Hybrid Response Sequential Experiment Design Method Based on Information Entropy
By using a mixture of Gaussian process models and an information entropy acquisition function, the problem of independently modeling continuous and classified responses in the design of multi-layer space protection structures was solved. This enabled the efficient finding of the optimal combination of protection parameters with a small number of experiments, reducing design costs and the number of simulations.
Patent Information
- Application Number
- CN202611439875.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-16
AI Technical Summary
Existing sequential experimental design methods cannot effectively utilize the coupling information between continuous and classified responses when evaluating multi-layered space protection structures, resulting in excessive simulations and experiments and high costs.
A hybrid response sequential experimental design method based on information entropy is adopted. The continuous response and the classified response are jointly modeled by a mixture Gaussian process model. Information entropy is used as a unified measurement tool to design a sampling function to balance the two types of sampling requirements and optimize the selection of test points.
It significantly reduces the testing cost and number of simulations required for spacecraft protection structure design, improves model prediction accuracy and convergence speed, and enables the finding of the optimal combination of protection parameters under limited testing conditions.
Smart Images

Figure CN122946927B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of structural simulation technology, and in particular to a hybrid response sequential experimental design method based on information entropy. Background Technology
[0002] With the increasing frequency of human space activities, the amount of space debris in the near-Earth orbit environment has increased dramatically, posing a severe threat to spacecraft from micrometeoroids and orbital debris impacts. To ensure the safety of spacecraft during on-orbit operation, multi-layered protective structures (such as Whipple shields and their derivatives) are typically employed. These structures consist of multiple layers of materials, from the outside in, including a buffer shield, a filler layer, and a rear wall. Their protective mechanism is as follows: when a high-speed impactor strikes the front wall, it generates a violent impact effect, causing the impactor and the shield material to break apart, melt, or even vaporize, forming a rapidly spreading debris cloud. Subsequently, the debris cloud further diffuses within the gap space, eventually impacting the rear wall. Through this mechanism, the concentrated high-kinetic-energy impact is dispersed into a large area of low-intensity load, effectively reducing the risk of the rear wall being breached.
[0003] To improve the protective performance of protective structures, the traditional method is to conduct multiple high-speed ground impact tests. However, ground tests are extremely expensive, with a single test costing hundreds of thousands of yuan. To evaluate the performance of protective structures, numerical simulations based on smoothed particle hydrodynamics (PH) or adaptive meshes have become common tools for assessing protective performance. Simulations can output rich physical field data, and their cost is far lower than that of physical tests. However, high-precision ultra-high-speed impact simulations require enormous computational resources; solving complex elastoplastic and phase transition coupling processes in a single run can take hours or even days of high-performance computing time. Therefore, finding the optimal combination of protective parameters with the fewest number of tests is a significant challenge currently facing spacecraft protection design.
[0004] To address computational cost constraints, sequential experimental design employs a dynamic strategy of learning and sampling simultaneously. The basic process involves: constructing an initial surrogate model based on a limited amount of existing initial experimental data; utilizing the uncertainty inherent in the model's current state; and actively selecting the next experimental point with the most information through a specific acquisition function. Simulation data for this point is then acquired and added to the training set to update the model. This process is repeated until a stopping condition is met. This sampling mechanism, utilizing existing sample information, allows sequential design to approach the target with higher efficiency.
[0005] Extensive research has been conducted in the field of sequential design, with various sampling functions proposed for different models and objectives. For continuous responses, different sampling functions exist for different sampling objectives. Classical methods include: if the goal is to achieve globally accurate prediction of the entire input space, it tends to seek unsampled regions with high uncertainty, such as methods based on prediction variance; if the goal is to find the input point that optimizes the quantitative response, the criterion focuses on regions where better function values may be found, requiring a balance between utilization and exploration, such as the Expected Improvement (EI) method. For classification responses, uncertainty sampling strategies tend to explore fuzzy regions where the model's predicted probability is close to 0.5, i.e., conducting experiments near the classification boundary. Commonly used sampling functions include Predictive Entropy or Classification Uncertainty.
[0006] However, in the actual assessment of multi-layered space protection structures, the simulation output of the damage process includes both quantitative (continuous) and qualitative (categorical) responses. Different types of responses often require the construction of different surrogate models for simulation. Continuous responses (such as the remaining mass of the protection structure, the diffusion radius of the debris cloud, or the maximum deformation of the rear wall) follow a smooth variation pattern, and continuous response models are usually established, such as regression models or neural networks. On the other hand, categorical responses (such as the number of layers of the multi-layered protection structure that are damaged or the classification of the damage mode of the protection structure) will have abrupt changes in category at the boundary, and discrete classification models or models that follow discrete probability distributions are usually established, such as support vector machines or logistic regression.
