A distributed decision-making method and system based on large model

Through adaptive chaos sampling and multi-view fusion fine-tuning methods, a distributed decision system is built, which solves the performance bottlenecks of traditional decision systems in large-scale data and complex model processing, and improves the decision-making efficiency of coordinated target capture of satellites and drones and Accuracy enhances environmental adaptability and resource utilization.

CN118657200BActive Publication Date: 2024-11-22NANJING ARTIFICIAL INTELLIGENCE CHIPS RES INST OF AUTOMATION CHINESE ACAD OF SCI
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202411152283.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-11-22
Estimated Expiration
2044-08-21

AI Technical Summary

Technical Problem

Traditional decision-making systems have performance bottlenecks when processing large-scale data and complex models, and it is difficult to effectively integrate satellite global vision and local fine observation of drones, resulting in insufficient decision-making efficiency and accuracy, especially in complex geographical environments and changing weather conditions. Improve environmental adaptability and decision-making robustness.

Method used

Adaptive chaos sampling and multi-view fusion fine-tuning methods are used to generate decision training data, build a distributed decision system, including dynamic resource orchestration and load prediction modules, and build a decision tree model using an adaptive multi-dimensional recursive partitioning method to generate visual decision paths.

Benefits of technology

Effectively solve the performance bottlenecks of traditional decision-making systems, improve the decision efficiency and accuracy of the model, improve the accuracy and efficiency of coordinated target capture of satellites and drones, and achieve decision-making robustness and resource efficiency in complex environments. use.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118657200B_ABST
    Figure CN118657200B_ABST
Patent Text Reader

Abstract

The present invention discloses a distributed decision-making method and system based on a large model, the method comprising obtaining system parameters, preprocessing the system parameters, and obtaining the preprocessed system parameters; based on the preprocessed system parameters, using an adaptive chaotic sampling method to generate system initial state parameters; calculating the capture success probability matrix of each target point at different time points; generating decision training data based on the capture success probability matrix and the system initial state parameters; using an adaptive multi-view fusion fine-tuning method to fine-tune the pre-configured large model to obtain a fine-tuned large model; constructing a distributed decision system to obtain a resource allocation strategy; using an adaptive multi-dimensional recursive partitioning method to construct a decision tree model, and using the decision tree model to generate a visual decision path. The present invention not only effectively solves the performance bottleneck of traditional decision systems when processing large-scale data and complex models, but also improves the decision-making efficiency and accuracy of the model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of big data processing, and in particular to a distributed decision-making method and system based on a large model. Background Art

[0002] With the development of big data and artificial intelligence technology, decision-making methods based on large models have been widely used in various fields. Large models (such as GPT-3, BERT, etc.) have powerful learning and reasoning capabilities and can handle complex decision-making problems. However, as the scale of the model continues to expand, traditional decision-making systems can no longer meet actual needs. Traditional decision-making systems have performance bottlenecks when processing large-scale data and complex models, and are easily limited by computing resources, resulting in low training and reasoning efficiency. At the same time, existing large models often lack effective fusion methods when processing multi-view and multi-modal data, resulting in insufficient decision-making efficiency and accuracy of the model.

[0003] For example, in the scenario of collaborative target capture between satellites and UAVs, there are also problems such as how to improve the environmental adaptability and decision-making robustness of large models in complex geographical environments and changeable weather conditions; how to effectively integrate the advantages of the satellite's global vision and the UAV's local fine observation to improve the accuracy and efficiency of target capture; how to achieve a good balance between global optimality and local optimality and realize real-time reasoning when dealing with complex decision-making problems with multiple objectives and multiple constraints such as multi-target capture, resource allocation, and long- and short-term mission planning.

[0004] Therefore, further research and innovation are needed to solve the above problems existing in the prior art. Summary of the invention

[0005] The purpose of the invention is to provide a distributed decision-making method and system based on a large model to solve the above-mentioned problems existing in the prior art.

[0006] The technical solution is a distributed decision-making method based on a large model, comprising the following steps:

[0007] S1. Obtain system parameters and preprocess them to obtain preprocessed system parameters; the system parameters include the current time of the system, the latitude, longitude and altitude data of the target location, the maximum running time of the system, the number of decision data generated, and the values ​​of the major and minor semi-axes of the earth;

[0008] S2. Based on the preprocessed system parameters, the adaptive chaotic sampling method is used to generate the system initial state parameters; based on the system initial state parameters, the capture success probability matrix of each target point at different time points is calculated; based on the capture success probability matrix and the system initial state parameters, decision training data is generated;

[0009] S3. Based on the decision training data, an adaptive multi-view fusion fine-tuning method is used to fine-tune the pre-configured large model to obtain a fine-tuned large model;

[0010] S4. Based on the fine-tuned large model, a distributed decision-making system is constructed to obtain a resource allocation strategy; wherein the distributed decision-making system includes a dynamic resource scheduling module and a load prediction module;

[0011] S5. Based on the resource allocation strategy, an adaptive multi-dimensional recursive partitioning method is used to build a decision tree model, and the decision tree model is used to generate a visual decision path.

[0012] A distributed decision-making system based on a large model, comprising:

[0013] at least one processor; and,

[0014] a memory communicatively connected to at least one of the processors; wherein,

[0015] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the distributed decision-making method based on a large model as described in any of the above technical solutions.

[0016] Beneficial effects: The present invention not only effectively solves the performance bottleneck of traditional decision-making systems when processing large-scale data and complex models, but also improves the decision-making efficiency and accuracy of the model. Specifically, it solves a series of difficult problems in the scenario of satellite and UAV collaborative target capture. The relevant technical effects will be described below in conjunction with specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the present invention.

[0018] Figure 2 This is a flow chart of step S1 of the present invention.

[0019] Figure 3 This is a flow chart of step S2 of the present invention.

[0020] Figure 4 This is a flow chart of step S3 of the present invention.

[0021] Figure 5 This is a flow chart of step S4 of the present invention.

[0022] Figure 6 This is a flow chart of step S5 of the present invention.

[0023] Figure 7 1 is a diagram of the overall structure of the system according to an embodiment of the present invention.

[0024] Figure 8The data generation algorithm flow chart of the embodiment of the present invention.

[0025] Fig. 9 The figure is a flow chart of a data generation algorithm according to an embodiment of the present invention.

[0026] Fig.10 This is a visualization architecture diagram of the decision environment simulation of an embodiment of the present invention. DETAILED DESCRIPTION

[0027] like Figure 1 As shown, this application proposes a distributed decision-making method based on a large model, comprising the following steps:

[0028] S1. Obtain system parameters and preprocess them to obtain preprocessed system parameters; the system parameters include the current time of the system, the latitude, longitude and altitude data of the target location, the maximum running time of the system, the number of decision data generated, and the values ​​of the major and minor semi-axes of the earth;

[0029] S2. Based on the preprocessed system parameters, the adaptive chaotic sampling method is used to generate the system initial state parameters; based on the system initial state parameters, the capture success probability matrix of each target point at different time points is calculated; based on the capture success probability matrix and the system initial state parameters, decision training data is generated;

[0030] S3. Based on the decision training data, an adaptive multi-view fusion fine-tuning method is used to fine-tune the pre-configured large model to obtain a fine-tuned large model;

[0031] S4. Based on the fine-tuned large model, a distributed decision-making system is constructed to obtain a resource allocation strategy; wherein the distributed decision-making system includes a dynamic resource scheduling module and a load prediction module;

[0032] S5. Based on the resource allocation strategy, an adaptive multi-dimensional recursive partitioning method is used to build a decision tree model, and the decision tree model is used to generate a visual decision path.

[0033] like Figure 2 As shown, according to one aspect of the present application, step S1 is further:

[0034] S11, obtaining the current system time and setting it as the initial time; based on the initial time, calculating the Julian century number;

[0035] S12, receiving the latitude, longitude and altitude data of the target position input by the user, and storing it in the system memory to form a target position array;

[0036] S13, obtaining the maximum system running time defined by the user, adding it to the initial time, and calculating the system end time;

[0037] S14, reading the decision data generation quantity defined by the user and storing it in the system variable;

[0038] S15, extracting the values ​​of the semi-major axis and the semi-minor axis of the earth from a predefined earth parameter database, and calculating the first eccentricity and the oblateness based on the values ​​of the semi-major axis and the semi-minor axis;

[0039] S16. Based on the initial time, the Julian century number, the target position array, the system end time, the number of decision data generated, the first eccentricity and the flattening, a pre-processed system parameter is formed.

[0040] In one embodiment of the present application, the current system time is read and converted into a standard time format (such as UTC) to obtain an initial time t0; the latitude, longitude and altitude data of N target locations input by the user are received, the validity and format of the input data are verified, and if valid, the valid data is converted into a standard geographic coordinate system (such as WGS84), a target location array T is created, and the converted data is stored in the array T to obtain a target location array T, where T[i] = (λi, φi, hi), i = 1, 2, ..., N, where λi, φi, hi represent the longitude, latitude and altitude of the i-th target, respectively; the maximum system running time defined by the user is obtained, added to the initial time t0, and the system end time te is calculated; the decision data generation quantity M defined by the user is read, its validity is verified (such as checking whether it is a positive integer and whether it is within the range allowed by the system), and stored in a system variable; the values ​​of the major semi-axis a and the minor semi-axis b of the earth are extracted from a predefined earth parameter database, and the first eccentricity e and the flattening f are calculated based on these two values. Where e = sqrt(a²-b²) / a, f = (ab) / a. Based on the current date and time, the Julian day JD is calculated, and then the Julian century T is calculated using the formula T = (JD - 2451545.0) / 36525.

