LEO lightweight adaptive image reasoning strategy

Abstract

The invention provides a lightweight adaptive image reasoning strategy based on a Markov Decision Process-Deep Q Network (MDP-DQN), aiming at the challenges of limited computing resources and limited energy faced by a Low Earth Orbit (LEO). The invention provides the lightweight adaptive image reasoning strategy based on the Markov Decision Process-Deep Q Network (MDP-DQN). According to the method, an AEE (Accuracy-Energy Efficiency) index is introduced, and the AEE index is used for evaluating the image reasoning performance. A simulation result shows that compared with a Markov Decision Process-Q Learning (Markov Decision Process-Q Learning, MDP-QL), the AEE of the strategy is improved by 6.09%, and compared with a Markov Decision Process-Policy Gradient (Markov Decision Process-Policy Gradient, MDP-PG), the AEE of the strategy is improved by 12.2%.

Description

Technical Field

[0001] The invention relates to the field of LEO low-orbit satellite communications, and in particular to a LEO lightweight adaptive image reasoning strategy. Background Art

[0002] Low Earth Orbit (LEO Satellite Networks, LSNs) is an important technology to achieve seamless global coverage of the sixth generation mobile communication network (6G). At the same time, artificial intelligence (AI) has been widely used in many fields. In this context, LSNs deployed with image reasoning services can be used to analyze images to support tasks such as natural disaster monitoring, resource management, and urban planning.

[0003] When deploying image inference services on LSNs, there are two main challenges: First, LEO satellites have limited computing resources due to their limited physical size. Second, LEO satellites rely on solar cells for power supply, which has limited and unstable energy resources. In addition, LEO satellites are alternately exposed to the sun and the shadow, and their energy supply changes accordingly, resulting in unstable energy availability.

[0004] In order to solve the problem of limited satellite computing resources, Shang et al. reduced the inference time to 3.26 milliseconds and achieved an accuracy of 93.56% for the real-time image classification task of meteorological satellites by reducing the neural network depth and model parameters. Pang et al. \cite{bib6} used depthwise separable convolution and global average pooling to reduce the network depth and model parameters, reducing the inference time to 2.45 milliseconds and achieving an accuracy of up to 99%.

[0005] In order to solve the problem of satellite energy limitation and unstable energy supply, Bai et al. reduced the inference energy consumption per unit time by quantizing the weights and activation values ​​in the convolution operation into low-precision representations. Tarasion et al. added position encoding to maintain spatial information and adopted a multi-head self-attention mechanism (MHSAM) to capture global features, significantly reducing computational energy consumption while maintaining high inference accuracy.

[0006] However, no comprehensive solution has been proposed for LEO satellites under the dual constraints of computing resources and energy resources. Summary of the invention

[0007] In view of the defects pointed out in the above background technology, this paper focuses on optimizing the reasoning accuracy and energy consumption of LEO satellite image reasoning, and proposes a lightweight adaptive image reasoning strategy that combines Markov Decision Process (MDP) and Deep Q Network (DQN). At the same time, the Accuracy-Energy Efficiency (AEE) indicator is proposed to evaluate the reasoning accuracy performance per unit energy. Finally, the simulation results show that the proposed strategy outperforms other benchmark strategies in terms of AEE indicator.

[0008] The technical solution for achieving the purpose of the present invention is:

[0009] A LEO lightweight adaptive image inference strategy, including:

[0010] (1) Building a LEO network model

[0011] The LEO satellites deployed with image inference services alternately pass through the sunlit area and the shadow area during orbital operation. During operation, the satellites capture images and perform inference tasks;

[0012] The orbit equation of a LEO satellite is expressed as:

[0013] r(t)=(R+h)[1-e·cosα(t)] (1)

[0014] Where R is the radius of the Earth, h is the orbital altitude, e is the orbital eccentricity, and α(t) is the eccentric anomaly angle;

[0015] LEO satellite orbit period T s It is expressed as:

[0016]

[0017] Where G is the gravitational constant, M e is the mass of the Earth;

[0018] (2) Building a satellite energy model

[0019] Assume that the instantaneous power of solar radiation energy received by the LEO satellite is P c (t), as shown in the following formula:

[0020]

[0021] Where η represents the conversion efficiency of the solar panel, D is the area of ​​the solar panel, P max is the solar radiation intensity at the point where the sun is directly shining, α is the proportion of the sunlit area in the orbital period, and σ represents the width of the distribution of controlled radiation intensity;

