Efficient artificial intelligence algorithm and system for satellite intelligent decision

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses an efficient artificial intelligence algorithm and system for satellite intelligent decision making, and the algorithm comprises the following steps: constructing a lightweight high-energy-efficiency edge intelligent frame based on a DNN model; training by using a preset loss function, taking the output of the teacher model as a hard tag, learning the network information of the teacher model by the output of a classification token, comparing the output of the classification token with a real tag to calculate the cross entropy loss # imgabs0 #, and reducing the size of the model; according to the method, an early-leaving branch module based on heterogeneous calculation is designed, generalization ability, stability and reliability are evaluated by using uncertainty of a model, a lightweight high-energy-efficiency edge intelligent framework based on a DNN model is designed for edge equipment with limited resources on the premise of not sacrificing model performance, and the framework adopts a frequency domain model GFNet as a backbone network, so that the framework can be applied to edge equipment with limited resources. The method can learn long-term spatial dependence in a frequency domain and is low in calculation complexity, so that the method is more suitable for edge equipment.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, and in particular relates to an efficient artificial intelligence algorithm and system for satellite intelligent decision-making. Background Art

[0002] With the continuous advancement of the trend of miniaturization and lightweighting of satellites, the computing power of satellites has experienced an unprecedented leap. This technological advancement has not only broadened the boundaries of satellite applications, but also spawned a cutting-edge computing paradigm called orbital edge computing (OEC). Under the OEC framework, the role of satellites has undergone a fundamental change. They are no longer just transit stations for signal transmission, but have become key edge nodes in the network architecture, directly participating in the execution of data processing and computing tasks. The core of this change is that by deploying computing resources closer to the data source, that is, in Earth orbit, OEC can instantly process massive amounts of data from various space sensors, significantly reducing data transmission delays and improving the overall network response speed and communication efficiency.

[0003] However, despite the many advantages brought by OEC, the energy challenges faced by satellite systems remain severe. Since satellites operate in a space environment far away from the earth, their energy supply is almost entirely dependent on the energy collected by solar panels and limited battery reserves. Changes in sunlight intensity and the specific conditions of the satellite's orbit make energy acquisition unstable and limited. In addition, due to the strict requirements of satellite platforms on weight and volume, the capacity of battery packs is strictly limited, further exacerbating the difficulty of energy management. Therefore, how to maximize energy savings while ensuring the efficient execution of computing tasks has become a key issue that needs to be urgently addressed in the development of satellite technology. Summary of the invention

[0004] The purpose of the present invention is to provide an efficient artificial intelligence algorithm and system for satellite intelligent decision-making to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solution: an efficient artificial intelligence algorithm and system for satellite intelligent decision-making, comprising the following steps: S1. Build a high-energy-efficiency edge intelligence framework based on lightweight DNN model; S2. Use the preset loss function for training, take the output of the teacher model as the hard label, the output of the distillation token to learn the network information of the teacher model, and compare the output of the class token with the true label to calculate the cross entropy loss , reduce the model size; S3. Design an early-withdrawal branching module based on heterogeneous computing and use the uncertainty of the model to evaluate the generalization ability, stability and reliability.

[0006] Preferably, the construction of a lightweight high-energy-efficiency edge intelligence framework based on a DNN model specifically includes the following steps: S11. Select GFNet as the backbone network; S12. Configure framework parameters to adapt to satellite intelligent decision-making scenarios.

[0007] Preferably, in S2, the formula of the loss function is: .

[0008] Preferably, the cross entropy loss The calculation formula is: ; ; Among them, z L Indicates The feature vector output of the layer, where , is the total number of layers in the model, is the true label.

[0009] Preferably, the design is based on an early-retirement branch module of heterogeneous computing, and using the uncertainty of the model to evaluate the generalization ability, stability and reliability specifically includes the following steps: S31. Construct the classification head module, uncertainty statistics module UCSTAT and gate module ; S32. Use model uncertainty to assess generalization, stability, and reliability; S33. Make a decision on whether to exit a branch based on the current sample characteristics.

[0010] Preferably, the The formula is: ; in, For the final result, .