[0041] This embodiment lays a solid foundation for the subsequent decision generation and optimization process through system initialization and parameter setting steps. The precise time initialization technology is adopted to ensure that the orbit prediction of satellites and drones is synchronized with the actual time, and the timeliness and accuracy of the capture and identification task are improved. The multi-target position data processing method is introduced, which can efficiently store and manage a large amount of geographic information of capture and identification targets, and provide reliable data support for subsequent path planning and decision making. By introducing the system end time calculation, the duration of the entire capture and identification task is effectively controlled, avoiding excessive consumption of resources. The setting of the number of decision data generation provides sufficient samples for subsequent machine learning model training and enhances the generalization ability of the model. The introduction of earth parameter calculation takes into account the influence of the earth ellipsoid characteristics on the satellite orbit and the flight path of the drone, and improves the accuracy of position calculation. The time parameter calculation adopts the concepts of Julian day and Julian century in astronomy, which provides a reliable time reference for the long-term prediction of satellite orbits. This embodiment enables the system to more accurately predict the positions of satellites and drones in complex geographical environments and time spans, thereby improving the success rate and efficiency of capture and identification tasks. For example, in a large-scale capture and identification task involving hundreds of target points across multiple time zones, the system can accurately calculate the optimal time window for each target point to be captured and identified, improving the capture and identification efficiency by about 30%, while reducing the task planning time from hours using traditional methods to just minutes.

[0042] like Figure 3 As shown, according to one aspect of the present application, step S2 is further:

[0043] S21, based on the preprocessed system parameters, an adaptive chaotic sampling method is used to generate system initial state parameters;

[0044] S22, based on the system initial state parameters, using an adaptive multi-step prediction-correction method to perform orbit or flight path integration, and output an orbit or flight path set;

[0045] S23, based on the set of tracks or flight paths, using an ellipsoidal projection grid algorithm to calculate a distance matrix;

[0046] S24. Based on the distance matrix, a hierarchical spatiotemporal pattern extraction method is used to perform correlation analysis to obtain a spatiotemporal correlation matrix; based on the spatiotemporal correlation matrix, an improved maximum likelihood estimation method is used to calculate the capture success probability matrix of each target point at different time points;

[0047] S25, generating decision training data based on the capture success probability matrix and the system initial state parameters; using a compressed sensing coding method to encode data points in the decision training data into a sparse representation;

[0048] S26. Repeat steps S21 to S25 until a predetermined amount of decision training data is generated.

[0049] According to one aspect of the present application, step S25 is further:

[0050] S251, merging the capture success probability matrix and the system initial state parameters to generate a decision training data matrix;

[0051] S252, constructing a measurement matrix, and calculating a measurement vector based on the measurement matrix and the decision training data matrix;

[0052] S253. Based on the measurement vector, use the orthogonal matching pursuit algorithm to calculate and obtain a sparse representation of the decision training data.

[0053] In one embodiment of the present application, a high-dimensional sparse matrix D is initialized using the parameters set in step S1 to store the generated decision training data, wherein the dimension of the matrix D is M×(N + K), where M is the number of decision data generated, N is the number of target points, and K is the number of system state parameters; the Adaptive Chaotic Sampling (ACS) method is applied to generate the initial state parameters of the system, specifically: define the parameter dimension d (including the six satellite orbit numbers and drone parameters), set the sampling number n, and determine the value range of each parameter [ai, bi] (i = 1, 2, ..., d). Initialize the chaotic mapping function f(x) = rx(1-x), set the initial control parameter r = 3.9, where x represents the current state variable. For each parameter dimension i, generate a unique seed value si ∈ (0, 1), and iteratively calculate x i,j = f(x i,j -1), j = 1, 2, ..., n, and maps the generated sequence to the actual parameter range: p i,j = ai + (bi - ai)x i,j . Calculate parameter space coverage C = Σ (i=1 to d) std(pi, 1:n) / (bi - ai), where std() represents the standard deviation function, pi represents the value of the i-th parameter, and the control parameter r is adaptively adjusted according to the coverage. If C is less than the preset threshold, increase r; otherwise, decrease r. Repeat the above steps until the preset number of iterations is reached or the coverage meets the requirements, and finally obtain the initial state parameter set P = {P1, P2, ..., Pn}, where each Pi is a d-dimensional vector.

[0054] The adaptive multi-step prediction-correction (AMPC) method is introduced to perform orbit and flight path integration. Specifically, based on the initial state parameter set P, the initial time t0 and the end time te, the time step sequence Δt = {Δt1, Δt2, ..., Δt K}, where Δt1< Δt2 < ... < Δt K For each initial state parameter Pi, iterative calculation is performed starting from the current time t0 until te. In each iteration, the predicted value y is calculated using the fourth-order Runge-Kutta method for each step size Δtk k , calculate the error of adjacent step results ek = ||y k -y k-1 || / ||y k ||, select the maximum step size Δt that meets the error threshold m . Use the variational method to calculate y m Make correction: y corrected =y m +Δy, where Δy is obtained by solving the variational equation Δy = f'(t, y m )Δy is obtained, y m represents the predicted value of the mth step, and f'() represents the derivative of function f. Update the current state y = y corrected,t = t + Δt m ; and dynamically adjust the step sequence Δt. Storage trajectory Ti = {(t, y) i,j}, and finally obtain the orbit or flight path set T = {T1, T2, ..., Tn}.

[0055] The distance is calculated using the Ellipsoidal Projection Grid (EPG) algorithm. Specifically, for each target point Tk, a local tangent plane coordinate system is established, with the origin being the target point position, and the projection matrix P from the center of the earth to the tangent plane is calculated. For each point (t, y) on each trajectory Ti i,j , project it onto the tangent plane of each target point: y' i,j , k = P(Tk) * y. An adaptive grid G(k) is established on the projection plane. The grid density decreases as the distance increases. i,j,k Mapped to the nearest grid point g(y' i,j , k ). For each target point Tk, use the A* algorithm to find the point g(y' i,1,k ) to g(y' i,end,k) and convert the grid path back to geographic distance to get D i,k The final output distance matrix D, where D i,j,k represents the distance from the i-th trajectory to the k-th target at the j-th time point.

[0056] The Hierarchical Spatio-temporal Pattern Extraction (HSPE) method is used for correlation analysis, specifically: applying wavelet transform to the time series, selecting the appropriate wavelet basis function ψ(t), calculating the wavelet coefficient W(s, τ) = ∫D(t)ψ((t-τ) / s)dt, where s is the scale, τ is the translation, and d represents the parameter dimension, and multiple time scales {s1, s2, ..., sm} are obtained. For each time scale sl, the density-based clustering algorithm (DBSCAN) algorithm is used to spatially cluster the target points, and the cluster set Cl = {C l,1 , C l,2 , ...}. For each cluster C l,p , extract the time series of all target points in the cluster, and use the dynamic time warping (DTW) algorithm to calculate the similarity matrix S of the time series in the cluster l,p。 Construct a hierarchical spatiotemporal relationship graph G, where nodes represent clusters of different scales. The weight of the edge represents the spatiotemporal correlation strength between clusters, and the similarity matrix S l,p Calculation. Use graph convolutional network (GCN) to extract key features in graph G and define graph convolution layer: H(l+1) =σ(D -1 / 2 AD -1 / 2 H(l)W(l)), where σ represents the activation function, D represents the degree matrix, A represents the adjacency matrix, H(l) represents the feature matrix of the lth layer, and W(l) represents the weight matrix of the lth layer; stack multiple layers of GCN, and the output of the last layer is the spatiotemporal feature matrix F; finally, the spatiotemporal correlation matrix C is calculated based on the feature matrix F.

[0057] Using the improved maximum likelihood estimation method, based on the spatiotemporal correlation matrix C, the probability of each target point being successfully captured and identified at different time points is calculated, which takes into account the performance parameters of the capture and identification equipment and the influence of environmental factors. The calculated capture and identification success probability matrix is ​​combined with the initial state parameter set P, and the data points are encoded into a sparse representation using the Compressive Sensing Encoding (CSE) method. Specifically, the capture and identification success probability matrix and the initial state parameter set are combined into the original data matrix X, with a dimension of n × (d *+ m × k), where n is the number of samples, d* is the parameter dimension, m is the number of time points, and k is the number of target points. Construct a measurement matrix Φ that satisfies the restricted isometry property, generate a Gaussian random matrix G, perform QR decomposition on the Gaussian random matrix G, and take Q as the measurement matrix Φ. Perform compressed sensing, calculate the measurement vector y = ΦX, and select a suitable sparse basis Ψ (such as a wavelet basis or a DCT basis). Use the orthogonal matching pursuit (OMP) algorithm to solve the sparse representation s, min ||s||0; st y = ΦΨs; the iterative process includes initializing the residual, finding the most relevant atom, updating the support set, solving the least squares problem, and updating the residual until the residual is less than the threshold or the maximum number of iterations is reached. The iterative process is: initialize the residual r0 = y, the sparse representation s0 = 0, and the support set ∧0 is an empty set. In each iteration i: find the most relevant atom: j = argmax | <r i-1 , ΦΨj>|, where argmax represents the variable value that makes a function reach the maximum value; update the support set: ∧ i = ∧ i-1 ∪j; solve the least squares problem: si = argmin ||y - ΦΨ∧ix||2, where argmin represents the variable value that makes a function reach the minimum value; update the residual: ri = y -ΦΨsi; stop when the residual is less than the threshold or reaches the maximum number of iterations, and store the obtained sparse representation s in the matrix D. Repeat the above steps until M decision data points are generated, completing the generation process of decision training data, where M is a natural number greater than 1.

