A method of applying an artificial intelligence model in a spaceborne environment

Abstract

The embodiment of the application discloses a method for applying an artificial intelligence model in a satellite-borne environment, and the embodiment of the application stores model parameters of a scaled artificial intelligence model in a satellite, and restores the model parameters of the scaled artificial intelligence model when the artificial intelligence model is applied, so that the accuracy of the artificial intelligence model when the artificial intelligence model is applied can be ensured, and the application accuracy and reliability of the artificial intelligence model when the artificial intelligence model is applied in subsequent satellite memories are ensured.

Description

TECHNICAL FIELD

[0001] The present application relates to the cross-technical field of satellites and artificial intelligence, and in particular to a method for applying an artificial intelligence model in a satellite-borne environment. BACKGROUND

[0002] With the increasing amount of data generated by satellites and other spacecraft, the transmission of a large amount of data such as high-resolution earth images occupies a large amount of communication bandwidth resources between the satellite and the ground station, resulting in some data being unable to be transmitted to the ground station in a timely manner. The data collected by the satellite, especially the image data collected, is often covered by clouds or uninhabited areas on the ground, making it difficult for the ground station to obtain useful information from the data even if the data is transmitted to the ground station. Therefore, before transmitting the collected data to the ground station, the satellite can run an artificial intelligence model stored in the satellite memory to process the data and obtain valuable data processing results, and then transmit the data processing results to the ground station.

[0003] In order to apply the above method, an artificial intelligence model needs to be stored in the memory of the satellite so that it can be run directly in the satellite memory when applied. However, the space where the satellite is located is not protected by the earth's atmosphere, and the hardware such as the memory in the satellite is easily affected by solar radiation, and single event effect is a typical radiation damage that causes one or more bits in the data stored in the satellite to flip between 0 and 1, resulting in unacceptable damage to the data in the satellite. In this case, if the model parameters in the artificial intelligence model stored in the satellite are damaged by radiation such as single event effect, the application accuracy of the artificial intelligence model will be reduced, resulting in inaccurate data processing results when the artificial intelligence model is used to process the data collected by the satellite, such as neglecting to process earth images containing important information or delaying the processing of important disaster reports.

[0004] Therefore, how to avoid the artificial intelligence model stored in the satellite from being affected by radiation damage and ensure its application accuracy is a problem to be solved. SUMMARY

[0005] Therefore, the present application provides a method and system for applying an artificial intelligence model in a satellite-borne environment, which can avoid the artificial intelligence model stored in the satellite from being affected by radiation damage and ensure the accuracy of the application of the artificial intelligence model.

[0006] One embodiment of the present application provides a method for applying an artificial intelligence model in a satellite-borne environment, comprising:

[0007] The scaled model parameters of the artificial intelligence model are stored in the satellite corresponding to the scaling weight;

[0008] When the satellite applies the artificial intelligence model, the stored scaled model parameters are recovered by using the corresponding scaling weight, to obtain the model parameters of the artificial intelligence model;

[0009] The model parameters of the artificial intelligence model are placed into the corresponding layer structure of the artificial intelligence model, and run to process the data collected by the satellite.

[0010] In the above method, before the storage in the satellite, further comprising:

[0011] The satellite receives the scaled model parameters corresponding to the scaling weight sent by the ground station, and the scaled model parameters are obtained by scaling processing based on the obtained scaling weight in the ground station using a preset scaling method.

[0012] In the above method, the scaling weight includes:

[0013] Obtain a model parameter of the artificial intelligence model, and the artificial intelligence model is trained;

[0014] For the model parameters belonging to different layers of the artificial intelligence model, the corresponding scaling weight is set respectively.

[0015] In the above method, the corresponding scaling weight of the model parameters belonging to different layers of the artificial intelligence model includes:

[0016] For the current layer of the artificial intelligence model, the floating point number of the model parameter of the layer is obtained, the difference between the exponent bit value of the floating point number and the maximum value of the exponent bit of the floating point number is taken as the maximum scaling weight of the model parameter, and the minimum scaling weight is selected from the maximum scaling weight of all model parameters of the layer as the scaling weight corresponding to the model parameter of the layer.

