Satellite-borne computer with fault-tolerant reinforcement capability and on-orbit reasoning system

Abstract

The invention provides a spaceborne computer with fault-tolerant reinforcement capability. The spaceborne computer is configured to execute a calculation task according to the following mode: a CPU (Central Processing Unit) preprocesses a to-be-processed image to obtain a plurality of slice image batches; the CPU schedules all the intelligent acceleration cards to load an on-orbit task inference model to process different slice image batches respectively until all the slice image batches are processed; wherein each intelligent acceleration card loads an on-orbit task reasoning model from the CPU to process the slice image batch, whether the reasoning result of a prior image in the slice image batch obtained through processing of the on-orbit task reasoning model is consistent with the known reasoning result of the prior image or not is compared after processing is completed, and if not, the intelligent acceleration card is started. If yes, the slice image batch is redistributed to other intelligent acceleration cards for processing, and the intelligent acceleration card does not participate in processing of other slice image batches any more.

Description

Technical Field

[0001] The present invention relates to the field of satellite payload calculation, specifically to on-board computers, and more specifically, to an on-board computer and an on-orbit inference system with fault-tolerant and reinforcement capabilities. Background Art

[0002] On-orbit processing refers to the ability of a satellite to perform real-time processing on the data collected in space. By obtaining remote sensing data from sensors and using the computing resources on the satellite (such as on-board computers and algorithms) for analysis, processing, and inference, on-orbit processing can significantly reduce the latency of data transmission, save bandwidth, and improve the response speed. Especially in scenarios such as disaster monitoring and environmental reconnaissance that require real-time decision-making, this ability is crucial. Remote sensing images have characteristics such as large image format and high resolution. By adopting a heterogeneous computing architecture, the CPU is responsible for task coordination and management, and the parallel processing ability of intelligent acceleration cards is used to accelerate the image processing process, thereby greatly improving the on-orbit processing efficiency.

[0003] However, on-board computers are vulnerable to radiation effects such as single-event upsets in the space environment, resulting in a decrease in the reliability of on-orbit calculation results. To address the impact of radiation effects such as single-event upsets, researchers have proposed fault-tolerant and reinforcement strategies for on-orbit processing. For example, at the hardware level, multiple redundant hardware modules can be added to ensure that the system can still operate even if a certain module fails; for example, at the data level, parity checks or error-correcting codes are used to correct memory errors; and for another example, at the algorithm level, weight adjustment is used to continue inference and decision-making to avoid partial node calculation errors.

[0004] Although the fault-tolerant and reinforcement strategies for on-orbit processing proposed in the prior art can improve the reliability of on-board computer processing, each strategy has certain limitations. For the strategy of adding redundant modules at the hardware level to achieve fault tolerance and reinforcement, although it can effectively improve the fault tolerance of on-board computers, it also means an increase in the complexity and cost of on-board computers, and in the space environment, the maintenance and replacement of hardware are extremely difficult, and excessive hardware redundancy may bring unnecessary burdens. For the fault-tolerant strategy at the data level, although it can correct single-bit memory errors caused by single-event effects, double-bit errors often cannot be directly corrected, and the uncorrectable errors will accumulate in the storage unit, ultimately affecting the reliability of the calculation results and the normal operation of the function. For the fault-tolerant strategy at the algorithm level, although it can avoid partial node calculation errors to a certain extent, this strategy may depend on specific algorithms and application scenarios, and different adjustment strategies may be required for different tasks.

[0005] It should be noted that: This background technology is only used to introduce relevant information of the present invention to facilitate understanding of the technical solution of the present invention, but it does not necessarily mean that the relevant information is prior art. Without evidence indicating that the relevant information was publicly available before the filing date of the present invention, the relevant information should not be regarded as prior art. Summary of the Invention

[0006] Therefore, the object of the present invention is to overcome the defects of the above-mentioned prior art and provide an on-board computer with fault-tolerant and reinforcement capabilities, an on-orbit inference system, and an on-orbit inference method applicable to a satellite platform.

[0007] The object of the present invention is achieved through the following technical solutions.

