Satellite-ground cooperative reasoning optimization method combining bandwidth allocation and unloading decision

Abstract

The invention discloses a satellite-ground cooperative reasoning optimization method combining bandwidth allocation and unloading decision making. The method comprises the steps that reasoning tasks generated by satellite equipment are unloaded to the local, an edge base station and a cloud end in sequence; decoupling the optimization problem of joint bandwidth allocation and unloading decision into an integer linear programming problem and a convex optimization problem; initializing an unloading strategy matrix and a bandwidth allocation strategy; solving an integer linear programming problem by using a branch and bound method, and updating an unloading strategy matrix; solving a convex optimization problem by using a gradient descent method; and recursively optimizing the integer linear programming problem and the traditional convex optimization problem until the result converges. According to the method provided by the invention, the satellite equipment can depend on the computing power of the ground and the cloud through collaborative reasoning, so that a more complex reasoning task is executed, and through a reasonable bandwidth allocation strategy, the data transmission cost among all levels (local, edge and cloud) can be reduced, and the overall communication efficiency is optimized.

Description

Technical Field

[0001] The present invention belongs to the field of communication technology, and in particular relates to a satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision. Background Art

[0002] With the rapid development of deep learning technology (DNN), it has achieved remarkable results in the fields of image processing, speech recognition and natural language processing. DNN technology has also been widely used in low-orbit satellites, such as remote sensing and environmental monitoring. However, these DNN applications usually require a lot of computing resources and data transmission. Low-orbit satellites are limited by their computing power and bandwidth resources. Running complex DNN tasks often leads to high latency and high energy consumption, making it difficult for traditional satellite processing methods to meet actual needs.

[0003] The current mainstream solutions are divided into three categories. The first is to offload all tasks to the ground for reasoning. Although this can achieve high reasoning accuracy, the resulting high latency cannot meet the real-time requirements of the task. In addition, it consumes a lot of space bandwidth resources. The second solution is to apply lightweight models on low-Earth orbit satellites for local reasoning, which can significantly reduce latency and energy consumption. However, this approach sacrifices the accuracy of reasoning, which is unacceptable for some mission scenarios. Therefore, researchers proposed a compromise. This method uses the hierarchical characteristics of the DNN network to partition the DNN tasks to achieve collaborative reasoning between satellites and the ground. This not only greatly reduces the computational burden and energy consumption of the satellite, but also improves the performance and efficiency of the entire system. However, the bandwidth resources available in space are extremely precious, and offloading tasks means consuming a lot of communication bandwidth resources. Simply saving bandwidth resources may lead to higher latency and energy consumption, resulting in the inability of DNN tasks to be processed in a timely manner. Therefore, it is more practical and valuable to integrate bandwidth resource optimization into offloading decisions.

[0004] Therefore, the present invention designs a satellite-ground collaborative reasoning method for joint bandwidth allocation and offloading decision-making, and jointly optimizes the offloading decision and bandwidth resource allocation strategy of low-orbit satellite DNN tasks. It improves reasoning efficiency and maximizes bandwidth resource savings, which has practical significance and good application prospects. Summary of the invention

[0005] The purpose of the present invention is to provide a satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision-making, so as to solve the problem that the existing technology proposed in the background technology often leads to high latency and high energy consumption in solving complex DNN tasks running on low-orbit satellites.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision-making includes the following steps:

[0008] Step S1, the DNN tasks generated by the satellite are offloaded to the local, edge base station and cloud in turn for collaborative reasoning, and the optimization problem of joint bandwidth allocation and offloading decision of DNN tasks is defined and established;

[0009] Step S2, by fixing the bandwidth allocation strategy and the task offloading strategy respectively, the optimization problem of joint bandwidth allocation and offloading decision is decoupled into an integer linear programming problem and a convex optimization problem;

[0010] Step S3, initializing the offloading strategy matrix H' and the bandwidth allocation strategy λ';

[0011] Step S4, fix the bandwidth allocation strategy λ', use the branch and bound method to solve the integer linear programming problem, and update the offloading strategy matrix H';

[0012] Step S5, fix the unloading strategy matrix H' and use the gradient descent method to solve the convex optimization problem;

[0013] Step S6, recursively optimize the integer linear programming problem and the traditional convex optimization problem until the results converge, and obtain the optimized unloading strategy matrix H and the bandwidth allocation strategy λ.

