource utilization.

[0080] Specifically, an initial compression method prediction model can be pre-built, and a sample compression training set can be obtained, wherein the sample compression training set includes: sample model attribute information of the sample large model, sample task type, sample configuration attribute information of the sample satellite terminal, and the actual compression method corresponding to the sample large model, and the actual compression method can meet various compression requirements; the sample initial compression method prediction model is trained using the sample compression data set to obtain a preset compression method prediction model. Further, the model attribute information of the task large model, the task type of the scheduled task, and the configuration attribute information of the target satellite terminal are input into the preset compression method prediction model for prediction, and the target compression method applicable to the task large model is obtained.

[0081] 104. The mission large model is compressed using the target compression method to obtain the initial onboard large model.

[0082] For the embodiment of the present invention, after determining the target compression method applicable to the task large model, before deploying the task large model to the satellite terminal, it is also necessary to use the target compression method to perform compression optimization processing on the task large model, so as to reduce the storage and computing requirements of the model. When the target compression method is the knowledge transfer compression method, step 104 specifically includes: obtaining a basic single model and a first sample compressed data set corresponding to the task attribute information, and determining a temperature parameter based on the number of model parameters of the basic single model, wherein the number of model parameters of the basic single model is less than a preset number threshold, the structural complexity of the model structure of the basic single model is less than a preset complexity threshold, and the first sample compressed data set includes first sample input data and actual annotation information corresponding to the first sample input data; the first sample input data is input into the task large model for information prediction to obtain the task The task large model prediction information is inputted into the basic single model, and the first sample input data is inputted into the basic single model for information prediction to obtain the single model prediction information; based on the temperature parameter, the task large model prediction information is adjusted to obtain the adjusted task large model prediction information; based on the actual annotation information corresponding to the first sample input data and the single model prediction information, the cross entropy loss function of the basic single model is determined, and based on the adjusted task large model prediction information and the single model prediction information, the information gain loss function of the basic single model is determined; the weight coefficients corresponding to the cross entropy loss function and the information gain loss function are determined respectively, and based on the weight coefficients, the cross entropy loss function and the information gain loss function are added to obtain the comprehensive loss function; based on the comprehensive loss function, the basic single model is iteratively trained to obtain the initial satellite-borne large model.

[0083] Among them, the temperature parameter refers to the parameter for adjusting the softness of the output distribution of the large task model. When the temperature parameter is large, the model output distribution becomes smoother, which helps to improve the compression effect of the large task model; when the temperature parameter is small, the model output distribution becomes sharper, which will improve the prediction accuracy of the compressed model; the preset quantity threshold and the preset complexity threshold are set according to actual needs. For example, if the task type in the task attribute information is a target recognition task, the first sample input data in the first sample compressed data set can be a sea area image, and the actual annotation information corresponding to the first sample input data can be a ship annotation box in the sea area image. That is, the first sample data set will be different depending on the scheduled task. The weight coefficient is set according to actual needs.

[0084] Specifically, first obtain or design a basic single model with a simple structure and fewer parameters (relative to the task large model). The structure of the basic single model should have a certain similarity with the structure of the task large model, so that the knowledge in the task large model can be effectively transferred to the basic single model. At the same time, the basic single model should have a certain degree of customizability so that it can be adjusted and optimized according to different tasks and application scenarios. For example, different neural network architectures, hyperparameter settings, etc. can be selected according to specific needs. Then select a first sample compression data set of the same type and similar to the ground-side sample data set used to train the task large model. In the process of selecting the temperature parameter, if the number of parameters of the basic single model is small, its learning ability is relatively weak. At this time, in order to enable the basic single model to better learn the knowledge of the task large model, a smaller temperature parameter can be selected to make the softened model output distribution more sharp, thereby increasing the sensitivity of the basic single model to the difference between the output data. If the number of parameters of the basic single model is large, its learning ability is relatively strong. At this time, a larger temperature parameter can be selected to make the softened model output distribution smoother, so that the basic single model can learn more detailed information and feature representations in the task large model. Furthermore, in the model compression process, the task large model is first used to forward propagate the first sample input data in the first sample compressed data set to obtain the prediction information output by the task large model, and at the same time, the prediction information output by the task large model is adjusted according to the temperature parameter. For example, in the process of using the task large model to forward propagate the first sample input data in the first sample compressed data set, the input data propagated to the last layer of the task large model can be divided by the temperature parameter, and the division result can be input to the last layer of the task large model, and the adjusted task large model prediction information can be output through the last layer of the task large model. At the same time, the basic single model is used to forward propagate the first sample input data in the first sample compressed data set to obtain the prediction information output by the basic single model, and then the loss function between the basic single model prediction information and the adjusted task large model prediction information is calculated (usually measured by KL divergence (Kullback-Leibler divergence, information gain)), and the cross entropy loss function between the actual annotation information corresponding to the first sample input data and the single model prediction information is calculated. The two losses are then weighted and summed to obtain a comprehensive loss function. The parameters of the basic single model are updated through the back propagation algorithm to minimize the comprehensive loss function. The above steps are repeated to continuously iterate the training of the basic single model until its performance reaches the expected level or the number of training rounds reaches the preset value, and the basic single model that finally meets the training requirements is determined as the initial onboard large model.Furthermore, in order to further improve the model performance of the initial onboard large model, the basic single model can also be evaluated using a validation data set or a test data set to check its performance indicators such as accuracy and generalization ability, and according to the evaluation results, the structure and parameters of the basic single model can be adjusted to further improve its performance. Different temperature parameters, loss functions and other hyperparameter settings can also be tried to find the optimal model configuration, so as to obtain the optimal initial onboard large model. The embodiment of the present invention transfers the knowledge of the task large model to the basic single model, so that the basic single model can significantly reduce the number of model parameters, model structure and calculation amount while maintaining a high accuracy. The basic single model has a significantly increased operating efficiency when the model parameters are small, the calculation amount is small and the structure is simple. The model performs better in terms of reasoning speed, response time and other aspects, and is particularly suitable for application scenarios with high real-time requirements. At the same time, before deploying the large mission model to the satellite terminal, the embodiment of the present invention can significantly reduce the storage space occupied by the large mission model and reduce the complexity and cost of data transmission by compressing and optimizing the large mission model, because the storage space on the satellite is limited and data transmission is limited by bandwidth and energy. That is, it can improve the deployment efficiency of the large model and reduce the deployment cost. The compressed and optimized model has fewer parameters and a simpler structure, which means that the amount of calculation required for reasoning will also be reduced accordingly, which not only helps to improve the running speed of the model, but also reduces energy consumption and extends the service life of the satellite terminal.

[0085] 105. Based on the configuration attribute information, the initial onboard large model is adapted and adjusted for the target satellite terminal to obtain the onboard large model, and the onboard large model is deployed to the target satellite terminal.