[0008] According to a first aspect of the present invention, there is provided an on-board computer with fault-tolerant and reinforcement capabilities. The on-board computer includes a CPU and multiple intelligent acceleration cards. An on-orbit task inference model for executing computing tasks and multiple prior images with known inference results are configured on the CPU. The on-board computer is configured to execute computing tasks in the following manner: The CPU preprocesses the acquired image to be processed to obtain multiple batches of sliced images and reads each batch of sliced images into the CPU memory; wherein, each batch of sliced images includes a prior image and multiple sub-images with the same image size, and the sub-images are images obtained by dividing the image to be processed into blocks; the CPU schedules all intelligent acceleration cards to load the on-orbit task inference model to process different batches of sliced images respectively until all batches of sliced images are processed, obtaining the inference result of the image to be processed; wherein, each intelligent acceleration card loads the on-orbit task inference model from the CPU to process a batch of sliced images, and after processing, compares whether the inference result of the prior image in the batch of sliced images processed by the on-orbit task inference model is consistent with the known inference result of the prior image. If not, the batch of sliced images is re-allocated to other intelligent acceleration cards for processing, and this intelligent acceleration card no longer participates in the processing of other batches of sliced images.

[0009] In some embodiments of the present invention, the on-board computer is configured to preprocess the image to be processed in the following manner: taking the data input size of the on-orbit task inference model as the window size, sliding and taking values in the image to be processed to obtain multiple sub-images with the same size; dividing the obtained sub-images into multiple batches of sliced images according to the order of sliding and taking values and the preset size of each batch of sliced images, and allocating a prior image to each batch of sliced images.

[0010] In some embodiments of the present invention, the on-board computer is configured to determine the preset size of each batch of sliced images in the following manner:

[0011]

[0012]

[0013] Among them, represents the slice image batch size; represents the data input size, represents the width of the image, represents the height of the image, represents the number of channels of the image; represents the number of bytes Bytes required for the data type Dtype, where Dtype represents the data type and Bytes represents the number of bytes; represents the storage space size occupied by the on-orbit mission inference model; represents the memory size occupied by the intermediate results generated during the operation of the on-orbit mission inference model; represents the video memory size of the intelligent acceleration card; represents the number of intelligent acceleration cards, represents the memory size of the CPU, " " represents that the product of the video memory size of the intelligent acceleration card and the number of intelligent acceleration cards is less than or equal to a preset ratio of the CPU memory.

[0014] In some embodiments of the present invention, the preset ratio is configured to be 80%.

[0015] In some embodiments of the present invention, the on-orbit mission inference model is a pre-trained object detection model.

[0016] According to a second aspect of the present invention, there is provided an on-orbit inference system for being carried on a satellite platform to perform on-orbit computing tasks, the system comprising: a data acquisition module for acquiring an image to be processed; and an on-board computer as described in the first aspect of the present invention for processing the image to be processed to obtain an on-orbit inference result of the image to be processed.

[0017] In some embodiments of the present invention, the system further comprises: a space-ground transmission module for transmitting the on-orbit inference result of the image to be processed back to a ground station.

[0018] According to a third aspect of the present invention, there is provided an on-orbit inference method applicable to a satellite platform, the method comprising: processing an image to be processed by using the on-orbit inference system as described in the second aspect of the present invention.

[0019] Compared with the prior art, the advantages of the present invention are as follows: (1) By calculating a reasonable slice image batch size and slicing the images, the memory resources and computing resources of the on-board computer can be efficiently utilized; (2) A prior image is introduced, and the working state of the intelligent acceleration card is analyzed based on the inference result obtained by processing the prior image according to the on-orbit task inference model loaded by the intelligent acceleration card. When the working state of the intelligent acceleration card is abnormal, the execution of the calculation task is stopped, thereby solving the problem of the decreased computing reliability of the on-board computer. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The following further describes the embodiments of the present invention with reference to the accompanying drawings, where:

[0021] Figure 1 FIG. is a schematic diagram of the composition of an on-board computer according to an embodiment of the present invention;

[0022] Figure 2 FIG. is a schematic diagram of the composition of an on-orbit inference system according to an embodiment of the present invention;

[0023] Figure 3 FIG. is a schematic diagram of the slice image batch processing flow according to an embodiment of the present invention;

[0024] Figure 4 FIG. is a schematic diagram of the slice image batch allocation according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0025] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below through specific embodiments with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0026] As mentioned in the background art section, although the fault-tolerant and reinforcement strategies proposed by the prior art for on-orbit processing can improve the reliability of on-board computer processing, each strategy has certain limitations. For the strategy of adding redundant modules at the hardware level to achieve fault tolerance and reinforcement, although it can effectively improve the fault tolerance of the on-board computer, this also means an increase in the complexity and cost of the on-board computer. Moreover, in the space environment, the maintenance and replacement of hardware are extremely difficult, and excessive hardware redundancy may bring unnecessary burdens. For the fault-tolerant strategy at the data level, although it can correct single-bit memory errors caused by single-event effects, double-bit errors often cannot be directly corrected, and the uncorrectable errors will accumulate in the storage unit, ultimately affecting the reliability of the calculation result and the normal operation of the function. For the fault-tolerant strategy at the algorithm level, although it can avoid some node calculation errors to a certain extent, this strategy may depend on specific algorithms and application scenarios, and different adjustment strategies may be required for different tasks.