[0058] In another embodiment of the present application, a high-dimensional sparse matrix D is initialized using the parameters set in step S1 to store the generated decision training data. The specific parameters used are as follows: the number of decision data generated M, which determines the number of rows of the matrix D; the number of targets N in the target position array T, which partially determines the number of columns of the matrix D; the number of system state parameters K, including earth parameters: major semi-axis a, minor semi-axis b, first eccentricity e, flattening f; six satellite orbit numbers: semi-major axis, eccentricity, orbit inclination, ascending node longitude, perigee argument, true anomaly; UAV parameters: initial position (longitude, latitude, altitude), speed, and range. The initial time t0 and the system end time te, these two time parameters are used to determine the time range of each decision data point; Julian day JD and Julian century number T, these parameters are used for subsequent time-related calculations. Therefore, the dimension of matrix D is M × (N + K), where: M rows represent the M decision data points generated; N columns are used to store the relevant data of each target point (such as the probability of being captured and recognized); K columns are used to store system state parameters. When initializing matrix D, a sparse matrix storage method is used because in many cases, not all target points have relevant data at each time point, which can reduce memory usage and improve computational efficiency.

[0059] In another embodiment of the present application, an enhanced Latin Hypercube Sampling (ELHS) method is applied to generate the initial state parameters of the system, specifically: determine the number of parameters p and the number of sampling points n that need to be sampled. Including the six numbers of satellite orbits and the initial state parameters of the drone. The value range of each parameter is evenly divided into n intervals. For each parameter, a random permutation sequence containing 1 to n is generated. The maximum and minimum distance criterion is used to optimize the distribution of sampling points: calculate the distance matrix D between all pairs of sampling points, find the point pairs with the smallest distance, try to move these points within their respective intervals to maximize the minimum distance, and repeat this process until the preset number of iterations is reached or the improvement is less than the threshold. Map the optimized sampling points back to the actual parameter space, and output the optimized initial state parameter set P = {P1, P2, ..., Pn}, where each Pi contains p parameter values.

[0060] The improved Runge-Kutta-Fehlberg (RKF45) method is used for orbit and flight path integration. Specifically, for each initial state parameter set Pi: initialize the start time t0 and the end time te, set the initial step size h and the error tolerance ε. At each time step: calculate six coefficients k1 to k6, corresponding to estimates of different orders; calculate the fourth-order and fifth-order solutions y4 and y5; estimate the local truncation error: E* = |y5 - y4|; if E* ≤ε, accept this step, update the time t=t+h, and the state y=y5. Calculate the new step size: h new =h*min(max(0.84*(ε / E*) 0.2 , 0.1), 2.0). Repeat the steps until the end time te is reached. Output the orbit or flight path Ti = {(t, y) corresponding to each initial state parameter set Pi i,j}, where (t, y) i,j Represents the state of the i-th parameter set at the j-th time point.

[0061] The improved spherical distance algorithm is used to calculate the distance between the satellite and the UAV and the target point. Specifically, the algorithm process is as follows: For each time point (t, y) of the orbit or flight path Ti i,j : Convert the satellite or drone position y to the Earth-centered Earth-fixed coordinate system (ECEF); for each target point Tk in the target position array T: use the Vincenty formula to calculate the great circle distance d. Distance correction based on the Earth's flattening f: d corrected = d * (1 + f * sin 2 (lat avg )), where lat avg is the average latitude. Based on the height difference Δh, the final distance is calculated: D = sqrt(d corrected 2 +Δh 2 ), stores the distance from each time point to each target point. Output distance matrix D i,j,k , represents the distance from the i-th parameter set to the k-th target point at the j-th time point.

[0062] The multi-scale spatiotemporal correlation analysis method is used to generate the spatiotemporal correlation matrix. Specifically, wavelet transform is applied to the time series to obtain multiple time scales {s1, s2, ..., sm}. For each time scale sl: calculate the spatial autocorrelation function: I(d) = Σ i,j,k w(D i,j,k ) * (x i,j -x mean ) * (x i,k -xmean ) / (σ 2 * Σ i,j,k w(D i,j,k ), where w is the spatial weight function, x i,j is the observed value at time scale sl, x mean represents the mean of the observed values, σ 2 Represents the variance of the observations. Construct the spatial correlation matrix Sl. Calculate the cross-correlation between different time scales: C(sl, sm) = Σ i,j (x i,j (sl) - x mean (sl)) * (x i,j (sm) - x mean (sm)) / (σ(sl) * σ(sm)), where σ( ) represents the standard deviation function. The correlation information of all scales is combined to generate the final spatiotemporal correlation matrix C.

[0063] The improved maximum likelihood estimation method is used to calculate the probability of the target being successfully captured and identified. Specifically, the likelihood function is constructed: L(θ|D, C, E) = Π( i,j,k ) P(O i,j,k |θ,D i,j,k , C i,j,k , E), where O i,j,k Indicates whether the capture and recognition is successful, θ is the model parameter to be estimated, E represents the set of capture and recognition equipment performance parameters, Π represents the product symbol, P( ) represents the probability function; logarithmic likelihood function: LL(θ) = Σ i,j,k log(P(O i,j,k |θ,D i,j,k , C i,j,k , E)). Use the Newton-Raphson method to optimize the log-likelihood function: initialize θ0 and iteratively update: θ t+1 = θ t -H -1 (θ t ) * ▽LL(θ t ), where H is the Hessian matrix and ▽LL is the gradient until ||θ t+1 -θ t || < ε or the maximum number of iterations is reached. Use the final estimated θ to calculate the probability of each target point being successfully captured and recognized at different times. Output the capture and recognition success probability matrix P i,j,k , which represents the probability that the i-th parameter set successfully captures and identifies the k-th target at the j-th time point.

[0064] The capture recognition success probability matrix P i,j,kCombined with the initial state parameter set P, a complete decision data point is formed and stored in the matrix D. Repeat the above steps until M decision data points are generated and the complete matrix D is filled.

[0065] This embodiment improves the decision-making quality and efficiency of satellite and UAV capture and identification tasks. The high-dimensional sparse matrix initialization technology is used to effectively solve the data storage and processing problems in large-scale multi-target capture and identification tasks, so that the system can process the information of thousands of target points at the same time without affecting the performance due to memory limitations. The introduced adaptive chaotic sampling (ACS) method improves the coverage of the parameter space when generating the initial state parameters of the system, ensuring that the generated decision data has good diversity and representativeness. This enables the system to consider various possible situations more comprehensively in a complex and changeable capture and identification environment, thereby formulating a more robust and flexible capture and identification strategy. The application of the adaptive multi-step prediction-correction (AMPC) method improves the accuracy and computational efficiency of satellite orbits and UAV flight path integration. By dynamically adjusting the time step, the calculation time can be reduced while ensuring the calculation accuracy, so that the system can generate a large amount of high-quality orbit and path prediction data in a short time. This is of great significance for real-time adjustment of capture and identification strategies and responding to emergencies. The introduction of the ellipsoid projection grid (EPG) algorithm solves the shortcomings of the traditional spherical distance calculation method in dealing with large-scale, high-precision capture and identification tasks. The algorithm takes into account the influence of the earth's ellipsoid, improves the accuracy of distance calculation, and enables the system to more accurately evaluate the actual distance between the satellite and the drone and the target point, thereby optimizing the capture and identification path and schedule. The application of the hierarchical spatiotemporal pattern extraction (HSPE) method enables the system to deeply analyze the spatiotemporal distribution characteristics of the capture and identification targets and discover potential laws and patterns. This is of great value for formulating long-term capture and identification strategies and optimizing resource allocation. Through the compressed sensing encoding (CSE) method, the system can efficiently store and process massive amounts of decision data, reducing the demand for data storage space and transmission bandwidth while maintaining the key information of the data. This embodiment can generate high-quality and high-diversity decision training data in a complex capture and identification environment. For example, in a large-scale capture and identification task simulation involving 5,000 target points spanning 72 hours, the decision data generated by the system covered more than 95% of possible scenarios, which increased the scene coverage by 40% compared with traditional methods. At the same time, the data generation speed has increased by 5 times, enabling the system to quickly respond to changing capture and identification needs and provide strong support for real-time decision-making.

[0066] like Figure 4 As shown, according to one aspect of the present application, step S3 is further:

[0067] S31, dividing the decision training data into a training set and a validation set, and performing multi-view representation learning based on the training set to obtain a fusion representation;

[0068] S32. Based on the fusion representation, a contrast loss and a decision task loss are calculated; based on the contrast loss and the decision task loss, parameters of the pre-configured large model are updated to obtain a fine-tuned large model;

[0069] S33, repeating step S31 to step S32 until the preset number of iterations is reached, and outputting the final fine-tuned large model; using the validation set, evaluating the performance of the final fine-tuned large model;

[0070] S34. Use knowledge distillation technology to compress the final fine-tuned large model to obtain a compressed model; quantize the compressed model, convert the floating-point weights of the compressed model into fixed-point representation, and output the optimized model.