[0014] According to the above technical solution, in step S1, the optimization problem of joint bandwidth allocation and offloading decision is defined and established as follows: the variables that need to be decided are the task offloading strategy and the bandwidth allocation strategy; the variables that need to be decided are converted into minimizing the system's delay, energy consumption and communication bandwidth consumption, and weight values ​​are assigned to the three values, and the branch and bound method and the gradient descent method are used alternately to obtain the best offloading decision and bandwidth allocation strategy.

[0015] According to the above technical solution, the system delay, energy consumption and communication bandwidth consumption are minimized, and the weight values ​​of the three values ​​are assigned as follows:

[0016] Let ω, ζ, and ψ be the coefficients of inference latency, inference energy consumption, and bandwidth resource consumption respectively (ω+ζ+ψ=1), and normalize the optimization target:

[0017]

[0018] h 1,k ≥h 1,k+1 h 3,k ≤h 3,k+1 h ik ∈{0,1}λ=(0,1]

[0019] Where T represents the total latency of the DNN task; T min , T max 、Emax 、E min All represent constants and are used for normalization;

[0020] By decoupling the problem into two sub-problems, the branch and bound method and gradient descent method can be used alternately to obtain the approximately optimal task offloading strategy and bandwidth allocation decision.

[0021] According to the above technical solution, the delay includes propagation delay, transmission delay and calculation delay; the calculation of the delay is specifically as follows:

[0022] Calculate the delay: Let δ k ,δ' k ,δ″ k They represent the latency of the satellite, ground station, and cloud data center computing the k-th layer of DNN data;

[0023] Transmission delay and propagation delay: Let t k and t k ' respectively represent the time when the k-th layer data is unloaded from the satellite to the ground station and from the ground station to the cloud center;

[0024]

[0025] In the formula, t k represents the transmission time of the intermediate data of the kth layer from the satellite to the ground station, t k ' represents the transmission time of the intermediate data of the kth layer from the ground station to the cloud data center; D represents the size of the original data generated by the satellite. The data size of each layer of DNN is usually predetermined and is calculated using the coefficient α k Indicates that the output data size of the kth layer is α k ·D, R GtoC represents the transmission rate from the ground station to the cloud center, B max is the maximum bandwidth available for the satellite-to-ground channel, and λ is defined as the bandwidth occupancy coefficient. The actual bandwidth consumption is λB max ;

[0026] Define the 3*k H' offloading matrix to represent the offloading strategy of DNN tasks:

[0027]

[0028] The total latency of a DNN task is represented by T:

[0029]

[0030] Among them, h 1,k 、h 2,k 、h 3,k is a binary variable, where h 1,k Indicates the task offloading status of the satellite end, h2,k represents the task offloading of the ground edge server, h 3,k Indicates the task offloading status of the cloud computing center; a value of 1 indicates that the k-th layer task is offloaded to the corresponding device, on the contrary, a value of 0 indicates that the k-th layer task is not offloaded to the corresponding device; the DNN network has a total of K layers, and the index k is the layer index of the DNN task; then (h 1,k-1 -h 1,k ) represents the layer where data is transmitted from the satellite to the ground station, (h 3,k -h 3,k-1 ) represents the layer where data is transmitted from the ground station to the cloud center.

[0031] According to the above technical solution, energy consumption includes transmission energy consumption and computing energy consumption; bandwidth consumption refers to the bandwidth resources consumed for task transmission between satellite and ground, which is specifically calculated as:

[0032]

[0033] In the formula, represents the computational cost of computing the kth layer on the satellite, represents the energy consumption of transmitting the k-th layer of data from the satellite to the ground, and the total energy consumption of the DNN task is represented by E.