[0086] For the embodiment of the present invention, the initial satellite-borne large model is adapted and adjusted according to the hardware attribute information, software environment information, model framework information that can be carried, and other information of the target satellite terminal, such as the CPU model, main frequency, number of cores, and other data of the satellite terminal to evaluate the computing requirements of the model; the memory size and storage type of the satellite terminal are obtained to determine the memory occupancy and storage requirements of the model. The operating system type and version used by the satellite terminal are determined to select compatible programming languages ​​and libraries. The programming languages ​​and common libraries supported by the satellite terminal are understood to determine the model framework supported by the satellite terminal, and a suitable model format is selected according to the model framework supported by the satellite terminal, and the model format of the initial satellite-borne large model is converted using a conversion tool to obtain a satellite-borne large model adapted to the satellite terminal, and finally the adapted satellite-borne large model is deployed to the satellite terminal. By adapting and adjusting the initial satellite-borne large model, the embodiment of the present invention can ensure that the model can make full use of the hardware and software resources of the satellite terminal to achieve the best performance, which helps to reduce the risk of model performance loss or failure due to compatibility issues. Satellite terminals usually have limited computing power and storage space. Adapting large models to these limitations can optimize the model's calculation process, reduce unnecessary resource consumption, and thus improve operational efficiency. The environment in which satellite terminals are located is complex and changeable, including temperature changes, radiation interference, etc. Adaptation can ensure that the model can still maintain stable operation in these harsh environments and reduce performance fluctuations or failures caused by environmental factors.

[0087] According to a method for deploying a large model on a satellite side provided by the present invention, compared with the current method of directly deploying a ground large model on a satellite, the present invention obtains configuration attribute information of the target satellite terminal and task attribute information of a predetermined task, and obtains an initial large model by responding to a deployment signal of a target satellite terminal large model; and obtains a ground-side sample data set corresponding to the task attribute information, and uses the ground-side sample data set to train the initial large model to obtain a task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data; at the same time, the model attribute information of the task large model is determined, and based on the model attribute information and the task type of the predetermined task, the target compression method applicable to the task large model is determined; then the task large model is compressed using the target compression method to obtain an initial satellite-borne large model; finally, based on the configuration attribute information, the initial satellite-borne large model is adapted and adjusted for the target satellite terminal to obtain a satellite-borne large model, and the satellite-borne large model is deployed to the target satellite terminal. Therefore, by training the large model on the ground side, since the computing resources on the ground side are relatively abundant, the abundant computing resources can support efficient training of the large model, thereby improving the training accuracy of the large model; at the same time, by comprehensively analyzing the model attribute information, task type, and configuration attribute information to select a suitable model compression method, the accuracy of the compression method can be improved, so that the compressed large model can maintain a high performance; before the trained large model is deployed on the satellite, by compressing the large model, the size of the model can be significantly reduced, and the demand for storage and computing resources can be reduced, so that the compressed model is easier to deploy and integrate on the satellite, reducing the difficulty and cost of deployment. At the same time, due to the reduction in the size of the model, the bandwidth and time of data transmission are also reduced, thereby improving the efficiency of deployment; before the large model is deployed on the satellite, by adapting and adjusting the large model, the stable and efficient operation of the large model on the hardware and operating system of the satellite can be ensured.

[0088] Further, in order to better illustrate the above process of classifying data, as a refinement and extension of the above embodiment, the embodiment of the present invention provides another method for deploying a large model on a satellite end, such as Figure 2 As shown, the method includes:

[0089] 201. In response to a deployment signal of a target satellite terminal large model, obtain configuration attribute information of the target satellite terminal and task attribute information of a scheduled task, and obtain an initial large model.

[0090] Specifically, when the deployment signal of the target satellite terminal large model is received, the configuration attribute information such as the hardware information and software information of the target satellite terminal and the task attribute information such as the task type, task scenario, and task characteristics of the task that the target satellite terminal wants to implement are obtained. At the same time, the initial large model of the appropriate architecture is selected based on the task type and other information.

[0091] 202. Obtain a ground-side sample data set corresponding to the task attribute information, and use the ground-side sample data set to train the initial large model to obtain the task large model.

[0092] Specifically, the ground-side sample data set is divided into a training data set and a test data set. The training data set is used to train the initial large model, and the test data set is used to test the trained initial large model. Finally, the trained initial large model that meets the test conditions is determined as the task large model.

[0093] 203. Determine model attribute information of the task model, and based on the model attribute information and the task type of the scheduled task, determine a target compression method applicable to the task model.

[0094] Specifically, in order to improve the accuracy of determining the target compression method, it is necessary to conduct a comprehensive analysis of information such as model attribute information and the task type of the scheduled task. For example, when a large and complex model needs to be deployed on a resource-constrained satellite terminal, the knowledge transfer compression method can transfer the knowledge of the large task model to a small model, thereby reducing the size and computational complexity of the model. When the large task model has a high accuracy requirement for the scheduled task, it is necessary to use the parameter removal compression method to compress the large task model. This method can remove the influence of noise in the model, reduce the risk of overfitting of the model, and improve the generalization ability of the model. In some complex machine learning models, such as deep learning models, there may be a large number of parameters and layers. Through the matrix transformation compression method, some layers or parameter matrices in the model can be reduced in dimension, thereby reducing the size and computational complexity of the model.

[0095] 204. The mission large model is compressed using a target compression method to obtain an initial onboard large model.

[0096] For the embodiment of the present invention, since the computing power of the satellite end is limited, it is difficult to deploy the full large model on the satellite end. Therefore, in order to reduce the difficulty of deploying the large model, it is necessary to compress the task large model using a target compression method. When the target compression method is a de-parameterization compression method, step 204 specifically includes: obtaining a second sample compressed data set corresponding to the task attribute information, wherein the second sample compressed data set includes second sample input data and actual annotation information corresponding to the second sample input data, inputting the second sample input data into the task large model for forward propagation, and determining the input of each network layer in the task large model during the forward propagation process. input data and output data of each network layer; determine the loss function of the task model based on the actual labeling information corresponding to the output data of the last network layer in the task model and the second sample input data; take any network layer except the last network layer in each network layer in the task model as a current network layer, determine the loss partial derivative of the loss function with respect to the output data of the next network layer corresponding to the current network layer, and determine the activation derivative of the activation function of the next network layer at the input data of the next network layer; multiply the loss partial derivative of the output data of the next network layer by the activation derivative at the input data of the next network layer, and obtain the activation derivative of the loss function with respect to the input data of the next network layer. The activation derivative of the activation function of the current network layer at the input data of the current network layer is determined, and the activation derivative at the input data of the current network layer is multiplied by the loss partial derivative of the input data of the subsequent network layer, and the product is multiplied by the transpose of the weight of the subsequent network layer to obtain the loss partial derivative of the loss function to the input data of the current network layer; the transpose of the output data of the current network layer is multiplied by the loss partial derivative of the input data of the subsequent network layer to obtain the sensitivity of the weight parameter in the subsequent network layer to the loss function, and the loss partial derivative of the input data of the subsequent network layer is determined as the sensitivity of the bias parameter in the subsequent network layer to the loss function. sensitivity; multiplying the transpose of the second sample input data by the loss partial derivative of the current network layer input data to obtain the sensitivity of the weight parameters in the current network layer to the loss function, and determining the loss partial derivative of the current network layer input data as the sensitivity of the bias parameters in the current network layer to the loss function; determining a target sensitivity less than a preset sensitivity threshold from the sensitivity of the weight parameters of each network layer in the task large model to the loss function and the sensitivity of the bias parameters to the loss function, determining the weight parameters and bias parameters corresponding to the target sensitivity as redundant parameters, and removing the redundant parameters in the task large model to obtain the initial onboard large model.