[0027] To solve the above problems, the inventors propose to introduce a prior image with known inference results into the on-board computer to analyze the working states of multiple intelligent acceleration cards in the on-board computer, thereby solving the problem of the decreased computing reliability of the on-board computer. Specifically, after introducing the prior image, each intelligent acceleration card processes a prior image to obtain the inference result of the prior image when executing a computing task, and determines the working state of the intelligent acceleration card by comparing whether the known inference result of the prior image is consistent with the inference result of the prior image processed by the intelligent acceleration card; if they are consistent, it indicates that the working state of the intelligent acceleration card is normal; if they are inconsistent, it indicates that the working state of the intelligent acceleration card is abnormal. At this time, it is necessary to use other intelligent acceleration cards to re-execute the computing task of this intelligent acceleration card, and the abnormal intelligent acceleration card is no longer used to execute subsequent computing tasks.

[0028] Generally speaking, as Figure 1 shown, the present invention proposes an on-board computer with fault tolerance and reinforcement capabilities. The on-board computer includes a CPU and multiple intelligent acceleration cards. An on-orbit task inference model for executing computing tasks and multiple prior images with known inference results are configured on the CPU. The on-board computer is configured to execute computing tasks in the following manner: The CPU preprocesses the acquired image to be processed to obtain multiple batches of sliced images, and reads each batch of sliced images into the CPU memory; wherein, each batch of sliced images includes a prior image and multiple sub-images with the same image size, and the sub-images are images obtained by block-processing the image to be processed; The CPU schedules all intelligent acceleration cards to load the on-orbit task inference model to process different batches of sliced images until all batches of sliced images are processed, obtaining the inference result of the image to be processed; wherein, each intelligent acceleration card loads the on-orbit task inference model from the CPU to process a batch of sliced images, and after processing, compares whether the inference result of the prior image in the batch of sliced images processed by the on-orbit task inference model is consistent with the known inference result of the prior image. If they are inconsistent, the batch of sliced images is reallocated to other intelligent acceleration cards for processing, and this intelligent acceleration card no longer participates in the processing of other batches of sliced images.

[0029] To better understand the present invention, the working principle of the on-board computer will be described below with specific embodiments.

[0030] According to an embodiment of the present invention, the on-board computer is configured to preprocess the image to be processed in the following manner: sliding and taking values in the image to be processed with the data input size of the on-orbit task inference model as the window size to obtain a plurality of sub-images of the same size; dividing the obtained sub-images into a plurality of sliced image batches according to the order of sliding and taking values and a preset sliced image batch size, and assigning a prior image to each sliced image batch. Wherein, the prior image is an image that conforms to the data input size of the on-orbit task inference model and has an interested target, and the known inference result of the prior image is obtained by inputting the prior image into a pre-trained on-orbit task inference model under the condition that there is no fault injection.

[0031] It should be noted that the prior image can be expressed as {image: 1.png, object: 1.txt}, {image: 2.png, object: 2.txt}, …, {image: n.png, object: n.txt}}, where {image: 1.png, object: 1.txt} indicates that the specified prior image 1 (image: 1.png) is associated with the text (object: 1.txt), that is, the known inference result of the prior image 1 is recorded in text 1; {image: 2.png, object: 2.txt} indicates that the specified prior image 2 (image: 2.png) is associated with the text (object: 2.txt), that is, the known inference result of the prior image 2 is recorded in text 2; {image: n.png, object: n.txt} indicates that the specified prior image n (image: n.png) is associated with the text (object: n.txt), that is, the known inference result of the prior image n is recorded in text n.

