Efficient artificial intelligence system for satellite intelligent decision

Abstract

The application belongs to the technical field of artificial intelligence, and discloses an efficient artificial intelligence system for satellite intelligent decision-making, which comprises the following steps: constructing a high-energy-efficiency edge intelligent framework based on DNN model lightening; training by using a preset loss function, taking the output of a teacher model as a hard label, learning the network information of the teacher model by the output of a distillation token, comparing the output of a class token with a real label to calculate cross-entropy loss, and reducing the model size; designing an early exit branch module based on heterogeneous computing, and evaluating the generalization ability, stability and reliability by using the uncertainty of the model; and under the premise of not sacrificing the model performance, designing a high-energy-efficiency edge intelligent framework based on DNN model lightening for the resource-limited edge device, which adopts a frequency domain model GFNet as a backbone network, and is more suitable for edge devices because it can learn long-term spatial dependence in the frequency domain and has low computational complexity.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to an efficient artificial intelligence system for satellite intelligent decision-making. Background Technology

[0002] With the continuous advancement of satellite miniaturization and lightweighting, satellite computing power has experienced an unprecedented leap. This technological advancement has not only broadened the boundaries of satellite applications but also given rise to the cutting-edge computing paradigm of Orbital Edge Computing (OEC). Within the OEC framework, the role of satellites has fundamentally changed. They are no longer merely relay stations for signal transmission but have become critical edge nodes in the network architecture, directly participating in the execution of data processing and computing tasks. The core of this transformation lies in the fact that by deploying computing resources closer to the data source—namely, in Earth orbit—OEC can process massive amounts of data from various space sensors in real time, significantly reducing data transmission latency and improving the overall network's response speed and communication efficiency.

[0003] However, despite the numerous advantages brought by OEC, the energy challenges facing satellite systems remain significant. Because satellites operate in the remote space environment, their energy supply relies almost entirely on energy collected by solar panels and limited battery reserves. Variations in solar radiation intensity and the specific conditions of satellite orbits make energy acquisition unstable and limited. Furthermore, the stringent weight and size requirements of satellite platforms severely limit battery capacity, further complicating energy management. Therefore, how to maximize energy conservation while ensuring efficient execution of computing tasks has become a critical issue that urgently needs to be addressed in the development of satellite technology.

[0004] This patent is based on research conducted under the National Key Research and Development Program (Project No.: 2022YFF0503900). Summary of the Invention

[0005] The purpose of this invention is to provide an efficient artificial intelligence system for satellite intelligent decision-making, in order to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a highly efficient artificial intelligence system for satellite intelligent decision-making, comprising the following steps:

[0007] S1. Construct a lightweight, energy-efficient edge intelligence framework based on a DNN model;

[0008] S2. Training is performed using a pre-defined loss function. 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, and the output of the class token is compared with the true label to calculate the cross-entropy loss. Reduce the model size;

[0009] S3. Design an early branching module based on heterogeneous computing, and use the uncertainty of the model to evaluate the generalization ability, stability and reliability.

[0010] Preferably, the construction of a lightweight, energy-efficient edge intelligence framework based on a DNN model specifically includes the following steps:

[0011] S11. Select GFNet as the backbone network;

[0012] S12. Configure framework parameters to adapt to satellite intelligent decision-making scenarios.

[0013] Preferably, in step S2, the formula for the loss function is:

[0014] .

[0015] Preferably, the cross-entropy loss The calculation formula is:

[0016] ;

[0017] ;

[0018] Among them, z L Indicates the first The layer's feature vector output, where , It is the total number of layers in the model. This is a real label.

[0019] Preferably, the design based on the early termination branch module of heterogeneous computing, which utilizes the uncertainty of the model to evaluate generalization ability, stability, and reliability, specifically includes the following steps:

[0020] S31. Construct the classification header module, the uncertainty statistics module UCSTAT, and the gate module. ;

[0021] S32. Utilize the model's uncertainty to assess its generalization ability, stability, and reliability;

[0022] S33. Make a decision on whether to exit a branch based on the current sample characteristics.