[0007] However, while existing sequential experimental design methods have achieved good results in single-type response problems, when directly applied to mixed-response scenarios of multi-layered protective structures, traditional methods typically establish separate surrogate models (such as Gaussian process regression and support vector machines) for continuous and categorical responses. This independent modeling approach severs the natural physical correlation between the two types of responses (e.g., the implicit correlation between debris cloud morphology and the degree of damage to the rear wall), failing to utilize the coupling information between responses to improve model prediction accuracy. Under conditions of extremely limited experimental samples, this waste of information directly reduces the model's convergence efficiency.
[0008] Sampling strategies for continuous responses typically favor regions in the input space with high prediction uncertainty, which are often distributed across a sparse global space. In contrast, sampling strategies for classification responses focus on decision boundary regions where class abrupt changes occur. These fundamentally different sampling preferences mean that current technologies lack a statistically unified measurement function for mixed response uncertainty that can dynamically balance conflicts between different sampling targets. This makes it difficult for sequential sampling processes to simultaneously achieve the global fitting accuracy required for continuous responses and the boundary recognition needs of classification responses.
[0009] Due to the aforementioned shortcomings in both modeling and sampling, existing methods require a significantly higher number of simulation experiments to achieve the same surrogate model accuracy or optimization goals. In engineering contexts where single-scale simulation costs can reach several hours or even days, and ground-based testing costs are even higher, this inefficient sampling strategy will lead to unacceptable development cycles and expenditures. Summary of the Invention
[0010] Therefore, it is necessary to provide an information entropy-based hybrid response sequential test design method that can efficiently find the optimal combination of protection parameters with the fewest high-speed impact simulations or physical tests, and significantly reduce the test cost of spacecraft protection design, in order to address the above-mentioned technical problems.
[0011] A hybrid response sequential experimental design method based on information entropy, the method comprising: Step 1: Obtain the combination of design parameters to be optimized for the multi-layered protection structure of the spacecraft as the input factor vector, determine the number of initial test samples, and use the space filling design method to generate the initial test point set in the design space; Step 2: For each test point in the initial test point set, obtain the corresponding continuous response vector and classified response vector through ultra-high speed impact simulation or physical experiment, and construct the initial training dataset based on the continuous response vector and classified response vector. Step 3: Construct a Gaussian mixture process model based on the initial training dataset. The Gaussian mixture process model jointly models continuous and categorical responses by introducing latent variables, and trains the model parameters using variational inference methods to obtain the trained Gaussian mixture process model. Step 4: Based on the trained Gaussian mixture process model, calculate the global average prediction entropy of the current model in the design space; Step 5: Determine whether the relative rate of change of the global average prediction entropy is less than the preset stopping threshold. If so, proceed to step 8; otherwise, continue to step 6. Step 6: Based on the posterior prediction distribution of the Gaussian mixture process model, construct an information entropy-based acquisition function, optimize the acquisition function within the design space, and select the experimental point that optimizes the acquisition function value as the new sample point. Step 7: Perform ultra-high-speed impact simulation or physical experiment at the newly added sample point, obtain the corresponding continuous response and classification response, add the newly added sample point and its response data to the training dataset, update the iteration counter, and return to step 3; Step 8: Output the optimized Gaussian mixture process model and the final training dataset to determine the optimal combination of protection parameters for the spacecraft's multi-layer protection structure.
[0012] The aforementioned information entropy-based hybrid response sequential experimental design method, through the construction of a latent variable-based Gaussian mixture process model, successfully integrates the continuous physical quantities and discrete results of the protective structure into a unified multi-output Gaussian process framework. This overcomes the limitations of traditional methods that independently model continuous and classified responses, and can uncover and utilize the inherent statistical correlations between different outputs of the protective structure. The information entropy-based acquisition function, using information entropy as a unified metric, eliminates the dimensional differences between the prediction variance of continuous responses and the boundary uncertainty of classified responses, automatically balancing the two types of sampling requirements and placing the next simulation test point in the region that contributes most to improving the overall model accuracy. Two acquisition function strategies adapted to different computational scenarios are designed, balancing online computational efficiency with the global optimality of sequential experiments. Using this method, the surrogate model can converge to the true protective performance response surface more quickly with a limited number of ultra-high-speed impact simulations, finding the optimal combination of protective parameters with the fewest number of experiments, significantly reducing the experimental cost of spacecraft protection design. Attached Figure Description
[0013] Figure 1 This is a flowchart illustrating a hybrid response sequential experimental design method based on information entropy in one embodiment. Figure 2 This is a schematic diagram of the hybrid response model structure in one embodiment; Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0015] The information entropy-based hybrid response sequential experimental design method provided in this application can be applied to the design and optimization of multi-layered protective structures for spacecraft. In this application scenario, designers need to determine an optimal combination of protective parameters (buffer thickness, protective spacing, infill material type, rear wall thickness, etc.) to maximize the protective structure's resistance to damage when subjected to ultra-high-speed debris impacts. Since single-scale simulations can cost several hours or even days, and ground-based testing is even more expensive, it is necessary to efficiently find the optimal parameter combination with the fewest possible tests.