[0032] According to an embodiment of the present invention, the on-board computer is configured to determine the preset sliced image batch size in the following manner:

[0033]

[0034]

[0035] Wherein, represents the sliced image batch size; represents the data input size, represents the width of the image, represents the height of the image, represents the number of channels of the image; Indicates the number of bytes required for the data type Dtype, where Dtype represents the data type and Bytes represents the number of bytes; Indicates the storage space size occupied by the on-orbit task inference model; Indicates the memory size occupied by the intermediate results generated during the operation of the on-orbit task inference model; Indicates the video memory size of the intelligent acceleration card; Indicates the number of intelligent acceleration cards, Indicates the memory size of the CPU, " " indicates that the product of the video memory size of the intelligent acceleration card and the number of intelligent acceleration cards is less than or equal to a preset ratio of the CPU memory. According to an embodiment of the present invention, the preset ratio is configured to be 80%. Among them, the preset ratio can also be set to other values according to actual situations, and the present invention does not make special restrictions. It should be noted that the purpose of dividing the sliced image batches is to enable the on-board computer to maximize the use of memory when performing computing tasks, while avoiding the collapse of computing tasks due to memory exhaustion.

[0036] According to an embodiment of the present invention, the on-orbit task inference model is a pre-trained object detection model. Among them, the on-orbit task inference model can not only be set as an object detection model to identify the targets of interest in the image to be processed, such as identifying airplanes and ships; it can also be set as an image classification model to classify different regions in the image to be processed, such as classifying rivers, lakes, forests, etc.; the on-orbit task inference model can be set according to actual needs, and the present invention does not make special restrictions.

[0037] The on-board computer described in the foregoing embodiment can be carried on a satellite platform to perform on-orbit computing tasks. Based on this, as Figure 2 shown, the present invention also proposes an on-orbit inference system for being carried on a satellite platform to perform on-orbit computing tasks. The system includes: a data acquisition module for acquiring the image to be processed; the on-board computer as described in the foregoing embodiment for processing the image to be processed to obtain the on-orbit inference result of the image to be processed.

[0038] According to an embodiment of the present invention, as Figure 2 shown, the system further includes: a space-ground transmission module for transmitting the on-orbit inference result of the image to be processed back to the ground station.

[0039] To better understand the on-orbit inference system, the following details how to use the system to perform on-orbit computing tasks.

[0040] First, the data acquisition module first acquires the image to be processed, where the image to be processed is a remote sensing image obtained in real time by a sensor on the satellite in orbit.

[0041] Then, the on-board computer is used to process the acquired remote sensing images. The specific processing process is as follows.

[0042] Sliced image batch division: The CPU preprocesses the acquired images to be processed to obtain multiple sliced image batches, and reads each sliced image batch into the CPU memory. Among them, the preprocessing process is as follows: taking the data input size W*H*C of the on-orbit task inference model as the window size, sliding and taking values in the remote sensing image (starting from the upper left corner or the lower right corner of the remote sensing image, sliding and taking values to the right / downward or left / upward) to obtain multiple sub-images of the same size; dividing the obtained sub-images into multiple sliced image batches according to the order of sliding and taking values and the preset sliced image batch size B, and allocating a prior image to each sliced image batch. At this time, a sliced image batch includes B-1 sub-images with a size of W*H*C and a prior image, and each sliced image batch corresponds to a batch number. For example, assume that a remote sensing image is divided into 15 sub-images (sub-image numbers are 0-14), and the sliced image batch size is 4. After preprocessing, there will be 5 sliced image batches for this remote sensing image. Among them, sliced image batch 1 includes sub-image 0, sub-image 1, sub-image 2, and a prior image, sliced image batch 2 includes sub-image 3, sub-image 4, sub-image 5, and a prior image, sliced image batch 3 includes sub-image 6, sub-image 7, sub-image 8, and a prior image, sliced image batch 4 includes sub-image 9, sub-image 10, sub-image 11, and a prior image, and sliced image batch 5 includes sub-image 12, sub-image 13, sub-image 14, and a prior image.

[0043] Sliced image batch processing: The CPU schedules all intelligent acceleration cards to load the on-orbit task inference model to process different sliced image batches respectively until all sliced image batches are processed, and the on-orbit inference result of the remote sensing image is obtained. Among them, each intelligent acceleration card loads the on-orbit task inference model from the CPU to process the sliced image batch, and after processing, compares whether the on-orbit inference result of the prior image in the sliced image batch processed by the on-orbit task inference model is consistent with the known inference result of the prior image. If not, the sliced image batch is reassigned to other intelligent acceleration cards for processing, and this intelligent acceleration card no longer participates in the processing of other sliced image batches.