[0097] The second sample compressed data set is a data set of the same type and similar to the ground-side sample data set used by the training task large model. If the predetermined task is to perform target recognition, the input data in the second sample compressed data set is an image adapted to the task scenario, and the actual annotation information corresponding to the second sample input data is annotation information such as the target annotation box and target type in the image.

[0098] Specifically, the input data in the second sample compressed data set is first input into the task large model, and the predicted data is output by the task large model. According to the difference between the predicted data and the actual annotation information corresponding to the second sample input data, the loss function of the task large model (such as mean square error loss function, cross entropy loss function, etc.) is determined. At the same time, in the process of inputting the second sample input data into the task large model for forward propagation, the input data and output data of each layer in the task large model are recorded, and according to the input data and output data and loss function of each layer, the gradient of the loss function with respect to the parameters of each layer is determined, that is, the sensitivity of the parameters of each layer to the loss function. For example, if the task model includes an input layer, a hidden layer and an output layer, the activation function is sigmoid, and the loss function is mean square error (MSE), during the forward propagation of the second sample input data in the task model, if the input data of the hidden layer is z1 and the output data is a1, and the input data of the output layer is z2 and the output data is y (the data output by the last layer of the task model, that is, the predicted data of the task model is y), the hidden layer can be used as the current network layer, and the output layer can be used as the next network layer corresponding to the current network layer. First, the loss partial derivative of the loss function with respect to the output data of the output layer is calculated. The specific calculation formula is as follows:

[0099]

[0100] Among them, δ y is the partial derivative of the loss function with respect to the output data of the output layer, y is the output data of the output layer, is the actual annotation information corresponding to the input data in the second sample compression data set, and then the loss partial derivative of the loss function to the output layer input data is calculated according to the following formula:

[0101] δ2=δ y ·σ′(z2)

[0102] Among them, δ2 is the loss partial derivative of the loss function with respect to the input data of the output layer, z2 is the input data of the output layer, σ′ is the derivative of the activation function, and σ′(z2) is the activation derivative of the activation function of the output layer at the input data of the output layer. Then, according to the loss partial derivative of the input data and output data of the output layer, backpropagation is performed to the hidden layer, and the loss partial derivative of the loss function with respect to the input data of the hidden layer is calculated according to the following formula:

[0103] δ1=(W2) T δ2·σ′(z1)

[0104] Among them, δ1 is the loss partial derivative of the loss function with respect to the hidden layer input data, W2 is the weight of the output layer, δ2 is the loss partial derivative of the loss function with respect to the output layer input data, σ′(z1) is the activation derivative of the hidden layer activation function at the hidden layer input data, and z1 is the hidden layer input data. Further, the sensitivity of the weight parameter in the output layer to the loss function is calculated according to the following formula:

[0105]

[0106] in, is the sensitivity of the weight parameter W2 in the output layer to the loss function L, δ2 is the loss partial derivative of the loss function to the input data of the output layer, a1 is the output data of the hidden layer, (a1) T is the transpose of the hidden layer output data. Further, the sensitivity of the bias parameter b2 in the output layer to the loss function L is calculated according to the following formula:

[0107]

[0108] in, is the sensitivity of the bias parameter b2 in the output layer to the loss function L. Further, the sensitivity of the weight parameter W1 in the hidden layer to the loss function L is calculated according to the following formula:

[0109]

[0110] in, is the sensitivity of the weight parameter W1 in the hidden layer to the loss function L, x is the second sample input data, (x) T is the transpose of the second sample input data. Further, the sensitivity of the bias parameter b1 in the hidden layer to the loss function L is determined according to the following formula:

[0111]

[0112] in, is the sensitivity of the bias parameter b1 in the hidden layer to the loss function L. Thus, according to the above method, the sensitivity of the weight parameters of each network layer in the task large model to the loss function and the sensitivity of the bias parameters of each network layer to the loss function can be determined. The weight parameter corresponding to the target sensitivity whose sensitivity is less than the preset sensitivity threshold (the preset sensitivity threshold is a value set according to actual needs) is determined in the sensitivity of each weight parameter to the loss function, and the weight parameter is determined as a redundant parameter. Similarly, the bias parameter corresponding to the target sensitivity whose sensitivity is less than the preset sensitivity threshold is determined in the sensitivity of each bias parameter to the loss function, and the bias parameter is determined as a redundant parameter. Finally, all redundant parameters in the task large model are removed, and the large model after removing the redundant parameters is determined as the initial onboard large model. It should be noted that the preset sensitivity threshold for judging the redundant bias parameter and the preset sensitivity threshold for judging the redundant weight parameter can be the same or different, and are set specifically according to actual needs. Removing redundant weight parameters and redundant bias parameters can reduce the model's dependence on specific parameters, thereby improving the model's robustness and enabling large models to better maintain stable performance when facing noisy data or abnormal inputs. Since redundant weight parameters and bias parameters increase the amount of computation during model inference, removing these redundant parameters can significantly reduce the consumption of computing resources, which is particularly important for satellite environments with limited computing resources, ensuring that large models run efficiently on satellite terminals. Models with redundant parameters removed are easier to deploy and update on satellites, reducing bandwidth and time consumption, and also helping to reduce satellite energy consumption and costs.

[0113] In another embodiment of the present invention, when the target compression method is a matrix conversion compression method, the method for compressing the task large model includes: taking any network layer in the task large model as a target network layer, determining the weight matrix of the target network layer, and determining the transposed matrix of the weight matrix; multiplying the weight matrix by the transposed matrix to obtain a first matrix, and multiplying the transposed matrix by the weight matrix to obtain a second matrix; determining the first eigenvalue and the first eigenvector of the first matrix, and determining the second eigenvalue and the second eigenvector of the second matrix; constructing a first orthogonal matrix based on the first eigenvector, and constructing a second orthogonal matrix based on the second eigenvector; determining the square root of the first eigenvalue or the second eigenvalue, and constructing a diagonal matrix based on the square root; multiplying the first orthogonal matrix, the second orthogonal matrix, and the diagonal matrix to obtain a low-rank matrix, and using the low-rank matrix to replace the weight matrix of the target network layer, and determining the model constituted by the target network layer with the low-rank matrix as the initial onboard large model.