[0034] According to the above technical solution, in step S2, by fixing the bandwidth allocation strategy and the task offloading strategy, the decoupled integer linear programming problem and the convex optimization problem are:

[0035] Integer Linear Programming Problem:

[0036]

[0037] h 1,k ≥h 1,k+1 h 3,k ≤h 3,k+1 h ik ∈{0,1}

[0038] Where ω, ζ, and ψ are the coefficients of inference latency, inference energy consumption, and bandwidth resource consumption, respectively; h 1,k 、h 2,k 、h 3,k is a binary variable; k ', δ″ k represents the time delay of computing the k-th layer DNN data at the ground station and the cloud data center respectively; t k and t k ' respectively represent the time for the k-th layer data to be unloaded from the satellite to the ground station and from the ground station to the cloud center; (h 1,k-1 -h 1,k ) represents the layer where data is transmitted from the satellite to the ground station, (h 3,k -h3,k-1 ) represents the layer where data is transmitted from the ground station to the cloud center, represents the computational cost of computing the kth layer on the satellite, represents the energy consumption of transmitting the k-th layer of data from the satellite to the ground;

[0039] Convex optimization problem:

[0040]

[0041] λ=(0,1]

[0042] Where D represents the size of the raw data generated by the satellite. The data size of each layer of the DNN is usually predetermined and is adjusted by the coefficient α. k Indicates that the output data size of the kth layer is α k D, represents the transmission rate from the ground station to the cloud center, P Offload Indicates the data transmission power of the satellite.

[0043] According to the above technical solution, in step S3, the offloading decision matrix H' and the bandwidth allocation strategy λ' are initialized as follows: the offloading decision matrix H' and the bandwidth allocation strategy λ' are initialized; all layers of the DNN are offloaded locally, that is, the elements of the first row of the H' matrix are all 1, and the remaining rows are all 0, and all bandwidth is allocated, that is, λ'=1.

[0044] According to the above technical solution, in step S4, the integer linear programming problem decoupled by the branch and bound method is specifically:

[0045] First, the task offloading strategy and system overhead are initialized, and the branch and bound function is called according to the established optimization problem;

[0046] For each variable in the policy matrix H', first solve the optimal solution of the relaxed problem, that is, there is no need to meet the integer condition. If the solution is an integer, directly update the policy matrix H', otherwise branch according to the value of the optimal solution. After branching, recursively call the branch and bound function with the updated policy matrix H' until the recursive boundary is reached to obtain all optimal solutions.

[0047] According to the above technical solution, in step S5, the convex optimization problem is solved by using the gradient descent method as follows:

[0048] λ * =argxminf(λ),λ k+1 =λ k -η▽f(λ k )

[0049] In the formula, λ * is the optimal bandwidth allocation strategy, f(λ) is the decoupled convex optimization problem, λ k+1 =λ k-η▽f(λ k ), k=0,1,2,... represents the gradient descent iterative update formula, η is the learning rate, when ||▽f(λ k )||≤ò or when the maximum number of iterations K is reached, the optimal solution λ * ≈λ k .

[0050] Compared with the prior art, the present invention has the following beneficial effects:

[0051] In the present invention, by decoupling the optimization problems of bandwidth allocation and offloading decision, and respectively using integer linear programming and convex optimization methods, it is possible to more effectively utilize available computing and transmission resources, reduce latency, and improve computing efficiency. Through collaborative reasoning, the method in the present invention allows satellite equipment to rely on the computing power of the ground and the cloud to perform more complex reasoning tasks. Through a reasonable bandwidth allocation strategy, the data transmission cost between various levels (local, edge, and cloud) can be reduced, and the overall communication efficiency can be optimized. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a schematic diagram of the low-orbit satellite cloud-edge collaborative reasoning structure of the present invention;

[0053] Figure 2 A schematic diagram of the satellite-ground collaborative reasoning method for joint bandwidth allocation and offloading decision-making of the present invention;

[0054] Figure 3 This is a performance comparison result diagram of the present invention under different neural network models and different algorithms;

[0055] Figure 4 This is a diagram showing the optimization performance of bandwidth resources according to the present invention;

[0056] Figure 5 This is a diagram showing the optimization performance of the unloading strategy of the present invention. DETAILED DESCRIPTION

[0057] 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.