[0044] Specifically, as Figure 3As shown in the figure, the processing of the slice image batch is as follows: the CPU schedules all intelligent acceleration cards to load the on-orbit task inference model to execute multiple computing tasks until all slice image batches are processed; among them, during each execution of the computing task, each intelligent acceleration card will load the on-orbit task inference model to process its corresponding slice image batch, and after each inference is completed, each intelligent acceleration card will compare whether the on-orbit inference result of the prior image in the slice image batch processed by the on-orbit task inference model is consistent with the known inference result of the prior image. If they are inconsistent, it means that the working state of the intelligent acceleration card is abnormal, and the slice image batch corresponding to the intelligent acceleration card with inconsistent comparison results is marked as Flag = 0 (Flag = 0 indicates that the slice image batch processing is abnormal, the processing result is not saved, and the working state of the corresponding intelligent acceleration card is abnormal); if they are consistent, it means that the working state of the intelligent acceleration card is normal, and the slice image batch corresponding to the intelligent acceleration card with consistent comparison results is marked as Flag = 1 (Flag = 1 indicates that the slice image batch processing is normal, the processing result is saved, and the working state of the corresponding intelligent acceleration card is normal). Among them, during each execution of the computing task, each intelligent acceleration card is allocated with a slice image batch as shown in Figure 4 shown in the figure.

[0045] It should be noted that when the working states of all intelligent acceleration cards are abnormal and there are still slice image batches not processed, at this time, restart the on-board computer and re-process the remaining slice image batches according to the processing process shown in Figure 4 the figure.

[0046] It should also be noted that when determining whether the on-orbit inference result of the prior image is consistent with the known inference result of the prior image, different judgment criteria will be set according to the actual on-orbit calculation tasks, and the present invention does not make special restrictions. For example, for the target recognition task, when the number of recognized targets in the on-orbit inference result of the prior image is less than 90% of the number of recognized targets in the known inference result of the prior image, it is considered inconsistent, and the working state of the corresponding intelligent acceleration card is abnormal; when the number of recognized targets in the on-orbit inference result of the prior image is greater than or equal to 90% of the known inference result of the prior image, then the mIOU between the on-orbit inference result of the prior image and the known inference result of the prior image is judged. If the mIOU between the recognized target boxes in the on-orbit inference result of the prior image and the recognized target boxes in the known inference result of the prior image is greater than or equal to 90%, it is considered consistent, and the working state of the corresponding intelligent acceleration card is normal, otherwise the working state of the corresponding intelligent acceleration card is abnormal. Another example is that for the image classification task, according to the classification mask corresponding to the on-orbit inference result of the prior image and the classification mask corresponding to the known inference result of the prior image, the mIOU between the classification result in the on-orbit inference result of the prior image and the classification result in the known inference result of the prior image is calculated. When the mIOU is greater than or equal to 90%, it is considered consistent, and the working state of the corresponding intelligent acceleration card is normal, otherwise the working state of the corresponding intelligent acceleration card is abnormal.

[0047] Finally, the on-orbit inference result of the remote sensing image is transmitted back to the ground station by using the space-ground transmission module.

[0048] Based on the on-orbit inference system described in the foregoing embodiments, the present invention also proposes an on-orbit inference method applicable to a satellite platform, and the method includes: processing the image to be processed by using the on-orbit inference system as described in the foregoing embodiments.

[0049] Different from the prior art, the present invention analyzes the working states of multiple intelligent acceleration cards in the on-board computer by introducing a prior image with a known inference result into the on-board computer, thereby solving the problem of the decreased calculation reliability of the on-board computer. Moreover, such a processing method can not only avoid excessive hardware redundancy and error accumulation, but also be applicable to different on-orbit calculation tasks.

[0050] The beneficial effects of the present invention are as follows: (1) By calculating a reasonable slice image batch size and slicing the image, the memory resources and computing resources of the on-board computer can be efficiently utilized; (2) A prior image is introduced, and the working state of the intelligent acceleration card is analyzed according to the inference result obtained by processing the prior image by the on-orbit task inference model loaded by the intelligent acceleration card. When the working state of the intelligent acceleration card is abnormal, the execution of the calculation task is stopped, thereby solving the problem of the decreased calculation reliability of the on-board computer.

[0051] It should be noted that although the above steps are described in a specific order, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently or even in a different order, as long as the required functions can be achieved.