[0114] Specifically, first convert the weight parameters or bias parameters of the task model into a matrix form. For example, convert each weight parameter in each network layer in the task model into the weight matrix A of the corresponding network layer. The transposed matrix of the weight matrix A is A T , then determine AA T The eigenvalues ​​and eigenvectors of T The eigenvectors of the first orthogonal matrix V T , by A T The eigenvectors of A form the second orthogonal matrix U. Further, calculate AA T or A T The square root of the eigenvalue of A is obtained by arranging the square roots in descending order. The target square root greater than the preset threshold (the value set according to actual needs) is determined from the square roots after descending order, and the diagonal matrix Σ is constructed by each target square root. Then, the low-rank matrix A corresponding to the weight matrix is ​​determined according to the following formula: D :

[0115] A D =UΣV T

[0116] Furthermore, the low-rank matrix A D The weight matrix A is replaced, so that the low-rank matrix corresponding to the weight matrix in each network layer in the task large model can be calculated in the above manner, and the low-rank matrix is ​​used to replace the weight matrix in the corresponding network layer, and finally the model composed of each network layer with the low-rank matrix is ​​determined as the initial satellite-borne large model. The embodiment of the present invention constructs a diagonal matrix by a target square root greater than a preset threshold, which can reduce model parameters, remove redundant information and noise in the model, improve the generalization ability and robustness of the model, and finally deploy the compressed initial satellite-borne large model to the satellite terminal, which can reduce the model's consumption of bandwidth and time, thereby improving deployment efficiency and reducing deployment difficulty.

[0117] 205. Based on the hardware attribute information, software environment information, and model framework information of the target satellite terminal, determine the target model file format adapted by the target satellite terminal.

[0118] 206. Determine whether the model file format of the initial satellite-borne large model matches the target model file format. If so, determine the initial satellite-borne large model as the satellite-borne large model. Otherwise, convert the model file format of the initial satellite-borne large model based on the target model file format, and determine the converted initial satellite-borne large model as the satellite-borne large model.

[0119] For the embodiment of the present invention, the final target model file format is determined by comprehensively considering factors such as hardware compatibility, software environment matching and model framework. For example, the target model file format supported by the satellite terminal is determined according to the hardware attributes such as processor type, memory capacity and storage type, software attributes such as operating system and programming language and model framework of the satellite terminal. The target model file format can be correctly loaded and parsed by the software tools or libraries on the satellite terminal. At the same time, the model file format of the initial satellite-borne large model is determined. If the model file format matches the target model file format, that is, they are the same, the initial satellite-borne large model is directly determined as the satellite-borne large model and deployed on the satellite terminal. If the model file format does not match the target model file format, that is, they are not the same, it is necessary to use a conversion tool to convert the model file format of the initial satellite-borne large model into the target model file format, and determine the model after the conversion format as the satellite-borne large model to ensure that the model file format of the satellite-borne large model can be adapted on the satellite terminal, thereby ensuring the stable operation of the satellite-borne large model on the satellite terminal. The model file format is a script format corresponding to a certain programming language, including h5 file format (Hierarchical Data Format version 5), .ckpt file format (a model file format in the TensorFlow framework (an open source machine learning framework)), .pt file format (a model file format in the PyTorch framework (an open source machine learning framework)), etc. In another embodiment of the present invention, hardware attribute information, software environment information, and model framework information can be input into a preset model format prediction model for prediction to obtain a target model file format adapted by the target satellite terminal. The preset model format prediction model is obtained by training based on a sample data set, which includes sample hardware attribute information, sample software environment information, sample model framework information, and the actual model file format adapted by the sample satellite. The initial model is trained using the training data set in the sample data set. During the training of the initial model, the sample hardware attribute information, sample software environment information, and sample model framework information are used as input data, and the actual model file format is used as output data. The trained initial model is tested using the test data set in the sample data set, and finally the trained initial model that meets the test conditions is determined as the preset model format prediction model. The embodiment of the present invention can avoid the problem of the inability to deploy or the failure of the large model to run due to incompatibility of hardware, software, etc. by adapting and adjusting the initial satellite-borne large model, thereby ensuring the smooth operation of the large model in the satellite terminal.

[0120] 207. Deploy the onboard large model to the target satellite terminal.

[0121] Specifically, after determining the onboard large model, in order to verify the prediction accuracy of the onboard large model, it is also necessary to verify the prediction accuracy of the onboard large model, and deploy the satellite terminal based on the verification result. Based on this, the method includes: obtaining a test data set corresponding to the task attribute information, wherein the test data set includes sample input test data and annotation information corresponding to the sample input test data; based on the test data set, testing the onboard large model to obtain the prediction accuracy of the onboard large model; judging whether the prediction accuracy of the onboard large model is greater than a preset accuracy threshold, if so, directly deploying the onboard large model to the target satellite terminal, otherwise, retraining, compressing and adapting the initial large model.

[0122] Among them, the test data set and the ground-side sample data set used by the training task large model are data sets of the same type and similar. The preset accuracy threshold is set according to actual needs. Specifically, if the sample input test data in the test data set is a regional image, the annotation information corresponding to the sample input test data is the annotation box and object category corresponding to the target object in the regional image. The sample input test data is input into the satellite-borne large model for information prediction to obtain the satellite-borne large model prediction information. Based on the satellite-borne large model prediction information and the annotation information corresponding to the sample input test data, the prediction accuracy of the satellite-borne large model is determined. If the prediction accuracy is greater than the preset accuracy threshold, the satellite-borne large model is directly deployed to the satellite terminal. If the prediction accuracy is less than or equal to the preset accuracy threshold, the initial large model needs to be retrained, compressed, adapted and adjusted until a satellite-borne large model with a prediction accuracy greater than the preset accuracy threshold is obtained. The embodiment of the present invention can ensure the prediction accuracy of the satellite-borne large model by testing the satellite-borne large model before deploying it to the satellite terminal.

[0123] Furthermore, after the onboard large model is deployed to the satellite terminal, in order to meet the needs of various tasks, it is also necessary to update the onboard large model of the satellite terminal. Based on this, the method includes: responding to the parameter update signal of the onboard large model on the target satellite terminal in real time to obtain the computing resource information of the target satellite terminal; judging whether the target satellite terminal has the ability to update the model parameters based on the hardware attribute information, software environment information, and computing resource information of the target satellite terminal; if so, obtaining a new satellite-side data set that is compatible with the task attribute information, and using the new satellite-side data set to update the model parameters of the onboard large model of the target satellite terminal, and obtaining the onboard large model after the parameters are updated. large model, otherwise, obtain a new ground-side data set that is compatible with the task attribute information, and use the new ground-side data set to train the initial large model to obtain an updated task large model, wherein the new satellite-side data set includes new input data from the satellite and actual annotation information corresponding to the new input data from the satellite, and the new ground-side data set includes new training data from the ground and actual annotation information corresponding to the new training data from the ground; compress the updated task large model to obtain an updated initial satellite-borne large model; based on the configuration attribute information, adapt the updated initial satellite-borne large model to the target satellite terminal to obtain an updated satellite-borne large model, and deploy the updated satellite-borne large model to the target satellite terminal.

[0124] Among them, computing resource information includes the satellite terminal's processor performance information (such as processor model, number of cores, number of threads, main frequency and turbo frequency, etc.), memory configuration (such as memory capacity, memory type and speed), storage space (such as storage capacity, storage type, etc.), network interface performance, etc.