[0011] Preferably, the selecting GFNet as the backbone network specifically comprises the steps of: Investigate and evaluate the performance of GFNet in satellite intelligent decision-making tasks, including accuracy, computational complexity, and memory usage. Based on the evaluation results, determine the specific version and configuration parameters of GFNet.

[0012] Preferably, the configuration framework parameters to adapt to the satellite intelligent decision-making scenario specifically includes the steps of: Analyze the characteristics and requirements of satellite intelligent decision-making tasks, including real-time performance, accuracy, and resource limitations. According to the mission characteristics, adjust GFNet parameters and optimize the hardware acceleration support of the framework.

[0013] Preferably, the GFNet parameters include input size, output category, and training strategy.

[0014] Preferably, the GFNet includes three operation groups, specifically including 2D discrete Fourier transform, element-by-element multiplication between frequency domain features and learnable global filters, and 2D inverse Fourier transform.

[0015] Compared with the prior art, the present invention has the following beneficial effects: Without sacrificing model performance, the present invention designs a high-efficiency edge intelligence framework based on a lightweight DNN model for resource-limited edge devices. The framework adopts the frequency domain model GFNet as the backbone network, because it can learn long-term spatial dependencies in the frequency domain and has low computational complexity, it is more suitable for edge devices. Through the frequency domain distillation method, the model can achieve efficient image classification with fewer parameters. At the same time, it integrates an early exit branch module based on heterogeneous computing, allowing confidently classified samples to exit early at the shallow layer of the network, realizing the real-time and high energy efficiency of edge intelligence, reducing the number of model parameters, improving energy efficiency and reducing latency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall flow chart of the framework of the present invention; Figure 2 It is a flow chart of the frequency domain distillation of the present invention; Figure 3 This is a workflow diagram of the Early Exit Branch module of the present invention; Figure 4 It is a composition diagram of the GFNetBlock module of the present invention; Figure 5 The energy consumption and latency comparison results of different frameworks on AGX Orin of the present invention are shown; Figure 6 The energy consumption and delay comparison results of different frames on the Orin Nano of the present invention. DETAILED DESCRIPTION

[0017] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0018] See also Figure 1-Figure 6 As shown, the present invention provides the following technical solutions: The efficient artificial intelligence algorithm and system for satellite intelligent decision-making includes the following steps: S1. Build a high-energy-efficiency edge intelligence framework based on lightweight DNN model; S2. Use the preset loss function for training, take the output of the teacher model as the hard label, the output of the distillation token to learn the network information of the teacher model, and compare the output of the class token with the true label to calculate the cross entropy loss , reduce the model size; S3. Design an early-withdrawal branching module based on heterogeneous computing and use the uncertainty of the model to evaluate the generalization ability, stability and reliability.

[0019] In addition, in the present invention, regarding the above S11, building a high-efficiency edge intelligence framework based on lightweight DNN model specifically includes the following steps: S11. Select GFNet as the backbone network; S12. Configure framework parameters to adapt to satellite intelligent decision-making scenarios.

[0020] Furthermore, in the present invention, regarding the above S2, the formula of the loss function is: .

[0021] In addition, in the present invention, the cross entropy loss The calculation formula is: ; ; Among them, z L Indicates The feature vector output of the layer, where , is the total number of layers in the model, is the true label.

[0022] And, regarding the above S3, in the present invention, the design of the early-retirement branch module based on heterogeneous computing, using the uncertainty of the model to evaluate the generalization ability, stability and reliability specifically includes the following steps: S31. Construct the classification head module, uncertainty statistics module UCSTAT and gate module ; S32. Use model uncertainty to assess generalization, stability, and reliability; S33. Make a decision on whether to exit a branch based on the current sample characteristics.

[0023] In addition, regarding the above , The formula is: ; in, For the final result, .

[0024] In addition, the specific steps for selecting GFNet as the backbone network include: Investigate and evaluate the performance of GFNet in satellite intelligent decision-making tasks, including accuracy, computational complexity, and memory usage. Based on the evaluation results, determine the specific version and configuration parameters of GFNet.