[0022] From formula (3), it can be seen that in the sunshine zone of one orbital period, that is, 0≤t≤α·T s When α·T s ≤t≤T s When the solar radiation energy is 0, the satellite-mounted battery serves as the energy source for the LEO satellite;

[0023] Assume that the initial battery energy of the satellite just entering the sunshine area is The maximum battery capacity of the satellite is The minimum tolerable energy of the satellite battery to maintain the basic operation of the satellite. Before entering the Earth's shadow, the LEO satellite needs to charge the battery to For use by LEO satellites in the Earth's shadow;

[0024] The LEO satellite orbit period is divided into N time slots of equal length. Δt is the unit time interval, whose value is the orbit period T. s The ratio of the orbital period time slot sequence to the total number of time slots N is expressed as where t i represents any time slot;

[0025] If the time slot t i In the sunshine area, that is, 1≤i≤α·N, the energy captured by the LEO satellite is:

[0026]

[0027] The charging energy of the battery in this time slot is:

[0028]

[0029] Among them, the same average energy is added to each time slot, so the available energy for the inference service in the time slot is:

[0030]

[0031] In the shadow area, a cyclic average energy mechanism is used: first, the energy of the initial time slot is averaged to determine the available energy for the image reasoning service in that time slot, then the reasoning energy of that time slot is deducted and averaged again to determine the available energy for the next time slot; this process is repeated until the last time slot in the shadow area;

[0032] If the time slot t i In the shadow region, that is, α·N<i≤N, the available energy for image reasoning is:

[0033]

[0034] Among them, E v is the energy consumption of any time slot from time slot α·N to i;

[0035] The available energy sequence of image inference of LEO satellite in the orbit period can be obtained as follows:

[0036]

[0037] (3) Building an inference model

[0038] The original image taken by the satellite is divided into several small-size images for preprocessing, and then the preprocessed image data is input into the convolutional neural network (CNN) model for feature extraction and analysis. A certain model is selected in a single time slot, and the inference performance set is output, including the inference result, inference energy consumption, and inference time. Finally, the useful information is transmitted back to the ground for subsequent decision-making;

[0039] Assume time slot t i The number of images to be processed follows a Poisson distribution as follows:

[0040]

[0041] Among them, n i Indicates time slot t i The number of pictures, λ i Indicates time slot t i The average number of pictures;

[0042] Time slot t i The set of images to be processed in is represented as ik is the time slot t i Any picture number of ;

[0043] The set of CNN models with different parameters for the deployment of M LEO satellites is Indicates that m j Represents any CNN model;

[0044] Time slot t i Using model m j The performance set for reasoning the kth image is:

[0045] P I ={p ijk ,t ijk ,y ijk} , 1≤i≤N,1≤j≤M (10)

[0046] Among them, p ijk is the inference power of the image, t ijk is the inference time of the image, y ijkis the inference label of the picture;

[0047] Time slot t i Select model m j The inference energy consumption is:

[0048]

[0049] If the inference label y ijk With the actual label x ijk If they are the same, it is counted as a correct inference; and in time slot t i Select the inference model m i The inference accuracy is:

[0050]

[0051] Among them, 1(x) is an indicator function, indicating whether the condition is met, 1 if it is met, and 0 otherwise;

[0052] (4) There is a coupling relationship between the inference accuracy and inference energy consumption of the CNN model carried by the satellite, that is, the improvement of inference accuracy will lead to an increase in inference energy consumption;

[0053] First, define the accuracy and energy efficiency index as:

[0054]

[0055] It indicates the accuracy of image reasoning that the system can obtain under unit energy consumption, and the unit is accuracy / joule.