[0125] Specifically, first, according to the hardware attribute information of the target satellite terminal, the computing power of the target satellite terminal is determined, and according to the software environment information, whether the target satellite terminal has an update tool and framework that can support the model parameters, and whether the target satellite terminal is installed with a programming language that supports the model parameter update, according to the computing resource information, whether the memory, resources, etc. of the satellite terminal can meet the parameter update requirements, if the computing power of the target satellite terminal is greater than a preset threshold (set according to actual needs), and the target satellite terminal has an update tool and framework that can support the model parameters, and the target satellite terminal is installed with a programming language that supports the model parameter update, and the memory, resources, etc. of the satellite terminal can meet the parameter update requirements, then it is determined that the target satellite terminal has the ability to update the model parameters, if any of the above conditions is not met, then it is determined that the target satellite terminal does not have the ability to update the model parameters. If the target satellite terminal has the ability to update the model parameters, at this time, a new satellite terminal data set corresponding to the mission attribute information is obtained at the target satellite terminal, and the new satellite terminal data set is a data set of the same type as the sample data set used to train the initial large model. The satellite-side new data set is used to directly iteratively train the satellite-borne large model, so as to update the model parameters of the satellite-borne large model. By directly using the satellite terminal data set to train the satellite-borne model, the time for transmitting the ground-side data to the satellite terminal is reduced, thereby improving the updating efficiency of the satellite-side large model. Furthermore, if the target satellite terminal does not have the ability to update the model parameters, it is necessary to obtain a new ground-side data set at the ground end (the ground-side data set is of the same type as the satellite-side new data set), and use the ground-side data set to update the satellite-borne large model. Alternatively, the initial large model is retrained, compressed, adapted and adjusted using the ground-side data set to obtain an updated satellite-borne large model, and the updated satellite-borne large model is deployed to the target satellite terminal. In the case where the satellite terminal does not have the model update capability, the embodiment of the present invention updates the model at the ground end, which can improve the updating accuracy of the model.

[0126] Furthermore, after the onboard large model is deployed to the target satellite terminal, it is necessary to use the deployed onboard large model to perform predetermined task processing. Based on this, the method includes: responding to a task execution instruction input by the ground end through a control interface, and transmitting the task execution instruction to the satellite service computing device of the target satellite terminal through the receiving device of the ground end; controlling the satellite service computing device to transmit the task execution instruction to the onboard computing device of the target satellite terminal, and using the onboard computing device to parse out task requirement information from the task execution instruction; based on the task requirement information, calling the onboard large model of the target satellite terminal to execute the task, and obtaining the task execution result of the onboard large model, and sending the task execution result to the control interface of the ground end for display.

[0127] Among them, the task requirement information refers to the task requirements that need to be performed by the satellite-borne large model, such as ship identification in a certain sea area, vehicle identification in a certain area, etc. Specifically, for example, the control center: inputs the task execution instruction through the satellite-borne large model web page: please shoot the sea area with ships in the current field of view; the control center transmits the task execution instruction to the satellite computer through the ground receiving station; after the satellite computer receives the task execution instruction, it transmits the task execution instruction to the satellite computer, the satellite computer parses the task execution instruction, calls the background task scheduling large model and the interpretation large model function module, and starts searching for and detecting ships in the current field of view. If there is a ship in the field of view, the current video screenshot is saved and then transmitted to the satellite computer; the satellite computer sends the screenshot of the ship to the ground receiving station, and the ground receiving station sends the screenshot to the control center; the control center displays the received screenshot in the dialog box of the satellite-borne large model web page. At this point, the task is completed.

[0128] According to another method for deploying a large model on a satellite side provided by the present invention, compared with the current method of directly deploying a ground large model on a satellite, the present invention obtains configuration attribute information of the target satellite terminal and task attribute information of a predetermined task, as well as an initial large model, in response to a deployment signal of a target satellite terminal large model; and obtains a ground side sample data set corresponding to the task attribute information, and uses the ground side sample data set to train the initial large model to obtain a task large model, wherein the ground side sample data set includes sample input data and actual annotation information corresponding to the sample input data; at the same time, the model attribute information of the task large model is determined, and based on the model attribute information and the task type of the predetermined task, the target compression method applicable to the task large model is determined; then the task large model is compressed using the target compression method to obtain an initial satellite-borne large model; finally, based on the configuration attribute information, the initial satellite-borne large model is adapted and adjusted for the target satellite terminal to obtain a satellite-borne large model, and the satellite-borne large model is deployed to the target satellite terminal. Therefore, by training the large model on the ground side, since the computing resources on the ground side are relatively abundant, the abundant computing resources can support efficient training of the large model, thereby improving the training accuracy of the large model; at the same time, by comprehensively analyzing the model attribute information, task type, and configuration attribute information to select a suitable model compression method, the accuracy of the compression method can be improved, so that the compressed large model can maintain a high performance; before the trained large model is deployed on the satellite, by compressing the large model, the size of the model can be significantly reduced, and the demand for storage and computing resources can be reduced, so that the compressed model is easier to deploy and integrate on the satellite, reducing the difficulty and cost of deployment. At the same time, due to the reduction in the size of the model, the bandwidth and time of data transmission are also reduced, thereby improving the efficiency of deployment; before the large model is deployed on the satellite, by adapting and adjusting the large model, the stable and efficient operation of the large model on the hardware and operating system of the satellite can be ensured.

[0129] Further, as Figure 1 In the specific implementation, the embodiment of the present invention provides a large model deployment device on the satellite end, such as Figure 3 As shown, the device includes: an acquisition unit 31, a training unit 32, a determination unit 33, a compression unit 34, and a deployment unit 35.

[0130] The acquisition unit 31 may be used to acquire configuration attribute information of the target satellite terminal and task attribute information of a predetermined task in response to a deployment signal of the target satellite terminal large model, and to acquire an initial large model.

[0131] The training unit 32 can be used to obtain a ground-end sample data set corresponding to the task attribute information, and use the ground-end sample data set to train the initial large model to obtain a task large model, wherein the ground-end sample data set includes sample input data and actual annotation information corresponding to the sample input data.

[0132] The determination unit 33 may be used to determine the model attribute information of the task large model, and determine the target compression method applicable to the task large model based on the model attribute information and the task type of the predetermined task.

[0133] The compression unit 34 can be used to compress the mission large model using the target compression method to obtain an initial onboard large model.

[0134] The deployment unit 35 may be configured to adapt the initial onboard large model to the target satellite terminal based on the configuration attribute information, obtain the onboard large model, and deploy the onboard large model to the target satellite terminal.

[0135] In specific application scenarios, in order to compress large task models, such as Figure 4 As shown, the compression unit 34 includes: an acquisition module 341, an information prediction module 342, an adjustment module 343, a first determination module 344, an addition module 345, and a training module 346.

[0136] The acquisition module 341 can be used to obtain a basic single model and a first sample compressed data set corresponding to the task attribute information, and determine the temperature parameter based on the number of model parameters of the basic single model, wherein the number of model parameters of the basic single model is less than a preset number threshold, the structural complexity of the model structure of the basic single model is less than a preset complexity threshold, and the first sample compressed data set includes first sample input data and actual annotation information corresponding to the first sample input data.