[0017] In the above method, the scaled model parameters obtained by scaling processing based on the obtained scaling weight using a preset scaling method include:

[0018] Determine the layer to which the model parameter to be scaled belongs in the artificial intelligence model, and determine the corresponding scaling weight according to the layer;

[0019] The model parameter to be scaled is multiplied by 2 放缩权重 Obtain a first calculation intermediate value;

[0020] performing a left shift operation on the exponential bits of the first calculation intermediate value by a set number of shifts.

[0021] In the above method, the stored model parameters after the scaling processing are restored by using the corresponding scaling weights, and the model parameters of the artificial intelligence model are obtained.

[0022] For a model parameter after the scaling processing, the layer to which the artificial intelligence model belongs is determined.

[0023] The scaling weight corresponding to the layer is determined.

[0024] The exponential bits of the model parameter after the scaling processing are right shifted by a set number of shifts to obtain a second calculation intermediate value.

[0025] The second calculation intermediate value is divided by 2 放缩权重 to obtain the model parameters of the artificial intelligence model.

[0026] In the above method, the set number of shifts is 1.

[0027] In another embodiment of the present application, an electronic device is provided, which is applied in a satellite. The electronic device comprises:

[0028] a processor;

[0029] a memory storing a program configured to implement the method for applying an artificial intelligence model in a satellite-borne environment according to any one of the above embodiments when executed by the processor.

[0030] As seen above, the present application stores the model parameters of the artificial intelligence model after scaling processing in the satellite, and restores the model parameters of the artificial intelligence model after scaling processing when applied, thereby avoiding the change of the artificial intelligence model stored in the satellite, especially the model parameters of the artificial intelligence model, caused by the influence of radiation damage, which cannot guarantee the accuracy of the artificial intelligence model when applied, and ensures the application accuracy and reliability of the artificial intelligence model applied in the satellite memory subsequently. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 A method flowchart for applying an artificial intelligence model in a satellite-borne environment is provided for the embodiments of the present application.

[0032] Figure 2 A method flowchart for obtaining the scaling weight corresponding to each layer of the artificial intelligence model is provided for the embodiments of the present application.

[0033] Figure 3A process flowchart for scaling model parameters of an artificial intelligence model is provided for embodiments of the present application.

[0034] Figure 4 A method flowchart for restoring stored scaled model parameters in a satellite-borne environment is provided for embodiments of the present application.

[0035] Figure 5 A schematic diagram of an electronic device is provided for another embodiment of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of protection of the present application.

[0037] The terms "first", "second", "third", "fourth" and the like (if any) in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device that includes a series of steps or units does not necessarily limit to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product or device.

[0038] The technical solutions of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and some embodiments may not be described again for the same or similar concepts or processes.

[0039] As can be seen from the background, in order to reduce the amount of data transmitted between the satellite and the ground station, an artificial intelligence model is stored in the satellite, the satellite runs the stored artificial intelligence model in the memory, and the satellite first performs intelligent processing on the data collected from the ground to obtain a data processing result, and then transmits the data processing result to the ground station. However, due to the different environments in space and on the ground, the model parameters in the artificial intelligence model stored in the satellite will be damaged by radiation such as single particle effect, resulting in a decrease in the application accuracy of the artificial intelligence model in application, and thus the data collected by the satellite using the artificial intelligence model is inaccurate.

[0040] To solve the above problems, the hardware of the satellite, such as the memory, can be improved to avoid the influence of radiation damage by using a triple modular redundancy mode. The triple modular redundancy mode is to set the same artificial intelligence model in three memory modules in the satellite, and when processing the data collected by the satellite, the satellite memory runs three artificial intelligence models stored in the three memory modules to obtain three data processing results, and takes the majority of the same data processing results as the output. In specific implementation, it is usually two out of three, that is, as long as two same error results do not appear at the same time in the three data processing results, the error of the faulty memory module can be masked, and the correct data processing result can be ensured. However, this method requires a large amount of hardware resources in the satellite, thereby increasing the power consumption, mass and volume of the satellite.

[0041] Although the triple modular redundancy mode avoids the influence of radiation damage on the artificial intelligence model stored in the satellite, it causes the defect of increasing the hardware resources in the satellite. In order to avoid the influence of radiation damage on the artificial intelligence model stored in the satellite without increasing the hardware resources in the satellite, the embodiment of the present application stores the model parameters of the scaled artificial intelligence model corresponding to the scaling weight when storing the artificial intelligence model in the satellite. When applying the artificial intelligence model, the model parameters of the scaled artificial intelligence model are restored by using the corresponding scaling weight, and then the restored model parameters are placed in the corresponding layer structure of the artificial intelligence model. The artificial intelligence model is run in the satellite memory to process the data collected by the satellite to obtain the data processing result and send it to the ground station.

