Spacecraft attitude control parameter optimization method considering dynamic partitioning and sorting model assistance
Through dynamic partitioning and RankNet proxy model assisted optimization, combined with PSO and fuzzy inference system, the uncertainty and local optimal solution problems in the design of spacecraft attitude control parameters are solved, and efficient and reliable attitude control parameters are achieved.
Patent Information
- Application Number
- CN202410968808.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-18
- Publication Date
- 2024-12-03
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
There is a high degree of uncertainty in the design of spacecraft attitude control parameters. It is difficult for existing intelligent optimization algorithms to effectively avoid the high unreliability of local optimal solutions and proxy models, and there is a lot of waste of computing resources, which affects optimization efficiency.
Using dynamic partitioning and sorting models to optimize the attitude control parameters, partition shrinking through recent better clustering and DBSCAN, combining the RankNet proxy model and PSO algorithm and fuzzy inference system, dynamically allocate search particles, improve computing efficiency and model reliability .
It significantly improves the efficiency and quality of the generation of pose control parameters, avoids local optimal solutions, reduces waste of computing resources, and improves the credibility and optimization speed of the model.
Smart Images

Figure CN119066944A_ABST
Abstract
Description
Technical field:
[0001] The present invention relates to the field of attitude control parameter design of spacecraft, and in particular to an intelligent optimization design of attitude control parameters of a large-scale controlled object. The attitude control of a spacecraft refers to the process of ensuring that it reaches a predetermined orbit or performs a specific task by adjusting its own attitude during flight. The parameters of the attitude control system determine whether the flight indicators of the spacecraft during flight can meet the design requirements, and whether it can operate stably and accurately enter the predetermined orbit after launch. Background technology:
[0002] For the large-scale controlled object model in the spacecraft control system, its attitude control parameters have a high degree of uncertainty. First, there is uncertainty in the number of each typical link that affects the basic structure of the correction network. Secondly, in each link, parameters such as time constant, natural frequency and damping ratio are also uncertain. These parameters determine the specific performance of each added link. Only by setting these parameters correctly can we ensure that the attitude control parameters meet the requirements of system design. The existing manual setting of attitude control parameters is difficult to achieve the optimal performance of the attitude control system, and it takes a lot of time and manpower costs. On the one hand, due to the large number of parameters, the solution space generated by the combination of different values of different parameters is extremely large. The range of solutions that can be covered by manual parameter setting is very limited relative to the entire solution space. On the other hand, manual setting of attitude control parameters requires a lot of practical design experience as a basis. For parameter setting personnel with insufficient design experience, repeated manual debugging is required to make the attitude control system performance meet the design requirements.
[0003] Based on the multi-constraint, high-dimensional and strong nonlinear characteristics of spacecraft attitude control parameter generation. Some research works use intelligent algorithms to find the optimal solution that meets the control system design requirements in the parameter solution space. The existing intelligent optimization algorithms for parameter generation can be divided into three categories: evolutionary algorithms, swarm intelligence algorithms and hybrid algorithms. Evolutionary algorithms are optimization methods that simulate the biological evolution mechanism in nature, such as genetic algorithms, differential evolution algorithms and immune algorithms. Swarm intelligence algorithms are optimization methods that simulate the group behavior of social animals, such as ant colony algorithms, particle swarm algorithms and fish school algorithms. Hybrid algorithms combine multiple optimization methods to improve search efficiency and solution capabilities, such as genetic particle swarm algorithms and simulated annealing particle swarm algorithms. When solving the attitude control parameter design of large-scale controlled object models in spacecraft, in addition to the multi-peak characteristics of the optimization problem itself, it is also necessary to consider the computational complexity of the stability margin index of the large-scale controlled object model in the optimization process, that is, the time-consuming fitness evaluation in the intelligent optimization algorithm, also known as expensiveness. Time-consuming fitness evaluation not only affects the algorithm's full exploration of the solution space, but also reduces the optimization efficiency and affects the practicality of the algorithm. In recent years, there are three main solutions to solve the problems of expensive evaluation and complex multi-peak target space: one is to use parallel computing, such as multi-core processors or multiple graphics cards for simultaneous operation; the second is to approximate the evaluation problem and use simplified functions or models instead; the third is to improve the heuristic algorithm framework, optimization operators and hybrid