[0137] The information prediction module 342 can be used to input the first sample input data into the task big model for information prediction to obtain the task big model prediction information, and to input the first sample input data into the basic single model for information prediction to obtain the single model prediction information.

[0138] The adjustment module 343 can be used to adjust the task large model prediction information based on the temperature parameter to obtain the adjusted task large model prediction information.

[0139] The first determination module 344 can be used to determine the cross entropy loss function of the basic single model based on the actual annotation information corresponding to the first sample input data and the single model prediction information, and to determine the information gain loss function of the basic single model based on the adjusted task large model prediction information and the single model prediction information.

[0140] The adding module 345 can be used to determine the weight coefficients corresponding to the cross entropy loss function and the information gain loss function respectively, and based on the weight coefficients, add the cross entropy loss function and the information gain loss function to obtain a comprehensive loss function.

[0141] The training module 346 can be used to iteratively train the basic single model based on the comprehensive loss function to obtain the initial onboard large model.

[0142] In a specific application scenario, in order to compress the large task model, the compression unit 34 also includes a multiplication module 347 and a parameter removal module 348.

[0143] The acquisition module 341 can also be used to obtain a second sample compressed data set corresponding to the task attribute information, wherein the second sample compressed data set includes second sample input data and actual annotation information corresponding to the second sample input data, and the second sample input data is input into the task large model for forward propagation, and the input data and output data of each network layer in the task large model are determined during the forward propagation process.

[0144] The first determination module 344 can also be used to determine the loss function of the task model based on the actual labeling information corresponding to the last network layer output data in the task model and the second sample input data.

[0145] The first determination module 344 can also be used to take any network layer except the last network layer in each network layer in the task large model as a current network layer, determine the loss partial derivative of the loss function with respect to the output data of the subsequent network layer corresponding to the current network layer, and determine the activation derivative of the activation function of the subsequent network layer at the input data of the subsequent network layer.

[0146] The multiplication module 347 can be used to multiply the loss partial derivative of the subsequent network layer output data with the activation derivative at the subsequent network layer input data to obtain the loss partial derivative of the loss function with respect to the subsequent network layer input data.

[0147] The multiplication module 347 can also be used to determine the activation derivative of the activation function of the current network layer at the input data of the current network layer, multiply the activation derivative at the input data of the current network layer by the loss partial derivative of the input data of the subsequent network layer, and multiply the product by the transpose of the weight of the subsequent network layer to obtain the loss partial derivative of the loss function with respect to the input data of the current network layer.

[0148] The multiplication module 347 can also be used to multiply the transpose of the current network layer output data with the loss partial derivative of the subsequent network layer input data to obtain the sensitivity of the weight parameter in the subsequent network layer to the loss function, and determine the loss partial derivative of the subsequent network layer input data as the sensitivity of the bias parameter in the subsequent network layer to the loss function.

[0149] The multiplication module 347 can also be used to multiply the transpose of the second sample input data by the loss partial derivative of the current network layer input data to obtain the sensitivity of the weight parameters in the current network layer to the loss function, and determine the loss partial derivative of the current network layer input data as the sensitivity of the bias parameters in the current network layer to the loss function.

[0150] The parameter removal module 348 can be used to determine the target sensitivity that is less than a preset sensitivity threshold among the sensitivity of the weight parameters of each network layer in the mission large model to the loss function and the sensitivity of the bias parameters to the loss function, determine the weight parameters and bias parameters corresponding to the target sensitivity as redundant parameters, and remove the redundant parameters in the mission large model to obtain the initial satellite-borne large model.

[0151] In a specific application scenario, in order to compress the large task model, the first determination module 344 can also be used to take any network layer in the large task model as a target network layer, determine the weight matrix of the target network layer, and determine the transposed matrix of the weight matrix.

[0152] The multiplication module 347 may also be used to multiply the weight matrix and the transposed matrix to obtain a first matrix, and to multiply the transposed matrix and the weight matrix to obtain a second matrix.

[0153] The first determination module 344 may also be used to determine a first eigenvalue and a first eigenvector of the first matrix, and to determine a second eigenvalue and a second eigenvector of the second matrix.

[0154] The first determination module 344 may also be configured to construct a first orthogonal matrix based on the first eigenvector, and to construct a second orthogonal matrix based on the second eigenvector.

[0155] The first determination module 344 may also be configured to determine a square root of the first eigenvalue or the second eigenvalue, and construct a diagonal matrix based on the square root.

[0156] The multiplication module 347 can also be used to multiply the first orthogonal matrix, the second orthogonal matrix, and the diagonal matrix to obtain a low-rank matrix, and use the low-rank matrix to replace the weight matrix of the target network layer, and determine the model constituted by the target network layer with the low-rank matrix as the initial satellite-borne large model.

[0157] In a specific application scenario, the configuration attribute information includes the hardware attribute information, software environment information, and model framework information that can be carried of the target satellite terminal. In order to adapt and adjust the initial onboard large model, the deployment unit 35 includes a second determination module 351 and a format conversion module 352.

[0158] The second determination module 351 may be configured to determine the target model file format adapted by the target satellite terminal based on the hardware attribute information, the software environment information, and the model framework information.

[0159] The format conversion module 352 can be used to determine whether the model file format of the initial satellite-borne large model matches the target model file format. If so, the initial satellite-borne large model is determined as the satellite-borne large model. Otherwise, based on the target model file format, the model file format of the initial satellite-borne large model is converted, and the initial satellite-borne large model after the conversion is determined as the satellite-borne large model.

[0160] In a specific application scenario, in order to test the large satellite model, the device further includes: a testing unit 36.

[0161] The testing unit 36 ​​can be used to obtain a test data set corresponding to the task attribute information, wherein the test data set includes sample input test data and annotation information corresponding to the sample input test data; based on the test data set, the satellite-borne large model is tested to obtain the prediction accuracy of the satellite-borne large model.

[0162] In a specific application scenario, in order to deploy the onboard large model to the target satellite terminal, the deployment unit 35 can be specifically used to determine whether the prediction accuracy of the onboard large model is greater than a preset accuracy threshold. If so, the onboard large model is directly deployed to the target satellite terminal; otherwise, the initial large model is re-trained, compressed and adapted.

[0163] In a specific application scenario, in order to update the deployed large satellite model, the device further includes: an updating unit 37 .

[0164] The updating unit 37 can be used to respond to the parameter update signal of the onboard large model on the target satellite terminal in real time, and obtain the computing resource information of the target satellite terminal; based on the hardware attribute information, software environment information, and computing resource information of the target satellite terminal, determine whether the target satellite terminal has the ability to update model parameters; if so, obtain a new satellite-side data set that is compatible with the mission attribute information, and use the new satellite-side data set to update the model parameters of the onboard large model of the target satellite terminal to obtain the onboard large model after the updated parameters; otherwise, obtain a new ground-side data set that is compatible with the mission attribute information The method comprises the following steps: obtaining a data set, and using the new data set on the ground side to train the initial large model to obtain an updated task large model, wherein the new data set on the satellite side includes new input data on the satellite side and actual annotation information corresponding to the new input data on the satellite side, and the new data set on the ground side includes new training data on the ground side and actual annotation information corresponding to the new training data on the ground side; compressing the updated task large model to obtain an updated initial satellite-borne large model; based on the configuration attribute information, adapting the updated initial satellite-borne large model to the target satellite terminal to obtain an updated satellite-borne large model, and deploying the updated satellite-borne large model to the target satellite terminal.

