Swarm intelligence optimization algorithm and system for optimizing aerodynamic parameters of aircraft

Abstract

The invention discloses a swarm intelligence optimization algorithm and system for optimizing aerodynamic parameters of an aircraft, and the algorithm comprises the steps: S1, generating an initial population in a continuous space based on the upper and lower bounds of the aerodynamic parameters of the aircraft; s2, updating the current speed of each individual, and calculating the total disturbance of each individual by adopting a gravity random walk strategy; s3, updating the current position by adopting the sum of the total disturbance of the individual, the current speed and the position of the previous moment; s4, performing fine adjustment on the large model, predicting the fitness function value of each individual after position updating by adopting the large model after fine adjustment, and selecting the optimal fitness function value in all the individuals; s5, whether the optimal fitness function value is larger than or equal to a preset threshold value or whether the number of iterations is larger than or equal to a preset number is judged, if yes, the step S6 is executed, and if not, the step S2 is executed again; and S6, taking an individual corresponding to the optimal fitness function value as an optimized aerodynamic parameter of the aircraft.

Description

Technical Field

[0001] This invention relates to a method for optimizing aircraft parameters, specifically to a swarm intelligence optimization algorithm and system for optimizing aircraft aerodynamic parameters. Background Technology

[0002] In aircraft design, aerodynamic design is a fundamental and crucial element, directly impacting the overall flight performance and characteristics of an aircraft. The design of aerodynamic characteristic parameters is a key aspect of aircraft design. In recent years, with the continuous maturation of computational fluid dynamics (CFD) simulation technology and the widespread application of large-scale parallel computing platforms, some progress has been made in research on optimizing aircraft aerodynamic parameters. However, traditional gradient methods and Lagrange multiplier methods rely on the differentiability of the flow field, making them prone to getting trapped in local optima when dealing with multi-peak, non-convex design spaces. They are also sensitive to hyperparameters, and their implementation heavily depends on expert experience. Furthermore, high-precision simulations often consume enormous computational resources and require lengthy iterations, further limiting optimization efficiency and effectiveness.

[0003] To improve the ability to search for global optima, the academic community has introduced various swarm intelligence optimization algorithms. For example, solutions based on genetic algorithms enhance population diversity through selection, crossover, and mutation operations, enabling the rapid discovery of multiple feasible regions. Particle swarm optimization methods explore the design space by simulating the cooperative behavior of "bird flocks," but it is difficult to balance convergence speed and accuracy, and problems such as premature convergence and sensitivity to the initial population still exist.

[0004] Subsequently, to reduce overall computational overhead, some studies introduced surrogate models or response surface methodology, combining high-fidelity CFD simulation with machine learning techniques such as radial basis function networks. This approach obtains an approximate response surface through training with a limited number of samples, and then performs multiple rounds of iterative optimization based on this surface. While these methods alleviate the high simulation costs to some extent, the fitting errors inherent in the surrogate models themselves often make it difficult to guarantee high accuracy in the optimization results, and they are prone to distortion in regions with sparse sample distribution.

[0005] In summary, although existing technologies employ swarm intelligence algorithms and surrogate models to varying degrees to improve global search capabilities and computational efficiency, they either struggle to balance convergence speed and solution accuracy, or rely too heavily on initial parameters and human experience. Consequently, they cannot provide a unified solution that is both efficient and accurate for optimizing aircraft aerodynamic parameters. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, the swarm intelligence optimization algorithm and system for optimizing aerodynamic parameters of aircraft provided by this invention solves the problem that existing methods struggle to balance convergence speed and solution accuracy.

[0007] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows: Firstly, a swarm intelligence optimization algorithm for optimizing aerodynamic parameters of an aircraft is provided, comprising the following steps: S1. Based on the upper and lower bounds of the aerodynamic parameters of the aircraft, an initial population is generated in a continuous space. Each individual in the initial population represents a set of aerodynamic parameters of the aircraft. S2. The current velocity of each individual is updated using a hiking optimization algorithm, a historical best guidance strategy, and a multi-group collaborative search mechanism, and the total perturbation of each individual is calculated using a gravity random walk strategy. S3. Update the current position of an individual by summing its total perturbation, current velocity, and position at the previous moment. S4. Fine-tune the large model and use the fine-tuned large model to predict the fitness function value of each individual after the position update, and select the optimal fitness function value among all individuals. S5. Determine whether the optimal fitness function value is greater than or equal to the preset threshold or whether the number of iterations is greater than or equal to the preset number. If yes, proceed to step S6; otherwise, return to step S2. S6. The individual corresponding to the optimal fitness function value is used as the optimized aerodynamic parameters of the aircraft.