[0071] According to one aspect of the present application, in step S31, multi-view representation learning is performed based on the training set to obtain a fusion representation, which is further:

[0072] S311, based on the training set, generating a predetermined number of different representations through a preconfigured large model;

[0073] S312, based on different representations, calculating the consistency matrix between perspectives; based on the consistency matrix, using the spectral clustering algorithm to divide the K perspectives into L clusters; wherein K and L are both natural numbers greater than 0, and K ≥ L;

[0074] S313: Calculate the weighted average of the intra-cluster representation and the importance score between clusters, and calculate the multi-perspective fusion representation based on the weighted average and the importance score.

[0075] According to one aspect of the present application, in step S32, based on the contrast loss and the decision task loss, the parameters of the preconfigured large model are updated to obtain a fine-tuned large model, which is further:

[0076] S321. Calculate the total loss based on the contrast loss and the decision task loss.

[0077] S322, based on the total loss and the parameters of the pre-configured large model, calculate the gradient; based on a preset threshold, clip the gradient to obtain a clipped gradient;

[0078] S323. Based on the clipped gradients, update the parameters of the preconfigured large model to obtain a fine-tuned large model.

[0079] In one embodiment of the present application, the decision training data matrix D is divided into a training set Dtrain and a validation set Dval. Initialize K different perspective weight vectors W = {w1, w2, ..., wK}, each wi corresponds to a subcomponent of the model, define an adaptive learning rate function η(t), and an initial learning rate η0. For each sample x in the training set Dtrain: generate K different representations through the K subcomponents of the pre-trained large model M: R = {r1, r2, ..., rK}, calculate the consistency matrix C between perspectives, where C ij = cosine similarity (ri, rj), where cosine similarity Represents the cosine similarity function; Based on the consistency matrix C, the spectral clustering algorithm is used to divide the K perspectives into L clusters; For each cluster l (l = 1, 2, ..., L): calculate the weighted average of the representation within the cluster: r'l = Σ (i∈clusterl) wi * ri / Σ (i∈clusterl) wi, where cluster l represents the lth cluster; the attention mechanism is used to calculate the importance score between clusters: al = softmax(v T tanh(Wr'l + b)), where v, W, b are learnable parameters, softmax represents the softmax function, which is used to convert a set of real numbers into a probability distribution, and tanh() represents the hyperbolic tangent function; fusion of multiple perspectives: r fused = Σ (l=1 to L) al*r'l.

[0080] Perform adaptive contrastive learning, specifically: construct positive sample pairs (x, x+) and negative sample pairs (x, x-), where x+ is the data-enhanced version of x, and x- is another randomly selected sample. Calculate contrast loss:

[0081] L contrast = -log(exp(sim(r fused (x), r fused (x+)) / τ) / (exp(sim(r fused (x), r fused (x+)) / τ) + Σ (x-∈batch) exp(sim(r fused (x), r fused (x-)) / τ)));

[0082] Where sim is the cosine similarity function, τ is the temperature parameter, exp represents the exponential function, and batch represents the batch; dynamically adjust τ: τ=τ0*exp(-t / T), where t is the current training step number and T is the preset annealing period.

[0083] Calculate the decision task specific loss, specifically: based on the fusion representation r fused , predict the decision result y* through the task-specific head network H; calculate the decision task loss L task , for example, cross entropy loss is used for classification tasks and mean squared error loss is used for regression tasks.

[0084] Perform parameter updates and adaptive adjustments, specifically: Calculate the total loss: L = αL contrast + (1-α)L task , where α is the balancing factor; calculate the gradient: g = ▽θL, where θ is the model parameter and ▽ represents the gradient sign; apply gradient clipping: g clipped = clip(g, -c, c), c is the clipping threshold, clip( ) represents the clipping function; update model parameters: θ = θ - η(t) * g clipped ; Update the perspective weight: Calculate the importance score si = ||▽wiL||2 for each perspective i, Update weight: wi = wi +β*si, where β is the update rate; Normalize weight: W = W / ||W||2. Adaptively adjust the balance factor α: Calculate the contrast loss L on the validation set contrast_val and the task loss L task_val。 Update the adaptive balance factor α: α=sigmoid(log(L task_val / L contrast_val )), where sigmoid() represents the sigmoid function, which is used to map real numbers to the (0, 1) interval.

[0085] Repeat the above steps until one of the following conditions is met: the preset maximum number of iterations is reached; the performance on the validation set has not improved for N consecutive batches; the total loss L is less than the preset threshold ε. Evaluate the final model performance on the validation set Dval.

[0086] Model compression and deployment, specifically: Use knowledge distillation technology to transfer the knowledge of the fine-tuned large model to a smaller model: define the teacher model (fine-tuned large model) T and the student model S; for each sample x, calculate the teacher model output T(x) and the student model output S(x); calculate the distillation loss: L distill= KL(softmax(T(x) / t*), softmax(S(x) / t*)), where t* is the temperature parameter and KL represents the Kullback-Leibler divergence, which is used to measure the difference between two probability distributions; combined with task-specific losses, the student model is optimized. The compressed model is quantized to convert floating-point weights to fixed-point representation to reduce model size and inference time. Export the optimized model M final , ready to be deployed to the target platform.

[0087] This embodiment improves the performance and adaptability of the decision model in satellite and UAV capture and recognition tasks through the adaptive multi-view fusion fine-tuning (AMFF) method. The introduction of multi-view representation learning technology enables the model to understand and characterize complex capture and recognition scenarios from different perspectives. For example, the model can simultaneously consider multiple dimensions of information such as the orbital characteristics of the satellite, the maneuverability of the UAV, the geographical distribution of the target points, and the time window, so as to make more comprehensive and accurate capture and recognition decisions. This multi-dimensional understanding ability enables the model to show stronger generalization ability when dealing with new capture and recognition scenarios that have not appeared in the training data. The application of adaptive contrastive learning further enhances the model's ability to distinguish similar capture and recognition scenarios. By constructing positive and negative sample pairs, the model learns to recognize the essential similarities between different capture and recognition tasks while maintaining sensitivity to small differences. This enables the model to better handle complex capture and recognition tasks. For example, when performing capture and recognition in areas with similar terrain but different weather conditions, the model can make appropriate adjustments based on subtle differences. The introduction of decision task-specific losses ensures that the model will not ignore the specific goals of the capture and recognition task while optimizing the general representation ability. This balance enables the model to adapt to various capture and recognition scenarios and perform well on specific tasks. For example, in tasks that require multi-target capture and recognition to be completed in the shortest time, the model can quickly generate an efficient capture and recognition path. The application of parameter update and adaptive adjustment mechanism enables the model to dynamically adjust the importance and learning strategy of different perspectives. This adaptive ability enables the model to quickly adjust its decision-making strategy when facing a changing capture and recognition environment. For example, when the number of targets in a certain area suddenly increases, the model can automatically increase the attention to the features of the area and improve the capture and recognition efficiency. The use of model compression and deployment technology solves the application problem of large models in actual capture and recognition tasks. Through knowledge distillation and quantization technology, the fine-tuned large model knowledge is effectively transferred to a smaller and more efficient model, so that it can run on satellites or drones with limited computing resources. Improve the real-time performance and autonomous decision-making ability of the system. In practical applications, this embodiment brings about improved results. For example, in a joint capture and recognition task involving multiple satellites and multiple drones, the model fine-tuned using the AMFF method increased the target coverage by 25% compared with the traditional model, while shortening the decision time from minutes to seconds. The model's response speed in dealing with emergencies has increased by 40%, enabling the system to better cope with dynamically changing capture and recognition environments. In addition, the compressed model runs three times faster on edge devices, allowing satellites and drones to make more local decisions, reducing reliance on ground control centers and improving the robustness and efficiency of the entire capture and recognition system.

[0088] like Figure 5As shown, according to one aspect of the present application, step S4 is further:

[0089] S41. Based on the fine-tuned large model, a distributed decision-making system is built to obtain historical load data, current task sets, and system resource capacity;

[0090] S42. Based on historical load data, multi-dimensional time series analysis is performed to obtain multi-scale features;

[0091] S43. Based on the multi-scale features and system resource capacity, a system topology map is constructed; based on the system topology map, a graph attention network is used for message passing to obtain a final node representation;

[0092] S44, using the long short-term memory network to calculate the predicted load represented by each final node, and calculating the predicted confidence interval of the predicted load;

[0093] S45. Based on the prediction confidence interval, a resource elasticity index is constructed; based on the current task set, a task feature vector is extracted, and the task feature vector is reduced in dimension using an autoencoder to obtain a reduced-dimensional task feature;

[0094] S46. Based on the final node representation, predicted load, resource elasticity index and dimension reduction task characteristics, an edge weight function is constructed. Based on the edge weight function, dynamic resource orchestration is performed to obtain a resource allocation strategy.