[0042] The embodiment of the present application stores the model parameters of the scaled artificial intelligence model in the satellite, and restores the model parameters of the scaled artificial intelligence model when applying, thereby avoiding the change of the artificial intelligence model stored in the satellite, especially the model parameters in the artificial intelligence model, caused by the influence of radiation damage, which cannot guarantee the accuracy of the artificial intelligence model when applying. Therefore, the method provided by the embodiment of the present application guarantees the application accuracy and reliability of the artificial intelligence model when subsequently applied in the satellite memory.

[0043] The embodiment of the application does not change the layer structure of the artificial intelligence model applied in the satellite, that is, does not change the running logic of the artificial intelligence model, but adopts a scaling processing mode to scale and store the model parameters required by the artificial intelligence model before running, and restores the model parameters to the layer structure corresponding to the artificial intelligence model in the application, that is, does not change the running logic of the artificial intelligence model. The satellite memory runs according to the running logic of the artificial intelligence model, processes the data collected by the satellite, and obtains a data processing result. In this way, the embodiment of the application solves the two technical problems described above: first, by scaling and restoring the model parameters, the scaled model parameters stored in the artificial intelligence model in the satellite greatly weaken the radiation damage such as single event effect; second, since the scaling and restoring of the model parameters are solved at the software level, the reliability of the artificial intelligence model applied in the satellite environment is improved, and the defect of additional hardware resources of the satellite caused by the hardware solution such as three-mode redundancy is avoided, saving the power consumption, reducing the mass and reducing the volume of the satellite.

[0044] The applicant adopts the scaling and restoring process of the model parameters to avoid the radiation damage such as single event effect on the artificial intelligence model stored in the satellite. Because the applicant has found through research that the model parameters in the artificial intelligence model are generally 32-bit bytes when stored according to the storage specification, and 8 bits are index bits. These index bits are most affected by the single event effect on the accuracy of the artificial intelligence model application, that is, when the index value corresponding to these index bits is affected by the single event effect and becomes larger, the accuracy of the artificial intelligence model application is greatly affected, and when the index value corresponding to these index bits is affected by the single event effect and becomes smaller, the accuracy of the artificial intelligence model application is less affected. Therefore, the applicant adopts the scaling and restoring of the model parameters, which can greatly weaken the single event effect during the storage of the satellite, thereby ensuring the accuracy of the artificial intelligence model application.

[0045] Figure 1 A method flowchart for applying an artificial intelligence model in a satellite environment is provided, and the specific steps include:

[0046] Step 101, store the scaled model parameters of the artificial intelligence model in the satellite corresponding to the scaling weight;

[0047] Step 102, when the satellite applies the artificial intelligence model, the corresponding scaling weight is adopted for the stored scaled model parameters, and after recovery, the model parameters of the artificial intelligence model are obtained.

[0048] Step 103, the model parameters of the artificial intelligence model are placed into the corresponding layer structure of the artificial intelligence model, and run to process the data collected by the satellite.

[0049] In the method, the storage in the satellite further includes:

[0050] The satellite receives the scaled model parameters corresponding to the scaling weight sent by the ground station, and the scaled model parameters are obtained by scaling processing based on the obtained scaling weight in the ground station using a preset scaling method.

[0051] Wherein, the scaling weight includes:

[0052] Obtain a model parameter of an artificial intelligence model, and the artificial intelligence model is trained.

[0053] For the model parameters of different layers of the artificial intelligence model, the corresponding scaling weight is set respectively.

[0054] Here, for each layer of the artificial intelligence model, the corresponding scaling weight of the model parameter of the layer includes:

[0055] For the current layer of the artificial intelligence model, the floating point number of a model parameter of the layer is obtained, the difference between the exponent bit value of the floating point number and the maximum value of the exponent bit of the floating point number is taken as the maximum scaling weight of the model parameter, and the minimum scaling weight is selected from the maximum scaling weight of all model parameters of the layer as the corresponding scaling weight of the model parameter of the layer.