[0023] Preferably, the The formula is:

[0024] ;

[0025] in, For the final result, .

[0026] Preferably, the selection of GFNet as the backbone network specifically includes the following steps:

[0027] 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.

[0028] Preferably, configuring the framework parameters to adapt to the satellite intelligent decision-making scenario specifically includes the following steps:

[0029] The characteristics and requirements of satellite intelligent decision-making tasks are analyzed, including real-time performance, accuracy, and resource constraints. Based on the task characteristics, GFNet parameters are adjusted and the hardware acceleration support of the framework is optimized.

[0030] Preferably, the GFNet parameters include input size, output category, and training strategy.

[0031] Preferably, the GFNet includes three operation groups, specifically including 2D discrete Fourier transform, element-wise multiplication between frequency domain features and learnable global filters, and 2D inverse Fourier transform.

[0032] Compared with the prior art, the beneficial effects of the present invention are:

[0033] This invention designs a lightweight, energy-efficient edge intelligence framework based on a DNN model for resource-constrained edge devices without sacrificing model performance. The framework uses 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, making it more suitable for edge devices. Through frequency domain distillation, the model achieves efficient image classification with fewer parameters. At the same time, it integrates an early exit branch module based on heterogeneous computing, allowing samples that are confident in their classification to exit early in the shallow layers of the network. This achieves real-time performance and high energy efficiency in edge intelligence, reduces the number of model parameters, improves energy efficiency, and reduces latency. Attached Figure Description

[0034] Figure 1 This is an overall flowchart of the framework of the present invention;

[0035] Figure 2 This is a flowchart of the frequency domain distillation process of the present invention;

[0036] Figure 3 This is a flowchart of the Early Exit Branch module of the present invention;

[0037] Figure 4 This is a diagram showing the composition of the GFNetBlock module of the present invention;

[0038] Figure 5 This is a comparison of the energy consumption and latency of different frames on the AGX Orin of this invention;

[0039] Figure 6 This presents a comparison of energy consumption and latency of different frames on the Orin Nano of this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Please see Figures 1-6 As shown, the present invention provides the following technical solution:

[0042] An efficient artificial intelligence system for satellite-based intelligent decision-making includes the following steps:

[0043] S1. Construct a lightweight, energy-efficient edge intelligence framework based on a DNN model;

[0044] S2. Training is performed using a pre-defined loss function. 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, and the output of the class token is compared with the true label to calculate the cross-entropy loss. Reduce the model size;

[0045] S3. Design an early branching module based on heterogeneous computing, and use the uncertainty of the model to evaluate the generalization ability, stability and reliability.

[0046] Furthermore, in this invention, regarding the above-mentioned S11, constructing a lightweight, energy-efficient edge intelligence framework based on a DNN model specifically includes the following steps:

[0047] S11. Select GFNet as the backbone network;

[0048] S12. Configure framework parameters to adapt to satellite intelligent decision-making scenarios.

[0049] Furthermore, in this invention, the formula for the loss function regarding S2 is:

[0050] .

[0051] Furthermore, in this invention, cross-entropy loss The calculation formula is:

[0052] ;

[0053] ;

[0054] Among them, z L Indicates the first The layer's feature vector output, where , It is the total number of layers in the model. This is a real label.

[0055] Furthermore, regarding the aforementioned S3, in this invention, the design of the early-retreat branch module based on heterogeneous computing, which utilizes the uncertainty of the model to evaluate generalization ability, stability, and reliability, specifically includes the following steps:

[0056] S31. Construct the classification header module, the uncertainty statistics module UCSTAT, and the gate module. ;

[0057] S32. Utilize the model's uncertainty to assess its generalization ability, stability, and reliability;

[0058] S33. Make a decision on whether to exit a branch based on the current sample characteristics.

[0059] In addition, regarding the above , The formula is:

[0060] ;

[0061] in, For the final result, .

[0062] Furthermore, the specific steps involved in selecting GFNet as the backbone network are as follows:

[0063] 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.