[0165] In a specific application scenario, in order to control the onboard large model of the target satellite terminal to perform a task, the device further includes: a task execution unit 38.

[0166] The task execution unit 38 can be used to respond to the task execution instruction input by the ground end through the control interface, and transmit the task execution instruction to the satellite computing device of the target satellite terminal through the receiving device of the ground end; control the satellite computing device to transmit the task execution instruction to the satellite computing device of the target satellite terminal, and use the satellite computing device to parse the task requirement information in the task execution instruction; based on the task requirement information, call the satellite large model of the target satellite terminal to execute the task, and obtain the task execution result of the satellite large model, and send the task execution result to the control interface of the ground end for display.

[0167] It should be noted that for other corresponding descriptions of the functional modules involved in the deployment device of a large model on a satellite end provided in an embodiment of the present invention, reference can be made to Figure 1 The corresponding description of the method shown will not be repeated here.

[0168] Based on the above Figure 1The method shown, accordingly, an embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the following steps are implemented: in response to a deployment signal of a target satellite terminal large model, obtaining configuration attribute information of the target satellite terminal and task attribute information of a predetermined task, and obtaining an initial large model; obtaining a ground-end sample data set corresponding to the task attribute information, and using the ground-end sample data set to train the initial large model to obtain a task large model, wherein the ground-end sample data set includes sample input data and actual annotation information corresponding to the sample input data; determining the model attribute information of the task large model, and based on the model attribute information and the task type of the predetermined task, determining a target compression method applicable to the task large model; compressing the task large model using the target compression method to obtain an initial satellite-borne large model; based on the configuration attribute information, adapting the initial satellite-borne large model to the target satellite terminal to obtain a satellite-borne large model, and deploying the satellite-borne large model to the target satellite terminal.

[0169] Based on the above Figure 1 The method shown and Figure 3 The embodiment of the device shown in the figure, the embodiment of the present invention also provides a physical structure diagram of a computer device, such as Figure 5 As shown, the computer device includes: a processor 41, a memory 42, and a computer program stored in the memory 42 and executable on the processor, wherein the memory 42 and the processor 41 are both arranged on a bus 43. When the processor 41 executes the program, the following steps are implemented: in response to a deployment signal of a target satellite terminal large model, configuration attribute information of the target satellite terminal and task attribute information of a predetermined task are obtained, and an initial large model is obtained; a ground-side sample data set corresponding to the task attribute information is obtained, and the initial large model is trained using the ground-side sample data set to obtain a task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data; model attribute information of the task large model is determined, and a target compression method applicable to the task large model is determined based on the model attribute information and the task type of the predetermined task; the task large model is compressed using the target compression method to obtain an initial satellite-borne large model; based on the configuration attribute information, the initial satellite-borne large model is adapted and adjusted for the target satellite terminal to obtain a satellite-borne large model, and the satellite-borne large model is deployed to the target satellite terminal.

[0170] Through the technical scheme of the present invention, the present invention obtains the configuration attribute information of the target satellite terminal and the task attribute information of the scheduled task, and obtains the initial large model in response to the deployment signal of the target satellite terminal large model; and obtains a ground-side sample data set corresponding to the task attribute information, and uses the ground-side sample data set to train the initial large model to obtain the task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data; at the same time, the model attribute information of the task large model is determined, and based on the model attribute information and the task type of the scheduled task, the target compression method applicable to the task large model is determined; then the task large model is compressed using the target compression method to obtain the initial satellite-borne large model; finally, based on the configuration attribute information, the initial satellite-borne large model is adapted and adjusted for the target satellite terminal to obtain the satellite-borne large model, and the satellite-borne large model is deployed to the target satellite terminal. Therefore, by training the large model on the ground side, since the computing resources on the ground side are relatively abundant, the abundant computing resources can support efficient training of the large model, thereby improving the training accuracy of the large model; at the same time, by comprehensively analyzing the model attribute information, task type, and configuration attribute information to select a suitable model compression method, the accuracy of the compression method can be improved, so that the compressed large model can maintain a high performance; before the trained large model is deployed on the satellite, by compressing the large model, the size of the model can be significantly reduced, and the demand for storage and computing resources can be reduced, so that the compressed model is easier to deploy and integrate on the satellite, reducing the difficulty and cost of deployment. At the same time, due to the reduction in the size of the model, the bandwidth and time of data transmission are also reduced, thereby improving the efficiency of deployment; before the large model is deployed on the satellite, by adapting and adjusting the large model, the stable and efficient operation of the large model on the hardware and operating system of the satellite can be ensured.

[0171] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0172] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for deploying a large model on a satellite terminal, characterized in that: include: In response to a deployment signal of a target satellite terminal large model, obtaining configuration attribute information of the target satellite terminal and task attribute information of a predetermined task, and obtaining an initial large model; Acquire a ground-side sample data set corresponding to the task attribute information, and use the ground-side sample data set to train the initial large model to obtain a task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data; Determine model attribute information of the task large model, and determine a target compression method applicable to the task large model based on the model attribute information and the task type of the predetermined task; Using the target compression method to compress the mission large model to obtain an initial onboard large model; Based on the configuration attribute information, the initial onboard large model is adapted and adjusted for the target satellite terminal to obtain the onboard large model, and the onboard large model is deployed to the target satellite terminal.

2. The method according to claim 1, characterized in that If the target compression method is a knowledge transfer compression method, the task large model is compressed using the target compression method to obtain an initial onboard large model, including: Acquire a basic single model and a first sample compressed data set corresponding to the task attribute information, and determine a temperature parameter based on the number of model parameters of the basic single model, wherein the number of model parameters of the basic single model is less than a preset number threshold, the structural complexity of the model structure of the basic single model is less than a preset complexity threshold, and the first sample compressed data set includes first sample input data and actual annotation information corresponding to the first sample input data; Inputting the first sample input data into the task big model for information prediction to obtain task big model prediction information, and inputting the first sample input data into the basic single model for information prediction to obtain single model prediction information; Based on the temperature parameter, the task large model prediction information is adjusted to obtain adjusted task large model prediction information; Determine the cross entropy loss function of the basic single model based on the actual annotation information corresponding to the first sample input data and the single model prediction information, and determine the information gain loss function of the basic single model based on the adjusted task large model prediction information and the single model prediction information; Determine the weight coefficients corresponding to the cross entropy loss function and the information gain loss function respectively, and based on the weight coefficients, add the cross entropy loss function and the information gain loss function to obtain a comprehensive loss function; Based on the comprehensive loss function, the basic single model is iteratively trained to obtain the initial satellite-borne large model.