[0056] The objective function and constraints are as follows:

[0057]

[0058] Among them, constraint (a) is the time slot t i The reasoning accuracy constraint is is the time slot t i The minimum inference accuracy of i The reasoning energy consumption constraint of t; constraint (c) is the time slot t i The reasoning time constraint, T i is the time slot t i The reasoning time, T i max is the time slot t i The maximum inference time of

[0059] (5) The satellite adaptive reasoning strategy problem is modeled as a Markov decision process to capture the changes in system state and the uncertainty in the decision-making process, and the Markov decision process is solved by using the Deep Q-Network (DQN) algorithm;

[0060] The Markov decision process consists of the tuple (S, A, T, r, γ);

[0061] is a state space where the state elements Indicates time slot t i Available energy for satellites, state element I i Indicates time slot t i The distribution of tasks;

[0062] is the action space, which is the selection object of the adaptive strategy;

[0063] T={(T i E ,T i I )} is the state transfer function, where T i E represents the energy state transfer function, T i I Represents the task state transfer function;

[0064] The energy state transfer function is:

[0065]

[0066] The task state transfer function is:

[0067]

[0068] r(t i ,m j ) is the reward function, which means that at time slot t i The CNN model m is adopted j The rewards obtained are as follows:

[0069]

[0070] Among them, δ 1 is the weight coefficient of precision energy efficiency, δ 2 is the coefficient of deviation from the minimum precision;

[0071] γ is the discount factor, which represents the discount rate of long-term rewards and is set between 0 and 1. The higher the setting, the better the expected performance of long-term planning.

[0072] The present invention adopts a deep Q network (DQN) to solve the problem: by introducing a target Q network and an experience replay buffer, data correlation is broken and the stability of decision making is improved. The ∈-greedy strategy is used to balance the selection of actions between exploration and utilization, and the Q network parameters are updated by minimizing the loss function through gradient descent, so as to learn the optimal strategy and obtain the optimal CNN model for each time slot. Simulation results show that compared with Markov Decision Process-Q Learning (MDP-QL), the AEE of this strategy is improved by 6.09%, and compared with Markov Decision Process-Policy Gradient (MDP-PG), the AEE is improved by 12.2%. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A model diagram of a LEO satellite system equipped with image reasoning services according to an embodiment of the present invention;

[0074] Figure 2 This is an image reasoning model diagram of an embodiment of the present invention;

[0075] Figure 3 This is a performance analysis comparison diagram of the average number of images per time slot and the reasoning accuracy energy efficiency AEE under different strategies of the embodiment of the present invention;

[0076] Figure 4 This is a performance analysis comparison diagram of the number of training rounds and the reasoning accuracy and energy efficiency AEE under different strategies of the embodiment of the present invention;

[0077] Figure 5 A performance analysis comparison diagram of learning rate / discount factor and reasoning accuracy energy efficiency AEE under different strategies of the embodiment of the present invention;

[0078] Figure 6 A comparison diagram of the inference available energy and average AEE value in different time slots according to an embodiment of the present invention;

[0079] Figure 7 This is a relationship diagram between the average number of images in each time slot and the average AEE value under different strategies in an embodiment of the present invention. DETAILED DESCRIPTION

[0080] The present invention is further described below in conjunction with the accompanying drawings and embodiments, but the contents of the present invention are not limited thereto.

[0081] Example:

[0082] A LEO lightweight adaptive image inference strategy, including:

[0083] 1. Building a system model

[0084] 1. Build LEO network model

[0085] like Figure 1 As shown in FIG. 1 , a LEO satellite deployed with an image inference service alternately passes through the sunlit area and the shadow area during its orbit. During the operation, the satellite captures images and performs inference tasks.

[0086] The orbit equation of a LEO satellite is expressed as:

[0087] r(t)=(R+h)[1-e·cosα(t)] (1)

[0088] Where R is the radius of the Earth, h is the orbital altitude, e is the orbital eccentricity, and α(t) is the eccentric anomaly angle. LEO satellite orbit period T s It is expressed as:

[0089]

[0090] Where G is the gravitational constant, M e is the mass of the Earth.

[0091] 2. Building a satellite energy model

[0092] Assume that the instantaneous power of solar radiation energy received by the LEO satellite is P c (t), as shown in the following formula:

[0093]

[0094] Where η represents the conversion efficiency of the solar panel, D is the area of ​​the solar panel, P max is the solar radiation intensity at the point where the sun is directly overhead, α is the proportion of the sunlit area in the orbital period, and σ represents the width of the radiation intensity distribution.

[0095] From formula (3), it can be seen that in the sunshine zone of one orbital period, that is, 0≤t≤α·T s When α·T s ≤t≤T s When the solar radiation energy is 0, the satellite carries a battery as the energy source for the LEO satellite.

[0096] Assume that the initial battery energy of the satellite just entering the sunshine area is The maximum battery capacity of the satellite is The minimum tolerable energy of the satellite battery to maintain the basic operation of the satellite. Before entering the Earth's shadow, the LEO satellite needs to charge the battery to For use by LEO satellites in the Earth's shadow.