parameter algorithms. In terms of parallel computing, the core is to use multi-processor acceleration of specific modules of the algorithm. Krityakierne et al. shortened the running time by using 8 processors to evaluate expensive objective functions in parallel. Ong et al. proposed a parallel evolutionary optimization algorithm, which judged the trust region and selected accurate calculation or surrogate model assistance to quickly obtain local optimal solutions under limited computing resources. Akinsolu et al. introduced a parallel surrogate model-assisted evolutionary algorithm in electromagnetic design, which achieved parallel operation by selecting mutation operators in a parallel environment and combining electromagnetic simulation experiments. In terms of approximate true evaluation, surrogate models are used to replace the true fitness function to reduce the complexity of the fitness evaluation process. Common ways to build surrogate models include Kriging, SVM, RBF, PR and neural networks. Cai Xiwen et al. proposed a differential evolution algorithm assisted by an agent model, in which each individual in the population manages a non-overlapping sub-region, builds a local agent model in the sub-region and performs iterative differential evolution, and finally makes a true evaluation of the individual. Liu et al. proposed an evolutionary optimization algorithm based on the assistance of the Kriging model, which improves the accuracy of the Kriging model through dimensionality reduction technology. In terms of optimization algorithms, the improvement strategies mainly include the use of multiple populations, improved operators, and the use of hybrid algorithms. For expensive single-objective high-dimensional optimization problems, most scholars tend to use classical meta-heuristic algorithms assisted by agent models, such as differential evolution algorithms, particle swarm algorithms, teaching and learning optimization algorithms, etc.Sun Chaoli et al. proposed a proxy model-assisted collaborative particle swarm algorithm, which uses the idea of collaborative evolution to alternately share the location information and optimal solution information of individuals in the population through two different types of particle swarm algorithms. Subsequently, Wang Xinjing et al. proposed a new evolutionary sampling-assisted optimization algorithm, which alternately performs a proxy model-assisted local difference algorithm and a proxy model-assisted global search.
[0004] At present, there are still some challenges to be solved for the optimization of spacecraft attitude control parameters, which is expensive to evaluate. The first is how to avoid falling into the local optimal solution. In the optimization of spacecraft attitude control parameters, the target space has complex multi-peak characteristics, and the optimization algorithm can easily fall into the local optimal solution. The second challenge is the selection of the proxy model and the definition of the model input and output. In the existing proxy model-assisted algorithms, the input of the model is usually a decision vector in the solution space, and the output is usually the value of the target space or the label of the solution in the solution space. That is, the proxy model can be divided into two categories: regression and classification models. The purpose of the regression model is to make a regression prediction of the fitness of the target solution, and the purpose of the classification model is to distinguish particles with better prospects. Such model construction methods all face the dilemma of few training samples. In the complex target space, the credibility of the model is low. In addition, some researchers have adopted the method of building a model pool to integrate multiple proxy models and select the most suitable proxy model according to the accuracy performance of the model during the optimization process. However, the cost of building and maintaining the model of this method is too high, which reduces the execution efficiency of the algorithm. Summary of the invention:
[0005] In view of the strong nonlinearity, multiple extreme values and time-consuming calculation of the spacecraft attitude control parameter generation problem and the shortcomings of existing optimization technologies, this paper proposes an attitude control parameter optimization method that considers dynamic partitioning and sorting model assistance. The solution space and the target space of the problem are considered, the solution space is partitioned, and a more efficient proxy model is used in the partition to assist optimization. In the task of generating attitude control parameters for real spacecraft, the parameter generation efficiency and parameter generation quality of this invention are better than other algorithms. The main contributions are as follows:
[0006] (1) A dynamic partitioning method based on nearest better clustering is proposed. In the partitioning process, both the knowledge space and the target space are considered, and DBSCAN is used to shrink each partition, making the boundaries between different partitions clearer.
[0007] (2) A proxy-assisted model based on RankNet is proposed. Compared with other proxy-assisted models, it can make full use of sample data, overcome the high unreliability of the model caused by the small number of real data samples, and has a better convergence speed.
[0008] (3) The PSO algorithm and fuzzy inference system are used to co-evolve between different partitions, and the search particles in the partitions are dynamically allocated to avoid wasting computing resources. Description of the drawings:
[0009] Figure 1 It is the overall flow chart of the present invention, including the overall optimization process assisted by dynamic partitioning and proxy model.