[0016] In one embodiment, such as Figure 1 As shown, a hybrid response sequential experimental design method based on information entropy is provided, including the following steps: Step 1: Obtain the combination of design parameters to be optimized for the multi-layered protection structure of the spacecraft as the input factor vector, determine the number of initial test samples, and use the space filling design method to generate the initial test point set in the design space.
[0017] In the experimental scenario of spacecraft multi-layer protection structure design, let the input factor vector to be investigated be... This represents the combination of protection parameters to be optimized, where The number of factors.
[0018] Space-filling design methods (such as Latin hypercube sampling) can uniformly cover the entire design space with fewer sample points, avoiding initial sample aggregation and providing a good initial model foundation for subsequent sequential design. This step directly solves the problem of lack of prior information in the initial experimental design, enabling subsequent modeling to obtain preliminary response surface trends on a global scale.
[0019] Step 2: For each test point in the initial test point set, obtain the corresponding continuous response vector and classified response vector through ultra-high speed impact simulation or physical experiment, and construct the initial training dataset based on the continuous response vector and classified response vector.
[0020] Continuous response Physical quantities including the maximum deformation of the rear wall, the radius of debris cloud diffusion, and the remaining mass of the protective structure, are classified into different responses. Discrete results, including the number of layers of the protective structure breached and the type of damage mode, are included. These data constitute the initial training dataset. This step transforms costly simulations or experiments into usable structured data, providing training samples for the surrogate model.
[0021] Step 3: Construct a Gaussian mixture process model based on the initial training dataset. The Gaussian mixture process model jointly models continuous and categorical responses by introducing latent variables, and trains the model parameters using variational inference methods to obtain the trained Gaussian mixture process model.
[0022] The core innovation of this step lies in: using latent variables The process of generating the categorical response is coupled with the process of regressing the continuous response within the same multi-output Gaussian process. For example... Figure 2 As shown, the input factor vector Mapped to continuous response respectively and latent variables Latent variables Generate classification response using the softmax function Simultaneous continuous response With latent variables The cross covariance matrix is used to distinguish them. Establish correlations. This joint modeling approach can uncover physical correlations between responses, such as the intrinsic relationship between the dispersion radius of the debris cloud (continuous) and whether the back wall is breached (classified), thereby improving prediction accuracy when the sample size is limited.
[0023] Specifically, latent variables are introduced for classification responses. ,in Given the number of categories, the latent variables are mapped to the category probability space using the softmax function: ; For including The dataset consists of several test points, with the input of all samples denoted as a matrix. The continuous response vector is The classification response vector is Accordingly, the latent variable components of all samples across all categories are organized into vector form: ; Its dimensions are .
[0024] To characterize the dependency structure between continuous and classified responses, this invention uses the continuous response vector With latent variables Consider as a A joint Gaussian process of dimension is used, and a multi-output Gaussian process regression model is constructed. Given input... Joint response vector Satisfies the following prior distribution: ; in For the continuous response, its own covariance matrix ( ), The covariance matrix of the latent variables themselves ( ), The cross-covariance matrix between the continuous response and the latent variables ( Each covariance matrix is determined by a selected kernel function (radial basis function, Matérn kernel function, etc.) and its hyperparameters. The decision can be made by using a linear common-region model (LMC) for parameterization to ensure positive definiteness.
[0025] Due to the likelihood function of the classification response Having a non-Gaussian form leads to edge likelihood Since analytical calculations are not possible, it is difficult to directly solve for model parameters using maximum likelihood estimation. This invention employs a variational inference method, introducing a Gaussian variational distribution. To approximate latent variables True posterior distribution Furthermore, the model parameters and variational parameters are jointly optimized by maximizing the lower bound of evidence (ELBO).
[0026] Define variational distribution It is a multivariate Gaussian distribution, i.e. , where the mean vector covariance matrix Its lower triangular Cholesky factor Parameterization is performed to ensure the positive definiteness of the covariance matrix.
[0027] The objective function for model training is the evidence lower bound (ELBO) of the log marginal likelihood, and its expression is derived as follows: ; in: First item For continuous response The marginal log-likelihood can be obtained from the properties of multi-output Gaussian processes: ; Second Expectation Item Because the softmax function is nonlinear and cannot be directly calculated analytically, this invention employs the Monte Carlo reparameterization method for estimation. Specifically, it extracts parameters from a standard Gaussian distribution. Independent and identically distributed samples Construct latent variable samples Then the expected term is approximately: ; in This represents the product of the softmax functions over all samples, i.e.: ; Third item Variational distribution with conditional priors The KL divergence between them. Based on the posterior formula of a multi-output Gaussian process, the conditional distribution... Also following a Gaussian distribution, its mean is The covariance matrix is The KL divergence between two Gaussian distributions has an analytical expression: ; Combining the above three terms, and merging the constant term and the logarithmic determinant term, we obtain the final lower bound of the evidence optimization objective function. for: ; The model training process involves applying gradient-based optimization algorithms (such as the Adam optimizer) to the aforementioned objective function. Regarding variational parameters and kernel function hyperparameters Maximize. Due to the introduction of the reparameterization technique, the gradients of all expectation terms can be computed efficiently.