[0025] Configuring framework parameters to adapt to satellite intelligent decision-making scenarios includes the following steps: Analyze the characteristics and requirements of satellite intelligent decision-making tasks, including real-time performance, accuracy, and resource limitations. According to the mission characteristics, adjust GFNet parameters and optimize the hardware acceleration support of the framework.

[0026] In addition, in the present invention, GFNet parameters include input size, output category, and training strategy.

[0027] And, regarding the above-mentioned GFNet, GFNet includes three operation groups, specifically including 2D discrete Fourier transform, element-by-element multiplication between frequency domain features and learnable global filters, and 2D inverse Fourier transform.

[0028] The present invention solves the problem of achieving computationally intensive tasks such as image analysis with high energy efficiency and low latency on resource-limited edge devices without sacrificing model performance. The present invention studies a high-energy-efficiency edge intelligence framework based on lightweight DNN models, which is based on the frequency domain model GFNet and is specially designed for energy-saving edge analysis on edge devices.

[0029] In addition, in the present invention, Figure 1 This is the overall process of the framework of this article. The framework selects GFNet as the backbone network. GFNet can learn long-term spatial dependencies in the frequency domain, and the computational complexity is O(nlog n), which makes it an ideal choice for the backbone network of this framework. In order to reduce the model size, the present invention also studies a frequency domain distillation method specifically for frequency domain models. This enables the model to achieve efficient image classification with fewer parameters. The present invention will study the incorporation of early exit branch modules based on heterogeneous computing into the distillation model. These branches allow samples that can be confidently classified by the shallow layer of the network to exit early. In general, the real-time and energy-efficient performance of edge intelligence is achieved by reducing the number of model parameters and allowing samples to exit early.

[0030] In frequency domain distillation, we introduced distillation token and class token, such as Figure 2 In the present invention, z L Indicates The feature vector output of the layer, where , is the total number of layers in the model. The distillation token and class token operate in a similar way, interacting globally with other patchtokens through the global filter layer to capture global information and being output by the network after the final layer. At the same time, a large, high-precision GFNet model is first pre-trained as the teacher model .

[0031] When training using the loss function (Formula 3), the output of the teacher model is used as the hard label , the output of the distillation token learns the network information of the teacher model to achieve higher efficiency. Its goal is to reproduce the hard label of the predicted teacher model (Formula 3), rather than the true label.

[0032] The class token learns the inherent characteristics and distribution of the data and compares its final output with the true label Compare to calculate the cross entropy loss .

[0033] ; ; The framework in this paper adopts the effective hard label distillation loss of DeiT as the total loss of the distillation stage, which is expressed as , where the hard decisions of the teacher model are used as the true labels.

[0034] ; The objective loss function used in the frequency domain distillation process yes: .

[0035] In addition, in the present invention, regarding the early exit branch module based on heterogeneous computing; The Early-Exit Branch mechanism consists of three modules, such as Figure 3 Shown: classification head module, uncertainty statistics module UCSTAT and gate module (Formula 5). The classifier module processes the intermediate features To obtain the intermediate prediction probability , as the input of the uncertainty statistics module. The Early-Exit Branch mechanism designed by the present invention utilizes the uncertainty of the model, that is, the confidence of the model in predicting new data. Calculating the uncertainty of the model helps to evaluate the generalization ability, stability and reliability of the model. Therefore, the introduction of model uncertainty makes The decision of whether to exit a branch based on the current sample features is more accurate.

[0036] In the present invention, the following four uncertainty metrics are selected as The uncertainty statistics module is responsible for calculating these uncertainty statistics: 1) is the maximum predicted probability; 2) is the entropy of the prediction; 3) ,in ; 4) ;, is the predicted probability The maximum value in is the second largest value.

[0037] The gating module concatenates these four uncertainty statistics into a one-dimensional vector and then uses a linear layer to get the final result. : ; In addition, in the present invention, an exit threshold is set in advance for each Early-Exit Branch module. , Range is between 0 and 1. For simplicity, the framework uses the same exit threshold for all branches. , the sample exits at this point. The larger the value, the fewer samples that exit early, and the higher the accuracy, and vice versa. In addition, in order to further improve energy efficiency and delay efficiency, the present invention will also study the use of heterogeneous computing to fully utilize the computing power of the CPU and GPU. Specifically, during the reasoning process, the parameters of the early exit branch module are loaded onto the CPU for calculation, while the parameters of other modules are loaded onto the GPU. This method can further improve the reasoning efficiency of the model by giving full play to the computing advantages of heterogeneous processors.