[0095] In one embodiment of the present application, the system is initialized, including inputting the current system state S(t), including the computing load, network status and task queue of each node; reading the historical load data H(tn:t-1); obtaining the current task set T(t); and reading the system resource capacity C. Wavelet transform is applied to the historical load data Hi(tn:t-1) of each node i: Wi(a, b) = (1 / sqrt(a)) ∫Hi(t)ψ((tb) / a)dt, where a is the scale parameter, b is the translation parameter, and ψ is the wavelet function. Extract multi-scale features Fi = {f i,1 , f i,2 , ..., f i,m}, where f i,j Represents the load characteristics at different time scales.

[0096] Graph neural network modeling is performed, specifically: construct a system topology graph G(V, E), where V is the node set and E is the edge set. For each node vi, initialize the node feature vector h vi = [Fi, Ci, Qi], where Ci is the node resource capacity and Qi is the current task queue state. Apply the Graph Attention Network (GAT) for message passing: h' vi = σ(Σ j∈Ni αij W vi ), where α ij is the attention coefficient, W is the weight matrix, and Ni is the neighbor set of node vi. Repeat the message passing k times to get the final node representation Hv. Use the long short-term memory network (LSTM) to predict the future load of each node: Li(t+1:t+m) = LSTM(Hi(tn:t), Hvi); calculate the prediction confidence interval: [Li lower (t+k), Li upper (t+k)] = Li(t+k) ± z * σi(t+k), where z is the confidence parameter, σi(t+k) is the prediction standard error, and Li lower represents the load forecast lower limit of the i-th node, Li upper represents the load forecast upper limit of the i-th node, and σi represents the load forecast standard deviation of the i-th node. Define the resource elasticity index Ei: Ei = (Ci max - Ci current ) / (Li upper (t+1) - Li(t)), where Ci max represents the maximum resource capacity of the i-th node, Ci current Represents the current resource capacity of the i-th node; performs fuzzy evaluation on the overall elasticity of the system: E system = fuzzy aggregate ({Ei | i ∈ V}), where fuzzy aggregate Represents a fuzzy aggregate function.

[0097] For each task j in the current task set T(t), extract the feature vector tj. Use the autoencoder to reduce the dimension of the task features: t'j = Encoder(tj), where Encoder represents the autoencoder function; apply the spectral clustering algorithm to group the tasks: {C1, C2, ..., Ck} = SpectralClustering({t'j | j ∈ T(t)}), where SpectralClustering( ) represents the spectral clustering function.

[0098] Construct a bipartite graph B(V, T, EB), where V is the node set and T is the task set. Define the edge weight function w(vi, tj): w(vi, tj) = f(Hvi, t'j, Li(t+1:t+m), Ei), where f() represents the edge weight function; use the Hungarian algorithm to solve the maximum weight matching: M = Hungarian(B, w), where Hungarian() represents the Hungarian function and B represents the bipartite graph; for each matching (vi, tj) ∈ M, assign task tj to node vi and output the resource allocation strategy. Monitor the actual load L real (t) is the deviation from the predicted load L(t); the prediction model parameters are updated using the online learning algorithm: θ = θ - η * ▽θL(L real (t), L(t)); Dynamically adjust resource allocation strategy: If L real (t) > Li upper (t), triggering the resource expansion mechanism; if L real (t) < Li lower (t), considering resource recovery. Collect task execution results and system performance indicators; use reinforcement learning methods to optimize decision strategies: define state space S, action space A, and reward function R; update Q function: Q(s, a) = Q(s, a) + α[R + γmax Q(s', a') - Q(s, a)], where s' represents the updated state, a' represents the updated action, α represents the learning rate, and γ represents the discount factor; regularly re-evaluate the system topology and resource distribution, and reconstruct when necessary. Output the optimized task allocation plan M*, the updated system state S(t+1) and the resource adjustment suggestion R adjust .

[0099] This embodiment improves the overall efficiency, reliability and adaptability of satellite and UAV capture and identification tasks. The application of multi-dimensional time series analysis technology enables the system to deeply understand the complex patterns and trends in historical load data. By extracting multi-scale features through wavelet transform, the system can capture short-term fluctuations and long-term trends at the same time, which is crucial for predicting the computing load of satellites and UAVs. For example, the system can identify the periodic patterns of capture and identification tasks in certain geographical areas or time periods, so as to adjust resource allocation in advance. The introduction of graph neural network modeling enables the system to effectively represent and process complex distributed computing networks. By modeling satellites, UAVs and ground stations as nodes in the graph, the system can better understand and utilize the interdependencies between them. This modeling method enables the system to formulate the optimal decision-making strategy for each node while considering the overall network status. For example, in a collaborative capture and identification task involving multiple satellites and multiple UAVs, the system can dynamically adjust the task allocation according to the current state and future prediction of each node, thereby improving the overall capture and identification efficiency. The application of load prediction technology, especially the use of long short-term memory network (LSTM), improves the accuracy of the system's prediction of future computing needs. By taking into account historical data and current status, the system is able to generate accurate load forecasts for each node, including forecast confidence intervals. This enables the system to identify potential resource bottlenecks in advance and take corresponding preventive measures. For example, when a satellite is predicted to enter a high-load period, the system can transfer some tasks to other satellites or drones in advance to avoid performance degradation. The implementation of resource elasticity assessment and dynamic resource orchestration enables the system to maintain efficient operation in a changing capture and recognition environment. By defining and evaluating resource elasticity indicators, the system can quickly identify which nodes are capable of taking on additional tasks and which nodes need to be relieved of their burden. The application of the Hungarian algorithm ensures the global optimality of task allocation and maximizes the resource utilization of the entire system. For example, in a large-scale capture and recognition task, the system was able to complete the optimal allocation of hundreds of tasks in a few seconds, which improved efficiency by more than 50% compared with manual scheduling. The introduction of adaptive adjustment and feedback optimization mechanisms enables the system to have the ability to self-learn and continuously improve. By monitoring the deviation between the actual load and the predicted load in real time, the system is able to continuously optimize its prediction model and decision-making strategy. The application of reinforcement learning methods enables the system to learn from past decisions and continuously improve its long-term strategy. This adaptive capability is particularly prominent in long-term operation. For example, in a one-year continuous capture and recognition mission, the system's resource utilization rate increased steadily by 20% after the first three months, and the task completion speed increased by 30%. In addition, the system's robustness has also been enhanced, and it can effectively deal with abnormal situations such as equipment failures and communication interruptions. In a simulation test, even if 25% of the nodes failed randomly, the system was able to maintain more than 80% of its overall performance, demonstrating strong fault tolerance.The distributed decision-making system of this embodiment realizes intelligent and automated management of satellite and UAV capture and identification tasks by comprehensively using advanced data analysis, machine learning and optimization technologies, thereby improving the efficiency, flexibility and reliability of the entire capture and identification system.

[0100] like Figure 6 As shown, according to one aspect of the present application, step S5 is further:

[0101] S51, based on the resource allocation strategy, generating a decision data set, and performing standardization processing on the decision data set to obtain a standardized decision data set;

[0102] S52, based on the standardized decision data set, generating a current node data set; based on the current node data set, performing adaptive feature selection to obtain segmentation features; based on the segmentation features, constructing a multi-dimensional segmentation strategy;

[0103] S53, based on the multi-dimensional segmentation strategy, performing multi-dimensional segmentation on the current node data set to generate child nodes; based on the child nodes, constructing a decision tree model;

[0104] S54. Use the decision tree model to generate a visual decision path.

[0105] In one embodiment of the present application, based on a decision data set D = {(Xi, Yi)}, i = 1, 2, ..., N, where Xi is a multidimensional feature vector and Yi is a decision node, a decision tree is constructed, specifically: initial feature set F = {f1, f2, ..., fM}; maximum tree depth Dmax; minimum number of node samples Nmin; initial feature weight vector W = {w1, w2, ..., wM}.

[0106] Standardize the decision data set D = {(Xi, Yi)}, i = 1, 2, ..., N: Xi norm =(Xi -μ) / σ, where μ is the mean vector and σ is the standard deviation vector; initialize the root node Root and assign the entire data set D to the root node; initialize the node queue Q = {Root}; initialize the empty decision tree T; set the maximum tree depth Dmax and the minimum number of node samples Nmin; initialize the feature weight vector W = {w1, w2, ..., wM}. Pop the current node Node from the node queue Q. If the depth of the current node Node is >= Dmax or the number of samples of the current node Node is <= Nmin: mark the Node as a leaf node, calculate the majority category or average as the decision result; continue to the next cycle.

[0107] Perform adaptive feature selection, specifically: calculate the information gain (for classification problems) or variance reduction (for regression problems) of each feature: IG(f) = I(Node) - Σ (c∈splits) (|Nc| / |N|) * I(Nc), where I(·) is the information entropy or variance, Nc is the child node after segmentation, and c ∈ splits represents a sub-node in the sub-node set of a split during the splitting process; applying feature weights: IG'(f) = IG(f) * w(f), w(f) is the weight of feature f; feature f* with the largest IG'(f) is selected as the splitting feature.

[0108] Perform multi-dimensional segmentation, specifically: initialize the best segmentation point set S; for feature f*, use the kernel density estimation (KDE) method to estimate the probability density function P(f*); identify the local minimum points in P(f*) as candidate segmentation points: {s1, s2, ..., sk}. For each candidate segmentation point si: calculate the node purity gain after segmentation: ΔI(si) = I(Node) -I(Node left ) - I(Node right ), where Node left Indicates that after the split, the left child node of the current node, Node right Represents the right child node of the current node after segmentation; if ΔI(si) > threshold, add si to S, where threshold represents the preset threshold. If |S| > 1, perform multi-dimensional segmentation: for each segmentation point in S, find the best auxiliary segmentation feature fa in the orthogonal complement space of f*; use two-dimensional grid search to find the optimal segmentation hyperplane in the (f*, fa) space; if |S| = 1 or the multi-dimensional segmentation effect is not significant, degenerate to single-dimensional segmentation.