[0008] The beneficial effects of this invention are as follows: This scheme enhances local exploitation capabilities through a historical optimal guidance strategy and adopts a multi-population collaborative search mechanism to divide population adaptation into global exploration and local exploitation, achieving "breadth-first search, then fine-grained convergence"; at the same time, it introduces random perturbations based on a gravity model to help individuals escape local optima; finally, it integrates a large model to quickly evaluate aerodynamic performance, reducing dependence on high-fidelity CFD. This scheme balances search efficiency and solution accuracy.

[0009] Furthermore, the method for updating the current velocity of each individual in step S2 includes: S21. The speed of each individual in the population is updated using the hiking optimization algorithm: in, and Let be the velocity of individual i in the t-th and t-1-th iterations, respectively; Let be the random perturbation coefficient of individual i in the t-th iteration; The optimal position of an individual in the population; Let i be the spatial position of individual i in the t-th iteration; S22. The update rate of each individual in the population is adjusted using a historical best-practice guidance strategy: in, To Corrected speed; and Both control the overall and local weights. ; For individuals Historically best position; and All are sweep factors; S23. Divide the population into a global exploration subgroup that searches in a continuous space and a local development subgroup that searches in a high-quality area that has been discovered. S24. For each individual in the two subgroups, a bias coefficient affecting its search behavior is dynamically generated using a skewed beta distribution. and ; S25, Using deviation coefficient and renew and To achieve Update: in, Let be the current velocity of individual i at the t-th iteration.

[0010] The beneficial effects of the above technical solution are as follows: The historical best guidance strategy of this solution records the optimal position of each individual in the evolutionary process (i.e., Personal Best), and guides individuals towards their historical best position in each iteration, thereby improving local exploitation capability and convergence speed. Multiple population collaboration mechanisms can enhance the algorithm's global search capability, maintain population diversity, and effectively balance the relationship between exploration and exploitation, further improving the population's search efficiency and solution accuracy.

[0011] Furthermore, step S24 further includes: S241, Beta distribution parameters of the global exploration subgroup and Adjustments will be made: in, and The initial and final parameters are preset to control the beta distribution pattern of the global exploration subgroup. and T represents the maximum number of iterations. S242, Beta distribution parameters of locally developed subgroups and Adjustments will be made: in, and The initial and final parameters are preset to control the beta distribution pattern of the local development subgroup. and ; S243. Each individual in both subgroups samples a search bias coefficient from its beta distribution: in, and Here are the beta distribution parameters adjusted at time t, when the individual belongs to the global exploration subgroup: , When an individual belongs to a locally developed subgroup: , .

[0012] The beneficial effects of the above technical solution are as follows: This solution introduces a skewed beta distribution to dynamically generate key parameters that affect its search behavior. and By adaptively adjusting these two parameters during the iteration process, the output bias of the random number generator can be controlled, thereby guiding the behavior of the subgroup. In the global exploration subgroup, in the early stage (t→0), the parameter combination causes the beta distribution to favor generating larger or more diverse values, which are subsequently used as multipliers to adjust the individual's step size or direction, thus achieving "stronger jumping ability". In the local development subgroup, in the later stage (t→T), the parameter combination causes the beta distribution to favor generating smaller or more concentrated values, which are used to fine-tune the individual's position, achieving "steady convergence".

[0013] Furthermore, for individuals in the global exploration subgroup, after each iteration, when they meet preset conditions, they are added to the local development subgroup. The preset conditions are: in, and The first The individual in the first Subsequent The fitness function value at the next iteration; The preset number of consecutive iterations; This represents the optimal fitness function value in the global exploration subgroup. For individuals in the local exploration subgroup, after each iteration, if the historical best position of an individual has not been updated for several consecutive generations, it is added to the global exploration subgroup.