[0058] Embodiment 1

[0059] like Figure 1 As shown, a satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision-making includes the following steps:

[0060] Step S1, the DNN tasks generated by the satellite are offloaded to the local, edge base station and cloud in turn for collaborative reasoning, and the optimization problem of joint bandwidth allocation and offloading decision of DNN tasks is defined and established;

[0061] Step S2, by fixing the bandwidth allocation strategy and the task offloading strategy respectively, the optimization problem of joint bandwidth allocation and offloading decision is decoupled into an integer linear programming problem and a convex optimization problem;

[0062] Step S3, initializing the offloading strategy matrix H' and the bandwidth allocation strategy λ';

[0063] Step S4, fix the bandwidth allocation strategy λ', use the branch and bound method to solve the integer linear programming problem, and update the offloading strategy matrix H';

[0064] Step S5, fix the unloading strategy matrix H' and use the gradient descent method to solve the convex optimization problem;

[0065] Step S6, recursively optimize the integer linear programming problem and the traditional convex optimization problem until the results converge, and obtain the optimized unloading strategy matrix H and the bandwidth allocation strategy λ.

[0066] In the present invention, by decoupling the optimization problems of bandwidth allocation and offloading decision, and respectively using integer linear programming and convex optimization methods, it is possible to more effectively utilize available computing and transmission resources, reduce latency, and improve computing efficiency. Through collaborative reasoning, the method in the present invention allows satellite equipment to rely on the computing power of the ground and the cloud to perform more complex reasoning tasks. Through a reasonable bandwidth allocation strategy, the data transmission cost between various levels (local, edge, and cloud) can be reduced, and the overall communication efficiency can be optimized.

[0067] Embodiment 2

[0068] This embodiment provides a specific implementation method.

[0069] The present invention is directed to a low-orbit satellite cloud-edge-end collaborative reasoning model, namely, a LEO satellite, a ground station and a cloud computing center, as shown in the attached Figure 1 . First, the limited cloud greatly restricts the direct offloading of reasoning tasks by LEO satellites. The long communication distance causes very large delays in the transmission of large amounts of data, which cannot meet the real-time requirements of satellite reasoning tasks. Secondly, due to the limited computing power of LEO satellites, reasoning tasks cannot all be completed on the satellite. Therefore, the reasoning tasks generated by satellite equipment can be offloaded to local, edge base stations, and the cloud in turn to jointly perform reasoning tasks. The present invention jointly optimizes offloading decisions and satellite communication bandwidth to achieve weighted and minimization of delay, energy consumption, and communication bandwidth consumption.

[0070] The delay mainly includes calculation delay and transmission delay:

[0071] δ k ,δ' k ,δ″ k Respectively represent the delay of satellite, ground station and cloud data center in processing the k-th layer DNN data. k and t k ′ represents the time when the k-th layer data is unloaded from the satellite to the ground station and from the ground station to the cloud center.

[0072] Define the 3*k H' offloading matrix to represent the offloading strategy of DNN tasks:

[0073]

[0074] The total latency of a DNN task is represented by T:

[0075]

[0076] Among them, h 1,k 、h 2,k 、h 3,k is a binary variable, where h 1,k Indicates the task offloading status of the satellite end, h 2,k represents the task offloading of the ground edge server, h 3,k Indicates the task offloading status of the cloud computing center; a value of 1 indicates that the k-th layer task is offloaded to the corresponding device, on the contrary, a value of 0 indicates that the k-th layer task is not offloaded to the corresponding device; the DNN network has a total of K layers, and the index k is the layer index of the DNN task; then (h 1,k-1 -h 1,k ) represents the layer where data is transmitted from the satellite to the ground station, (h 3,k -h 3,k-1 ) represents the layer where data is transmitted from the ground station to the cloud center.

[0077] Energy consumption mainly includes computing energy consumption and transmission energy consumption:

[0078]

[0079] In the formula, represents the computational cost of computing the kth layer on the satellite, It represents the energy consumption of transmitting the k-th layer data from the satellite to the ground, and the total energy consumption of the DNN task is represented by E.