[0056] Here, the calculation method of the floating point number of a model parameter of the layer is calculated by IEEE binary floating point number arithmetic standard (IEEE 754), which can be represented in binary. For example, a model parameter is-0.75, and the binary representation of the IEEE 754 floating point number corresponding to the model parameter is 10111111010000000000000000000000, the second to eighth bits of which are the exponent bit of the floating point number, and the decimal representation of the exponent bit value is 126, therefore, the maximum scaling weight of the model parameter is the difference between 127 (the decimal representation of the maximum value of the exponent bit) and 126, which is 1.

[0057] It can be seen that each layer of the artificial intelligence model has different scaling weights for its model parameters. The scaling weights of each layer in the artificial intelligence model can be stored in the form of a dictionary or similar format, corresponding to the layers of the artificial intelligence model. The scaling weights of the model parameters can then be determined based on their respective layers within the artificial intelligence model.

[0058] In this method, the scaling process based on the acquired scaling weights and using a preset scaling method to obtain the scaled model parameters includes:

[0059] The model parameters to be scaled are determined to be in the layer to which the artificial intelligence model belongs, and the corresponding scaling weights are determined based on the layer to which the model belongs.

[0060] Multiply the model parameters to be scaled by 2 放缩权重 The first intermediate value is obtained by adding scaling weights to the exponent bits of the floating-point number of the model parameters.

[0061] A left shift operation with a set shift number is performed on the exponent of the first calculated intermediate value. The set shift number can be 1, that is, when the exponent of the first calculated intermediate value is 8 bits, the last 7 exponent bits of the first calculated intermediate value become the first 7 exponent bits, and the eighth exponent bit becomes 0, thus obtaining the model parameters after scaling.

[0062] Here, a 1-bit left shift is performed on the 8-bit exponent in the intermediate calculation value to further protect the storage of the model parameters. For example, suppose a model parameter is -0.75, whose corresponding binary representation in IEEE 754 floating-point numbers is 101111110100000000000000000000000, with a scaling weight of 1. This parameter is then scaled. That is, the model parameter is multiplied by 2. 1 That is, the 8 exponent bits become 01111111, then the exponent bits are shifted left by 1 bit to become 11111110, and the final binary representation of the scaled model parameters of the IEEE754 floating-point number is 111111110100000000000000000000000.

[0063] In this method, the model parameters of the artificial intelligence model obtained after restoring the stored scaled model parameters by applying the corresponding scaling weights include:

[0064] For a model parameter that has undergone scaling, determine the layer to which the artificial intelligence model belongs;

[0065] Determine the scaling weight corresponding to the layer to which it belongs;

[0066] The exponent bits of the scaled model parameters are shifted to the right by a set shift number, which can be 1, to obtain a second intermediate value. That is, when the exponent bits of the scaled model parameters are 8 bits, the first 7 exponent bits of the compressed artificial intelligence model parameters become the last 7 exponent bits, while the first exponent bit becomes 0.

[0067] Divide the second intermediate value by 2 放缩权重 The model parameters of the artificial intelligence model are obtained by subtracting the corresponding scaling weight from the exponent of the second intermediate value.

[0068] To illustrate, suppose a model parameter is -0.75, whose corresponding IEEE 754 floating-point binary representation is 101111110100000000000000000000000. Its scaling weight is 1. The scaled model parameter's IEEE 754 floating-point binary representation is 111111110100000000000000000000000. To restore it, first shift the 8 exponent bits one bit to the right, resulting in 01111111, then divide by 2. 1 That is, the 8-bit exponent becomes 01111110, so the binary representation of the IEEE754 floating-point number of the restored model parameter is 10111111010000000000000000000000, which is the same as the binary representation of the IEEE754 floating-point number corresponding to the model parameter -0.75.

[0069] As can be seen from the above scheme, the method provided in this application includes three parts: the first part: obtaining the scaling weights corresponding to each layer of the artificial intelligence model; the second part: scaling the model parameters of the artificial intelligence model and uploading them to the satellite for storage; the third part: restoring the scaled model parameters in the spaceborne environment and then applying them. The first and second parts are completed at a ground station, while the third part is implemented in the satellite, i.e., in the spaceborne environment. These three parts will be described in detail below.