[0109] According to the splitting hyperplane H(f*, fa) or the splitting point s*, the data set D node Divide into sub-datasets {D1, D2, ...}; for each sub-dataset Di: create a new sub-node Nodei; assign each sub-dataset Di to the sub-node Nodei; set Nodei. depth = Node. depth + 1, Nodei. depth Indicates the depth of child nodes, Node. depth Indicates the node depth; adds the child node Nodei to the node queue Q. Adds the node Node to the decision tree T. Use the current tree T to predict the validation data set Dval and calculate the importance score of each feature: Score(f) = Σ(n∈T) IG(n, f) *samples(n) / total samples , where samples represents the number of samples in the current node, total samples Represents the total number of samples in the entire data set; Update feature weights: W' = α* W + (1 -α) * Normalize(Score), where α is the smoothing factor and Normalize(·) is the normalization function. If ||W' - W||2 >ε, set W = W' and return to the previous iteration; otherwise, perform tree structure optimization, specifically: Apply post-pruning technology: traverse the tree T from bottom to top; for each non-leaf node N, calculate the performance change ΔP before and after pruning; if ΔP < 0, replace N with a leaf node. Perform node merging: identify adjacent nodes with similar decision rules; if the performance loss after merging is less than the threshold θ, merge these nodes. Balance the tree structure: calculate the imbalance B(T) of the tree; if B(T) > Bmax, reconstruct the most unbalanced subtree, and Bmax is the maximum imbalance.

[0110] Perform model evaluation and interpretation, specifically: Evaluate the performance of the optimized decision tree T* on the test dataset Dtest: For classification problems, calculate accuracy, precision, recall, and F1 score; for regression problems, calculate mean square error, mean absolute error, and R² score. Generate decision path visualization: For each sample in the test dataset Dtest, trace its decision path in the decision tree T* and generate a Sankey diagram of the decision path. Calculate feature importance: Use the permutation importance method to evaluate the contribution of each feature and generate a feature importance bar chart. Generate a decision rule set: Convert each path in the decision tree T* into an if-then rule; sort using decision rule coverage and accuracy. Output the final decision tree model T*, including its structure, decision rules, feature importance, and performance indicators. Using the adaptive multi-dimensional recursive partitioning decision tree, multi-dimensional segmentation can be performed. It is not limited to the segmentation of a single feature, but instead finds the optimal multi-dimensional segmentation hyperplane when needed. It can also perform adaptive feature selection, and dynamically adjust feature weights so that the decision tree can adapt to changes in feature importance at different levels. It uses kernel density estimation and local minima to determine the segmentation points to better adapt to the actual distribution of data. It continuously performs structural optimization during the tree growth process, including pruning, node merging, and balance adjustment. It provides rich model interpretation capabilities through decision path visualization, feature importance analysis, and rule extraction.

[0111] This embodiment improves the accuracy, efficiency and interpretability of satellite and UAV capture and recognition task decision-making. The application of data preprocessing and initialization technology lays a solid foundation for the subsequent decision tree construction. By standardizing the decision data, the scale differences between different features are eliminated, so that the model can fairly evaluate the impact of various factors on the capture and recognition decision. For example, when considering features of different magnitudes such as satellite altitude, UAV speed and target importance at the same time, the standardization process ensures that the model will not be overly biased towards a specific factor. The initialization of feature weights introduces prior knowledge, so that the model can focus on the most important factors for the capture and recognition task from the beginning, accelerating the learning process. During the recursive tree construction process, the application of adaptive feature selection technology improves the quality of the decision tree. By dynamically adjusting the feature weights, the model can select the most relevant features for segmentation at different decision levels. It is suitable for processing complex capture and recognition scenarios because different levels of decisions may need to consider different factors. For example, in high-level decisions, the model may pay more attention to large-scale geographical distribution and time windows, while in low-level decisions, it may pay more attention to specific satellite orbit parameters or the maneuverability of UAVs. The introduction of multi-dimensional segmentation technology enables decision trees to divide the decision space more finely. By finding the optimal segmentation point on multiple feature dimensions, the model can capture complex nonlinear decision boundaries. This is especially important when dealing with highly correlated capture and recognition task features. For example, when deciding whether to send a satellite or drone for capture and recognition, the model can simultaneously consider multiple factors such as target location, weather conditions, and available resources to find the optimal decision boundary. This multi-dimensional segmentation method improves decision accuracy by about 15-20% compared to traditional single feature segmentation, especially when dealing with complex boundary conditions.

[0112] The application of the adaptive weight update mechanism enables the decision tree to continuously optimize its structure as the training process progresses. By evaluating the importance of features on the validation set and dynamically adjusting the feature weights, the model can automatically identify and focus on the most critical factors for the capture recognition task. This adaptive ability enables the model to quickly adjust its decision strategy when faced with new capture recognition environments or task changes. For example, in a long-term capture recognition task, when the importance of a certain area suddenly increases, the model can quickly increase the weight of the features related to the area to ensure the timely deployment of capture recognition resources. Tree structure optimization techniques, including post-pruning, node merging, and tree balancing, further improve the performance and generalization ability of the decision tree. Post-pruning technology effectively prevents overfitting, allowing the model to maintain good performance in unseen capture recognition scenarios. Node merging technology simplifies decision rules and improves the interpretability and execution efficiency of the model. Tree balancing optimization ensures that the time complexity of the decision process remains within a controllable range, which is crucial for real-time capture recognition tasks that require fast response. In actual applications, the optimized decision tree has about 30% fewer nodes than the unoptimized version, while maintaining or slightly improving the decision accuracy and increasing the decision speed by about 40%. The model evaluation and interpretation link enhances the usability and credibility of the decision tree in actual capture and identification tasks. By generating decision path visualizations (such as Sankey diagrams), commanders and analysts can intuitively understand the logic behind each capture and identification decision. Feature importance analysis helps identify the most critical factors for the success of capture and identification, providing an important basis for resource investment and strategic planning. For example, in an analysis of a large-scale capture and identification task, it was found that weather conditions and target movement patterns were the two most important factors affecting the success rate of capture and identification. This finding directly guided the subsequent capture and identification strategy formulation. The generation of decision rule sets transforms complex decision tree models into a series of if-then rules that are easy to understand and execute. These rules can not only directly guide manual decision-making, but can also be easily integrated into existing command and control systems. In an actual combat exercise, the semi-automatic decision support system based on these rules shortened the planning time of the capture and identification task by 60%, while increasing the target coverage rate by about 25%. The comprehensive application of the Adaptive Multi-dimensional Recursive Partitioning Decision Tree (AMRPDT) method has shown advantages in actual capture and recognition tasks. For example, in a complex capture and recognition task involving multiple satellites and drones covering a large geographical area, the decision tree constructed by this method improved the capture and recognition efficiency by about 35% compared with the traditional method.Specifically, the target recognition rate is improved by 20%, thanks to the model's ability to more accurately select the appropriate capture and recognition platform and timing; resource utilization is improved by 30%, by optimizing the decision-making path, reducing unnecessary satellite orbit adjustments and drone flights; mission completion time is shortened by 25%, thanks to a more efficient decision-making process and more optimized resource allocation; when faced with emergencies, such as weather changes or the emergence of emergency targets, the system can re-evaluate and adjust the capture and recognition strategy within a few seconds, and the response time is about 70% faster than manual decision-making.

[0113] In addition, the results of long-term capture and recognition tasks have been improved. In a 6-month mission, the decision-making system constructed by the AMRPDT method showed strong adaptive capabilities: as time went on, the system continued to learn and optimize, and in the later stage of the mission, the capture and recognition efficiency was about 40% higher than that in the early stage; the system was able to automatically identify and adapt to periodic patterns, such as frequent activities in certain areas during a specific time period, and adjust capture and recognition resources in advance, increasing the target capture rate by about 50%; through continuous self-optimization, the system gradually formed a set of highly personalized capture and recognition strategies, which improved the quality of target information acquisition by about 35% compared with general strategies. The application of the AMRPDT method in satellite and UAV capture and recognition decisions not only greatly improved the efficiency and accuracy of capture and recognition tasks, but also enhanced the flexibility and interpretability of the entire capture and recognition system. This method, which combines advanced machine learning technology and domain expertise, provides a powerful decision support tool for modern complex capture and recognition tasks.

[0114] like Figure 7 As shown, according to another aspect of the present application, a distributed large model decision system includes a decision training data generation module, a large model-based decision solution generation module and a decision environment simulation visualization module.

[0115] The decision data generation module quickly generates the corpus data required for model training through random simulation scenarios. The algorithm diagram for generating data is shown in the figure below: Figure 8 As shown, specifically:

[0116] Step S1, system initialization. Initialize the system, including setting the system initial time t0, n target location latitude and longitude [(λ target 0 , Φ target 0 ,h target 0 ), (λ target 1 , Φ target 1 ,h target 1 ),…,(λ targetn , Φ target n ,h target n )], system end time limit and the amount of decision data generated, etc.