[0038] In addition, in order to achieve energy-efficient edge intelligence, this paper will study an energy-efficient edge intelligence framework based on lightweight DNN model, which uses GFNet as the backbone network. GFNet follows the trend of removing the inductive bias in the visual model and has The computational characteristics of complexity. The basic idea of ​​GFNet is to learn the interactions between spatial positions in the frequency domain. Unlike the self-attention mechanism in Vision Transformer and the fully connected layer in the MLP model, the interactions between Patchtokens are modeled as a set of learnable global filters that are applied to the frequency domain of the input features. Since the global filters can cover all frequency domains, GFNet can capture long-term and short-term interactions. Moreover, these filters are learned directly from the raw data without the introduction of artificial prior knowledge.

[0039] Given the high computational efficiency of GFNet compared to traditional deep neural networks (DNNs), it is an ideal backbone network for this framework. GFNet mainly consists of three key operations: 2D discrete Fourier transform, frequency domain features, and learnable global filters. , and 2D inverse Fourier transform. The present invention assumes that GFNet consists of The GFNet consists of a stack of GFNet blocks, each of which consists of a global filter layer (Global Filter Layer) and a feedforward network (FFN) (Formula 6, 7). In GFNet, the first step is to perform a two-dimensional discrete Fourier transform on the input spatial features to convert them to the frequency domain; the second step is to combine the frequency domain features with the global filter layer. Element-by-element multiplication is performed between them; finally, a two-dimensional inverse Fourier transform is performed to map the features back to the spatial domain. Since the Fourier transform is used to mix the information of different tokens, the global filter is more efficient than self-attention and MLP, thanks to the fast Fourier transform algorithm (FFT). With this in mind, the proposed global filter layer has a high efficiency on token length is less sensitive and is therefore compatible with larger feature maps and convolutional neural network (CNN) style layered architectures without modification.

[0040] ; ; Formulas 6 and 7 describe the operations within each GFNet block, where Representation layer normalization, represents a multi-layer perceptron, Represents element-by-element multiplication. Similar to Vit, the present invention converts the input image Reshape into a series of 2D patches ,in is the resolution of the original image, is the resolution of each image patch, and is the number of image patches. GFNet uses a constant embedding dimension D in all layers, so the present invention flattens the patch and maps it to D dimensions using Patch Embedding Module (PEM). .

[0041] In order to distill knowledge into the original GFNet model, the model introduces a distillation token, similar to the classtoken in ViT. We concatenate the class token and distillation token into the sequence of embedding patches and add position encoding (Formula 8). The student model is trained using the loss function in Formula 4. Then, an early exit branch (EEB) is added to the distilled student model. If the input sample exits at the last layer, the final prediction is obtained by averaging the interacted class token and distillation token: ; If the sample is dropped early, the prediction is derived only from the interacting class tokens: ; ; A large number of experimental results show that the framework of the present invention achieves up to 64.74% energy consumption reduction and 38.87% latency reduction on heterogeneous CPU-GPU edge devices, while the accuracy loss of the model is controlled within 3%.

[0042] In addition, in the present invention, regarding the experimental settings: We have used the NVIDIA Jetson AGX Orin 64GB development board and the NVIDIA Jetson OrinNano 8GB development board to simulate resource-constrained edge devices. The former can achieve an AI performance of 275 TOPS, while the latter has an AI performance of only 40 TOPS. We have also evaluated the inference performance on these two edge devices. In order to give full play to the computing advantages of heterogeneous processors, when loading model parameters, the parameters of the Early-Exit Branchy module are loaded on the CPU, and the other parameters of the model are loaded on the GPU, realizing simple heterogeneous computing. In addition, we use jetson-stats to monitor the energy consumption of the device when running DNN inference.