[0028] The specific training process can be summarized as Algorithm 1: The steps of Algorithm 1 include: Input: Sample Quantitative response Qualitative response Number of categories Sample size ; 1. Initialize the kernel function hyperparameters of the multi-output Gaussian process model. ; 2. Calculate the prior covariance matrix and conditional covariance ; 3. Initialize variational distribution parameters and ; 4. Draw random samples from a standard Gaussian distribution. ; 5. Calculate the lower bound of evidence ; 6. Iteratively update parameters using the Adam optimizer. Continue until the convergence condition is met; Output: Gaussian process model, variational distribution
[0029] Step 4: Based on the trained Gaussian mixture process model, calculate the global average prediction entropy of the current model in the design space.
[0030] Global average prediction entropy is used to quantify the uncertainty of the current model over the entire design space. A higher entropy value indicates greater uncertainty in the model's predictions for that region. By calculating the average entropy, it can be determined whether the model has been sufficiently learned. The specific steps are as follows: Generated using Latin hypercube sampling within the design space test points For each test point Calculate joint information entropy Then calculate the average value. Joint information entropy is defined as a continuous response. With classification response At the input point Joint entropy at: ; Among them, continuous response edge entropy It has an analytical form; classification conditional entropy
[0031] Approximation using Monte Carlo integration.
[0032] This entropy value can uniformly measure continuous uncertainty and classification uncertainty, providing a foundation for the design of subsequent acquisition functions.
[0033] Step 5: Determine whether the relative rate of change of the global average prediction entropy is less than the preset stopping threshold. If so, proceed to step 8; otherwise, continue to step 6.
[0034] The relative rate of change of the global average prediction entropy is calculated as follows: ; This reflects the convergence of model uncertainty with iteration. When When the sampling rate reaches a certain level, it indicates that adding new samples can no longer significantly reduce the uncertainty of the model, the model has become stable, the sequential design process has converged, and stopping sampling at this point can avoid invalid experiments, thereby saving computational resources.
[0035] Step 6: Based on the posterior prediction distribution of the Gaussian mixture process model, construct an information entropy-based acquisition function, optimize the acquisition function within the design space, and select the experimental point that optimizes the acquisition function value as the new sample point.
[0036] For a new set of inputs It needs to be based on observed data. The trained model provides a continuous response vector. Predicted values and classification responses The class probability.
[0037] Based on the conditional distribution properties of multi-output Gaussian processes, under known... Under the condition, the joint response corresponding to the new sample Satisfies a posterior Gaussian distribution: ; Due to latent variables during the training phase The true value is unknown and requires the use of variational posterior. Approximate it. Substitute the variational distribution into the above conditional distribution and... Integrating, we obtain the predicted approximate distribution. It also follows a Gaussian distribution. The mean and covariance matrices are calculated as follows: ; ; Based on the above joint distribution, the continuous response vectors can be obtained respectively. With latent variables The marginal prediction distribution.
[0038] Continuous response prediction: Take the predicted mean vector The middle corresponds to The component is used as a point estimate of the continuous response, i.e. The prediction variance can be derived from the covariance matrix. The corresponding sub-block is given to measure the uncertainty of the prediction.
[0039] Classification response prediction: Take the predicted mean vector The middle corresponds to The amount As the mean of the distribution, the prediction variance can be derived from the covariance matrix. Corresponding sub-block Given the latent variable distribution, we obtain the distribution of the latent variables. To obtain the classification of new samples. The probability of belonging to each category, and the distribution of the predicted latent variables. Apply the softmax function and calculate the expectation: ; This integral is also approximated using the Monte Carlo method: from Extraction For each sample, calculate the softmax probability and average it, i.e.: ; Finally, the category with the highest posterior probability is selected as the predicted value for the classification response: ; The above prediction steps can be summarized as shown in Algorithm 2.
[0040] The steps of Algorithm 2 include: Input: Sample Quantitative response Qualitative response Gaussian process model (kernel function) Variational distribution parameters and New sample ; 1. Calculate the covariance matrix between the training input and the predicted input. ; 2. Utilize data saved during the training phase Calculate the predicted mean With covariance ; 3. From the Gaussian distribution Mid-sampling and estimation of classification probability via the softmax function ; 4. Select to use The largest ; Output: Continuous response prediction Compared with the predicted value of the classification response .
[0041] The above scheme enables unified modeling of mixed responses containing both continuous and categorical variables, effectively utilizing the correlation between responses to improve prediction accuracy. Simultaneously, variational inference reduces computational complexity, demonstrating good practicality and scalability. This application constructs an information entropy-based acquisition function based on the posterior prediction distribution of the mixed model. Two different criteria are proposed to construct acquisition functions based on information entropy to solve the sequential design problem in mixed response scenarios. Users can choose the appropriate criterion based on computing resources and optimization objectives: Regarding the continuous response involved in this invention With classification response The mixed observation variables, at the input points The joint information entropy at a location is defined as: ; The larger the entropy value, the higher the uncertainty of the joint response at that point. According to the chain rule of information theory, joint entropy can be decomposed into the sum of the continuous partial entropy and the conditional entropy of the classified response given the continuous response: ; Among them, continuous response edge entropy Can be predicted by distribution Calculate directly.