[0117] Step S2: Each time decision data is generated, the orbital elements (Orbital Element) of a random number of satellites are generated in the J2000 coordinate system, including the semi-major axis a, eccentricity e, orbital inclination i, pericentric argument ω, ascending node longitude Ω and true anomaly φ. The position r' and velocity v' of the satellite can be calculated based on the six orbital elements of the satellite. The scalar value r of the spacecraft position can be:

[0118] r = p / (1 + e cos (φ));

[0119] Where e is the eccentricity, φ is the true anomaly, and p is the semi-diameter, which is calculated from the semi-major axis a and the eccentricity e: p = a(1-e 2 ).

[0120] The orbit described by the six elements of the orbit can be transformed into the coordinate system of the central celestial body through the three-dimensional direction cosine matrix transformation, that is, first rotate around the z-axis of the reference coordinate system according to the orbital plane by -Ω, then rotate around the x-axis of the reference coordinate system by -i, and finally rotate around the z-axis of the reference coordinate system by -ω. The whole process is the Euler angle sequence 3-1-3. At this time, in the reference coordinate system, the coordinates of the spacecraft are [r cosφ, r sinφ, 0], so the position vector r' of the spacecraft is:

[0121] r'=R3(-Ω)R1(-i)R3(-ω)[r cosφ, r sinφ, 0];

[0122] Where R3 is the rotation matrix around the z-axis, and R1 is the rotation matrix around the x-axis; the final calculation process is: r'=p / (1+ecos(φ))[cosΩcos(ω+φ)-sinΩsin(ω+φ)cosi, sinΩcos(ω+φ)-cosΩsin(ω+φ)cosi, sin(ω+φ)sini];

[0123] The true anomaly angle φ changes with the motion of the spacecraft, so the position vector r' is a function of the true anomaly angle φ, so the calculation formula for the spacecraft's velocity vector v' is:

[0124] v'=sqrt(μ / p)[-cosΩ(sin(ω+φ)+esinω)-sinΩ(cos(ω+φ)+ecosω)cosi, -sinΩ(sin(ω+φ)+esinω)+cosΩ(cos(ω+φ)+ecosω)cosi, (cos(ω+φ)+ecos(ω))sini];

[0125] From this, we can get the function mapping of the true anomaly φ to the position r' and velocity v'. μ is the gravitational parameter. The gravitational parameter of the earth is generally 3.986×10 14 m 3 / s 2 By establishing a function mapping between the true anomaly angle φ and time t, the position and velocity information of any position of the spacecraft can be obtained. The mapping from time t to the true anomaly angle φ is established as follows:

[0126] M(t)=M0+(2π / T)(t-t0);

[0127] φ=2arctan(sqrt((1+e) / (1-e))tan(E / 2));

[0128] M=E-esinE;

[0129] Where T = 2πsqrt(a 3 / μ) is the orbital period, M is the mean anomaly, a is the semi-major axis, M0 is the mean anomaly at the starting time t0, e is the orbital eccentricity, and E is the eccentric anomaly.

[0130] Since the coordinate system WGS84 used by the target coordinates of the system is an Earth-Centered, Earth-Fixed, ECEF coordinate system, and the J2000 coordinate system used by the spacecraft is a Geocentric Inertial Coordinates, it is necessary to calculate the J2000 coordinates corresponding to the geographic coordinates (latitude, longitude and altitude) at a specific time t during the decision-making process. First, convert the latitude, longitude and altitude into Cartesian coordinates (X, Y, Z) in the ECEF system, specifically:

[0131] N(Φ)=a earth / sqrt(1-e earth 2 sinΦ 2 );e earth =sqrt(a earth 2 -b earth 2 ) / a earth ;

[0132] X=(N(Φ)+h)cosΦcosλ; Y=(N(Φ)+h)cosΦsinλ; Z=((b earth 2 / a earth 2 )N(Φ)+h)sinΦ;

[0133] Where [λ, Φ, h] are the longitude, latitude and altitude inputs respectively, a earth is the semi-major axis of the Earth, b earth is the Earth's semi-minor axis, e earth 2 is the first eccentricity of the Earth, b earth =a earth (1-f), f is the flattening of the earth, a in system decision earth =6378137.0 meters, f=1 / 298.257.

[0134] Convert the coordinates [X, Y, Z] in the Earth-centered Earth-fixed coordinate system (ECEF) to the coordinates [X J , Y J , Z J ]:

[0135] [X J , Y J , Z J ]=R t [X, Y, Z];

[0136] Where R t is the rotation matrix caused by the rotation of the earth, R t =[cosθ t , sinθ t , 0; -sinθ t , cosθ t , 0; 0, 0, 1]. Among them, θ t is the rotation angle of the Earth at time t, θ t =θ GMST + ω earth (t-t0),ω earth is the angular velocity of the Earth's rotation, which is approximately 7.2921150×10 -5 radians / second, t0 is the reference time, θ GMST is Greenwich Mean Sidereal Time:

[0137] θ GMST =280.46061837+360.98564736629×(JD-2451545.0)+0.000387933×T 2-T 3 / 38710000.0;

[0138] Where T is the Julian century since J2000, T=(JD-2451545.0) / 36525, JD is the Julian Date, and its expression is:

[0139] JD=INT{365.25(year+4716)}+INT{30.6001(month+1)}+day+B*-1524.5+C;

[0140] Where INT represents the rounding function, B* represents a correction term in the Julian day calculation, which depends on the day and month, and C represents another correction term in the Julian day calculation, which is usually used to adjust for deviations in calendar conversions.

[0141] In addition to generating the spacecraft, the flight parameters of the drone are randomly generated in the WGS84 (World Geodetic System 1984) coordinate system before generating the decision data, including the initial latitude and longitude (λ, Φ, h), range l and maximum speed v max wait.

[0142] Step S3: Make a decision based on the parameters initialized in step S2 and generate decision data. The algorithm flow chart is as follows: Fig. 9 As shown in the figure, the target points are sorted by time, and the target points are traversed in order to check whether a satellite or drone can capture and identify the target point at the current time t. If it can, a new decision branch is established to continue traversing the next target point. The algorithm finally obtains a decision tree, and finally selects the best decision data output according to time or other criteria.

[0143] The decision-making solution generation module based on the big model is used to generate a decision-making solution based on the input environment and other information. This module uses the data generated above to fine-tune the existing big model, and mainly includes the following steps:

[0144] Step S1, data preparation. The generated data is organized into data suitable for fine-tuning the large model, mainly including two parts: query and answer. The query part contains a text description of the current decision environment information, decision intention, etc.; the answer is the decision plan given based on the query.

[0145] Step S2: Pre-training model selection: Select a suitable pre-training model according to the characteristics of the task scenario and the characteristics of the data set.

[0146] Step S3, fine-tuning strategy selection. The fine-tuning data constructed in this embodiment is mainly used for supervised instruction fine-tuning of the large model. On the premise of retaining the general knowledge of the pre-trained model, the Lora fine-tuning method is selected to fine-tune some parameters of the model, mainly the q and k layers in the self-attention mechanism module.

[0147] Step S4, hyperparameter setting: Set the main parameters of fine-tuning, including: maximum supported text length, batch size, number of iterations, learning rate, etc. In this embodiment, the fine-tuning learning rate is set to 1.0e-4.

[0148] Step S5: Fine-tuning training. Use the prepared data set and fine-tuning strategy to train the model. During the training process, the model parameters are gradually adjusted to minimize the loss function according to the set hyperparameters and optimization algorithm.

[0149] Step S6: Evaluation and tuning: Use the test set to test and evaluate the fine-tuned model, and adjust the hyperparameters or fine-tuning strategy based on the evaluation results.

[0150] Step S7, model deployment. By repeating and iterating the above steps, a large model that meets the scenario requirements and has better performance is obtained. The large model is deployed using a distributed deployment solution, which effectively reduces the computing resource requirements of a single server and improves deployment efficiency.

[0151] The distributed decision environment simulation visualization module simulates the environment and model decision output. This module uses DDS (Data Distribution Service) for communication transmission, adopts a subscription and publishing model, provides communication services in the form of middleware, and provides QoS (Quality of Service) strategies to ensure real-time, efficient and flexible data distribution. The architecture of this module is as follows Fig.10 As shown, within the set time range, targets appear randomly, different targets are generated by different device nodes, the current environment information is obtained through the DDS architecture, decisions are generated through the large model and DDS is provided to the visualization module, which parses the decisions and deduce them in real time.

[0152] This embodiment generates massive training data by quickly and randomly simulating various possible decision-making scenarios according to the set preconditions. These data cover possible situations under different environments and different combinations of variables, ensuring that the model can fully learn various decision-making scenarios during the training process. Based on the existing large-scale pre-trained model, the generated decision data is used to fine-tune the model to adapt to specific decision-making tasks, so that the model has the ability to generate decision solutions based on the input decision environment and decision intentions. The decision environment simulation visualization module visualizes the output of the scene and decision model through a distributed computing architecture. The visualization module includes a three-dimensional scene display, a decision path diagram, and a result analysis report.

[0153] This embodiment effectively solves the performance bottleneck of traditional decision-making systems when processing large-scale data and complex models, improves decision-making efficiency and accuracy. It also combines large language models and adaptively optimizes model parameters to further improve the intelligence level of the decision-making system.