[0043] In the present invention, the specific configuration of the edge device is as shown in Table 1: Table 1 Specific configuration of edge devices In addition, in this invention, we use CIFAR10, UC Merced Land-Use

[35] and NWPU-RESISC45

[36] The framework is evaluated on datasets such as CIFAR-10. CIFAR-10 is a color image dataset that is closer to universal objects. It contains 10 classes, each image is 32×32 in size, each class has 6,000 images, and there are 50,000 training images and 10,000 test images in the dataset. The UC Merced Land-Use remote sensing dataset is a public dataset for remote sensing image scene classification published by the UC Merced Computer Vision Laboratory. It contains 21 categories of scenes, each category contains 100 remote sensing images, and the entire dataset contains 2,100 remote sensing images. The size of each remote sensing image is 256×256. The NWPU-RESISC45 remote sensing dataset is a large-scale public dataset for remote sensing image scene classification published by Northwestern Polytechnical University. It contains 45 categories of scenes, each category contains 700 remote sensing images, and the entire dataset contains 31,500 images. The size of each image is 256×256. It has the characteristics of large scale and rich information contained in the image. We use four NVIDIA GeForce GTX4090 GPUs (batch size is 128) to complete all model training processes on the edge server.

[0044] Furthermore, in the present invention, regarding performance evaluation: Comparisons of energy consumption, latency, and parameter count were conducted on these two edge devices. The results showed that on Jetson AGX Orin, compared with the baseline, the proposed framework significantly improved the energy saving rate on three data sets, achieving a 64.74% to 69.11% energy reduction rate. On Jetson Orin Nano, the proposed framework achieved a 64.21% to 71.03% energy reduction rate. Although the frequency domain distillation process greatly reduces the model's inference energy consumption, the addition of early exit branches to allow samples to exit early further reduces the model's energy consumption, and reduces energy consumption by about 12% to 17% on the basis of distillation, which is more conducive to achieving the goal of high-efficiency inference. In terms of inference latency, the proposed framework can achieve a latency reduction rate of up to 38.87%, while keeping the accuracy loss within 3%.

[0045] In general, the framework of this paper first obtains a small student model through frequency domain distillation. The distillation process greatly reduces the number of model parameters. Then, the Early-ExitBranch based on heterogeneous computing is added to the distilled student model. By allowing simple samples to exit early, the number of model layers is reduced in disguise, which reduces the power consumption during model reasoning. Even if the reasoning delay will increase, the overall energy consumption trend is still decreasing, and it is reduced by about 12% based on the student model after distillation.

[0046] In terms of model parameters, the student model Distil after distillation has been greatly reduced, with a parameter reduction rate of about 71.37%~72.75%, and the accuracy loss of the model after distillation is kept within 2%. Since the Early-Exit Branch is added to the distilled model in the early exit stage, the overall parameter volume of the model has increased, but only the classification head and the gating module have parameters in the Early-Exit Branch module, and the classification head and the gating module are designed to be as lightweight as possible. For example, the gating module is only a 4×1 linear layer, and the intermediate reasoning output only uses the output of the class token. Therefore, in the early exit stage, the increased parameters account for about 6%, which is acceptable compared with the energy consumption reduction brought by the early exit technology. In addition, in the present invention, the comparison of the parameter volume of different frameworks is shown in Table 2: Table 2 Comparison of parameters of different frameworks Furthermore, in the present invention, regarding the effect of frame rate on accuracy: In addition, in order to simulate the scenario where the computing resources of edge devices are extremely limited (i.e., equipped with only CPU but not GPU), the present invention also conducted experiments on CPU-only reasoning, in which all model parameters were loaded on the CPU. Unlike GPU, CPU cannot accelerate the calculation of various modules in the DNN model