3. The method according to claim 1, characterized in that If the target compression method is a parameter removal compression method, the target compression method is used to compress the mission large model to obtain an initial onboard large model, including: Acquire a second sample compressed data set corresponding to the task attribute information, wherein the second sample compressed data set includes second sample input data and actual annotation information corresponding to the second sample input data, input the second sample input data into the task large model for forward propagation, and determine the input data of each network layer and the output data of each network layer in the task large model during the forward propagation process; Determine the loss function of the task large model based on the actual annotation information corresponding to the last network layer output data in the task large model and the second sample input data; Taking any network layer except the last network layer in each network layer in the task large model as a current network layer, determining the loss partial derivative of the loss function with respect to the output data of the subsequent network layer corresponding to the current network layer, and determining the activation derivative of the activation function of the subsequent network layer at the input data of the subsequent network layer; Multiplying the loss partial derivative of the output data of the subsequent network layer by the activation derivative at the input data of the subsequent network layer to obtain the loss partial derivative of the loss function with respect to the input data of the subsequent network layer; Determine the activation derivative of the activation function of the current network layer at the input data of the current network layer, multiply the activation derivative at the input data of the current network layer by the loss partial derivative of the input data of the subsequent network layer, and multiply the product by the transpose of the weight of the subsequent network layer to obtain the loss partial derivative of the loss function with respect to the input data of the current network layer; Multiplying the transpose of the current network layer output data by the loss partial derivative of the subsequent network layer input data to obtain the sensitivity of the weight parameter in the subsequent network layer to the loss function, and determining the loss partial derivative of the subsequent network layer input data as the sensitivity of the bias parameter in the subsequent network layer to the loss function; Multiplying the transpose of the second sample input data by the loss partial derivative of the current network layer input data to obtain the sensitivity of the weight parameter in the current network layer to the loss function, and determining the loss partial derivative of the current network layer input data as the sensitivity of the bias parameter in the current network layer to the loss function; A target sensitivity less than a preset sensitivity threshold is determined among the sensitivity of the weight parameters of each network layer in the mission large model to the loss function and the sensitivity of the bias parameters to the loss function, the weight parameters and bias parameters corresponding to the target sensitivity are determined as redundant parameters, and the redundant parameters in the mission large model are removed to obtain the initial satellite-borne large model.

4. The method according to claim 1, characterized in that: If the target compression method is a matrix conversion compression method, the target compression method is used to compress the mission large model to obtain an initial onboard large model, including: Taking any network layer in the task model as a target network layer, determining a weight matrix of the target network layer, and determining a transposed matrix of the weight matrix; Multiplying the weight matrix by the transposed matrix to obtain a first matrix, and multiplying the transposed matrix by the weight matrix to obtain a second matrix; Determine a first eigenvalue and a first eigenvector of the first matrix, and determine a second eigenvalue and a second eigenvector of the second matrix; Based on the first eigenvector, construct a first orthogonal matrix, and based on the second eigenvector, construct a second orthogonal matrix; Determine a square root of the first eigenvalue or the second eigenvalue, and construct a diagonal matrix based on the square root; The first orthogonal matrix, the second orthogonal matrix and the diagonal matrix are multiplied to obtain a low-rank matrix, and the low-rank matrix is ​​used to replace the weight matrix of the target network layer, and the model constituted by the target network layer with the low-rank matrix is ​​determined as the initial onboard large model.

5. The method according to claim 1, characterized in that The configuration attribute information includes hardware attribute information, software environment information, and model framework information that can be carried of the target satellite terminal; The step of adapting the initial onboard large model to the target satellite terminal based on the configuration attribute information to obtain the onboard large model includes: Determine the target model file format adapted by the target satellite terminal based on the hardware attribute information, the software environment information, and the model framework information; Determine whether the model file format of the initial onboard large model matches the target model file format; if so, determine the initial onboard large model as the onboard large model; otherwise, convert the model file format of the initial onboard large model based on the target model file format, and determine the initial onboard large model after the conversion as the onboard large model; After the initial onboard large model is adapted and adjusted for the target satellite terminal based on the configuration attribute information to obtain the onboard large model, the method further includes: Acquire a test data set corresponding to the task attribute information, wherein the test data set includes sample input test data and annotation information corresponding to the sample input test data; Based on the test data set, the onboard large model is tested to obtain the prediction accuracy of the onboard large model; The step of deploying the onboard large model to the target satellite terminal comprises: Determine whether the prediction accuracy of the onboard large model is greater than a preset accuracy threshold; if so, directly deploy the onboard large model to the target satellite terminal; otherwise, retrain, compress and adapt the initial large model.

6. The method according to claim 1, characterized in that After deploying the onboard large model to the target satellite terminal, the method further includes: Responding in real time to a parameter update signal of the onboard large model on the target satellite terminal, acquiring computing resource information of the target satellite terminal; Based on the hardware attribute information, software environment information, and computing resource information of the target satellite terminal, determining whether the target satellite terminal has the ability to update model parameters; If available, a new satellite-side data set adapted to the mission attribute information is obtained, and the model parameters of the onboard large model of the target satellite terminal are updated by using the new satellite-side data set to obtain the onboard large model after the parameters are updated; otherwise, a new ground-side data set adapted to the mission attribute information is obtained, and the initial large model is trained by using the new ground-side data set to obtain an updated mission large model, wherein the new satellite-side data set includes new satellite-side input data and actual annotation information corresponding to the new satellite-side input data, and the new ground-side data set includes new ground-side training data and actual annotation information corresponding to the new ground-side training data; Compressing the updated mission large model to obtain an updated initial onboard large model; Based on the configuration attribute information, the updated initial onboard large model is adapted and adjusted for the target satellite terminal to obtain an updated onboard large model, and the updated onboard large model is deployed to the target satellite terminal.

7. The method according to claim 1, characterized in that After deploying the onboard large model to the target satellite terminal, the method further includes: Responding to a task execution instruction inputted by the ground terminal through a control interface, and transmitting the task execution instruction to the satellite service computing device of the target satellite terminal through a receiving device of the ground terminal; Controlling the satellite service computing device to transmit the task execution instruction to the satellite-borne computing device of the target satellite terminal, and using the satellite-borne computing device to parse the task requirement information from the task execution instruction; Based on the task requirement information, the onboard large model of the target satellite terminal is called to execute the task, and the task execution result of the onboard large model is obtained, and the task execution result is sent to the control interface of the ground end for display.

8. A large model deployment device on a satellite, characterized in that: include: an acquisition unit, configured to acquire configuration attribute information of the target satellite terminal and task attribute information of a predetermined task in response to a deployment signal of the target satellite terminal large model, and to acquire an initial large model; A training unit, used to obtain a ground-side sample data set corresponding to the task attribute information, and use the ground-side sample data set to train the initial large model to obtain a task large model, wherein the ground-side sample data set includes sample input data and actual annotation information corresponding to the sample input data; A determination unit, used to determine model attribute information of the task large model, and determine a target compression method applicable to the task large model based on the model attribute information and the task type of the predetermined task; A compression unit, used for compressing the mission large model using the target compression method to obtain an initial onboard large model; A deployment unit is used to adapt the initial onboard large model to the target satellite terminal based on the configuration attribute information to obtain the onboard large model, and deploy the onboard large model to the target satellite terminal.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images