[0097] The LEO satellite orbit period is divided into N time slots of equal length. Δt is the unit time interval, whose value is the orbit period T. s The ratio of the orbital period time slot sequence to the total number of time slots N is expressed as where t i Represents an arbitrary time slot.

[0098] If the time slot t i In the sunshine area, that is, 1≤i≤α·N, the energy captured by the LEO satellite is:

[0099]

[0100] The charging energy of the battery in this time slot is:

[0101]

[0102] Among them, the same average energy is added to each time slot, so the available energy for the inference service in the time slot is:

[0103]

[0104] In the shadow area, the present invention adopts a cyclic average energy mechanism: first, the energy of the initial time slot is averaged to determine the available energy of the image reasoning service in the time slot, then the reasoning energy of the time slot is deducted, and then the average calculation is performed again to determine the available energy of the next time slot. This process is repeated until the last time slot in the shadow area.

[0105] If the time slot t i In the shadow region, that is, α·N<i≤N, the available energy for image reasoning is:

[0106]

[0107] Among them, E v is the energy consumption of any time slot from time slot α·N to i.

[0108] The available energy sequence of image inference of LEO satellite in the orbit period can be obtained as follows:

[0109]

[0110] 3. Build an inference model

[0111] Figure 2For the image reasoning model, the first is image segmentation, where the original satellite image is segmented into several small-size images; the second is image reasoning, where the pre-processed image data is input into the convolutional neural network (CNN) model for feature extraction and analysis, and a certain model is selected in a single time slot to output a reasoning performance set, including reasoning results, reasoning energy consumption, and reasoning time. Finally, information is sent back to the ground for subsequent decision-making.

[0112] Assume time slot t i The number of images to be processed follows a Poisson distribution as follows:

[0113]

[0114] Among them, n i Indicates time slot t i The number of pictures, λ i Indicates time slot t i The average number of images.

[0115] Time slot t i The set of images to be processed in is represented as ik is the time slot t i Any picture number.

[0116] The set of CNN models with different parameters for the deployment of M LEO satellites is Indicates that m j Represents an arbitrary CNN model.

[0117] Time slot t i Using model m j The performance set for reasoning the kth image is:

[0118] P I ={p ijk ,t ijk ,y ijk} , 1≤i≤N,1≤j≤M (10)

[0119] Among them, p ijk is the inference power of the image, t ijk is the inference time of the image, y ijk is the inferred label of the image.

[0120] Time slot t i Select model m j The inference energy consumption is:

[0121]

[0122] If the inference label y ijk With the actual label x ijkIf they are the same, it is counted as a correct inference. i Select the inference model m i The inference accuracy is:

[0123]

[0124] Among them, 1(x) is an indicator function, indicating whether the condition is met, 1 if it is met, and 0 otherwise.

[0125] 2. Problem Modeling

[0126] There is a coupling relationship between the inference accuracy and inference energy consumption of the CNN model carried by the satellite, that is, the improvement of inference accuracy will lead to the increase of inference energy consumption.

[0127] First, define the accuracy and energy efficiency index as:

[0128]

[0129] It indicates the accuracy of image reasoning that the system can obtain under unit energy consumption, and the unit is Accuracy / Joule.

[0130] The objective function and constraints are as follows:

[0131]

[0132] Among them, constraint (a) is the time slot t i The reasoning accuracy constraint is is the time slot t i The minimum inference accuracy of i The reasoning energy consumption constraint of t; constraint (c) is the time slot t i The reasoning time constraint, T i is the time slot t i The reasoning time, T i max is the time slot t i The maximum inference time.

[0133] 3. Adaptive Strategy Based on MDP-DQN

[0134] First, the satellite adaptive reasoning strategy problem is modeled as a Markov decision process (MDP), which can effectively capture the changes in system state and the uncertainty in the decision-making process. Secondly, the Deep Q-Network (DQN) algorithm is used to solve the Markov decision process. As a reinforcement learning algorithm, DQN can effectively approximate the Q-value function in high-dimensional state space, thereby achieving the optimal strategy.

[0135] The Markov decision process consists of the tuple (S, A, T, r, γ).

[0136] is a state space, where the state element E i u Indicates time slot t i Available energy for satellites, state element I i Indicates time slot t i distribution of tasks.