[21] . Due to the lack of ability to accelerate matrix operations, the reasoning time of each layer of the model is about ten times that of the Early-ExitBranch module. In addition, since most input samples can exit at the middle layer, the evaluation performance of CPU-only reasoning should also be good. A large number of experimental results show that when only CPU is available, the framework can still achieve a maximum energy consumption reduction rate of 63.18% and a latency reduction rate of 58.88%. In addition, in the present invention, the experimental evaluation results of CPU-only reasoning are shown in Table 3: Table 3 Experimental evaluation results of CPU-only reasoning Moreover, regarding the ablation experiment, in the present invention, the present invention also compares the effects of frequency domain distillation and early exit branch on energy consumption reduction rate and delay reduction, as shown in Table 5-3. Distil means that only the frequency domain distillation method is used, while Dynn means that the original model adds an early exit branch but does not perform frequency domain distillation. Distil is superior to Dynn in terms of high energy efficiency, low latency and model parameters, while Dynn only shows advantages in accuracy. Due to frequency domain distillation, the model size is significantly reduced, and the parameters are reduced by two times compared with Dynn. In addition, the energy consumption reduction rate of Dynn ranges from 18.35% to 35.29%, while that of Distil ranges from 56.7% to 61.51%. In terms of reducing energy consumption and delay, frequency domain distillation performs better than the current early exit branch mechanism, and in the present invention, the effects of frequency domain distillation and early exit branch on framework energy consumption and delay are shown in Table 4: Table 4 Effects of frequency domain distillation and early-retirement branching on framework energy consumption and delay In addition, we implemented the framework of the present invention using Python 3.9 on two heterogeneous NVIDIA Jetson edge devices. In addition, we used jetson-stats to monitor the energy consumption of the device when running DNN inference. We evaluated the framework on datasets such as CIFAR10, UC Merced Land-Use, and NWPU-RESISC45. Four early exit points were placed for each DNN model. We trained the early exit DNN model offline on an edge server using four NVIDIA GeForce GTX 4090 GPUs with a mini-batch size of 128. All networks were trained using the Adam optimizer with a learning rate of 10-4.

[0047] The present invention solves the problem of achieving computationally intensive tasks such as image analysis with high energy efficiency and low latency on resource-limited edge devices without sacrificing model performance. The present invention studies a high-energy-efficiency edge intelligence framework based on lightweight DNN models, which is based on the frequency domain model GFNet and is specially designed for energy-saving edge analysis on edge devices. Figure 1The overall process of the framework in this paper is explained. The framework selects GFNet as the backbone network. GFNet can learn long-term spatial dependencies in the frequency domain, and the computational complexity is O(n log n), which makes it an ideal choice for the backbone network of this framework. In order to reduce the model size, the present invention also studies a frequency domain distillation method specifically for frequency domain models. This enables the model to achieve efficient image classification with fewer parameters. The present invention will study the incorporation of early exit branch modules based on heterogeneous computing into the distillation model. These branches allow samples that can be confidently classified by the shallow layer of the network to exit early. In general, the real-time and high energy efficiency of edge intelligence are achieved by reducing the number of parameters of the model and allowing samples to exit early.

[0048] In addition, the following implementation modes are provided: S1. Building a high-efficiency edge intelligence framework based on lightweight DNN models S11. Select GFNet as the backbone network: Investigate the performance of GFNet on satellite image classification tasks, evaluate its accuracy, computational complexity (especially the complexity advantage of O(n log n)), and memory usage. Based on the evaluation results, select GFNet-V2, configure its input size to 224x224, output categories to 10 (assuming there are 10 categories), and adopt a step-by-step training strategy to optimize model performance.

[0049] S12. Configure framework parameters: In view of the characteristics of satellite intelligent decision-making tasks, such as high real-time requirements and limited resources, adjust the hardware acceleration support of GFNet, such as using the CUDA acceleration library to optimize the running efficiency of the model on NVIDIA GPU, while ensuring the compatibility of the model on low-power edge devices.

[0050] S2. Training using a preset loss function The loss function uses a composite loss function that combines cross entropy loss and distillation loss. Cross entropy loss is used to guide the matching of the output of the class token with the true label, while distillation loss learns the network information of the teacher model through distillation token. The specific formula is as follows: Cross Entropy Loss: LCE =−∑ i =1 Cyi log( pi ),in yi is the one-hot encoding of the true label, pi is the probability distribution predicted by the model.