[0014] The beneficial effects of the above technical solution are: for individuals in the global exploration subgroup, when they meet the preset conditions... If a region is deemed unlikely to yield significant benefits in its current location, it can be migrated to a locally developed subgroup for refined searching and localized development. This process is initiated when preset conditions are met. When an individual is considered to have reached a "prospective region" near the global optimum, it can be transferred to a local subgroup for fine-grained local optimization.

[0015] This scheme balances the algorithm's optimization and exploration capabilities by dynamically migrating individuals within two subgroups, achieving "broad-area mining first, followed by local refinement." This effectively balances exploration and utilization, significantly improving the efficiency and stability of the algorithm in finding the global optimum.

[0016] Furthermore, the ratio of the number of individuals in the global exploration subgroup to the number of individuals in the local development subgroup is 2:1.

[0017] Furthermore, methods for fine-tuning large models include: S41. A pre-trained large model is used to predict the aerodynamic performance of each individual after the position update, and its fitness function value is obtained. The fitness function value is the lift-to-drag ratio of the aircraft. S42. Use an uncertainty estimation algorithm to estimate the uncertainty of the individual after the location update. If it is in uncertainty, proceed to step S43; otherwise, add it and the corresponding fitness function value to the training set. S43. Perform CFD simulation on the individual to obtain its lift-to-drag ratio, and use it as the individual's fitness function value. Add the individual and its fitness function value to the training set. S44. Based on the training set, the Qlore algorithm is used to fine-tune the pre-trained large model to obtain the fine-tuned large model.

[0018] Furthermore, methods for estimating the uncertainty of individuals after location updates using uncertainty estimation algorithms include: S421. Insert several Dropout layers before and after the inference network of the large model, and enable Dropout during the inference phase to simulate the uncertainty of the Bayesian neural network. S422. Use the individuals with updated positions as input parameters to form a training set. Perform multiple forward propagations on the same input parameters, randomly discarding some neurons each time, to obtain a set of predicted values. S423. Calculate the approximate uncertainty of the predicted mean and predicted variance: in, For the first The lift-to-drag ratio predicted by the inference of the next step; The mean of a set of predicted values; To reflect the uncertainty of the representation of the large model; For input parameters; S424. Calculate the residual based on the mean of a set of predicted values ​​corresponding to the input parameters and the lift-to-drag ratio: in, and Let be the residual and the lift-to-drag ratio of the i-th input parameter; The mean of a set of predicted values ​​corresponding to the i-th input parameter; S425, Variance of Statistical Residues The calibration coefficients are solved on the training set using the least squares method. and bias , so that: Where N is the total number of individuals; S426, According to the calibration coefficient and bias and the uncertainty reflecting the characterization of the large model Determine whether the conditions for triggering active learning are met: in, This is the trigger threshold; S427. When the conditions for triggering active learning are met, the prediction of the current input parameters is in uncertainty; when the conditions for triggering active learning are not met, the prediction of the current input parameters is not in uncertainty.

[0019] The beneficial effects of the above technical solution are as follows: In regions with sparse samples or low confidence, large models will give large prediction variances. This solution obtains these samples by proposing uncertainty estimation, determines whether learning is needed by confidence interval calibration (steps S24 to S27), and then actively learns from these samples that need learning. By iteratively supplementing key area data, the prediction blind spot is continuously reduced.

[0020] Furthermore, the large model is the Qwen2.5-7B model.

[0021] Furthermore, the aerodynamic parameters of the aircraft are area, longitudinal length, lateral length, and angle of attack, totaling four dimensions. Each individual in the initial population is: in, For the i-th individual One dimension, In order to be in Uniformly distributed random numbers within an interval; and The first The lower and upper bounds of each dimension; N is the total number of individuals in the initial population.