[0010] Figure 2 The overall process of auxiliary optimization based on the RankNet proxy model within the partition.
[0011] Figure 3 Fuzzy sets and fuzzy rules defined in fuzzy reasoning systems are used to regulate the co-evolution between partitions. Specific implementation method:
[0012] The overall structure of the present invention is shown in the attached Figure 1 As shown. Before executing the optimization process, the fitness function is designed according to the margin to be optimized of the controlled object. The large-scale controlled object is divided into multiple working points, each of which contains multiple controlled objects. The robustness index of a single working point is defined as j in formula (1)
[0013]
[0014] Where: V k,i represents the i-th index value of the k-th controlled object in a single working point, N is the number of controlled objects in a single working point, and I i (V k,i ) represents the mapping function between margin index and fitness, M represents the number of indicators to be optimized, P i is the mean value of the ith indicator at a single working point, w i YesP i The weighted weight of the indicator, j is the weighted value of the M indicators at a single working point. Mapping function I i According to the optimization objective defined as formula (2), it is ensured that the piecewise function reaches the maximum value 1 at the ideal index value.
[0015]
[0016] In summary, the optimization objective function of the attitude control parameter generation problem is defined as formula (3)
[0017]
[0018] Where: x is the attitude control parameter to be optimized, H is the number of working points, and bounds is the optimization boundary of the attitude control parameter.
[0019] During the execution of the optimization algorithm, N samples are uniformly sampled in the solution space of the attitude control parameters by using the Latin hypercube sampling method, and the fitness is accurately evaluated according to formula (3). Secondly, the initial graph of all search particles is constructed based on the nearest-better clustering, and different subgraphs are divided into different subgraphs considering the information of nodes and edges in the graph, and the partitions are shrunk by removing edge particles using DBSCAN. Then, the RankNet model is constructed inside each partition, and different partitions are co-evolved. Specifically, the archived particles in the partition are marked as levels 1 to R according to the fitness. Any batch of particles is sampled to form a set of training samples. In a pair of particles, if the ranking rank of the first particle is higher than that of the second particle, the training label of the pair of particles is set to 1. The total training label is a 0, 1 vector of length batch, and the training loss is the binary cross entropy loss of the true label and the predicted label. After the partition and proxy model are constructed, the particles inside the partition are searched with the assistance of the ranking model, and the particles of each partition are allocated according to the fuzzy inference system. Finally, the stagnation threshold of the optimal solution of the partition is used to decide whether to re-update the partition. More specific steps are:
[0020] S1. Set the number of initial sampled solutions to N and the maximum number of fitness evaluations to FE. max ;
[0021] S2. Use Latin hypercube sampling LHS to generate the initial sampling solution in the solution space, evaluate the true fitness of the particles in the initial sampling solution, and save the particles and their fitness values in the archive. The individual with the best fitness value in the initial sampling solution is defined as the optimal solution X gbest ;
[0022] S3, using the nearest better clustering algorithm to add directed edges between particles and record the length information of the edges;
[0023] S4, considering the in-degree information of each particle and the average length of all edges, setting different edge length thresholds for particles with different in-degrees, and deleting edges that do not meet the conditions according to this threshold;
[0024] S5, based on the multiple subgraphs obtained in step S4, the DBSCAN algorithm is used to shrink the partition of each subgraph, wherein the clustering radius and the minimum number of cluster particles of the algorithm are obtained according to the number of nodes and the average number of edge lengths of the subgraph and the weight factors θ1 and θ2;
[0025] S6. Based on the subgraph obtained in step S5, the particles in each partition are divided into levels 1 to R according to fitness, and the higher the level, the more promising the particle.