[0051] Distillation loss: Calculate the difference between the student model and the teacher model based on the output of the teacher model. The specific formula is omitted (because it is not provided in the original text).

[0052] S3. Design of early-retirement branch module based on heterogeneous computing S31. Build modules: Design an early exit branch including a classification head module, an uncertainty statistics module UCSTAT, and a gate module. The classification head is responsible for outputting the classification results, UCSTAT is used to evaluate the uncertainty of the model, and the gate module decides whether to exit early based on the uncertainty.

[0053] S32. Evaluate uncertainty: The uncertainty of the current sample is calculated through the UCSTAT module, including indicators such as entropy and variance of the predicted probability distribution, which are used to evaluate the generalization ability, stability and reliability of the model.

[0054] S33. Decision exit: According to the uncertainty threshold, if the uncertainty of the current sample is lower than the threshold, the model is considered to have sufficient confidence in the sample and can exit early in the shallow branch to reduce unnecessary computational overhead.

[0055] Furthermore, the present invention further provides an implementation mode: S1. Building the framework S11. Select GFNet: In view of the characteristics of satellite remote sensing data, GFNet-V3 version is selected because it has stronger feature extraction capabilities and lower computational complexity. The input size is configured to be 512x512, the output category is set according to the specific task (such as land cover classification, disaster monitoring, etc.), and the training strategy adopts a combination of transfer learning and fine-tuning.

[0056] S12. Parameter configuration: Consider the real-time processing requirements of satellite remote sensing data and optimize GFNet's hardware acceleration support, such as using TensorRT for model optimization to improve operating efficiency on edge devices.

[0057] S2. Training and Optimization The loss function also uses a composite loss function, combining cross entropy loss and distillation loss to improve the accuracy and generalization ability of the model.

[0058] S3. Early-withdrawal branch design S31. Module construction: Design an early exit branch that includes a multi-level classification head, an uncertainty statistics module, and a dynamic gating mechanism. The multi-level classification head is used to output classification results at different levels, the uncertainty statistics module evaluates the uncertainty of the model, and the dynamic gating mechanism dynamically adjusts the exit strategy according to the uncertainty.

[0059] S32. Uncertainty assessment: Use a variety of uncertainty assessment methods, such as Bayesian networks and Monte Carlo dropout, to comprehensively assess the uncertainty of the model.

[0060] S33. Decision and exit: Based on the uncertainty threshold and dynamic gating strategy, decide whether to exit the calculation of the current sample in advance. For high-confidence samples, exit in advance at the shallow branch to reduce calculation delay and resource consumption.

[0061] Furthermore, the present invention further provides an implementation mode: S1. Building the framework S11. Select GFNet: In view of the real-time and accuracy requirements of the satellite navigation system, select the GFNet-Lite version, which has lower computational complexity and memory usage while ensuring accuracy. Configure the input size to 128x128, the output category to be the number of categories of navigation instructions, and the training strategy adopts a combination of online learning and incremental learning.

[0062] S12. Parameter configuration: Optimize GFNet's hardware acceleration support, such as using OpenCL to optimize the model's running efficiency on heterogeneous computing platforms, while ensuring the model's compatibility on low-power embedded devices.

[0063] S2. Training and Optimization The loss function adopts a combination of cross entropy loss and distillation loss to improve the accuracy and efficiency of the model.

[0064] S3. Early-withdrawal branch design S31. Module construction: Design an early exit branch that includes a fast classification head, an uncertainty monitoring module, and an adaptive exit mechanism. The fast classification head is used to quickly output classification results, the uncertainty monitoring module monitors the uncertainty of the model in real time, and the adaptive exit mechanism dynamically adjusts the exit strategy according to the uncertainty.

[0065] S32. Uncertainty monitoring: Use real-time monitoring techniques, such as sliding window method, online learning algorithm, etc. to monitor the uncertainty of the model in real time.

[0066] S33. Decision and exit: Dynamically adjust the exit timing of samples according to the uncertainty threshold and adaptive exit strategy to achieve efficient and accurate navigation-assisted decision making.