[0042] The two criteria are as follows: Strategy 1: Approximate Maximum Information Entropy Criterion. The core idea of this criterion is to directly select the point with the highest joint information entropy under the current posterior prediction distribution as the next sampling point. The criterion is defined as follows: ; in This criterion is designed for space. It relies only on the predictive statistics of the current model, without simulating future observations, resulting in relatively low computational overhead.
[0043] During the calculation process, for The estimate can be fixed. The value of is the conditional expectation. This replaces the calculation of integrals. ; Substituting the above equation into... The middle can also be fixed The value is To replace points ; Therefore, the final acquisition function is: ; This criterion can quickly identify areas in the input space with high model uncertainty, enabling exploratory coverage of the design space.
[0044] In engineering practice, it is suitable for quickly generating the next test point between simulation tasks, rapidly exploring high uncertainty areas in the input space, and quickly learning the approximate safety boundaries and deformation trends of multi-layered protective structures.
[0045] Strategy Two: Information Entropy Reduction Criterion. This criterion evaluates whether, at candidate points... After conducting experiments and observing the response, the expected decrease in information entropy of the model across the entire design space is equivalent to minimizing the expected posterior entropy. This criterion can be equivalently written as: ; The key to calculating this criterion lies in estimating the expected posterior entropy, and the specific steps are as follows: 1. For candidate points Add it as a new data point to the existing dataset, while maintaining the current model hyperparameters. Without changing the conditional distribution update formula of the Gaussian process, we can quickly calculate the posterior mean and covariance matrix after adding new data points.
[0046] 2. Utilizing Latin hypercube sampling in the design space Generate a set of representative test points. Under the updated model, for each test point Calculate the variational posterior distribution The Monte Carlo reparameterization method is used for estimation, drawing parameters from a standard Gaussian distribution. Independent and identically distributed samples Constructing samples ,as well as under conditions samples ; in Prediction points The variance of the continuous response and the covariance of the continuous response with the latent variable.
[0047] 3. Calculate the posterior entropy of the samples under the new model, i.e., the joint information entropy. ; .
[0048] 4. Average the entropy values of all test points to obtain the global posterior entropy for this virtual observation implementation. Then, the average of all virtual observations is calculated to obtain the final expected posterior entropy estimate: ; After the above steps, the final acquisition function is: ; Although the computational complexity of the information entropy reduction criterion is significantly higher than that of the maximum entropy criterion, it has a global perspective and can effectively avoid the problem of oversampling in local areas.
[0049] Although this criterion involves a large amount of computation per step, it has global foresight, can effectively avoid local oversampling, and ensure long-term optimality. It is suitable for pursuing the globally optimal combination of protection parameters when computational resources allow. Select a data acquisition function according to a preset strategy. (maximum information entropy) Or the information entropy reduction criterion () Employing multi-starting-point gradient descent algorithms or evolutionary strategies in continuous design space To find the global optimal solution.
[0050] Step 7: Perform ultra-high-speed impact simulation or physical experiment at the newly added sample point to obtain the corresponding continuous response and classification response. Add the newly added sample point and its response data to the training dataset, update the iteration counter, and return to step 3.
[0051] At the selected point Experiments were conducted at the location to obtain new continuous responses. With classification response New data points Add to dataset: ; Update iterative counter Return to step 3.
[0052] Each time a new sample is obtained, the model is retrained, gradually approximating the true input-output mapping. Because the model selects the point with the most information each time, its convergence speed is much faster than passive sampling (such as uniform sampling or random sampling).
[0053] Step 8: Output the optimized Gaussian mixture process model and the final training dataset to determine the optimal combination of protection parameters for the spacecraft's multi-layer protection structure.
[0054] The final output model can be directly used to predict the protection performance (continuous response and classification probability) under any combination of input parameters. Designers can use the model to optimize parameters, such as finding the optimal combination of buffer screen thickness and spacing that minimizes the deformation of the rear wall and prevents it from being penetrated, thereby obtaining a high-performance protection structure without the need for a large number of simulation tests.
[0055] The pseudocode for the above steps is shown in Algorithm 3. The steps of Algorithm 3 include: Input: Initial sample size , Monte Carlo sample size Design Space ; 1. Constructing the initial design using space-filling design. ; 2. Conduct experiments to obtain responses. and Constructing a dataset , ; 3. Using Algorithm 1, through the dataset Training a Gaussian mixture model; 4. Generate sampling points using Latin squares ; 5. Calculate the predicted average mixed information entropy ; 6. WHILE DO: 7. Minimize using gradient descent get ; 8. According to The experiment received a response and ; 9. , ; 10. Using Algorithm 1, through the dataset Training a Gaussian mixture model; 11. Generate sampling points using Latin squares ; 12. Calculate the predicted average mixed information entropy. ; Output: Gaussian mixture model.