[0022] Secondly, a system for a swarm intelligence optimization algorithm used for optimizing aerodynamic parameters of aircraft is provided, comprising: The initial population generation module is used to generate an initial population in a continuous space based on the upper and lower bounds of the aerodynamic parameters of the aircraft. Each individual in the initial population represents a set of aerodynamic parameters of the aircraft. The velocity update module is used to update the current velocity of each individual using a walking optimization algorithm, a historical best guidance strategy, and a multi-group collaborative search mechanism, and to calculate the total perturbation of each individual using a gravity random walk strategy; the current position of an individual is updated by summing its total perturbation, current velocity, and position at the previous moment. The optimal fitness function value generation module is used to fine-tune the large model, and use the fine-tuned large model to predict the fitness function value of each individual after the position update, and select the optimal fitness function value among all individuals. The judgment module is used to determine whether the optimal fitness function value is greater than or equal to a preset threshold or whether the number of iterations is greater than or equal to a preset number. If so, it enters the aircraft aerodynamic parameter determination module; otherwise, it returns to the velocity update module. The aerodynamic parameter determination module for aircraft is used to select the individual corresponding to the optimal fitness function value as the optimized aerodynamic parameters of the aircraft.

[0023] In summary, the beneficial effects of the present invention are as follows: 1. This scheme introduces the hiking optimization algorithm into aircraft aerodynamic parameter tuning. Through a two-level search framework of "historical optimal guidance" combined with "global exploration subgroup - local development subgroup", it achieves a complete process of first large-scale leapfrog exploration and then fine-tuning local convergence. In the early stage of the search, the global subgroup, with its large jump coefficient generated by the skewed beta distribution, quickly covers the design space with higher randomness. In the later stage, the local subgroup is fine-tuned based on a more concentrated small step size factor to improve convergence accuracy. Throughout the process, individuals are adaptively and dynamically migrated between the two, which not only ensures the discovery rate of the global optimum but also significantly accelerates the convergence speed, effectively overcoming the limitation of a single group being prone to getting trapped in local optima.

[0024] 2. To address the premature convergence issue of traditional swarm intelligence algorithms in complex multimodal design spaces, this solution proposes a random walk strategy based on a "gravity model." Each individual calculates its "mass" based on its relative fitness and spatial position, and applies an iteratively decreasing gravitational constant to other individuals, generating a dynamic perturbation vector that is added to position updates. Through this mechanism, the population can effectively escape local optima traps while maintaining overall cooperation, improving the algorithm's robustness and exploration depth, and significantly enhancing its ability to overcome obstacles and find better solutions compared to existing technologies.

[0025] 3. This scheme utilizes a pre-trained Qwen2.5-7B large-scale model to predict the lift-to-drag ratio of aircraft aerodynamic parameters, replacing the computationally intensive traditional high-fidelity CFD simulation. Through multiple dropout uncertainty estimations, confidence interval calibration, and active learning strategies, it identifies high-uncertainty regions in real time and supplements key samples and fine-tunes the model, achieving significant savings in computational resources while ensuring evaluation accuracy. This method not only significantly reduces simulation costs but also maintains the continuous differentiability of the optimization results, ensuring accuracy and efficiency in the optimization process. Attached Figure Description

[0026] Figure 1 This is a flowchart of a swarm intelligence optimization algorithm used for optimizing aerodynamic parameters of aircraft.

[0027] Figure 2 This is a block diagram illustrating the principle of swarm intelligence optimization for optimizing aerodynamic parameters of aircraft. Detailed Implementation

[0028] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0029] refer to Figure 1 , Figure 1 A flowchart of a swarm intelligence optimization algorithm for optimizing aircraft aerodynamic parameters is shown; for example... Figure 1 As shown, the method S includes steps S1 to S6.

[0030] In step S1, based on the upper and lower bounds of the aerodynamic parameters of the aircraft, an initial population is generated in a continuous space, and each individual in the initial population represents a set of aerodynamic parameters of the aircraft. During implementation, the preferred aerodynamic parameters for the aircraft in this scheme are area, longitudinal length, lateral length, and angle of attack, totaling four dimensions. Each individual in the initial swarm is: in, For the i-th individual One dimension, In order to be in Uniformly distributed random numbers within an interval; and The first The lower and upper bounds of each dimension; N is the total number of individuals in the initial population.