[0137] It is the action space, which is the selection object of the adaptive strategy.

[0138] T={(T i E ,T i I )} is the state transfer function, where T i E represents the energy state transfer function, T i I Represents the task state transition function.

[0139] The energy state transfer function is:

[0140]

[0141] The task state transfer function is:

[0142]

[0143] r(t i ,m j ) is the reward function, which means that at time slot t i The CNN model m is adopted j The rewards obtained are as follows:

[0144]

[0145] Among them, δ 1 is the weight coefficient of precision energy efficiency, δ 2 is the coefficient of deviation from the minimum precision.

[0146] γ is the discount factor, which represents the discount rate of long-term rewards and is set between 0 and 1. The higher the setting, the better the expected performance of long-term planning.

[0147] In order to solve the above MDP problem, the present invention adopts a deep Q network (DQN) to solve it: by introducing a target Q network and an experience replay buffer, the data correlation is broken and the stability of the decision is improved. The ∈-greedy strategy is used to balance the selection of actions between exploration and utilization, and the Q network parameters are updated by minimizing the loss function through gradient descent, so as to learn the optimal strategy and obtain the optimal CNN model for each time slot.

[0148] 4. Simulation

[0149] 1. Simulation parameters

[0150] Table 1LEO satellite network parameters

[0151]

[0152] The simulation scenario is a single LEO satellite with an altitude of 500 km, an orbital period of 5670 s, which is divided into 30 time slots, and the number of CNN models carried is 3. The specific LEO satellite network parameters are shown in Table 1.

[0153] Table 2 CNN model and calculation parameters

[0154]

[0155] Table 2 shows the performance parameters of the CNN model set. These models have low computational complexity and are suitable for the resource-constrained LEO satellite environment. At the same time, these models are highly competitive and complementary in different satellite scenarios and conditions.

[0156] Table 3DQN algorithm parameters

[0157]

[0158] Table 3 shows the parameters of the DQN algorithm, where Episodes is the number of training rounds, β is the learning rate, γ is the discount factor, ∈ max is the initial exploration rate, ε is the exploration rate attenuation coefficient, ∈ min is the minimum exploration rate.

[0159] 2. Simulation results and analysis

[0160] In order to verify the performance of the MDP-DQN strategy proposed in the present invention, the present invention considers the following benchmark strategies for comparative experiments, which are all solutions to the MDP problem.

[0161] Q-learning strategy (MDP-QL): Q-learning learns the optimal strategy by updating the Q value of the state-action pair. In the present invention, Q-learning continuously adjusts the numerical values ​​in the Q table by interacting with the environment in order to find the best action for each state of the satellite.

[0162] Policy Gradient Policy (MDP-PG): Policy gradient methods directly optimize policy parameters to maximize cumulative rewards. In this invention, policy gradient generates trajectories by sampling, calculates gradients, and updates the best action according to the gradient ascent method.

[0163] from Figure 3 It can be seen that as the number of images increases, the AEE values ​​of all strategies gradually decrease. This is because processing more images with the same energy supply will lead to reduced accuracy performance, thereby reducing the AEE value. When the number of images is small, the performance gap between MDP-DQN and MDP-QL is not obvious, but when processing a large number of images, the MDP-DQN strategy can estimate the Q value more accurately, thereby performing better under high-load tasks. The AEE values ​​of MDP-PG under all image numbers are lower than those of the other two algorithms, indicating that its performance is relatively poor.

[0164] from Figure 4 It can be seen that all strategies tend to be stable after 600 training rounds, indicating that the algorithm performance has been fully optimized at this time. MDP-DQN significantly outperforms the other two strategies in AEE value, especially after reaching convergence, its cumulative AEE value is the highest. MDP-QL is second, while MDP-PG performs relatively poorly. MDP-PG converges fastest in the early rounds, but its long-term performance is not as good as MDP-DQN and MDP-QL strategies. This is because MDP-PG can quickly find a better strategy in the initial stage of the strategy, but it will fall into a local optimal solution.

[0165] from Figure 5 It can be seen that MDP-DQN always has an advantage over other benchmark strategies in the changes of learning rate and discount factor. MDP-DQN performs best when the learning rate is 0.001 and the discount rate is 0.99. MDP-QL is insensitive to learning rate and discount rate. This is because MDP-QL has a simple structure and does not need to rely on a large amount of data for complex network training, so it is insensitive to parameter changes. The performance of MDP-PG fluctuates greatly with the changes of learning rate and discount factor, indicating that it is sensitive to parameter selection.