[0056] This application not only provides a precise method for unified modeling of mixed data containing continuous and categorical responses, but also designs a smart sequential experiment strategy based on information entropy theory. This combined approach can significantly reduce the total number of experiments required to obtain a high-precision surrogate model, and has important application value in the simulation experiments of multi-layered space protection structures.
[0057] In the aforementioned information entropy-based hybrid response sequential experimental design method, this application successfully incorporates continuous physical quantities of the protective structure (maximum deformation of the rear wall, debris cloud diffusion radius, etc.) and discrete results (number of layers of the protective structure penetrated, etc.) into a unified multi-output Gaussian process framework by introducing a latent variable mapping mechanism. This joint modeling approach overcomes the limitations of traditional methods that independently model continuous and classified responses, and can explore and utilize the inherent statistical correlations between different outputs of the protective structure. This mechanism enables the generated surrogate model to reflect the physical changes of the multi-layered protective structure under hypervelocity impact, improving the prediction accuracy of replacing high-cost numerical simulations. Based on the essence of information theory, information entropy is used as a unified metric to eliminate the dimensional differences between the prediction variance of continuous response and the uncertainty of the boundary of classified response. During the sequential experiment, the information entropy-based acquisition function can automatically balance the two types of sampling requirements, and according to the uncertainty of the current surrogate model, the next simulation test point is placed in the region that contributes the most to improving the overall model accuracy. This mechanism ensures that, under a limited number of hypervelocity impact simulations, the surrogate model can converge to the true protective performance response surface more quickly. To address the pain point of long single-step solution time in ultra-high-speed impact simulations, two acquisition function strategies adapted to different computational scenarios were designed. The maximum information entropy criterion based on analytical approximation and gradient descent has extremely low computational overhead, making it suitable for rapidly generating the next test point during simulation intervals. It can quickly learn the approximate safety boundary and deformation trend of the multi-layered protective structure input space. While the global optimization criterion based on expected information entropy reduction has slightly higher single-step computational cost, it possesses forward-looking decision-making capabilities, effectively avoiding the sampling process from getting trapped in local extrema, thus ensuring the long-term optimality of the sequential experimental design and the efficiency of searching for the optimal combination of protective parameters. This hierarchical strategy design allows the invention to meet the requirements of rapid model updates while ensuring approximation of the global optimum under limited computational resources and experimental budgets, making it highly practical in the field of spacecraft multi-layered protective structure optimization design.
[0058] In one embodiment, the input factor vector is a combination of protection parameters of the spacecraft's multi-layer protection structure, including but not limited to one or more of the following: buffer screen thickness, protection spacing, filling layer material type, and rear wall thickness; the continuous response includes one or more of the following: maximum deformation of the rear wall, debris cloud diffusion radius, and remaining mass of the protection structure; and the classified response includes one or more of the following: number of layers of the protection structure that have been penetrated and damage mode category.
[0059] Specifically, these parameters and responses directly correspond to the input and output of ultra-high-speed impact simulations and have clear physical meanings. By unifying the modeling of continuous and discrete quantities of interest in engineering, this method can directly serve practical protective design tasks, such as predicting whether the rear wall will be penetrated and the amount of deformation after penetration, given the thickness and spacing of the buffer screen. This unified modeling capability is something that traditional methods cannot achieve.
[0060] In one embodiment, a Gaussian mixture process model is constructed based on an initial training dataset, including: Introducing latent variables for classification responses ,in Let the number of categories of the latent variable be the number of categories. The softmax function is used to map the latent variable to the category probability space: ; and continuous response vector With latent variables Construct it as a multi-output Gaussian process, given input Joint response vector Satisfying the joint prior distribution is ; in, A continuous response vector Its own covariance matrix has dimensions of , as latent variables Its own covariance matrix has dimensions of , A continuous response vector With latent variables The cross covariance matrix between them has dimensions of Its transpose is denoted as , The number of test sites.
[0061] Specifically, the classification problem is transformed into a regression problem using latent variables, and the correlation between continuous and classification responses is automatically learned through the covariance structure of a multi-output Gaussian process. For example, the larger the dispersion radius of the debris cloud, the higher the probability of the back wall being breached. This implicit association is demonstrated through the cross-covariance matrix. The model automatically captures these information, thereby improving accuracy when predicting continuous responses by leveraging classification information, and vice versa.
[0062] In one embodiment, the model parameters are trained using a variational inference method, including: Introducing the variational distribution in Gaussian form To approximate latent variables The true posterior distribution of , where Let be the variational mean vector. Cholesky factor of the lower triangle Parameterization; by maximizing the lower bound of evidence Joint optimization of variational parameters and kernel function hyperparameters yields the following evidence lower bound: ; Among them, the first item A continuous response vector The marginal log-likelihood, the second term is the log-likelihood of the classification response. The expectation under the variational distribution, the third term is the variational distribution. with conditional priors The KL divergence between them.
