A method and apparatus for predicting aerodynamic forces of a large transport vehicle

Abstract

The application discloses a kind of aerodynamic force prediction method and equipment of large transport equipment, it is first to obtain the structural parameter and historical flight data of large transport equipment, construct aerodynamic force prediction knowledge graph based on cloud server, and construct adaptive CFD numerical simulation model, adaptive optimization is carried out using adaptive local alternate genetic algorithm to adaptive CFD numerical simulation model, and iteration update of aerodynamic force prediction knowledge graph is carried out.Then, aerodynamic force prediction is carried out using deep reinforcement learning model, and knowledge reasoning is carried out in combination with aerodynamic force prediction knowledge graph;Solve is that the accuracy of CFD method in large transport equipment aerodynamic force prediction is limited by physical model and calculation grid, prone to cause the calculation failure under complex aerodynamic force phenomenon and flow characteristic, the present application can comprehensively consider various physical phenomena and flow characteristics of large transport equipment when running at high speed, improve the accuracy and reliability of prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aerodynamic force prediction, in particular to an aerodynamic force prediction method and device for large-scale transportation equipment. BACKGROUND

[0002] With the rapid development of industrialization and urbanization, there is a strong demand for large-scale logistics and freight transportation, and the importance of large-scale transportation equipment is increasingly prominent. How to ensure the reliability and safety of large-scale transportation equipment during high-speed operation is of great significance to the development of modern economy. For large-scale transportation equipment, aerodynamic performance directly affects the flight efficiency, stability and safety of the equipment. Therefore, it is crucial to accurately predict the aerodynamic force acting on large-scale transportation equipment.

[0003] As a most economical, efficient and accurate numerical calculation method, computational fluid dynamics (CFD) has been widely used in the field of aerodynamic force prediction in recent years. For example, the aerodynamic selection method for aircraft based on the shock wave simulation capability of CFD software is disclosed in Chinese patent document CN111859817A; and the establishment method and device of an aircraft aerodynamic model database based on CFD technology are disclosed in Chinese patent document CN114611437A. However, the accuracy of the CFD method depends on the physical model and the calculation grid. Large-scale transportation equipment involves multiple physical phenomena and flow characteristics during high-speed operation, and the complexity of aerodynamic force prediction is high, which may lead to failure of CFD calculation.

[0004] Therefore, there is a need for an aerodynamic force prediction method and device for large-scale transportation equipment to solve the above problems. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application discloses an aerodynamic force prediction method and device for large-scale transportation equipment, which combines CFD numerical simulation, aerodynamic force prediction knowledge graph, self-adaptive optimization algorithm and deep reinforcement learning model, and forms a comprehensive aerodynamic force prediction system. This system can more comprehensively consider various physical phenomena and flow characteristics of large-scale transportation equipment during high-speed operation, and improve the accuracy and reliability of prediction.

[0006] The present application adopts the following technical solutions:

[0007] An aerodynamic force prediction method for large-scale transportation equipment, comprising the following steps:

[0008] Step 1: Obtain the structural parameters of the target large-scale transportation equipment, and the historical flight data of large-scale transportation equipment with different structural parameters under different flight condition parameters and meteorological condition parameters;

[0009] Step two, a cloud server is used to build an aerodynamic force prediction knowledge graph, the cloud server extracts the structural parameters, flight condition parameters, meteorological condition parameters and historical flight data of various large transport equipment from the data obtained in step one as nodes of the aerodynamic force prediction knowledge graph, and integrates the aerodynamic force prediction discipline knowledge and aerodynamic force formula to form the network structure of the aerodynamic force prediction knowledge graph;

[0010] Step three, based on the aerodynamic force prediction knowledge graph and the structural parameters of the target large transport equipment, an adaptive CFD numerical simulation model is built, and an adaptive optimization algorithm is used to optimize the adaptive CFD numerical simulation model;

[0011] Step four, the aerodynamic performance of the target large transport equipment at high speed under different flight conditions and meteorological conditions is simulated by using the optimized adaptive CFD numerical simulation model, and the aerodynamic force prediction knowledge graph is iteratively updated based on the simulation process data;

[0012] Step five, a deep reinforcement learning model is used to predict the aerodynamic performance of the large transport equipment, and the updated aerodynamic force prediction knowledge graph is used to train the deep reinforcement learning model.