[0052] The present invention may be a system, a method, and / or a computer program product. The computer program product may include a computer-readable storage medium having thereon computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0053] The computer-readable storage medium may be a tangible device that retains and stores instructions for use by an instruction execution device. The computer-readable storage medium may include, for example, but is not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), a static random access memory (SRAM), a portable compact disc read-only memory (CD-ROM), a digital versatile disc (DVD), a memory stick, a floppy disk, a mechanically encoded device such as a punched card or raised structures in grooves having instructions stored thereon, and any suitable combination of the foregoing.

[0054] The embodiments of the present invention have been described above. The above description is exemplary and not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art in the technical field without departing from the scope and spirit of the described embodiments. The selection of the terms used herein is intended to best explain the principles of the embodiments, the practical application, or the improvement of technologies in the market, or to enable other ordinary skilled persons in the technical field to understand the embodiments disclosed herein.

Claims

1. A spaceborne computer with fault-tolerant and reinforcement capabilities, the spaceborne computer includes a CPU and multiple intelligent acceleration cards, and an on-orbit task inference model for executing computing tasks and multiple prior images with known inference results are configured on the CPU, characterized in that, The on-board computer is configured to execute computing tasks in the following manner: The CPU preprocesses the acquired image to be processed to obtain multiple batches of sliced images, and reads each batch of sliced images into the CPU memory; wherein, each batch of sliced images includes a prior image and multiple sub-images with the same image size, and the sub-images are images obtained by dividing the image to be processed into blocks. The CPU schedules all intelligent acceleration cards to load the on-orbit task inference model to process different batches of sliced images respectively until all batches of sliced images are processed, and obtains the inference result of the image to be processed; wherein, each intelligent acceleration card loads the on-orbit task inference model from the CPU to process a batch of sliced images, and after the processing is completed, compares whether the inference result of the prior image in the batch of sliced images processed by the on-orbit task inference model is consistent with the known inference result of the prior image. If not, the batch of sliced images is re-allocated to other intelligent acceleration cards for processing, and this intelligent acceleration card no longer participates in the processing of other batches of sliced images.

2. The on-board computer according to claim 1, characterized in that, The on-board computer is configured to preprocess the image to be processed in the following manner: Taking the data input size of the on-orbit task inference model as the window size, sliding and taking values in the image to be processed to obtain multiple sub-images with the same size. Dividing the obtained sub-images into multiple batches of sliced images according to the order of sliding and taking values and the preset size of the batch of sliced images, and allocating a prior image to each batch of sliced images.

3. The on-board computer according to claim 2, wherein The on-board computer is configured to determine the preset size of the batch of sliced images in the following manner: Among them, represents the slice image batch size; represents the data input size, represents the width of the image, represents the height of the image, represents the number of channels of the image; represents the number of bytes Bytes required for the data type Dtype, where Dtype represents the data type and Bytes represents the number of bytes; represents the storage space size occupied by the on-orbit mission inference model; represents the memory size occupied by the intermediate results generated during the operation of the on-orbit mission inference model; represents the video memory size of the intelligent acceleration card; represents the number of intelligent acceleration cards, represents the memory size of the CPU, " " means that the product of the video memory size of the intelligent acceleration card and the number of intelligent acceleration cards is less than or equal to the preset ratio of the CPU memory.

4. The on-board computer according to claim 3, characterized in that, The preset ratio is configured to 80%.

5. The on-board computer according to claim 4, characterized in that, The on-orbit task inference model is a pre-trained object detection model.

6. An on-orbit inference system for being carried on a satellite platform to perform on-orbit computing tasks, characterized in that The system includes: A data acquisition module, configured to acquire the image to be processed. The on-board computer according to any one of claims 1-5, configured to process the image to be processed to obtain the on-orbit inference result of the image to be processed.

7. The system according to claim 6, wherein The system further includes: A space-ground transmission module, configured to transmit the on-orbit inference result of the image to be processed back to the ground station.

8. An on-orbit reasoning method applicable to a satellite platform, characterized in that, The method includes: processing the image to be processed by using the on-orbit inference system according to any one of claims 6-7.

9. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and the computer program can be executed by a processor to implement the steps of the method according to claim 8.

10. An electronic device, characterized in that, Including: One or more processors, and A memory, wherein the memory is used to store executable instructions; The multiple processors are configured to implement the steps of the method according to claim 8 by executing the executable instructions.

Images