[0080] For bandwidth, B max is the maximum bandwidth that can be used in the satellite-to-ground channel, and λ is defined as the bandwidth occupancy coefficient (0≤λ≤1). The actual bandwidth consumption is λB max .

[0081] The satellite-ground collaborative reasoning problem to be solved weights latency, energy consumption, and bandwidth, and lets ω, ζ, and ψ be the coefficients of reasoning latency, reasoning energy consumption, and bandwidth resource consumption, respectively (ω+ζ+ψ=1), and normalizes the optimization target:

[0082]

[0083] h 1,k ≥h 1,k+1 h 3,k ≤h 3,k+1 h ik ∈{0,1}λ=(0,1]

[0084] Where T represents the total latency of the DNN task; T min , T max 、E max 、E min All represent constants and are used for normalization;

[0085] By decoupling the problem into two sub-problems, the branch and bound method and gradient descent method can be used alternately to obtain the approximately optimal task offloading strategy and bandwidth allocation decision.

[0086] In step S2, by fixing the bandwidth allocation strategy and task offloading strategy, the above formula can be decoupled into an integer linear programming problem and a convex optimization problem:

[0087] Integer linear programming problem (subproblem 1):

[0088]

[0089] h 1,k ≥h 1,k+1 h 3,k ≤h 3,k+1 h ik ∈{0,1}

[0090] Where ω, ζ, and ψ are the coefficients of inference latency, inference energy consumption, and bandwidth resource consumption, respectively; h 1,k 、h 2,k 、h 3,k is a binary variable; δ′ k ,δ″ k represents the time delay of computing the k-th layer DNN data at the ground station and the cloud data center respectively; t k and t k ' respectively represent the time for the k-th layer data to be unloaded from the satellite to the ground station and from the ground station to the cloud center; (h 1,k-1 -h 1,k ) represents the layer where data is transmitted from the satellite to the ground station, (h 3,k -h 3,k-1 ) represents the layer where data is transmitted from the ground station to the cloud center, represents the computational cost of computing the kth layer on the satellite, represents the energy consumption of transmitting the k-th layer of data from the satellite to the ground; this subproblem is solved using the branch and bound method.

[0091] Convex optimization problem:

[0092]

[0093] λ=(0,1]

[0094] Where D represents the size of the raw data generated by the satellite. The data size of each layer of the DNN is usually predetermined and is adjusted by the coefficient α. k Indicates that the output data size of the kth layer is α k D, represents the transmission rate from the ground station to the cloud center, P Offload represents the data transmission power of the satellite. This subproblem is solved using the gradient descent method.

[0095] The specific solution of convex optimization problem using gradient descent method is:

[0096] λ * =argxminf(λ),λ k+1 =λ k -η▽f(λ k )

[0097] In the formula, λ * is the optimal bandwidth allocation strategy, f(λ) is the decoupled convex optimization problem, λ k+1 =λ k -η▽f(λ k ), k=0,1,2,... represents the gradient descent iterative update formula, η is the learning rate, when ||▽f(λ k )||≤ò or when the maximum number of iterations K is reached, the optimal solution λ * ≈λ k .

[0098] The present invention obtains the optimal offloading decision and bandwidth allocation scheme of DNN tasks by minimizing the weighted problem.

[0099] This is a mixed integer nonlinear programming problem. The satellite-ground collaborative reasoning method for joint bandwidth allocation and offloading decision-making includes the following steps:

[0100] (1) By fixing the bandwidth allocation strategy and task offloading strategy, the original problem is decoupled into two sub-problems (0-1 integer linear programming problem and convex optimization problem).

[0101] (2) Initialize the offloading strategy matrix H' and bandwidth allocation strategy λ'.

[0102] (3) Fix the bandwidth allocation strategy λ', use the branch and bound method to solve subproblem 1 (0-1 integer linear programming problem with integer constraints), and update the offloading strategy matrix H'.