[0013] Further, in the step one, the structural parameters at least include the fuselage length, the fuselage width, the fuselage height, the fuselage shape, the wing shape, the wing area, the vertical tail area, the center of gravity position, the engine thrust, the Mach number and the angle of attack along the curved surface, the meteorological condition parameters at least include the wind speed, the wind direction, the temperature, the humidity and the air pressure, the flight condition parameters at least include the flight height, the flight speed, the flight attitude and the maneuvering action, and the flight data at least include the lift, the drag, the lateral force and the aerodynamic moment.

[0014] Further, in step two, the cloud server includes a data preprocessing unit, a domain classification unit, an entity recognition and linking unit, a relation extraction unit, a knowledge organization and representation unit, and an incremental update unit. The data preprocessing unit uses a natural language processor to extract structural parameters, flight condition parameters, meteorological condition parameters, and historical flight data of the large transportation equipment, and cleans and organizes the extracted parameters. The domain classification unit uses a text domain mining model (LDA) to classify the extracted parameters according to their subject area distribution. The entity recognition and linking unit automatically identifies aerodynamic entities in the extracted parameters using the Named Entity Recognition (NER) task, and links them using the Entity Linking (EL) task. The dynamic entities are linked to the corresponding entities in the aerodynamic prediction knowledge graph. The relation extraction unit automatically extracts the relationships between entities from the aerodynamic prediction discipline knowledge and aerodynamic formulas using a relation extraction algorithm, and stores them in the aerodynamic prediction knowledge graph to form a network structure. The relationships between entities include at least the relationship between structural parameters and aerodynamic performance, the relationship between flight condition parameters and aerodynamic changes, and the relationship between meteorological condition parameters and aerodynamic changes. The knowledge organization and representation unit uses a graph database or triplet storage format to store the extracted parameters, aerodynamic prediction discipline knowledge, and aerodynamic formulas. The incremental update unit updates the aerodynamic prediction knowledge graph in real time through a time window mechanism.

[0015] Furthermore, in step three, constructing the adaptive CFD numerical simulation model includes the following steps:

[0016] S301. Receive the structural parameters of the target large transportation equipment, and extract the structural parameters, flight condition parameters and meteorological condition parameters that match the target large transportation equipment from the aerodynamic prediction knowledge graph. Based on the extracted parameters, initialize the CFD numerical simulation model. The initialization settings include at least geometric model settings, mesh generation and boundary condition settings.

[0017] S302. Based on the structural parameters of the target large transportation equipment and the simulation target, an adaptive optimization algorithm is set, and the initialization settings of the CFD numerical simulation model are optimized using the adaptive optimization algorithm. The adaptive optimization algorithm sets the aerodynamic performance simulation as the optimization target and sets constraints based on the knowledge of aerodynamic prediction and aerodynamic formulas.

[0018] S303. Select a physical model based on the characteristics of the target large transport equipment and the expected flight environment to reflect the aerodynamic performance of the target large transport equipment when it is running at high speed.

[0019] S304, integrate the adaptive optimization algorithm and the CFD numerical simulation model to form an adaptive CFD numerical simulation model, and perform numerical simulation on the aerodynamic performance of the target large transport equipment under different flight conditions and weather conditions according to the selected physical model and the adaptive CFD numerical simulation model;

[0020] S305, compare the historical flight data with the numerical simulation, and calibrate the adaptive CFD numerical simulation model according to the verification comparison result.