[0166] Figure 6 The curves of available energy for inference, adaptive model selection, and AEE performance at different time slots in the orbital cycle are shown. In the sunshine area, the available energy for inference gradually increases and reaches a maximum value of about 5.8×10 6 Joules, and then began to decline. In the shadow area, the inferred available energy is relatively stable, about 3.2×10 6 joule.

[0167] Under high energy conditions (i.e. 3.7×106 ~6×10 6 Joules), MobileNet_v3 is preferred, supplemented by MobileNet_v2, with MobileNet_v3 accounting for 70% and MobileNet_v2 accounting for 30%. Under low energy conditions (i.e. 0.4×10 6 ~3.7×10 6 Joule), MobileNet_v2 and ResNet18 perform competitively.

[0168] Overall, the AEE value is positively correlated with the available energy for inference. However, under high energy conditions, the AEE of the model fluctuates greatly, with a maximum value of about 3.94 and a minimum value of 3.42. This fluctuation is mainly attributed to the change in the number of images and the tendency to select models with complex parameters and high-dimensional features, which makes them more susceptible to randomness and noise.

[0169] according to Figure 7 As shown in Figure 2, as the average number of images per time slot increases, the AEE values ​​of all strategies gradually decrease. This is because processing more images with the same energy supply will reduce the inference accuracy, resulting in a decrease in the AEE value. This shows that when energy is limited, it is necessary to find a balance between task load and model complexity.

[0170] Under low load conditions (i.e. 0.5×10 5 ~1.5×10 5 ), the performance gap between the MDP-DQN and MDP-QL strategies is smaller. However, under high load conditions (i.e., 1.5×10 5 ~5×10 5 ), the MDP-DQN strategy estimates the Q value more accurately and performs better. The AEE value of MDP-PG is always lower than that of the other two strategies, indicating that its performance is poor.

[0171] Specifically, under all load conditions, the overall average AEE values ​​of MDP-DQN, MDP-QL, and MDP-PG are 3.31, 3.12, and 2.95, respectively.

[0172] Table 4 Algorithm time complexity analysis

[0173]

[0174] Table 4 shows the time complexity of the three strategies, where L is the number of network layers, N e is the number of neurons in each layer, |S| is the size of the state space, and |A| is the size of the action space. Both MDP-DQN and MDP-PG use deep neural networks, so they are similar in complexity, both O(L×N e2 ). When the network structure is more complex, the time complexity is higher. In contrast, the time complexity of MDP-QL is O(|S|×|A|), which is directly related to the size of the state and action space. This algorithm performs better when the state and action space is small, but the amount of computation will increase significantly when the space is large.

[0175] V. Conclusion

[0176] The MDP-DQN strategy proposed in this paper shows high adaptive decision-making ability in the reasoning service deployed on LEO satellites. Compared with the MDP-QL and MDP-PG strategies, the MDP-DQN strategy improves the AEE index by 6.09% and 12.2%, respectively. By achieving efficient image reasoning, the development of green computing in low-orbit satellite networks (LSNs) is promoted.