[0031] In step S2, the current velocity of each individual is updated using a hiking optimization algorithm, a historical best guidance strategy, and a multi-group collaborative search mechanism, and the total perturbation of each individual is calculated using a gravity random walk strategy. In one embodiment of the present invention, the method for updating the current speed of each individual in step S2 includes: S21. The speed of each individual in the population is updated using the hiking optimization algorithm: in, and Let be the velocity of individual i in the t-th and t-1-th iterations, respectively; Let be the random perturbation coefficient of individual i in the t-th iteration; The optimal position of an individual in the population; Let i be the spatial position of individual i in the t-th iteration; S22. The update rate of each individual in the population is adjusted using a historical best-practice guidance strategy: in, To Corrected speed; and Both control the overall and local weights. ; For individuals Historically best position; and All are sweep factors; S23. Divide the population into a global exploration subgroup that searches within a continuous space and a local development subgroup that searches within discovered prime areas. Individuals in the global exploration subgroup are encouraged to make larger, more random movements. Individuals in the local development subgroup are encouraged to make smaller, more deterministic movements. The preferred ratio of individuals in the global exploration subgroup to those in the local development subgroup is 2:1.

[0032] S24. For each individual in the two subgroups, a bias coefficient affecting its search behavior is dynamically generated using a skewed beta distribution. and ; S25, Using deviation coefficient and renew and To achieve Update: in, Let be the current velocity of individual i at the t-th iteration.

[0033] During implementation, priority step S24 of this solution further includes: S241, Beta distribution parameters of the global exploration subgroup and Adjustments will be made: in, and The initial and final parameters are preset to control the beta distribution pattern of the global exploration subgroup. and T is the maximum number of iterations; in the initial stage (t→0), the parameter combination causes the beta distribution to be biased towards generating larger or more diverse values, which are then used as multipliers to adjust the individual's step size or direction, thereby achieving "stronger jumping ability".

[0034] S242, Beta distribution parameters of locally developed subgroups and Adjustments will be made: in, and The initial and final parameters are preset to control the beta distribution pattern of the local development subgroup. and In the later stages (t→T), the parameter combinations cause the beta distribution to tend to generate smaller or more concentrated values, which are used to fine-tune individual positions and achieve "steady convergence".

[0035] S243. Each individual in both subgroups samples a search bias coefficient from its beta distribution: in, and Here are the beta distribution parameters adjusted at time t, when the individual belongs to the global exploration subgroup: , When an individual belongs to a locally developed subgroup: , .

[0036] For each individual in the global exploration subgroup, after each iteration, if it meets a preset condition, it is moved to the local development subgroup. The preset condition is: in, and The first The individual in the first Subsequent The fitness function value at the next iteration; The preset number of consecutive iterations; This represents the optimal fitness function value in the global exploration subgroup. For individuals in the local exploration subgroup, after each iteration, if the historical best position of an individual has not been updated for several consecutive generations, it is added to the global exploration subgroup.

[0037] In step S3, the total disturbance of the individual, the current velocity, and the position at the previous moment are used. and update its current position: in, For the first The individual in the first The position at the next iteration, i.e., the current position; For the first The individual in the first Total perturbation at the next iteration.

[0038] Through gravitational random walks, individuals can dynamically perceive the distribution of the group and respond to the adjustment of the global gravitational field, thus possessing a stronger ability to break out of local constraints and overcome obstacles in complex high-dimensional design spaces.

[0039] This scheme is based on a gravity-based random walk strategy using a gravity model, which introduces a dynamic perturbation vector for each individual to enhance the randomness of the search. The strategy consists of the following three parts: 1) Individual mass calculation: The "mass" of each individual is calculated based on its relative fitness and spatial location, starting with energy. , Measuring individuals The difference in fitness relative to the worst individual ensures that superior individuals have higher "energy"; then, the relative height is calculated, simulating the difference in "height" based on the relative distance between an individual and the best and worst individuals: Finally, calculate the mass of each individual: in , It is a minimal constant that prevents division by zero. Through the above formula, individuals that are excellent and far from the optimal solution are given greater "mass" and will exert a stronger influence in the next gravitational action.

[0040] 2) Adaptive Gravitational Parameter Update: Gravitational strength decreases with the number of iterations to balance global jumps and local convergence. in As a preset gravitational constant, is the decay factor, and T is the maximum number of iterations.