[0070] Figure 2 The flowchart of the method for obtaining the scaling weights corresponding to each layer of the artificial intelligence model provided in this application embodiment includes the following specific steps:

[0071] Step 201: Provide an artificial intelligence model, which is pre-trained;

[0072] Step 202: Determine whether the scaling weights for each layer in the artificial intelligence model have been calculated. If yes, proceed to step 203; otherwise, proceed to step 204.

[0073] Step 203: For each layer of the artificial intelligence model, store the corresponding scaling weights and end the entire process;

[0074] Step 204: Select a layer of the artificial intelligence model with uncalculated scaling weights;

[0075] Step 205: Initialize the scaling weight of this layer to 127;

[0076] In this step, the weights are initialized to the maximum value of the exponent digits of the floating-point model parameters for that layer.

[0077] Step 206: Determine if the maximum scaling weights have been calculated for all model parameters in this layer. If yes, return to step 202; otherwise, proceed to step 207.

[0078] Step 207: Select a model parameter for this layer whose maximum scaling weight has not been calculated, and calculate its maximum scaling weight;

[0079] In this step, calculating the maximum scaling weight includes: obtaining the floating-point number of the model parameter, and taking the difference between the value of the exponent digit of the floating-point number and the maximum value of the exponent digit of the floating-point number as the maximum scaling weight of the model parameter.

[0080] Step 208: Select the smaller of the existing scaling weight of the layer and the maximum scaling weight obtained in step 207 as the updated scaling weight of the layer, and return to step 206 to continue execution.

[0081] Figure 3 The flowchart for scaling the model parameters of an artificial intelligence model provided in this embodiment of the invention includes the following specific steps:

[0082] Step 301: Provide an artificial intelligence model, which is pre-trained, and obtain the model parameters of each layer in the artificial intelligence model, as well as the scaling weights corresponding to each layer of the artificial intelligence model.

[0083] Step 302: Determine if the model parameters of each layer of the artificial intelligence model have been scaled. If yes, proceed to step 303; otherwise, proceed to step 304.

[0084] Step 303: Send the scaled model parameters, along with their corresponding scaling weights, to the satellite for storage, thus ending the entire process.

[0085] Step 304: Select a model parameter that has not been scaled and place it in the layer of the artificial intelligence model;

[0086] Step 305: Obtain the scaling weight of the layer to which it belongs;

[0087] Step 306: Multiply one of the unscaled model parameters of the layer by 2. 放缩权重 Obtain the first intermediate value;

[0088] Step 307: Determine if all model parameters of the layer have been shifted left by 1 bit. If yes, return to step 302; otherwise, proceed to step 308.

[0089] Step 308: Select an unshifted model parameter in the layer to which it belongs, shift the exponent of the first intermediate value of the model parameter to the left by 1 bit, and return to step 307 to continue execution.

[0090] Figure 4 This application provides a flowchart of a method for restoring scaled model parameters stored in a spaceborne environment, which includes the following steps:

[0091] Step 401: Obtain the stored scaled model parameters and the scaling weights based on the corresponding layers in the artificial intelligence model;

[0092] Step 402: Determine if the scaled model parameters of all layers of the artificial intelligence model have been restored. If yes, proceed to step 403; otherwise, proceed to step 404.

[0093] Step 403: Return the recovered model parameters to be placed into the layer of the artificial model and apply the artificial intelligence model;

[0094] Step 404: Select one of the unrecovered scaled model parameters in the layer to which the artificial intelligence model belongs;

[0095] Step 405: Determine if all unrecovered scaled model parameters of the layer have been right-shifted by 1 bit to obtain the second intermediate value. If yes, proceed to step 406; otherwise, proceed to step 408.

[0096] Step 406: Obtain the scaling weight corresponding to the layer to which it belongs;

[0097] Step 407: Select a second intermediate value for the corresponding layer, and divide the second intermediate value by 2. 放缩权重 The recovered model parameters of the layer to which the layer belongs are obtained until all the second intermediate values ​​of the layer to which the layer belongs have been processed, and all the recovered model parameters of the layer are obtained. Then, return to step 402 for execution.

[0098] Step 408: Select an unshifted, unrecovered, scaled model parameter from the layer to which it belongs;

[0099] Step 409: Shift the exponent of the unshifted, unrecovered, scaled model parameter one bit to the right, and return to step 408 for execution.