[0063] Specifically, it solves the problem of being unable to analytically calculate marginal likelihood under non-Gaussian likelihood conditions. Furthermore, through reparameterization techniques, it enables gradient propagation, allowing the entire model to be optimized end-to-end with manageable computational complexity. This makes the Gaussian mixture process model applicable to real-world engineering scenarios with hundreds of samples.
[0064] In one embodiment, the expectation term in the lower bound of evidence is approximated using a Monte Carlo reparameterization method, including: Extract from a standard Gaussian distribution Independent and identically distributed samples Construct latent variable samples Then the expected term is approximately: ; in, This represents the product of the softmax function over all samples.
[0065] Specifically, by reparameterizing sampling, low-variance gradient estimates can be obtained, ensuring stable convergence of model training.
[0066] In one embodiment, based on the trained Gaussian mixture process model, the global average prediction entropy of the current model in the design space is calculated, including: Generated using Latin hypercube sampling within the design space test points For each test point Calculate joint information entropy Then, the average of the joint information entropy of all test points is calculated as the global average prediction entropy; joint information entropy Defined as continuous response With classification response At the input point Joint entropy at: ; ; ; in, Represents the edge entropy of continuous response. Represents the classification conditional entropy. express The corresponding continuous response, express Corresponding continuous response variance Indicates the predicted distribution. By latent variables The conditional distribution is obtained by substituting it into the softmax function and integrating. This represents the number of categories for the latent variable.
[0067] Specifically, continuous entropy and classification entropy are unified into joint entropy, providing a single numerical indicator for subsequent data collection functions and avoiding the difficulty of parameter tuning caused by manually weighting the two uncertainties.
[0068] In one embodiment, constructing the information entropy-based acquisition function includes: The point with the highest joint information entropy under the current model is selected as the next sampling point using the approximate maximum information entropy criterion. The approximate maximum information entropy criterion is as follows: ; in, For design space.
[0069] Specifically, the criterion is simple to calculate, does not require simulating future data, is suitable for rapid online decision-making, and can quickly cover the entire design space in the early stages of sequential design, preventing the model from falling into local biases.
[0070] In one embodiment, an information entropy-based acquisition function is constructed, comprising: The information entropy reduction criterion is used to evaluate the candidate points. The experiment was conducted and the response was observed. Then, generate within the design space. For each test point, calculate the sample posterior entropy; The average of the posterior entropy of all test points is used as the global posterior entropy under the virtual observation implementation. Then, the average of all virtual observation implementations is used to obtain the acquisition function.
[0071] Specifically, this criterion has a global optimization perspective, which can avoid local oversampling caused by greedy strategies. For example, in the design of protective structures, although certain parameter regions have high current uncertainty, even sampling cannot significantly improve the global model. The information entropy reduction criterion will automatically avoid such regions and invest the limited experimental budget in parameter combinations that can truly reduce global uncertainty.
[0072] In one embodiment, calculating the sample posterior entropy for each test point includes: Calculate test points The sample posterior entropy is ; in, For test points The prediction variance of continuous responses, For Monte Carlo sample size, The number of categories for the latent variable. For the first The softmax probability of each category. For the test point The latent variable samples generated at the location.
[0073] Specifically, the above process of calculating the posterior entropy of the sample achieves an unbiased estimate of the expected posterior entropy through double Monte Carlo sampling of the inner latent variable and the outer response value. Although the computational cost is slightly higher, the accuracy is high and it is suitable for key decision points.
[0074] In one embodiment, the relative rate of change of the global average prediction entropy is calculated as follows: ; in, For the first The global average prediction entropy of the next iteration. For the first The global average prediction entropy of the next iteration, when Less than the preset stop threshold At that time, the sequential design process converges.
[0075] Specifically, there is no need to preset a fixed number of iterations; the model automatically stops based on its convergence, avoiding unnecessary simulation experiments and directly reducing experimental costs.
[0076] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0077] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a hybrid response sequential experimental design method based on information entropy. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.
[0078] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0079] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (RAM), dynamic RAM (DRAM), synchronous DRAM (DRAM), dual data rate DRAM (DDRDRAM), enhanced DRAM (EDRAM), synchronous link DRAM (LDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.