[0064] Configuring framework parameters to adapt to satellite intelligent decision-making scenarios includes the following steps:

[0065] The characteristics and requirements of satellite intelligent decision-making tasks are analyzed, including real-time performance, accuracy, and resource constraints. Based on the task characteristics, GFNet parameters are adjusted and the hardware acceleration support of the framework is optimized.

[0066] In addition, in this invention, the GFNet parameters include input size, output class, and training strategy.

[0067] Furthermore, regarding the aforementioned GFNet, GFNet includes three operation groups, specifically including 2D Discrete Fourier Transform, element-wise multiplication between frequency domain features and learnable global filters, and 2D Inverse Fourier Transform.

[0068] This invention addresses the challenge of achieving computationally intensive tasks such as image analysis with high energy efficiency and low latency on resource-constrained edge devices without sacrificing model performance. The invention investigates a lightweight, energy-efficient edge intelligence framework based on a DNN model. This framework, grounded in the frequency domain model GFNet, is specifically designed for energy-efficient edge analysis on edge devices.

[0069] Furthermore, in this invention, Figure 1 This outlines the overall workflow of the framework presented in this paper. The framework selects GFNet as its backbone network. GFNet can learn long-term spatial dependencies in the frequency domain and has a computational complexity of O(n log n), making it an ideal choice for the backbone network of this framework. To reduce the model size, this invention also investigates a frequency domain distillation method specifically for frequency domain models. This enables the model to achieve efficient image classification with fewer parameters. This invention will investigate incorporating early exit branch modules based on heterogeneous computing into the distillation model. These branches allow samples that can be confidently classified through shallow layers of the network to exit early. In summary, real-time performance and high energy efficiency of edge intelligence are achieved by reducing the number of model parameters and allowing samples to exit early.

[0070] In frequency domain distillation, we introduce distillation tokens and class tokens, such as Figure 2 As shown. In this invention, z L Indicates the first The layer's feature vector output, where , This represents the total number of layers in the model. Distillation tokens and class tokens operate in a similar manner, interacting globally with other patch tokens through a global filtering layer to capture global information, and are output by the network after the final layer. Simultaneously, a large, high-precision GFNet model is pre-trained as a teacher model. .

[0071] When training using the loss function (Equation 3), the output of the teacher model is used as a hard label. The output of the distillation token learns network information from the teacher model to achieve higher efficiency. Its goal is to reproduce the hard labels of the predictive teacher model. (Formula 3), instead of the actual label.

[0072] The class token learns the inherent characteristics and distribution of the data itself, and then compares its final output with the real labels. Compare them to calculate the cross-entropy loss. .

[0073] ;

[0074] ;

[0075] This paper's framework uses the effective hard-labeled distillation loss of DeiT as the total loss of the distillation stage, expressed as: In this model, the hard decisions of the teacher model are used as the true labels.

[0076] ;

[0077] The target loss function used in the frequency domain distillation process yes:

[0078] .

[0079] Additionally, in this invention, there is a module for early termination based on heterogeneous computing;

[0080] The Early-Exit Branch mechanism consists of three modules, such as... Figure 3 As shown: Classification Header Module, Uncertainty Statistics Module UCSTAT, and Gate Module (Formula 5). The classification head module processes intermediate features. To obtain intermediate prediction probability The Early-Exit Branch mechanism designed in this invention utilizes model uncertainty, i.e., the confidence level of the model's predictions for new data, as input to the uncertainty statistics module. Calculating model uncertainty helps evaluate the model's generalization ability, stability, and reliability. Therefore, introducing model uncertainty makes... Making decisions about whether to exit a branch based on the current sample characteristics is more accurate.

[0081] In this invention, the following four uncertainty measures are selected as... The input is the uncertainty statistics module, which is responsible for calculating these uncertainty statistics:

[0082] 1) It is the maximum prediction probability;

[0083] 2) It is the predicted entropy;

[0084] 3) ,in ;

[0085] 4) ;, It is the predicted probability. The maximum value in, It is the second largest value.