[0100] Figure 5 This is a schematic diagram of an electronic device provided for another embodiment of this application. This electronic device is used in a satellite, such as... Figure 5 As shown, it may include a processor 501, wherein the processor 501 is used to perform the steps of the method described above for applying an artificial intelligence model in a spaceborne environment. From Figure 5 It can also be seen that the electronic device provided in the above embodiments further includes a non-transitory computer-readable storage medium 502, on which a computer program is stored, and the computer program is executed by the processor 501 to perform the steps of the above-described method for applying an artificial intelligence model in a spaceborne environment.

[0101] Specifically, the non-transitory computer-readable storage medium 502 can be a general-purpose storage medium, such as a mobile disk, hard disk, FLASH, read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), or portable compact disk read-only memory (CD-ROM), etc. When the computer program on the non-transitory computer-readable storage medium 502 is run by the processor 502, it can cause the processor 501 to execute the various steps of the above-described method for applying an artificial intelligence model in a spaceborne environment.

[0102] In practical applications, the non-transitory computer-readable storage medium 502 may be included in the device / apparatus / system described in the above embodiments, or it may exist independently without being assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, can perform the various steps of the method for applying an artificial intelligence model in a spaceborne environment.

[0103] Another embodiment of this application provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method for applying an artificial intelligence model in a spaceborne environment.

[0104] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments disclosed in this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings. For example, two blocks shown connectedly may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0105] Those skilled in the art will understand that the features described in the various embodiments and / or claims of this disclosure can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, without departing from the spirit and teachings of this application, the features described in the various embodiments and / or claims of this application can be combined and / or combined in various ways, and all such combinations and / or combinations fall within the scope of this application.

[0106] This document uses specific embodiments to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this application, and are not intended to limit this application. For those skilled in the art, changes can be made to the specific implementation methods and application scope based on the ideas, spirit and principles of this application. Any modifications, equivalent substitutions, improvements, etc., made should be included within the scope of protection of this application.

Claims

1. A method for applying artificial intelligence models in a spaceborne environment, characterized in that, include: The scaled model parameters of the artificial intelligence model, along with their corresponding scaled weights, are stored in the satellite. The scaled weights are set separately for the different layers of the trained artificial intelligence model. When the artificial intelligence model is applied to the satellite, the stored scaled model parameters are restored using the corresponding scaling weights to obtain the model parameters of the artificial intelligence model. The model parameters of the artificial intelligence model are placed into the corresponding layer structure of the artificial intelligence model, and the process is performed to process the data collected by the satellite. The parameters for different layers of the trained artificial intelligence model are set as follows: For the current layer of the artificial intelligence model, obtain the floating-point number of a model parameter of the layer, and take the difference between the value of the exponent of the floating-point number and the maximum value of the exponent of the floating-point number as the maximum scaling weight of the model parameter. From the maximum scaling weights of all model parameters of the layer, select the minimum scaling weight as the scaling weight corresponding to the model parameter of the layer. The scaled-down model parameters of the artificial intelligence model are obtained in the following manner: The model parameters to be scaled are determined to be in the layer to which the artificial intelligence model belongs, and the corresponding scaling weights are determined based on the layer to which the model belongs. Multiply the model parameters to be scaled by 2 放缩权重 Obtain the first intermediate value; Perform a left shift operation on the exponent of the first calculated intermediate value by a set number of shifts.

2. The method as described in claim 1, characterized in that, The storage prior to the satellite also includes: The satellite receives the scaled model parameters corresponding to the scaling weights sent by the ground station. The scaled model parameters are obtained at the ground station.

3. The method as described in any one of claims 1 to 2, characterized in that, The process of applying the corresponding scaling weights to the stored scaled model parameters to restore the model results in the following model parameters for the artificial intelligence model: For a model parameter that has undergone scaling, determine the layer to which the artificial intelligence model belongs; Determine the scaling weight corresponding to the layer to which it belongs; The exponent of the scaled model parameters is shifted to the right by a set number of shifts to obtain the second intermediate value. Divide the second intermediate value by 2 放缩权重 The model parameters of the artificial intelligence model are obtained.

4. The method as described in claim 3, characterized in that, The set shift number is 1.

5. An electronic device, characterized in that, The electronic device is used in a satellite, and the electronic device includes: processor; A memory storing a program configured to implement, when executed by the processor, the method of applying an artificial intelligence model in a spaceborne environment as described in any one of claims 1 to 4.

Images