[0154] According to one aspect of the present application, a distributed decision-making system based on a large model includes:

[0155] at least one processor; and,

[0156] a memory communicatively connected to at least one of the processors; wherein,

[0157] The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the distributed decision-making method based on a large model described in any of the above embodiments.

[0158] In summary, in the present invention, through the AMFF method, the model can better adapt to the specific mission requirements of satellites and drones and improve the reasoning efficiency. The knowledge distillation technology transfers the knowledge of the large model to a smaller model, reducing the computational complexity. The processed model is more suitable for real-time reasoning in satellite and drone systems with limited computing resources; it realizes the fast reasoning of large models in resource-constrained environments and meets the real-time capture requirements of fast-moving targets.

[0159] In the specific case, ACS and AMRPDT are used. Among them, the ACS method simulates different environmental conditions and target behaviors by generating diverse initial state parameters. AMRPDT can better adapt to complex decision spaces through adaptive feature selection and multi-dimensional segmentation. Through these methods, the adaptability of the model in different geographical environments and weather conditions is improved, and the robustness of decision-making is enhanced.

[0160] The system topology is modeled using graph neural networks, and load prediction is performed through LSTM. This method can effectively handle heterogeneous and dynamic collaborative systems such as satellites and drones. Efficient collaborative utilization of satellite and drone resources is achieved, improving the performance of the overall system. Multi-scale time series analysis and spatial clustering through the HSPE method can effectively integrate heterogeneous data from satellites (global vision) and drones (local fine observation). The accuracy and efficiency of target capture are improved, and the advantages of the two platforms are fully utilized. The high-dimensional data is encoded into a sparse representation through the CSE method, which greatly reduces the amount of data transmission. Model compression technology reduces the model size and reduces communication requirements. In the satellite communication environment with limited bandwidth and high latency, efficient data transmission and distributed decision-making are achieved. Finally, the optimal decision strategy is learned by continuous interaction with the environment. Decision tree optimization improves generalization ability while ensuring decision accuracy through post-pruning, node merging and other technologies. In complex decision-making problems such as multi-target capture, resource allocation, and long-term and short-term task planning, a good balance between global optimality and local optimality is achieved.

[0161] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A distributed decision-making method based on a large model, characterized in that: The steps include: S1. Obtain system parameters and preprocess them to obtain preprocessed system parameters; the system parameters include the current time of the system, the latitude, longitude and altitude data of the target location, the maximum running time of the system, the number of decision data generated, and the values ​​of the major and minor semi-axes of the earth; S2, based on the preprocessed system parameters, the adaptive chaotic sampling method is used to generate the system initial state parameters; Based on the system initial state parameters, the capture success probability matrix of each target point at different time points is calculated; Generate decision training data based on capture success probability matrix and system initial state parameters; S3. Based on the decision training data, an adaptive multi-view fusion fine-tuning method is used to fine-tune the pre-configured large model to obtain a fine-tuned large model; S4. Based on the fine-tuned large model, a distributed decision-making system is built to obtain the resource allocation strategy; The distributed decision-making system includes a dynamic resource scheduling module and a load prediction module; S5. Based on the resource allocation strategy, an adaptive multi-dimensional recursive partitioning method is used to build a decision tree model, and a visual decision path is generated using the decision tree model; Step S3 is further as follows: S31, dividing the decision training data into a training set and a validation set, and performing multi-view representation learning based on the training set to obtain a fusion representation; S32. Based on the fusion representation, the contrast loss and the decision task loss are calculated; Based on the contrast loss and the decision task loss, update the parameters of the pre-configured large model to obtain the fine-tuned large model; S33, repeating step S31 to step S32 until the preset number of iterations is reached, and outputting the final fine-tuned large model; using the validation set, evaluating the performance of the final fine-tuned large model; S34. Use knowledge distillation technology to compress the final fine-tuned large model to obtain a compressed model; quantize the compressed model, convert the floating-point weights of the compressed model into fixed-point representation, and output the optimized model; In step S31, multi-view representation learning is performed based on the training set to obtain a fusion representation, which is further: S311, based on the training set, generating a predetermined number of different representations through a preconfigured large model; S312, calculating the consistency matrix between perspectives based on different representations; Based on the consistency matrix, the spectral clustering algorithm is used to divide the K perspectives into L clusters; where K and L are both natural numbers greater than 0, and K ≥ L; S313, calculating the weighted average of the intra-cluster representation and the importance score between clusters, and calculating the multi-view fusion representation based on the weighted average and the importance score; In step S32, based on the contrast loss and the decision task loss, the parameters of the pre-configured large model are updated to obtain a fine-tuned large model, which is further: S321, calculating the total loss based on the contrast loss and the decision task loss; S322, calculating a gradient based on the total loss and the parameters of the preconfigured large model; Based on the preset threshold, the gradient is clipped to obtain the clipped gradient; S323. Based on the clipped gradients, update the parameters of the preconfigured large model to obtain a fine-tuned large model.

2. The distributed decision-making method based on a large model according to claim 1, characterized in that: Step S1 is further as follows: S11, obtaining the current system time and setting it as the initial time; based on the initial time, calculating the Julian century number; S12, receiving the latitude, longitude and altitude data of the target position input by the user, and storing it in the system memory to form a target position array; S13, obtaining the maximum system running time defined by the user, adding it to the initial time, and calculating the system end time; S14, reading the decision data generation quantity defined by the user and storing it in the system variable; S15, extracting the values ​​of the semi-major axis and the semi-minor axis of the earth from a predefined earth parameter database, and calculating the first eccentricity and the oblateness based on the values ​​of the semi-major axis and the semi-minor axis; S16. Based on the initial time, the Julian century number, the target position array, the system end time, the number of decision data generated, the first eccentricity and the flattening, a pre-processed system parameter is formed.

3. The distributed decision-making method based on a large model according to claim 2, characterized in that: Step S2 is further as follows: S21, based on the preprocessed system parameters, an adaptive chaotic sampling method is used to generate the system initial state parameters; S22, based on the system initial state parameters, using an adaptive multi-step prediction-correction method to perform orbit or flight path integration, and output an orbit or flight path set; S23, based on the set of tracks or flight paths, using an ellipsoidal projection grid algorithm to calculate a distance matrix; S24, based on the distance matrix, using the hierarchical spatiotemporal pattern extraction method to perform correlation analysis and obtain the spatiotemporal correlation matrix; Based on the spatiotemporal correlation matrix, the improved maximum likelihood estimation method is used to calculate the capture success probability matrix of each target point at different time points. S25, generating decision training data based on the capture success probability matrix and the system initial state parameters; using a compressed sensing coding method to encode data points in the decision training data into a sparse representation; S26. Repeat steps S21 to S25 until a predetermined amount of decision training data is generated.

4. The distributed decision-making method based on a large model according to claim 3 is characterized in that: Step S4 is further as follows: S41. Based on the fine-tuned large model, a distributed decision-making system is built to obtain historical load data, current task sets, and system resource capacity; S42. Based on historical load data, multi-dimensional time series analysis is performed to obtain multi-scale features; S43, constructing a system topology map based on multi-scale features and system resource capacity; Based on the system topology graph, the graph attention network is used for message passing to obtain the final node representation; S44, using the long short-term memory network to calculate the predicted load represented by each final node, and calculating the predicted confidence interval of the predicted load; S45. Based on the prediction confidence interval, a resource elasticity index is constructed; based on the current task set, a task feature vector is extracted, and the task feature vector is reduced in dimension using an autoencoder to obtain a reduced-dimensional task feature; S46. Based on the final node representation, predicted load, resource elasticity index and dimension reduction task characteristics, an edge weight function is constructed. Based on the edge weight function, dynamic resource orchestration is performed to obtain a resource allocation strategy.

5. The distributed decision-making method based on a large model according to claim 4 is characterized in that: Step S5 is further as follows: S51, based on the resource allocation strategy, generating a decision data set, and performing standardization processing on the decision data set to obtain a standardized decision data set; S52, based on the standardized decision data set, generate the current node data set; based on the current node data set, perform adaptive feature selection to obtain segmentation features; based on Segmentation features and construct multi-dimensional segmentation strategies; S53, based on the multi-dimensional segmentation strategy, performing multi-dimensional segmentation on the current node data set to generate child nodes; based on the child nodes, constructing a decision tree model; S54. Use the decision tree model to generate a visual decision path.

6. The distributed decision-making method based on a large model according to claim 5 is characterized in that: Step S25 is further as follows: S251, merging the capture success probability matrix and the system initial state parameters to generate a decision training data matrix; S252, constructing a measurement matrix, and calculating a measurement vector based on the measurement matrix and the decision training data matrix; S253. Based on the measurement vector, use the orthogonal matching pursuit algorithm to calculate and obtain a sparse representation of the decision training data.

7. A distributed decision-making system based on a large model, characterized in that: include: at least one processor; as well as, a memory communicatively connected to at least one of the processors; wherein, The memory stores instructions that can be executed by the processor, and the instructions are used to be executed by the processor to implement the distributed decision-making method based on a large model as described in any one of claims 1 to 6.

Citation Information

Patent Citations

  • Telemetry data driven unmanned plane flight state identification method

    CN106197424A

  • Satellite polarization PM2.5 estimation method and system based on machine learning fusion model

    CN115356249A

  • Dynamic neural distribution function machine learning architecture

    WO2024145079A1