[0026] S7, the partition uses the ranking data constructed in the archive to train the RankNet model, and uses early stopping to reduce resource consumption;
[0027] S8, calculating the stagnation value of the optimal solution of the current partition, taking the execution probability and stagnation value of the follow operator and the transfer operator as the input of the fuzzy system, and inferring the execution probability of the follow operator and the transfer operator of the next iteration, the initial execution probabilities of the follow operator and the transfer operator are 1 and 0 respectively;
[0028] S9, randomly perform a follow-up operation according to the execution probability of the follow-up operator. Specifically, each particle in the partition updates the velocity vector according to its most recent better particle and the current best particle in the partition, and generates a child;
[0029] S10, performing a transfer operation according to the execution probability of the transfer operator. Specifically, reducing the number of particles in the partition with the worst optimal solution, and increasing the corresponding number of particles in the partition with the best optimal solution;
[0030] S11, using the proxy model obtained in step S7 to perform level prediction on the obtained offspring particles, and selecting particles with high predicted levels of the model for real fitness evaluation; the information of the accurately evaluated particles is saved in the archive;
[0031] S12: Determine whether the maximum number of evaluations has been reached. If so, the optimization process is terminated and output X gbest ;
[0032] S13, if the maximum number of evaluations is not reached, continue to determine whether the number of stagnations of the optimal solution of the partition reaches a threshold. If so, consider all the particles in the partitions comprehensively and re-perform the partitioning operation. Otherwise, continue the search strategy within each partition.
[0033] S14, the maximum number of evaluations to determine whether the fitness is reached is FE max , then output the X in the archive gbest As the global optimal solution, otherwise, repeat S3 to S12;
[0034] The detailed process of adding directed edges in step S3 is as follows: i Take adding a directed edge as an example, For all the i The particle subscript set, namely X i Point to X j , D i,j Defined as particle X i With X j The Euclidean distance between j The choice of is equivalent to taking D i,j The particle corresponding to the minimum value of ;
[0035] In step S4, particle X i The edge length threshold is defined as θ i , calculated as:
[0036]
[0037] Where Indegree(X i ) represents particle X i The in-degree.
[0038] The unsupervised clustering method DBSCAN used in step S5 is a density-based spatial clustering algorithm. The algorithm divides regions with sufficient density into clusters and can find clusters of any shape in noisy data. A "cluster" is defined as the largest set of points connected by density. This method can effectively remove noise points between partitions.
[0039] The RankNet model constructed in step S7 models the item pair (X i , X j ), that is, X i Ranked X j The previous probability:
[0040]
[0041] True probability, i.e. label Defined as:
[0042]
[0043] S i,j The value of particle X i With X j The true rank order is obtained.
[0044] The loss function of the model is defined as:
[0045] By minimizing the loss function, the model can learn how to sort objects correctly. Specifically, the neural network in the model includes three fully connected layers. Each fully connected layer consists of a linear transformation and a nonlinear activation function ReLU. A Dropout regularization layer is added after each hidden layer to reduce the risk of overfitting. The output dimension of the first hidden layer is 256, and the output dimension of the second hidden layer is 128.
[0046] The fuzzy inference system in step S8 adopts the Mamdani type fuzzy inference system. Figure 3The fuzzy sets and inference rules in the fuzzy inference system are shown. Three fuzzy sets and eight inference rules are defined to fine-tune the execution probability of the operator.
[0047] The velocity update vector in step S9 is defined as:
[0048]
[0049] Relative to the velocity update operator of PSO, the velocity vector pointing to the individual optimal particle is updated to the nearest better particle pointing to the current particle.
[0050] In step S10, the operations of increasing and decreasing particles in the partition are as follows: when decreasing particles, particles with low in-degree are ignored first; when increasing particles, a corresponding number of particles with high in-degree are selected to regenerate children according to the speed update formula.