[0041] 3) Calculation of gravitational perturbation: First, calculate the gravitational force between any two individuals: Then calculate the gravitational forces exerted by all other individuals on the current individual, and form the total perturbation: in, For random disturbance coefficients, It is a minimal constant to prevent division by zero, and this perturbation term is added to the individual position update.

[0042] In step S4, the large model is fine-tuned, and the fine-tuned large model is used to predict the fitness function value of each individual after the position update, and the optimal fitness function value among all individuals is selected; the preferred large model in this scheme is the Qwen2.5-7B model.

[0043] During implementation, this scheme prioritizes fine-tuning the large model using the following methods: S41. A pre-trained large model is used to predict the aerodynamic performance of each individual after the position update, and its fitness function value is obtained. The fitness function value is the lift-to-drag ratio of the aircraft. S42. Use an uncertainty estimation algorithm to estimate the uncertainty of the individual after the location update. If it is in uncertainty, proceed to step S43; otherwise, add it and the corresponding fitness function value to the training set. S43. Perform CFD simulation on the individual to obtain its lift-to-drag ratio, and use it as the individual's fitness function value. Add the individual and its fitness function value to the training set. S44. Based on the training set, the Qlore algorithm is used to fine-tune the pre-trained large model to obtain the fine-tuned large model.

[0044] In step S5, it is determined whether the optimal fitness function value is greater than or equal to a preset threshold or whether the number of iterations is greater than or equal to a preset number. If so, proceed to step S6; otherwise, return to step S2. In step S6, the individual corresponding to the optimal fitness function value is used as the optimized aerodynamic parameters of the aircraft.

[0045] In one embodiment of the present invention, a method for estimating the uncertainty of an individual after location update using an uncertainty estimation algorithm includes: S421. Insert several Dropout layers before and after the inference network of a large model (such as after the output of each layer in the Transformer's multi-head self-attention and feedforward network), and enable Dropout during the inference phase to simulate the uncertainty of the Bayesian neural network. S422. Use the individuals with updated positions as input parameters to form a training set. Perform multiple forward propagations on the same input parameters, randomly discarding some neurons each time, to obtain a set of predicted values. S423. Calculate the approximate uncertainty of the predicted mean and predicted variance: in, For the first The lift-to-drag ratio predicted by the inference of the next step; The mean of a set of predicted values; To reflect the uncertainty of the representation of the large model; For input parameters; The uncertainty in the characterization of the reaction model, however This only reflects the model's internal estimation of weight uncertainty and may not necessarily reflect the one-to-one correspondence between the model's predicted values ​​and the true CFD values. Therefore, this scheme proposes a confidence interval calibration algorithm (steps S424 to S427). The "confidence interval calibration" aims to transform the "prediction variance" obtained from multiple dropouts into a confidence measure that is more consistent with the true error (CFD calculation and model prediction residuals).

[0046] S424. Calculate the residual based on the mean of a set of predicted values ​​corresponding to the input parameters and the lift-to-drag ratio: in, and Let be the residual and the lift-to-drag ratio of the i-th input parameter; The mean of a set of predicted values ​​corresponding to the i-th input parameter; S425, Variance of Statistical Residues The calibration coefficients are solved on the training set using the least squares method. and bias , so that: Where N is the total number of individuals; S426, According to the calibration coefficient and bias and the uncertainty reflecting the characterization of the large model Determine whether the conditions for triggering active learning are met: in, This is the trigger threshold; S427. When the conditions for triggering active learning are met, the prediction of the current input parameters is in uncertainty; when the conditions for triggering active learning are not met, the prediction of the current input parameters is not in uncertainty.