[0086] The gating module concatenates these four uncertainty statistics into a one-dimensional vector, and then uses a linear layer to obtain the final result. :

[0087] ;

[0088] In addition, in this invention, an exit threshold is pre-set for each Early-Exit Branch module. , The range is between 0 and 1. For simplicity, the framework uses the same exit threshold for all branches. If If the sample exits at that point, then the sample will be removed. The larger the value, the fewer samples exit early, and the higher the accuracy, and vice versa. Furthermore, to further improve energy efficiency and latency efficiency, this invention will also explore the use of heterogeneous computing to fully utilize the computing power of CPUs and GPUs. Specifically, during inference, the parameters of the early-leaving branch module are loaded onto the CPU for calculation, while the parameters of other modules are loaded onto the GPU. This method, by fully leveraging the computing advantages of heterogeneous processors, can further improve the inference efficiency of the model.

[0089] Furthermore, to achieve energy-efficient edge intelligence, this invention will investigate a lightweight, energy-efficient edge intelligence framework based on a DNN model, using GFNet as the backbone network. GFNet follows the trend of removing inductive bias from visual models, while also possessing... The computational complexity of GFNet is a key feature. The fundamental idea behind GFNet is to learn the interactions between spatial locations in the frequency domain. Unlike the self-attention mechanism in Vision Transformer and the fully connected layers in MLP models, the interactions between patch tokens are modeled as a set of learnable global filters applied to the frequency domain of the input features. Because global filters can cover the entire frequency domain, GFNet can capture both long-term and short-term interactions. Furthermore, these filters are learned directly from the raw data, without requiring any prior human knowledge.

[0090] Given GFNet's high computational efficiency compared to traditional deep neural networks (DNNs), it is the ideal backbone network for the framework presented in this paper. GFNet mainly consists of three key operations: 2D Discrete Fourier Transform, frequency domain features, and a learnable global filter. Element-wise multiplication between them, and 2D inverse Fourier transform. This invention assumes that GFNet is composed of... The GFNet consists of stacked blocks, each block comprising a global filter layer and a feedforward network (FFN) (Equations 6, 7). In GFNet, the input spatial features are first subjected to a two-dimensional discrete Fourier transform to convert them to the frequency domain; then, the frequency domain features are compared with the global filter... Element-wise multiplication is performed between the elements; 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 information from different tokens, global filters are more efficient than self-attention and MLP, thanks to the Fast Fourier Transform (FFT) algorithm. Complexity. Leveraging this, the proposed global filter layer reduces token length... Its low sensitivity allows it to be compatible with large feature maps and hierarchical architectures in the style of convolutional neural networks (CNNs) without modification.

[0091] ;

[0092] ;

[0093] Equations 6 and 7 describe the operations within each GFNet block, where Representation layer normalization, This represents a multilayer perceptron. This represents element-wise multiplication. Similar to Vit, this invention uses the input image... Remodeled into a series of 2D patches ,in It is the resolution of the original image. It is the resolution of each image patch, and This refers to the number of image patches. GFNet uses a constant embedding dimension D across all layers; therefore, this invention flattens the patches and maps them to dimension D using the Patch Embedding Module (PEM). .

[0094] To distill knowledge into the original GFNet model, a distillation token, similar to the class token in ViT, is introduced. The class token and distillation token are concatenated into the embedding patch sequence, and positional encoding is added (Equation 8). The student model is trained using the loss function in Equation 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.

[0095] ;

[0096] If the sample exits early, the prediction can be derived solely from the class token of the interaction:

[0097] ;

[0098] ;

[0099] Extensive experimental results show that the framework of this invention achieves a maximum energy consumption reduction of 64.74% and a latency reduction of 38.87% on heterogeneous CPU-GPU edge devices, while keeping the accuracy loss of the model within 3%.

[0100] Furthermore, regarding the experimental setup in this invention:

[0101] We have used the NVIDIA Jetson AGX Orin 64GB development board and the NVIDIA Jetson Orin Nano 8GB development board to simulate resource-constrained edge devices. The former achieved an AI performance of 275 TOPS, while the latter only achieved 40 TOPS. Inference performance was evaluated on both edge devices. To fully leverage the computational advantages of heterogeneous processors, when loading model parameters, the parameters of the Early-Exit Branchy module were loaded onto the CPU, while other model parameters were loaded onto the GPU, achieving simple heterogeneous computing. Furthermore, we used jetson-stats to monitor the energy consumption of the devices during DNN inference.