Claims
1. A spacecraft attitude control parameter optimization method considering dynamic partitioning and ranking model assistance, including dynamic solution space partitioning, inter-partition co-evolution and partition search assisted by RankNet model, characterized in that Dynamic solution space partitioning includes the following steps: (1) Initially, Latin hypercube sampling is used to generate an initial solution in the decision space. The individuals in the initial sampling solution are evaluated for their true fitness. The individuals in the initial sampling solution and their fitness values are stored in the archive. At the same time, the individual with the best fitness in the initial sampling solution is defined as the optimal solution X gbest . (2) For each particle, add a directed edge pointing to the nearest better particle. The nearest refers to the closest Euclidean distance of the particle in the solution space, and the better refers to the particle pointed to by the current particle has a better fitness. The process of constructing the directed graph also considers the relationship between the particles in the solution space and the target space. (3) Calculate the in-degree and edge length of all nodes in the graph, set different edge length thresholds for particles with different in-degrees, and delete edges that do not meet the conditions based on this threshold. (4) Based on the multiple subgraphs obtained by deleting edges, the DBSCAN algorithm is used to shrink the partitions of each subgraph. (5) In the subsequent partition search process, if the optimal solution of the partition stagnates and reaches a threshold, the archived data of the partition is merged and the process is repeated from step (2). Inter-partition co-evolution includes the following steps: (6) Calculate the stagnation value of the optimal solution of the current partition, take the execution probabilities of the follow operator and the transfer operator and the stagnation value of the optimal solution as the input of the fuzzy system, and infer the execution probabilities of the follow operator and the transfer operator for the next iteration. The initial execution probabilities of the follow operator and the transfer operator are 1 and 0, respectively. (7) Randomly perform the following operation according to the execution probability of the following operator. Specifically, update the velocity vector in the partition according to the nearest better particle of the particle and the best particle of the partition, and generate offspring. (8) Randomly perform the transfer operation according to the execution probability of the transfer operator. Specifically, reduce the number of particles in the partition with the worst optimal solution, and increase the corresponding number of particles in the partition with the best optimal solution. The partition search based on the RankNet model includes the following steps: (9) The particles in each partition are divided into levels from 1 to R according to their fitness, and the higher the level, the more promising the particle. (10) The partition uses the ranking data constructed in the archive to train the RankNet model, and uses early stopping to reduce resource consumption. (11) Use the trained proxy model to predict the level of the generated offspring particles, and select particles with high model prediction levels for real fitness evaluation; update the accurately evaluated particle positions and fitness to the archive.
2. The method for optimizing spacecraft attitude control parameters with consideration of dynamic partitioning and sorting model assistance according to claim 1 is characterized in that The detailed description of the directed edges added in the dynamic solution space partitioning step (3) is as follows: i Take adding a directed edge as an example, For all the i The particle subscript set, namely X i point to D i,j Defined as particle X i With X j The Euclidean distance between j The choice of is equivalent to taking D i,j The particle that reaches the minimum value, The edge length threshold in step (4) is defined as θ i , calculated as follows: Where Indegree(X i ) represents particle X i The in-degree.
3. The method for optimizing spacecraft attitude control parameters with consideration of dynamic partitioning and sorting model assistance according to claim 1 is characterized in that The fuzzy inference system used in the inter-partition co-evolution is a Mamdani type fuzzy inference system. FIG3 shows the fuzzy sets and inference rules of the fuzzy inference system. The speed update vector of the following operator is defined as: Among them, ω is the inertia factor, c1 is the individual factor, c2 is the group factor, r1 and r2 are random numbers, is the individual better particle at time t, is the optimal particle in the partition at time t; the execution process of the transfer operator is: when reducing particles in the partition, particles with low in-degree are ignored first; when adding particles, a corresponding number of particles with high in-degree are selected to regenerate offspring according to the speed update formula.
4. The method for optimizing spacecraft attitude control parameters with consideration of dynamic partitioning and sorting model assistance according to claim 1 is characterized in that In the partition search step assisted by the RankNet model, the training data is obtained by dividing the fitness of the particles into levels and constructing training samples to model the item pairs (x i , x j ), x i Ranked x j The previous probability is defined as: True probability label Defined as: S i,j The value of particle X i With X j The true rank order is obtained. The loss function of the model is defined as: By minimizing the loss function, the model can learn how to sort objects correctly. Specifically, the neural network in the model includes three fully connected layers. Each fully connected layer consists of a linear transformation and a nonlinear activation function ReLU. A Dropout regularization layer is added after each hidden layer to reduce the risk of overfitting. The output dimension of the first hidden layer is 256, and the output dimension of the second hidden layer is 128.
Citation Information
Patent Citations
Method for optimizing wide-angle attitude control parameters of spacecraft
CN104331083A
Method for determining parameters of PWPF modulator of spacecraft attitude control system
CN112596390A
Integrated modeling evaluation method for attitude control reliability of multi-source uncertain spacecraft
CN114326772A
Pneumatic optimization design method suitable for aircraft three-dimensional shape high-dimensional design variable
CN117634050A
Cited By
Wireless sensor network communication link construction method of redundant nodes
CN119997036A