[0021] Further, the adaptive optimization algorithm optimizes the initialization setting of the CFD numerical simulation model based on a multi-agent proximal policy model and a local alternating genetic model, and the working method of the adaptive optimization algorithm comprises the following steps:

[0022] S3021, the initialization setting variable in the CFD numerical simulation model is regarded as an agent, the running environment of the agent is the CFD numerical simulation model, the strategy function and the value function of the agent are defined by using aerodynamic force prediction discipline and aerodynamic force formula, the strategy function adjusts the value of the current initialization setting variable according to the current state of the CFD numerical simulation model and the historical experience of the initialization setting variable, and the value function evaluates the value of the current initialization setting variable according to the current state of the CFD numerical simulation model and the strategy function;

[0023] S3022, the agent executes the adjustment action in the running environment according to the strategy function, and updates the strategy function parameters and the value function parameters of the agent based on a proximal policy optimization method, the proximal policy optimization method restricts the update step of the strategy and the value function through a strategy loss function and a value loss function;

[0024] S3023, in the update iteration process of the strategy function parameters and the value function parameters, the agent is divided into multiple groups, and a local alternating genetic model is used for local cooperation and overall competition among multiple agents;

[0025] S3024, the initialization setting variable value after the update iteration is input into the CFD numerical simulation model for numerical calculation, and the calculation result is fed back to the agent as the current state of the CFD numerical simulation model;

[0026] S3025, repeat steps S3022 and S3023 until a predetermined aerodynamic performance simulation performance is reached.

[0027] Further, the expression of the strategy loss function is:

[0028]

[0029] In formula (1), L P(θ) represents the update loss function of the policy function parameter θ, which measures the difference between the new policy and the old policy, E t represents the expected value of all states and actions at time step t, r t (θ) represents the policy function parameter at time step t in state s t takes action a t the probability ratio of the policy, represents the advantage estimate at time step t, which measures the advantage of taking action in state relative to the average action selection, ∈ represents the hyperparameter, which limits the probability ratio r t (θ), clip(r t (θ), 1-∈, 1+∈) represents limiting the probability ratio r t (θ) between 1-∈ and 1+∈, and the expression of the value loss function is:

[0030]

[0031] In formula (2), represents the update loss function of the value function parameter , represents the value function of the agent in state s t , represents the return value obtained at time step t, E t represents the expected value of all states and actions at time step t.

[0032] Further, the deep reinforcement learning model extracts aerodynamic force prediction features from adaptive CFD numerical simulation results as aerodynamic force prediction input features, and adopts a graph neural network to perform knowledge reasoning of the aerodynamic force prediction knowledge graph, so as to mine the aerodynamic characteristics inside the large transport equipment.

[0033] Further, an aerodynamic force prediction device for a large transport equipment comprises:

[0034] A data acquisition module acquires structural parameters, flight condition parameters, meteorological condition parameters and historical flight data of a target large transport equipment, and transmits them to a cloud server for processing and storage;

[0035] A knowledge graph construction module adopts the cloud server to construct an aerodynamic force prediction knowledge graph, the cloud server extracts information from the structural parameters, flight conditions, meteorological conditions and historical flight data as nodes of the aerodynamic force prediction knowledge graph, and integrates aerodynamic discipline knowledge, aerodynamic force formulas and physical models to form a multi-dimensional prediction framework;

[0036] A numerical simulation module constructs an adaptive CFD numerical simulation model based on the aerodynamic force prediction knowledge graph and the structural parameters of the target large transport equipment, simulates the aerodynamic performance, and iteratively updates the aerodynamic force prediction knowledge graph based on the simulation process data;

[0037] An aerodynamic force prediction module predicts the aerodynamic performance of the large transport equipment using a deep reinforcement learning model, and trains the model using the updated aerodynamic force prediction knowledge graph;

[0038] A user interaction module views the real-time state of the adaptive CFD numerical simulation model and the aerodynamic performance prediction results through a graphical interface, adjusts the flight conditions and weather conditions, and then automatically generates an aerodynamic performance report for the large transport equipment based on the simulation and prediction results.

[0039] The beneficial effects of the present application are:

[0040] 1. The present application can more accurately reflect the aerodynamic performance of the equipment during high-speed operation by constructing an adaptive CFD numerical simulation model and selecting a physical model according to the characteristics of the target large transport equipment and the expected flight environment. In addition, the use of an adaptive optimization algorithm to optimize the CFD numerical simulation model can further improve the accuracy and efficiency of the calculation.