[0102] In this invention, the specific configuration of the edge device is shown in Table 1:

[0103]

[0104] Table 1 Specific configuration of edge devices

[0105] Additionally, in this invention, we use CIFAR10 and UC Merced Land-Use.

[35] and NWPU-RESISC45

[36] We will evaluate the framework using other datasets. CIFAR-10 is a color image dataset that more closely approximates general objects. It contains 10 classes, each image is 32×32 pixels, and there are 6000 images per class. The dataset contains 50,000 training images and 10,000 test images. The UC Merced Land-Use remote sensing dataset is a public dataset for scene classification of remote sensing images released by the UC Merced Computer Vision Lab. It contains 21 scene classes, with 100 images per class, for a total of 2100 images. Each image is 256×256 pixels. The NWPU-RESISC45 remote sensing dataset is a large-scale public dataset for scene classification of remote sensing images released by Northwestern Polytechnical University. It contains 45 scene classes, with 700 images per class, for a total of 31,500 images. Each image is 256×256 pixels and is characterized by its large scale and rich information content. We used four NVIDIA GeForce GTX4090 GPUs (batch size 128) to complete the entire training process of the model on an edge server.

[0106] Furthermore, regarding performance evaluation in this invention:

[0107] A comparison of energy consumption, latency, and parameter count was conducted through extensive experimental evaluation on two edge devices. Results show that on the Jetson AGX Orin, compared to the baseline, the proposed framework significantly improves energy savings across three datasets, achieving a 64.74%–69.11% energy reduction. On the Jetson Orin Nano, the framework achieves a 64.21%–71.03% energy reduction. While frequency domain distillation significantly reduces inference energy consumption, adding early termination branches to allow samples to exit prematurely further reduces energy consumption, resulting in an additional 12%–17% reduction on top of distillation, which is more conducive to achieving high-energy-efficiency inference. Regarding inference latency, the proposed framework achieves a maximum latency reduction of 38.87% while maintaining accuracy loss within 3%.

[0108] In summary, the framework of this paper first obtains a small student model through frequency domain distillation, which greatly reduces the number of model parameters. Then, an Early-Exit Branch based on heterogeneous computing is added to the distilled student model. By allowing simple samples to exit early, the number of layers in the model is reduced, which reduces the power consumption during model inference. Even if some inference latency increases, the overall energy consumption trend is still downward, and it is about 12% lower than that of the distilled student model.

[0109] Regarding model parameters, the distilled student model Distil saw a significant reduction, with a parameter reduction rate of approximately 71.37% to 72.75%, while the accuracy loss remained within 2%. Because an Early-Exit Branch was added to the distilled model during the early termination phase, the overall number of parameters increased. However, only the classification head and gating module in the Early-Exit Branch module have parameters, and these modules were designed to be as lightweight as possible (e.g., the gating module is only a 4×1 linear layer, and the intermediate inference output only uses the class token output). Therefore, the increase in parameters during the early termination phase accounts for approximately 6%, which is acceptable compared to the energy reduction brought by the early termination technique. Furthermore, a comparison of the number of parameters for different frameworks in this invention is shown in Table 2.

[0110]

[0111] Table 2 Comparison of parameter counts for different frameworks

[0112] Furthermore, regarding the impact of frame rate on accuracy in this invention:

[0113] In addition, in order to simulate the scenario where edge devices have extremely limited computing resources (i.e., only CPU is provided and no GPU is provided), this invention also conducted experiments on CPU-only inference, in which all model parameters are loaded onto the CPU. Unlike GPU, CPU cannot accelerate the computation of various modules in DNN model

[21] . Due to the lack of ability to accelerate matrix operations, the inference time of each layer of the model is about ten times that of the Early-Exit Branch module. In addition, since most input samples can exit in the middle layer, the evaluation performance of CPU-only inference 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, the experimental evaluation results of CPU-only inference in this invention are shown in Table 3:

[0114]

[0115] Table 3. Experimental evaluation results of CPU-only inference

[0116] Furthermore, regarding the ablation experiment, this invention also compared the effects of frequency domain distillation and early withdrawal branching on energy consumption reduction and latency reduction, as shown in Table 5-3. Distil represents the method using only frequency domain distillation, while Dynn represents the original model with early withdrawal branching but without frequency domain distillation. Distil outperforms Dynn in terms of high energy efficiency, low latency, and model parameters, while Dynn only shows an advantage in accuracy. Due to frequency domain distillation, the model size is significantly reduced, with parameters reduced by a factor of two compared to Dynn. In addition, Dynn's energy consumption reduction rate ranges from 18.35% to 35.29%, while Distil's ranges from 56.7% to 61.51%. In terms of reducing energy consumption and latency, frequency domain distillation outperforms the current early withdrawal branching mechanism. Furthermore, in this invention, the effects of frequency domain distillation and early withdrawal branching on frame energy consumption and latency are shown in Table 4:

[0117]

[0118] Table 4. The impact of frequency domain distillation and early branching on frame energy consumption and delay.

[0119] Furthermore, we implemented the framework of this invention using Python 3.9 on two heterogeneous NVIDIA Jetson edge devices. Additionally, we utilized jetson-stats to monitor the energy consumption of the devices during DNN inference. We evaluated the framework on datasets such as CIFAR10, UC Merced Land-Use, and NWPU-RESISC45. Each DNN model had four early exit points. We trained the early exit DNN models 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.

[0120] This invention addresses the challenge of achieving computationally intensive tasks such as image analysis with high energy efficiency and low latency on resource-constrained edge devices without sacrificing model performance. The invention investigates a lightweight, energy-efficient edge intelligence framework based on a DNN model. This framework, grounded in the frequency domain model GFNet, is specifically designed for energy-efficient edge analysis on edge devices. Figure 1This paper describes the overall process of the framework. The framework selects GFNet as the backbone network. GFNet can learn long-term spatial dependencies in the frequency domain and has a computational complexity of O(n log n), making it an ideal choice for the backbone network of this framework. To reduce the model size, this invention also investigates a frequency domain distillation method specifically for frequency domain models. This enables the model to achieve efficient image classification with fewer parameters. This invention will investigate incorporating early exit branch modules based on heterogeneous computing into the distillation model. These branches allow samples that can be confidently classified through shallow layers of the network to exit early. In summary, real-time performance and high energy efficiency of edge intelligence are achieved by reducing the number of model parameters and allowing samples to exit early.

[0121] Alternatively, the following implementation method is provided:

[0122] S1. Construct a lightweight, energy-efficient edge intelligence framework based on DNN models.

[0123] S11. Select GFNet as the backbone network: Investigate the performance of GFNet on satellite image classification tasks, and evaluate its accuracy, computational complexity (especially its O(n log n) complexity advantage), and memory usage. Based on the evaluation results, select the GFNet-V2 version, configure its input size to 224x224, output classes to 10 (assuming there are 10 classes), and adopt a stepwise training strategy to optimize model performance.

[0124] 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 GPUs, while ensuring the compatibility of the model on low-power edge devices.

[0125] S2. Training using a preset loss function.

[0126] The loss function employs a composite loss function combining cross-entropy loss and distillation loss. Cross-entropy loss guides the matching of the output class token with the true label, while distillation loss learns network information from the teacher model using distillation tokens. The specific formula is as follows:

[0127] Cross-entropy loss: LCE =−∑ i =1 Cyi log( pi ),in yi One-hot encoding of the real label. pi This represents the probability distribution predicted by the model.

[0128] 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 (as it is not provided in the original text).

[0129] S3. Design an early departure branch module based on heterogeneous computing.

[0130] S31. Building Modules: Design an early termination branch that includes 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 determines whether to terminate early based on the uncertainty.