[0047] like Figure 2 As shown, this solution also provides a system for a swarm intelligent optimization algorithm applied to the optimization of aerodynamic parameters of aircraft, which includes: The initial population generation module is used to generate an initial population in a continuous space based on the upper and lower bounds of the aerodynamic parameters of the aircraft. Each individual in the initial population represents a set of aerodynamic parameters of the aircraft. The velocity update module is used to update the current velocity of each individual using a walking optimization algorithm, a historical best guidance strategy, and a multi-group collaborative search mechanism, and to calculate the total perturbation of each individual using a gravity random walk strategy; the current position of an individual is updated by summing its total perturbation, current velocity, and position at the previous moment. The optimal fitness function value generation module is used to fine-tune the large model, and use the fine-tuned large model to predict the fitness function value of each individual after the position update, and select the optimal fitness function value among all individuals. The judgment module is used to determine whether the optimal fitness function value is greater than or equal to a preset threshold or whether the number of iterations is greater than or equal to a preset number. If so, it enters the aircraft aerodynamic parameter determination module; otherwise, it returns to the velocity update module. The aerodynamic parameter determination module for aircraft is used to select the individual corresponding to the optimal fitness function value as the optimized aerodynamic parameters of the aircraft.

[0048] In summary, this scheme integrates the walking optimization algorithm framework from swarm intelligence, introduces historical optimal strategies, multiple swarm collaboration mechanisms, and gravity random walk perturbation strategies, and combines them with a large model for fitness evaluation, thereby achieving high-precision optimization of the aerodynamic parameters of the aircraft.

Claims

1. A swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft, characterized in that, Including the following steps: S1. Based on the upper and lower bounds of the aerodynamic parameters of the aircraft, an initial population is generated in a continuous space. Each individual in the initial population represents a set of aerodynamic parameters of the aircraft. S2. The current velocity of each individual is updated using a hiking optimization algorithm, a historical best guidance strategy, and a multi-group collaborative search mechanism, and the total perturbation of each individual is calculated using a gravity random walk strategy. S3. Update the current position of an individual by summing its total perturbation, current velocity, and position at the previous moment. S4. Fine-tune the large model and use the fine-tuned large model to predict the fitness function value of each individual after the position update, and select the optimal fitness function value among all individuals. S5. Determine whether the optimal fitness function value is greater than or equal to the preset threshold or whether the number of iterations is greater than or equal to the preset number. If yes, proceed to step S6; otherwise, return to step S2. S6. The individual corresponding to the optimal fitness function value is used as the optimized aerodynamic parameters of the aircraft.

2. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to claim 1, characterized in that, The methods for updating the current velocity of each individual in step S2 include: S21. The speed of each individual in the population is updated using the hiking optimization algorithm: in, and Let be the velocity of individual i in the t-th and t-1-th iterations, respectively; Let be the random perturbation coefficient of individual i in the t-th iteration; The optimal position for an individual in the population; Let i be the spatial position of individual i in the t-th iteration; S22. The update rate of each individual in the population is adjusted using a historical best-practice guidance strategy: in, To Corrected speed; and Both control the overall and local weights. ; For individuals Historically best position; and All are sweep factors; S23. Divide the population into a global exploration subgroup that searches in a continuous space and a local development subgroup that searches in a high-quality area that has been discovered. S24. For each individual in the two subgroups, a bias coefficient affecting its search behavior is dynamically generated using a skewed beta distribution. and ; S25, Using deviation coefficient and renew and To achieve Update: in, Let be the current velocity of individual i at the t-th iteration.

3. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to claim 2, characterized in that, Step S24 further includes: S241, Beta distribution parameters of the global exploration subgroup and Adjustments will be made: in, and The initial and final parameters are preset to control the beta distribution pattern of the global exploration subgroup. and T represents the maximum number of iterations. S242, Beta distribution parameters of locally developed subgroups and Adjustments will be made: in, and The initial and final parameters are preset to control the beta distribution pattern of the local development subgroup. and ; S243. Each individual in both subgroups samples a search bias coefficient from its beta distribution: in, and Here are the beta distribution parameters adjusted at time t, when the individual belongs to the global exploration subgroup: , When an individual belongs to a locally developed subgroup: , .

4. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to claim 2, characterized in that, For each individual in the global exploration subgroup, after each iteration, if it meets a preset condition, it is moved to the local development subgroup. The preset condition is: in, and The first The individual in the first Subsequent The fitness function value at the next iteration; The preset number of consecutive iterations; This represents the optimal fitness function value in the global exploration subgroup. For individuals in the local exploration subgroup, after each iteration, if the historical best position of an individual has not been updated for several consecutive generations, it is added to the global exploration subgroup.

5. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to any one of claims 2-4, characterized in that, The ratio of the number of individuals in the global exploration subgroup to the number of individuals in the local development subgroup is 2:

1.

6. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to claim 2, characterized in that, Methods for fine-tuning large models include: S41. A pre-trained large model is used to predict the aerodynamic performance of each individual after the position update, and its fitness function value is obtained. The fitness function value is the lift-to-drag ratio of the aircraft. S42. Use an uncertainty estimation algorithm to estimate the uncertainty of the individual after the location update. If it is in uncertainty, proceed to step S43; otherwise, add it and the corresponding fitness function value to the training set. S43. Perform CFD simulation on the individual to obtain its lift-to-drag ratio, and use it as the individual's fitness function value. Add the individual and its fitness function value to the training set. S44. Based on the training set, the Qlore algorithm is used to fine-tune the pre-trained large model to obtain the fine-tuned large model.

7. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to claim 6, characterized in that, Methods for estimating the uncertainty of an individual after its location update using uncertainty estimation algorithms include: S421. Insert several Dropout layers before and after the inference network of the large model, and enable Dropout during the inference phase to simulate the uncertainty of the Bayesian neural network. S422. Use the individuals with updated positions as input parameters to form a training set. Perform multiple forward propagations on the same input parameters, randomly discarding some neurons each time, to obtain a set of predicted values. S423. Calculate the approximate uncertainty of the predicted mean and predicted variance: in, For the first The lift-to-drag ratio predicted by the inference of the next step; The mean of a set of predicted values; To reflect the uncertainty of the representation of the large model; For input parameters; S424. Calculate the residual based on the mean of a set of predicted values ​​corresponding to the input parameters and the lift-to-drag ratio: in, and Let be the residual and the lift-to-drag ratio of the i-th input parameter; The mean of a set of predicted values ​​corresponding to the i-th input parameter; S425, Variance of Statistical Residues The calibration coefficients are solved on the training set using the least squares method. and bias , so that: Where N is the total number of individuals; S426, According to the calibration coefficient and bias and the uncertainty reflecting the characterization of the large model Determine whether the conditions for triggering active learning are met: in, This is the trigger threshold; S427. When the conditions for triggering active learning are met, the prediction of the current input parameters is in uncertainty; when the conditions for triggering active learning are not met, the prediction of the current input parameters is not in uncertainty.

8. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to any one of claims 1-4 and 6-7, characterized in that, The large model is the Qwen2.5-7B model.

9. The swarm intelligence optimization algorithm for optimizing aerodynamic parameters of aircraft according to any one of claims 1-4 and 6-7, characterized in that, The aerodynamic parameters of the aircraft are area, longitudinal length, lateral length, and angle of attack, totaling four dimensions. Each individual in the initial population is: in, For the i-th individual One dimension, In order to be in Uniformly distributed random numbers within an interval; and The first The lower and upper bounds of each dimension; N is the total number of individuals in the initial population.

10. A system for applying the swarm intelligent optimization algorithm for optimizing aerodynamic parameters of an aircraft as described in any one of claims 1-9, characterized in that, include: The initial population generation module is used to generate an initial population in a continuous space based on the upper and lower bounds of the aerodynamic parameters of the aircraft. Each individual in the initial population represents a set of aerodynamic parameters of the aircraft. The speed update module is used to update the current speed of each individual using a walking optimization algorithm, a historical best guidance strategy, and multiple group cooperative search mechanisms, and to calculate the total perturbation of each individual using a gravity random walk strategy. The current position of an individual is updated by summing its total perturbation, current velocity, and position at the previous moment. The optimal fitness function value generation module is used to fine-tune the large model, and use the fine-tuned large model to predict the fitness function value of each individual after the position update, and select the optimal fitness function value among all individuals. The judgment module is used to determine whether the optimal fitness function value is greater than or equal to a preset threshold or whether the number of iterations is greater than or equal to a preset number. If so, it enters the aircraft aerodynamic parameter determination module; otherwise, it returns to the velocity update module. The aerodynamic parameter determination module for aircraft is used to select the individual corresponding to the optimal fitness function value as the optimized aerodynamic parameters of the aircraft.

Images