[0067] To address this challenge, our research focuses on optimizing deep neural network (DNN) models, aiming to provide a low-latency, energy-efficient and high-precision solution for satellite edge computing by reducing the number of model parameters and introducing a dynamic sample exit mechanism. Specifically, we use model compression techniques such as weight pruning, quantization, and knowledge distillation to significantly reduce the complexity and memory usage of DNN, thereby reducing energy consumption during the computation process. At the same time, we design an intelligent sample processing strategy that allows some samples to exit the computation process in early neural network layers under the premise of meeting a certain accuracy threshold, avoiding unnecessary deep processing and further saving energy. This optimization strategy has far-reaching significance for satellite technology and the low-Earth orbit (LEO) satellite industry. With the rapid deployment of LEO satellite constellations and the gradual reduction of satellite launch costs, more and more commercial and scientific applications have begun to rely on these "space base stations" in low-Earth orbit. Our work not only helps to improve the energy efficiency of satellites when performing edge intelligence tasks (such as image analysis, environmental monitoring, etc.), but also promotes the popularization and commercialization of satellite technology, laying a solid foundation for building a more intelligent, flexible and sustainable space information network.

[0068] In addition, these technical optimizations have also inspired new ideas for future satellite design, encouraging developers to pay more attention to energy efficiency and intelligence when designing satellite missions and system architectures, and promoting satellite technology to develop in a greener, more efficient and autonomous direction. In short, by deeply optimizing DNN models to adapt to satellite energy constraints, we are gradually unlocking the unlimited potential of satellite edge computing, opening up new possibilities for human society to explore the universe, protect the environment, and promote economic development. Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An efficient artificial intelligence algorithm and system for satellite intelligent decision-making, characterized by: The steps include: S1. Build a high-energy-efficiency edge intelligence framework based on lightweight DNN model; S2. Use the preset loss function for training, take the output of the teacher model as the hard label, the output of the distillation token to learn the network information of the teacher model, and compare the output of the class token with the true label to calculate the cross entropy loss , reduce the model size; S3. Design an early-withdrawal branching module based on heterogeneous computing and use the uncertainty of the model to evaluate the generalization ability, stability and reliability.

2. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 1, characterized in that: The construction of a lightweight high-energy-efficiency edge intelligence framework based on a DNN model specifically includes the following steps: S11. Select GFNet as the backbone network; S12. Configure framework parameters to adapt to satellite intelligent decision-making scenarios.

3. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 1, characterized in that: In S2, the formula of the loss function is: ； in, For the teacher model, For hard tags.

4. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 1, characterized in that: The cross entropy loss The calculation formula is: ； ； Among them, z L Indicates The feature vector output of the layer, where , is the total number of layers in the model, is the true label.

5. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 1, characterized in that: The design is based on the early-withdrawal branch module of heterogeneous computing, and uses the uncertainty of the model to evaluate the generalization ability, stability and reliability. The specific steps include: S31. Construct the classification head module, uncertainty statistics module UCSTAT and gate module ; S32. Use model uncertainty to assess generalization, stability, and reliability; S33. Make a decision on whether to exit a branch based on the current sample characteristics.

6. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 5, characterized in that: Said The formula is: ； in, For the final result, .

7. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 2, characterized in that: The selection of GFNet as the backbone network specifically includes the following steps: Investigate and evaluate the performance of GFNet in satellite intelligent decision-making tasks, including accuracy, computational complexity, and memory usage. Based on the evaluation results, determine the specific version and configuration parameters of GFNet.

8. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 2, characterized in that: The configuration framework parameters to adapt to the satellite intelligent decision-making scenario specifically include the following steps: Analyze the characteristics and requirements of satellite intelligent decision-making tasks, including real-time performance, accuracy, and resource limitations. According to the mission characteristics, adjust GFNet parameters and optimize the hardware acceleration support of the framework.

9. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 8, characterized in that: The GFNet parameters include input size, output category, and training strategy.

10. The efficient artificial intelligence algorithm and system for satellite intelligent decision-making according to claim 7, characterized in that: The GFNet includes three operation groups, specifically including 2D discrete Fourier transform, element-by-element multiplication between frequency domain features and learnable global filters, and 2D inverse Fourier transform.

Images