[0041] 2. The present application forms an aerodynamic force prediction knowledge graph by integrating aerodynamic force prediction discipline knowledge and aerodynamic force formulas. This method can extract structural parameters, flight condition parameters, and weather condition parameters that match the target large transport equipment, providing more accurate and comprehensive initialization settings for CFD numerical simulation. At the same time, the iterative update of the knowledge graph can continuously optimize the accuracy of the CFD numerical simulation.

[0042] 3. The present application uses a deep reinforcement learning model to predict the aerodynamic performance of the large transport equipment, and trains the model using the updated aerodynamic force prediction knowledge graph. This combination of deep learning and reinforcement learning can explore the aerodynamic characteristics of the large transport equipment, improving the accuracy and generalization ability of the prediction.

[0043] 4. The present application uses a multi-agent proximal policy model and a local alternating genetic model for local collaboration and overall competition among agents, which can more effectively explore and optimize the initialization settings of the CFD numerical simulation model, thereby improving the accuracy and efficiency of the calculation. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 The overall method flowchart of the present application is shown in the figure;

[0045] Figure 2 The workflow diagram for constructing an adaptive CFD numerical simulation model in the present application is shown in the figure;

[0046] Figure 3 The schematic diagram of the workflow of the adaptive optimization algorithm in the application is shown in the figure;

[0047] Figure 4 The schematic diagram of the overall device architecture of the application is shown in the figure. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the accompanying drawings of the embodiments of the application. Figure 1 to the accompanying drawings of the embodiments of the application. Figure 4 It should be apparent that the described embodiments are only a part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0049] The embodiments of the application disclose a method for predicting aerodynamic force of large transport equipment, as shown in the accompanying drawings, comprising the following steps: Figure 1

[0050] Step one, data acquisition

[0051] The structural parameters of the target large transport equipment are obtained by consulting technical documents or design drawings, and the structural parameters at least include the length, width, height, shape of the fuselage, shape of the wing, wing area, vertical tail area, center of gravity position, engine thrust, Mach number and attack angle along the curved surface. The historical flight data of the large transport equipment with different structural parameters under different flight condition parameters and meteorological condition parameters are recorded by flight recorders, data recorders, flight simulators and other equipment. The collected structural parameters and historical flight data are stored on a cloud server, and a database management system is used to manage and maintain the data. The relevant data is automatically or manually imported through the data interface and the database, cloud storage and other data sources. The meteorological condition parameters at least include wind speed, wind direction, temperature, humidity and air pressure. The flight condition parameters at least include flight altitude, flight speed, flight attitude and maneuvering action. The flight data at least include lift, drag, lateral force and aerodynamic moment.

[0052] During the data collection process, the accuracy, integrity and consistency of the data are ensured. The data is preliminarily screened and cleaned to remove abnormal values, repeated values or invalid data. The data is standardized to ensure that the data from different sources is consistent in format, unit and precision. The collected data is stored in a safe and reliable database for subsequent analysis and processing. Data indexing and metadata are established to quickly retrieve and access data. Data is backed up regularly to prevent data loss or damage.

[0053] Step two, building an aerodynamic force prediction knowledge graph​

[0054] The cloud server integrates the various data types obtained in step one to build nodes of the aerodynamic force prediction knowledge graph. Each node represents a structural parameter, flight condition parameter, or meteorological condition parameter, and the association between each parameter (for example, the relationship between flight speed and lift, drag) will be connected as an edge. On the cloud server, based on existing aerodynamics theory, aerodynamic formulas (such as calculation formulas for lift, drag, and aerodynamic load), and empirical formulas extracted from historical data, the network structure of the knowledge graph is constructed. The edges in the graph can be physical models, empirical formulas, or even mathematical relationships obtained by historical data regression. The nodes are structural parameters, flight conditions, meteorological conditions, etc. The edges are aerodynamic relationships, mechanical formulas, historical data regression results, etc. The network structure of the knowledge graph is used to realize data transmission and learning between different nodes, and to establish an aerodynamic force prediction knowledge graph. All data and calculations will be performed on the cloud server, which can meet the large-scale data processing and calculation requirements, and facilitate the iterative update in the later stage.