[0080] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0081] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A hybrid response sequential experimental design method based on information entropy, characterized in that, The method includes: Step 1: Obtain the combination of design parameters to be optimized for the multi-layered protection structure of the spacecraft as the input factor vector, determine the number of initial test samples, and use the space filling design method to generate the initial test point set in the design space; Step 2: For each test point in the initial test point set, obtain the corresponding continuous response vector and classified response vector through ultra-high speed impact simulation or physical experiment, and construct the initial training dataset based on the continuous response vector and classified response vector; Step 3: Construct a Gaussian mixture process model based on the initial training dataset. The Gaussian mixture process model jointly models continuous and categorical responses by introducing latent variables, and trains the model parameters using variational inference methods to obtain the trained Gaussian mixture process model. Step 4: Based on the trained Gaussian mixture process model, calculate the global average prediction entropy of the current model in the design space; Step 5: Determine whether the relative rate of change of the global average prediction entropy is less than the preset stopping threshold. If so, proceed to step 8; otherwise, continue to step 6. Step 6: Based on the posterior prediction distribution of the Gaussian mixture process model, construct an information entropy-based acquisition function, optimize the acquisition function within the design space, and select the test point that makes the acquisition function value optimal as the new sample point. Step 7: Perform ultra-high-speed impact simulation or physical experiment at the newly added sample point, obtain the corresponding continuous response and classification response, add the newly added sample point and its response data to the training dataset, update the iteration counter, and return to step 3; Step 8: Output the optimized Gaussian mixture process model and the final training dataset to determine the optimal combination of protection parameters for the spacecraft's multi-layer protection structure; The training of model parameters using variational inference methods includes: Introducing the variational distribution in Gaussian form To approximate latent variables The true posterior distribution, where Let be the variational mean vector. Cholesky factor of the lower triangle Parameterization; by maximizing the lower bound of evidence The lower bound of the evidence is expressed as follows: (This is a joint optimization of the variational parameters and kernel function hyperparameters.) Among them, the first item A continuous response vector The marginal log-likelihood, the second term is the log-likelihood of the classification response. The expectation under the variational distribution, the third term is the variational distribution. with conditional priors KL divergence between them; The expectation term in the lower bound of evidence is approximated using the Monte Carlo reparameterization method, including: Extract from a standard Gaussian distribution Independent and identically distributed samples Construct latent variable samples Then the expected term is approximately: in, This represents the product of the softmax function over all samples.
2. The method according to claim 1, characterized in that, The input factor vector is a combination of protection parameters of the spacecraft's multi-layer protection structure, including but not limited to one or more of the following: buffer screen thickness, protection spacing, filling layer material type, and rear wall thickness; the continuous response includes one or more of the following: maximum deformation of the rear wall, debris cloud diffusion radius, and remaining mass of the protection structure; the classification response includes one or more of the following: number of layers of the protection structure that have been penetrated and damage mode category.
3. The method according to claim 2, characterized in that, A Gaussian mixture process model is constructed based on the initial training dataset, including: Introducing latent variables for classification responses ,in Let the number of categories of the latent variable be the number of categories. The softmax function is used to map the latent variable to the category probability space: ; and continuous response vector With latent variables Construct it as a multi-output Gaussian process, given input Joint response vector Satisfying the joint prior distribution is in, A continuous response vector Its own covariance matrix has dimensions of , as latent variables Its own covariance matrix has dimensions of , A continuous response vector With latent variables The cross covariance matrix between them has dimensions of Its transpose is denoted as , The number of test sites.
4. The method according to claim 1, characterized in that, Based on the trained Gaussian mixture process model, the global average prediction entropy of the current model in the design space is calculated, including: Generated using Latin hypercube sampling within the design space test points For each test point Calculate joint information entropy Then, the average of the joint information entropy of all test points is calculated as the global average prediction entropy; the joint information entropy Defined as continuous response With classification response At the input point Joint entropy at: in, Represents the edge entropy of continuous response. Represents the classification conditional entropy. express Continuous response, express Continuous response variance Indicates the predicted distribution. By latent variables The conditional distribution is obtained by substituting it into the softmax function and integrating. This represents the number of categories for the latent variable.
5. The method according to claim 4, characterized in that, Constructing an information entropy-based acquisition function includes: The point with the highest joint information entropy under the current model is selected as the next sampling point using the approximate maximum information entropy criterion. The approximate maximum information entropy criterion is as follows: in, For design space.
6. The method according to claim 4, characterized in that, Constructing an information entropy-based acquisition function includes: The information entropy reduction criterion is used to evaluate the candidate points. The experiment was conducted and the response was observed. Then, generate within the design space. For each test point, calculate the sample posterior entropy; The average of the posterior entropy of all test points is used as the global posterior entropy under the virtual observation implementation. Then, the average of all virtual observation implementations is used to obtain the acquisition function.
7. The method according to claim 6, characterized in that, Calculate the sample posterior entropy for each test point, including: Calculate test points The sample posterior entropy is in, For test points The prediction variance of continuous responses, For Monte Carlo sample size, The number of categories for the latent variable. For the first The softmax probability of each category. For the test point The latent variable samples generated at the location.
8. The method according to claim 1, characterized in that, The calculation process for the relative change rate of the global average prediction entropy is as follows: in, For the first The global average prediction entropy of the next iteration. For the first The global average prediction entropy of the next iteration, when Less than the preset stop threshold At that time, the sequential design process converges.
Citation Information
Patent Citations
Sequential test design method based on maximum entropy search criterion
CN118070480A
Sandy soil liquefaction discrimination method based on CPTU and weighted nonlinear similarity
CN120974283A