[0131] S32. Assess Uncertainty: Calculate the uncertainty of the current sample using the UCSTAT module, including metrics such as the entropy and variance of the predicted probability distribution, to evaluate the model's generalization ability, stability, and reliability.

[0132] S33. Decision Exit: Based on the uncertainty threshold, if the uncertainty of the current sample is lower than the threshold, it is considered that the model has enough confidence in the sample and can exit early in the shallow branch to reduce unnecessary computational overhead.

[0133] Furthermore, this invention also provides an embodiment:

[0134] S1. Building the Framework

[0135] S11. Select GFNet: Considering the characteristics of satellite remote sensing data, GFNet-V3 is selected due to its stronger feature extraction capabilities and lower computational complexity. The input size is configured as 512x512, and the output category is set according to the specific task (e.g., land cover classification, disaster monitoring, etc.). The training strategy combines transfer learning and fine-tuning.

[0136] S12. Parameter Configuration: Considering the real-time processing requirements of satellite remote sensing data, optimize the hardware acceleration support of GFNet, such as using TensorRT for model optimization to improve the running efficiency on edge devices.

[0137] S2. Training and Optimization

[0138] The loss function also adopts a composite loss function, combining cross-entropy loss and distillation loss to improve the accuracy and generalization ability of the model.

[0139] S3. Early Departure Branch Design

[0140] S31. Module Construction: The design includes a multi-level classification head, an uncertainty statistics module, and an early exit branch with 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 based on the uncertainty.

[0141] S32. Uncertainty Assessment: Utilize multiple uncertainty assessment methods, such as Bayesian networks and Monte Carlo dropout, to comprehensively assess the uncertainty of the model.

[0142] S33. Decision and Exit: Based on the uncertainty threshold and dynamic gating strategy, decide whether the current sample should exit the computation early. For samples with high confidence, exit early in shallow branches to reduce computational latency and resource consumption.

[0143] Furthermore, this invention also provides an embodiment:

[0144] S1. Building the Framework

[0145] S11. Selecting GFNet: Considering the real-time and accuracy requirements of satellite navigation systems, the GFNet-Lite version is selected. This version maintains accuracy while having lower computational complexity and memory usage. The input size is configured as 128x128, and the output category is the number of navigation command categories. The training strategy combines online learning and incremental learning.

[0146] 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.

[0147] S2. Training and Optimization

[0148] The loss function uses a combination of cross-entropy loss and distillation loss to improve the accuracy and efficiency of the model.

[0149] S3. Early Departure Branch Design

[0150] S31. Module Construction: The design includes an early exit branch comprising 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 model's uncertainty in real time, and the adaptive exit mechanism dynamically adjusts the exit strategy based on the uncertainty.

[0151] S32. Uncertainty Monitoring: Utilize real-time monitoring techniques, such as sliding window method and online learning algorithms, to monitor the uncertainty of the model in real time.

[0152] S33. Decision-making and Exit: Based on the uncertainty threshold and adaptive exit strategy, the exit timing of the sample is dynamically adjusted to achieve efficient and accurate navigation-assisted decision-making.

[0153] To address this challenge, our research focuses on optimizing deep neural network (DNN) models. The aim is 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 employ model compression techniques, such as weight pruning, quantization, and knowledge distillation, to significantly reduce the complexity and memory footprint of DNNs, thereby reducing energy consumption during computation. Simultaneously, we designed an intelligent sample processing strategy that allows some samples to exit the computation process earlier in the neural network layers, provided a certain accuracy threshold is met. This avoids unnecessary deep processing and further saves energy. This optimization strategy has profound implications for satellite technology and the low Earth orbit (LEO) satellite industry. With the rapid deployment of LEO satellite constellations and the gradual reduction in satellite launch costs, more and more commercial and scientific applications are relying on these "space base stations" in low Earth orbit. Our work not only helps improve the energy efficiency of satellites performing edge intelligence tasks (such as image analysis and environmental monitoring) 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.

[0154] Furthermore, these technological optimizations have inspired new ideas for future satellite design, encouraging developers to pay greater attention to energy efficiency and intelligence when designing satellite missions and system architectures, thus driving satellite technology towards 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 infinite potential of satellite edge computing, opening up new possibilities for human society to explore the universe, protect the environment, and promote economic development.