[0055] In the step two, the cloud server includes a data preprocessing unit, a domain classification unit, an entity recognition and linking unit, a relationship extraction unit, a knowledge organization representation unit, and an incremental updating unit.

[0056] Data preprocessing unit: using natural language processing (NLP) technology, the structural parameters, flight condition parameters, meteorological condition parameters and historical flight data of large transportation equipment are extracted from the original data sources (such as databases, documents, log files, etc.). NLP tools include word segmenter, part-of-speech tagger, named entity recognizer, etc. to accurately identify and extract relevant parameters. The extracted data is processed to remove duplicates, fill in missing values, correct error values, etc. to ensure the accuracy and integrity of the data. The data is formatted into a uniform format for subsequent processing and storage.

[0057] Domain classification unit: using LDA (Latent Dirichlet Allocation) text domain mining model, the parameter text is subjected to topic modeling. By analyzing the topic distribution of the text, the parameters are classified according to the discipline field (such as aerodynamics, mechanical engineering, meteorology, etc.).

[0058] Entity Recognition and Linking Unit: Using Named Entity Recognition (NER) technology, automatically identify and extract aerodynamic-related entities in the parameters, such as equipment names, component names, parameter names, etc. NER tools are based on deep learning models such as LSTM-CRF, BERT, etc. to improve the accuracy of identification. Through Entity Linking (EL) technology, link the identified aerodynamic entities to the corresponding entities already existing in the aerodynamic prediction knowledge graph. If there is no corresponding entity in the knowledge graph, create a new entity and add it to the graph.

[0059] Relationship Extraction Unit: Apply relationship extraction algorithms (such as rule-based methods, machine learning algorithms, deep learning models, etc.) to extract relationships between entities from aerodynamic prediction discipline knowledge and aerodynamic formulas. Relationship extraction algorithms involve natural language understanding, semantic analysis, etc. Store the extracted relationships in the aerodynamic prediction knowledge graph to form a network structure. Relationships include at least the relationship between structural parameters and aerodynamic performance, the relationship between flight condition parameters and aerodynamic changes, the relationship between meteorological condition parameters and aerodynamic changes, etc.

[0060] Knowledge Organization Representation Unit: Choose a graph database (such as Neo4j, JanusGraph, etc.) or a triple storage format (such as RDF, OWL, etc.) to store the extracted parameters, aerodynamic prediction discipline knowledge, and aerodynamic formulas. Graph databases are suitable for storing complex network structures, while triple storage formats facilitate semantic queries and reasoning. Store the extracted parameters and their attribute values. Store the structured representation of aerodynamic prediction discipline knowledge and aerodynamic formulas. Store the relationships between entities and their attributes.

[0061] Incremental Update Unit: Set a time window mechanism to regularly obtain new data from data sources. The time window can be set to a fixed time interval (such as daily, weekly, monthly, etc.), or dynamically adjusted according to the frequency of data updates. Preprocess, entity recognition and linking, relationship extraction, etc. on new data. Integrate new knowledge into the aerodynamic prediction knowledge graph, update the nodes and relationships of the graph. Verify the updated knowledge graph to ensure data accuracy and consistency. If errors or inconsistent data are found, correct and adjust them in a timely manner.

[0062] Through the above units, the cloud server can build an accurate and complete aerodynamic prediction knowledge graph, providing strong support for subsequent aerodynamic prediction. At the same time, through the continuous updating and improvement of the incremental update unit, the knowledge graph can maintain consistency with the actual situation, improving the accuracy and reliability of the prediction.

[0063] Step Three, Building an Adaptive CFD Numerical Simulation Model Based on Knowledge Graph

[0064] The structural parameters and historical flight data are input into the aerodynamic force prediction knowledge graph to obtain learning parameters in the network structure. Adaptive grid division and fixed calculation domain management are performed using the learning parameters. Based on the aerodynamic force prediction knowledge graph and the structural parameters of the target large transport equipment, an aerodynamic simulation model thereof is constructed by a computational fluid dynamics (CFD) method. The CFD simulation will consider the changes in the air flow field during flight and calculate the aerodynamic response of the equipment under different conditions. To improve the accuracy and computational efficiency of the CFD numerical model, an adaptive optimization algorithm is used to optimize the CFD model. This algorithm can automatically adjust parameters such as grid density and flow field simulation accuracy, so that the CFD simulation can reduce the calculation time while ensuring the calculation accuracy.