Claims

1. A LEO lightweight adaptive image inference strategy, characterized by: include: (1) Building a LEO network model The LEO satellites deployed with image inference services alternately pass through the sunlit area and the shadow area during orbital operation. During operation, the satellites capture images and perform inference tasks; The orbit equation of a LEO satellite is expressed as: r(t)=(R+h)[1-e·cosα(t)] (1) Where R is the radius of the Earth, h is the orbital altitude, e is the orbital eccentricity, and α(t) is the eccentric anomaly angle; LEO satellite orbit period T s It is expressed as: Where G is the gravitational constant, M e is the mass of the Earth; (2) Building a satellite energy model Assume that the instantaneous power of solar radiation energy received by the LEO satellite is P c (t), as shown in the following formula: Where η represents the conversion efficiency of the solar panel, D is the area of ​​the solar panel, P max is the solar radiation intensity at the point where the sun is directly shining, α is the proportion of the sunlit area in the orbital period, and σ represents the width of the distribution of controlled radiation intensity; From formula (3), it can be seen that in the sunshine zone of one orbital period, that is, 0≤t≤α·T s When α·T s ≤t≤T s When the solar radiation energy is 0, the satellite-mounted battery serves as the energy source for the LEO satellite; Assume that the initial battery energy of the satellite just entering the sunshine area is The maximum battery capacity of the satellite is The minimum tolerable energy of the satellite battery to maintain the basic operation of the satellite. Before entering the Earth's shadow, the LEO satellite needs to charge the battery to For use by LEO satellites in the Earth's shadow; The LEO satellite orbit period is divided into N time slots of equal length. Δt is the unit time interval, whose value is the orbit period T. s The ratio of the orbital period time slot sequence to the total number of time slots N is expressed as where t i represents any time slot; If the time slot t i In the sunshine area, that is, 1≤i≤α·N, the energy captured by the LEO satellite is: The charging energy of the battery in this time slot is: Among them, the same average energy is added to each time slot, so the available energy for the inference service in the time slot is: In the shadow area, a cyclic average energy mechanism is used: first, the energy of the initial time slot is averaged to determine the available energy for the image reasoning service in that time slot, then the reasoning energy of that time slot is deducted and averaged again to determine the available energy for the next time slot; this process is repeated until the last time slot in the shadow area; If the time slot t i In the shadow region, that is, α·N<i≤N, the available energy for image reasoning is: Among them, E v is the energy consumption of any time slot from time slot α·N to i; The available energy sequence of image inference of LEO satellite in the orbit period can be obtained as follows: (3) Building an inference model The original image taken by the satellite is divided into several small-size images for preprocessing, and then the preprocessed image data is input into the convolutional neural network (CNN) model for feature extraction and analysis. A certain model is selected in a single time slot, and the inference performance set is output, including the inference result, inference energy consumption, and inference time. Finally, the useful information is transmitted back to the ground for subsequent decision-making; Assume time slot t i The number of images to be processed follows a Poisson distribution as follows: Among them, n i Indicates time slot t i The number of pictures, λ i Indicates time slot t i The average number of pictures; Time slot t i The set of images to be processed in is represented as ik is the time slot t i Any picture number of ; The set of CNN models with different parameters for the deployment of M LEO satellites is Indicates that m j Represents any CNN model; Time slot t i Using model m j The performance set for reasoning about the kth image is: P I ={p ijk ,t ijk ,y ijk } , 1≤i≤N,1≤j≤M (10) Among them, p ijk is the inference power of the image, t ijk is the inference time of the image, y ijk is the inference label of the picture; Time slot t i Select model m j The inference energy consumption is: If the inference label y ijk With the actual label x ijk If they are the same, it is counted as a correct inference; and in time slot t i Select the inference model m i The inference accuracy is: Among them, 1(x) is an indicator function, indicating whether the condition is met, 1 if it is met, and 0 otherwise; (4) There is a coupling relationship between the inference accuracy and inference energy consumption of the CNN model carried by the satellite, that is, the improvement of inference accuracy will lead to an increase in inference energy consumption; First, define the accuracy and energy efficiency index as: It indicates the accuracy of image reasoning that the system can obtain under unit energy consumption, and the unit is accuracy / joule. The objective function and constraints are as follows: Among them, constraint (a) is the time slot t i The reasoning accuracy constraint is is the time slot t i The minimum inference accuracy of i The reasoning energy consumption constraint of t; constraint (c) is the time slot t i The reasoning time constraint, T i is the time slot t i The reasoning time, is the time slot t i The maximum inference time of (5) The satellite adaptive reasoning strategy problem is modeled as a Markov decision process to capture the changes in system state and the uncertainty in the decision-making process, and the Markov decision process is solved by using the Deep Q-Network (DQN) algorithm; The Markov decision process consists of the tuple (S, A, T, r, γ); is a state space where the state elements Indicates time slot t i Available energy for satellites, state element I i Indicates time slot t i The distribution of tasks; is the action space, which is the selection object of the adaptive strategy; is the state transition function, where represents the energy state transfer function, Represents the task state transfer function; The energy state transfer function is: The task state transfer function is: r(t i ,m j ) is the reward function, which means that at time slot t i The CNN model m is adopted j The rewards obtained are as follows: Among them, δ1 is the weight coefficient of accuracy energy efficiency, and δ2 is the deviation coefficient of minimum accuracy; γ is the discount factor, which represents the discount rate of long-term rewards and is set between 0 and 1. The higher the setting, the better the expected performance of long-term planning.

Images