[0103] (4) Fix the offloading strategy matrix H', use the gradient descent method to solve subproblem 2 (traditional convex optimization problem), and update the optimal bandwidth allocation strategy λ'.

[0104] (5) Recursively optimize subproblems 1 and 2 until convergence. After convergence, H' and λ' are obtained, i.e., the optimized offloading strategy matrix H and bandwidth allocation strategy λ.

[0105] Subproblem 1 is a 0-1 integer linear programming problem with integer constraints, which leads to a combinatorial explosion in the discrete solution space. Therefore, the present invention uses the branch and bound method to approximate the optimal solution. The branch and bound method continuously divides the problem space and applies bounding conditions to each subproblem to approximate the optimal solution and reduce the complexity of the problem.

[0106] This embodiment provides a specific optimization method. Specifically,

[0107] First, initialize the task offloading strategy and system overhead, and call the branch and bound function according to the established optimization problem to obtain the optimal offloading strategy. The branch and bound function is defined as follows: When the recursive boundary is reached, the result is returned. If all constraints are satisfied and the current system overhead is less than the recorded value, update the current strategy and recorded value, otherwise the strategy is not feasible and prune.

[0108] For each variable in the policy matrix H, first use the traditional linear programming problem solution to find the optimal solution to the relaxed problem, that is, without satisfying the integer condition. If the solution is an integer, directly update the policy matrix, otherwise branch according to the value of the optimal solution. After branching, recursively call the branch and bound function with the updated matrix until the recursive boundary is reached to obtain all optimal solutions.

[0109] The method of the present invention effectively reduces the computational complexity by eliminating the consideration of numerous subsets.

[0110] Subproblem 2 is a traditional convex optimization problem, which can be solved using the gradient descent method to obtain the optimal value of the bandwidth occupancy coefficient λ;

[0111] λ*=argxminf(λ),λ k+1 =λ k -η▽f(λ k )

[0112] In the formula, λ * is the optimal bandwidth allocation strategy, f(λ) is the specific representation of the convex optimization problem, λ k+1 =λ k -η▽f(λk ), k=0,1,2,... represents the gradient descent iterative update formula, η is the learning rate, when ||▽f(λ k )||≤ò or when the maximum number of iterations K is reached, the optimal solution λ * ≈λ k .

[0113] The present invention optimizes the offloading strategy in subproblem 1 and the bandwidth allocation strategy in subproblem 2 alternately, and finally converges to obtain an approximate optimal solution to the original problem.

[0114] The performance of the proposed algorithm is evaluated by comparing it with the following three algorithms:

[0115] LEO-Only: All layers of the mission are processed onboard the satellite and not offloaded to the ground.

[0116] Cloud-Only: All layers of the task are offloaded to a powerful cloud computing center for processing.

[0117] ILPB: Inference is only done by LEO satellites and distant cloud data centers.

[0118] Figure 3 It shows the trend of system overhead for task reasoning of vgg16 network and resnet50 network as the amount of input data increases. It can be seen that the system overhead increases with the increase of input data volume, and the overhead of VGG16 network is greater than that of resnet50 network. These are all in line with expectations. We found that the offloading strategy method we proposed has significantly better performance than LEO-Only and Cloud-Only. For the ILPB algorithm, the performance of the present invention (AO-SA) is also better than ILPB at conventional input data volumes. When the initial data size is 10GB, 50GB, and 500GB, the system overhead is saved by 31%, 27%, and 21%. When the input data volume is larger, the superiority of the present invention is weakened, but it is still higher than the ILPB method. This is because the present invention takes into account the optimization of bandwidth, and the increase in data volume has a greater demand for bandwidth. In order to maintain the principle of saving bandwidth resources, the system overhead is relatively increased.

[0119] Based on the method designed by the present invention, the changes in bandwidth allocation strategy under different input data sizes are evaluated. Figure 4 It shows the bandwidth occupancy rate obtained by the present invention under different input data sizes. It can be seen that the present invention effectively saves valuable bandwidth resources. Specifically, since the amount of data transmitted in the middle layer of the VGG16 network is larger than that of the ResNet50 network, it occupies more bandwidth under the same input data size. Our analysis shows that with the exponential growth of the input data volume, the bandwidth resources consumed will also increase significantly, gradually approaching the maximum available bandwidth we set.