[0065] In the third step, an adaptive CFD numerical simulation model is constructed, as shown in FIG. 4, including the following steps: Figure 2

[0066] S301, initialization setting of the CFD numerical simulation model

[0067] The structural parameters of the target large transport equipment are received, which usually include the size, shape, mass distribution, etc. of the equipment, and are the basis for establishing the geometric model. The parameters are extracted from the aerodynamic force prediction knowledge graph, which stores various parameters related to the target large transport equipment, including structural parameters, flight condition parameters (such as speed, altitude, angle of attack, etc.) and meteorological condition parameters (such as wind speed, wind direction, temperature, etc.). These parameters are used for more accurate initialization setting of the CFD numerical simulation model.

[0068] The CFD numerical simulation model is initialized based on the extracted parameters, including at least geometric model setting, grid division and boundary condition setting. The geometric model setting is to construct a three-dimensional model of the equipment according to the structural parameters; the grid division is to divide the geometric model into multiple small units for numerical calculation; and the boundary condition setting is to set the boundary conditions of the simulation model according to the flight conditions and meteorological conditions, such as inlet velocity, outlet pressure, etc.

[0069] S302, optimization of the initialization setting of the CFD numerical simulation model

[0070] ​According to the structural parameters of the target large transport equipment and the simulation model of the target, the related parameters of the adaptive optimization algorithm are set, such as the maximum number of grid lines and the number of multiple grid iterations, etc., which are used to optimize the initialization settings of the CFD numerical simulation model. The aerodynamic performance simulation is taken as the optimization target, such as lift coefficient, drag coefficient, lift-drag ratio, etc. During the simulation process, aerodynamic performance is the key indicator that the model needs to achieve. According to the aerodynamic force prediction knowledge and physical formula, the constraint conditions in the optimization process are set. Common constraints include: flight safety constraints (such as avoiding excessive aerodynamic load), structural strength constraints (such as the carrying capacity of the wing), range, fuel consumption, and other economic constraints. The adaptive optimization algorithm will adjust the settings of the CFD model (such as grid density, physical model parameters, etc.) through repeated iterations until the optimization target is improved.

[0071] S303, selecting a physical model

[0072] According to the characteristics of the target large transport equipment and the expected flight environment, a suitable physical model is selected. Common physical models include: turbulence model: select a suitable turbulence model (such as k-ε model, k-ω model, large eddy simulation (LES), etc.) according to the characteristics of the flow field. Heat transfer model: if temperature effects need to be simulated, select a suitable heat transfer model, such as convection-conduction model. Shock wave and high-speed flow model: if the equipment involves high-speed flight, a suitable aerodynamic model for supersonic or subsonic speed needs to be used, considering phenomena such as shock waves and shock waves. Structure-aerodynamic coupling model: for equipment involving aeroelastic effects, a model that couples structure and fluid needs to be selected.

[0073] S304, numerical simulation of the aerodynamic performance of the target large transport equipment

[0074] The adaptive optimization algorithm is integrated with the CFD numerical simulation model to form a closed-loop system. The optimization algorithm adjusts the simulation parameters through iteration and evaluates the optimization target after each simulation result. Using the selected physical model and the optimized CFD model, the aerodynamic performance simulation under different flight conditions and weather conditions is carried out. This process includes: changes in flight speed, changes in altitude, changes in weather conditions (such as different wind speeds, temperatures, etc.). The optimized aerodynamic performance data, such as lift, drag, stability coefficients, etc., are output for subsequent design and verification.

[0075] S305, calibration of the adaptive CFD numerical simulation model

[0076] The historical flight data (such as the aerodynamic performance data measured during actual flight) is compared with the numerical simulation results. The purpose of this step is to verify the accuracy of the CFD numerical simulation model. According to the results of the verification comparison, the accuracy of the CFD model is evaluated. If there is a large deviation between the numerical simulation results and the actual flight data, the CFD numerical simulation model needs to be calibrated. This can include modifying the physical model parameters, optimizing the weights of the algorithm, or re-adjusting the boundary conditions, etc. The calibrated model can be re-input into the optimization process to form a more accurate simulation model to improve the simulation accuracy and optimization results.