[0155] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An efficient artificial intelligence system for satellite intelligent decision-making, characterized in that: The method comprises the following steps: S1. Constructing a high-energy-efficiency edge intelligent framework based on DNN model lightweight, the input of the high-energy-efficiency edge intelligent framework being an image; The step of constructing a high-energy-efficiency edge intelligent framework based on DNN model lightweight specifically comprises the following steps: S11. Selecting GFNet as a backbone network; S12. Configuring framework parameters to adapt to a satellite intelligent decision-making scenario; The GFNet comprises three operation groups, specifically comprising 2D discrete Fourier transform, element-by-element multiplication between frequency domain features and learnable global filters, and 2D inverse Fourier transform; GFNet is composed of stacked GFNet blocks, each block consisting of a global filter layer and a feedforward network; in GFNet, first, a two-dimensional discrete Fourier transform is performed on the input spatial features to convert them to the frequency domain; second, element-wise multiplication is performed between the frequency domain features and the global filter ; finally, a two-dimensional inverse Fourier transform is performed to map the features back to the spatial domain; S2. Pre-training a GFNet model as a teacher model , using a preset loss function for training, taking the output of the teacher model as a hard label, in order to distill knowledge into the original GFNet model, the model introduces a distillation token, the output of the distillation token learns the network information of the teacher model, and the output of the class token is compared with the real label to calculate the cross-entropy loss , reducing the model size; S3. Designing an early exit branch module based on heterogeneous computing, and evaluating generalization ability, stability and reliability of the model by using uncertainty of the model; The step of designing an early exit branch module based on heterogeneous computing and evaluating generalization ability, stability and reliability of the model by using uncertainty of the model specifically comprises the following steps: S31. Constructing a classification head module, an uncertainty statistics module UCSTAT and a gate module , the classification head module processing the intermediate features to obtain intermediate prediction probabilities as input to the uncertainty statistics module; S32. Evaluating generalization ability, stability and reliability of the model by using uncertainty of the model; S33. Making a decision on whether to exit in a certain branch according to current sample characteristics; Four uncertainty metrics are selected as the input of the uncertainty statistics module, which is responsible for calculating the uncertainty statistics. The four uncertainty metrics are: maximum predicted probability, predicted entropy, , where , is the maximum value in the predicted probability , is the second largest value. The gating module concatenates the four uncertainty metrics into a one-dimensional vector, and then uses a linear layer to obtain the final result .

2. The satellite intelligent decision-oriented efficient artificial intelligence system according to claim 1, characterized in that: In the S2, a formula of the loss function is: ； wherein, is a teacher model, is a hard label.

3. The satellite intelligent decision-oriented efficient artificial intelligence system according to claim 1, characterized in that: the cross-entropy loss The formula for calculating the cross-entropy loss is: ； ； where z L represents the characteristic vector output of the layer, where , is the total number of layers in the model, is the true label.

4. The satellite intelligent decision-oriented efficient artificial intelligence system according to claim 3, characterized in that: The The formula is: ； wherein is the final result, .

5. The satellite intelligent decision oriented efficient artificial intelligence system according to claim 1, wherein: The step of selecting GFNet as a backbone network specifically comprises the following steps: Investigating and evaluating performance of GFNet in a satellite intelligent decision-making task, including accuracy, computational complexity and memory occupation, and determining a specific version and configuration parameters of GFNet according to an evaluation result.

6. The satellite intelligent decision-oriented efficient artificial intelligence system according to claim 1, characterized in that: The step of configuring framework parameters to adapt to a satellite intelligent decision-making scenario specifically comprises the following steps: Analyzing characteristics and requirements of a satellite intelligent decision-making task, including real-time performance, accuracy, resource limitation, adjusting GFNet parameters according to task characteristics, and optimizing hardware acceleration support of the framework.

7. The satellite intelligent decision-oriented efficient artificial intelligence system according to claim 6, characterized in that: The GFNet parameters include input size, output category and training strategy.

Images