[0120] The relationship between the computing power of satellite equipment and the task offloading strategy was evaluated (taking the VGG16 network as an example). Figure 5 As shown in the figure. When the computing power of the satellite increases linearly from 1G FLOPS to 10G FLOPS, the optimal offloading strategy tends to offload more layers for local processing on the satellite. When the computing power of the satellite reaches 7G FLOPS, the proportion of tasks assigned to the cloud computing center drops from 56% to 19%. The line graph shows that when performing the same task, the stronger the satellite computing power, the lower the total task delay.

[0121] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0122] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision, characterized by: The following steps are involved: Step S1, the DNN tasks generated by the satellite are offloaded to the local, edge base station and cloud in turn for collaborative reasoning, and the optimization problem of joint bandwidth allocation and offloading decision of DNN tasks is defined and established; Step S2, by fixing the bandwidth allocation strategy and the task offloading strategy respectively, the optimization problem of joint bandwidth allocation and offloading decision is decoupled into an integer linear programming problem and a convex optimization problem; Step S3, initializing the offloading strategy matrix H' and the bandwidth allocation strategy λ'; Step S4, fix the bandwidth allocation strategy λ', use the branch and bound method to solve the integer linear programming problem, and update the offloading strategy matrix H'; Step S5, fix the unloading strategy matrix H' and use the gradient descent method to solve the convex optimization problem; Step S6, recursively optimize the integer linear programming problem and the traditional convex optimization problem until the results converge, and obtain the optimized unloading strategy matrix H and the bandwidth allocation strategy λ.

2. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 1 is characterized in that: In step S1, the optimization problem of joint bandwidth allocation and offloading decision is defined and established as follows: the variables that need to be decided are the task offloading strategy and the bandwidth allocation strategy; the variables that need to be decided are converted into minimizing the system's delay, energy consumption and communication bandwidth consumption, and weight values ​​are assigned to the three values, and the branch and bound method and the gradient descent method are used alternately to obtain the best offloading decision and bandwidth allocation strategy.

3. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 2 is characterized by: Minimize the system's latency, energy consumption, and communication bandwidth consumption, and assign weights to the three values ​​as follows: Let ω, ζ, and ψ be the coefficients of inference latency, inference energy consumption, and bandwidth resource consumption respectively (ω+ζ+ψ=1), and normalize the optimization target: h 1,k ≥h 1,k+1 h 3,k ≤h 3,k+1 h ik ∈{0,1}λ=(0,1] Where T represents the total latency of the DNN task; T min , T max 、E max 、E min All represent constants and are used for normalization; By decoupling the problem into two sub-problems, the branch and bound method and gradient descent method can be used alternately to obtain the approximately optimal task offloading strategy and bandwidth allocation decision.

4. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 3 is characterized by: Delay includes propagation delay, transmission delay and calculation delay. The calculation of delay is as follows: Calculate the delay: Let δ k ,δ' k ,δ” k They represent the latency of the satellite, ground station, and cloud data center computing the k-th layer of DNN data; Transmission delay and propagation delay: Let t k and t k ' respectively represent the time when the k-th layer data is unloaded from the satellite to the ground station and from the ground station to the cloud center; Where, t k represents the transmission time of the intermediate data of the kth layer from the satellite to the ground station, t k ' represents the transmission time of the intermediate data of the kth layer from the ground station to the cloud data center; D represents the size of the original data generated by the satellite. The data size of each layer of DNN is usually predetermined and is calculated using the coefficient α k Indicates that the output data size of the kth layer is α k ·D, R GtoC represents the transmission rate from the ground station to the cloud center, B max is the maximum bandwidth available for the satellite-to-ground channel, and λ is defined as the bandwidth occupancy coefficient. The actual bandwidth consumption is λB max ; Define the 3*k H' offloading matrix to represent the offloading strategy of DNN tasks: The total latency of a DNN task is represented by T: Among them, h 1,k 、h 2,k 、h 3,k is a binary variable, where h 1,k Indicates the task offloading status of the satellite end, h 2,k represents the task offloading of the ground edge server, h 3,k Indicates the task offloading status of the cloud computing center; a value of 1 indicates that the k-th layer task is offloaded to the corresponding device, on the contrary, a value of 0 indicates that the k-th layer task is not offloaded to the corresponding device; the DNN network has a total of K layers, and the index k is the layer index of the DNN task; then (h 1,k-1 -h 1,k ) represents the layer where data is transmitted from the satellite to the ground station, (h 3,k -h 3,k-1 ) represents the layer where data is transmitted from the ground station to the cloud center.

5. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 4 is characterized in that: Energy consumption includes transmission energy consumption and computing energy consumption; bandwidth consumption refers to the bandwidth resources consumed for task transmission between satellite and ground, which is specifically calculated as: In the formula, represents the computational cost of computing the kth layer on the satellite, represents the energy consumption of transmitting the k-th layer of data from the satellite to the ground, and the total energy consumption of the DNN task is represented by E.

6. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 5 is characterized by: In step S2, by fixing the bandwidth allocation strategy and task offloading strategy, the decoupled integer linear programming problem and convex optimization problem are: Integer Linear Programming Problem: h 1,k ≥h 1,k+1 h 3,k ≤h 3,k+1 h ik ∈{0,1} Where ω, ζ, and ψ are the coefficients of inference latency, inference energy consumption, and bandwidth resource consumption, respectively; h 1,k 、h 2,k 、h 3,k is a binary variable; δ' k ,δ” k represents the time delay of the ground station and the cloud data center to calculate the k-th layer of DNN data; t k and t k ' respectively represent the time for the k-th layer data to be unloaded from the satellite to the ground station and from the ground station to the cloud center; (h 1,k-1 -h 1,k ) represents the layer where data is transmitted from the satellite to the ground station, (h 3,k -h 3,k-1 ) represents the layer where data is transmitted from the ground station to the cloud center, represents the computational cost of computing the kth layer on the satellite, represents the energy consumption of transmitting the k-th layer of data from the satellite to the ground; Convex optimization problem: λ=(0,1] Where D represents the size of the raw data generated by the satellite. The data size of each layer of the DNN is usually predetermined and is adjusted by the coefficient α. k Indicates that the output data size of the kth layer is α k D, represents the transmission rate from the ground station to the cloud center, P Offload Indicates the data transmission power of the satellite.

7. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 6 is characterized by: In step S3, the offloading decision matrix H' and the bandwidth allocation strategy λ' are initialized as follows: the offloading decision matrix H' and the bandwidth allocation strategy λ' are initialized; all layers of the DNN are offloaded locally, that is, the first row of the H' matrix is ​​all 1, and the remaining rows are all 0, and all bandwidth is allocated, that is, λ'=1.

8. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 7 is characterized by: In step S4, the integer linear programming problem decoupled by the branch and bound method is specifically: First, the task offloading strategy and system overhead are initialized, and the branch and bound function is called according to the established optimization problem; For each variable in the policy matrix H', first solve the optimal solution of the relaxed problem, that is, there is no need to meet the integer condition. If the solution is an integer, directly update the policy matrix H', otherwise branch according to the value of the optimal solution. After branching, recursively call the branch and bound function with the updated policy matrix H' until the recursive boundary is reached to obtain all optimal solutions.

9. The satellite-ground collaborative reasoning optimization method for joint bandwidth allocation and offloading decision according to claim 8 is characterized by: In step S5, the convex optimization problem is solved by using the gradient descent method as follows: l * =argxminf(λ),λ k+1 =λ k -η▽f(λ k ) In the formula, λ * is the optimal bandwidth allocation strategy, f(λ) is the decoupled convex optimization problem, λ k+1 =λ k -η▽f(λ k ), k=0,1,2,... represents the gradient descent iterative update formula, η is the learning rate, when ||▽f(λ k )||≤ò or when the maximum number of iterations K is reached, the optimal solution λ * ≈λ k .

Images