[0077] The adaptive optimization algorithm optimizes the initialization settings of the CFD numerical simulation model based on the multi-agent proximal policy model and the local alternating genetic model. The working method of the adaptive optimization algorithm is shown in FIG. 1, which includes the following steps: Figure 3

[0078] S3021, initialize the agent and the policy function

[0079] The initialization setting variables in the CFD numerical simulation model are regarded as multiple agents, and each agent corresponds to a variable in the CFD simulation (such as initial velocity, initial pressure, boundary condition, etc.). The task of the agent is to optimize these initialization variables by interacting with the environment to improve the simulation results. Each agent has a policy function and a value function. The policy function is used to determine the action (i.e. adjust the value of the initialization variable) that the agent should take in a specific state, while the value function is used to evaluate the "good" or "bad" degree of the current state. According to the current state of the CFD numerical simulation model (for example, the temperature, pressure, and velocity distribution of the current flow field) and the historical experience of the initialization setting variables, the value of the current initialization variable is adjusted. The role of the value function is to evaluate whether the strategy taken by the current agent helps to improve the quality of the simulation results, which is usually measured by the return value.

[0080] S3022, optimize the policy function and the value function of the agent by using the proximal policy optimization method

[0081] The agent executes the adjustment action in the running environment according to the policy function, and observes the results (i.e. the new state or the return) generated thereby. The proximal policy optimization method constrains the update step of the policy and value functions by maximizing the policy loss function and minimizing the value loss function, so as to ensure the stability of the learning process.

[0082] The policy loss function is used to limit the step size of policy update to avoid excessive changes that lead to unstable learning. The expression of the policy loss function is:

[0083]

[0084] In formula (1), L​P (θ) represents the update loss function of the policy function parameter θ, which measures the difference between the new policy and the old policy, E t represents the expected value of all states and actions at time step t, r t (θ) represents the policy function parameter at time step t in state s t takes action a t The probability ratio of the policy is the ratio of the new and old policies (i.e., the probability ratio of the new and old policies), which is used to calculate the difference between the new and old policies, represents the advantage estimate at time step t, which measures the advantage of taking action in state relative to the average action selection, ∈ represents the hyperparameter, which limits the probability ratio r t (θ) is within a certain range, ensuring that the change between the new and old policies does not exceed a certain threshold, clip(r t (θ),1-v,1+∈) represents the probability ratio r t (θ) is limited between 1-∈ and 1+∈.

[0085] The value loss function is used to optimize the value function so that it can better predict the expected return of each state, and the expression of the value loss function is:

[0086]

[0087] In formula (2), represents the update loss function of the value function parameter , represents the value function of the agent in state s t , represents the return value obtained at time step t, E t represents the expected value of all states and actions at time step t, which is used to estimate the average return of all actions in a certain state.

[0088] S3023, update the policy function parameters and value function parameters using the local alternating genetic model

[0089] According to the size and complexity of the problem, as well as the number and ability of the agents, the agents are reasonably divided into multiple groups. Ensure that the number of agents in each group is moderate to facilitate effective cooperation and competition. Set the initial parameter values for each agent, including the policy function parameters and the value function parameters. Initialize the parameters of the genetic algorithm, such as population size, crossover probability, mutation probability, etc.

[0090] Within each group, agents collaborate by sharing information and coordinating their actions to solve the problem. Communication protocols are used to ensure the reliability and real-time nature of information transmission, as well as task allocation and coordination mechanisms to optimize cooperation between agents. The fitness of the agents (i.e., their effectiveness in solving the problem) is used to select superior agents as parents. Within the group, the selected parent agents are subjected to crossover operations to produce new child agents. Crossover operations can include single-point crossover, multi-point crossover, etc. With a certain probability, the genes of the child agents are subjected to mutation operations to increase the diversity of solutions.

[0091] Between different groups, competition is used to drive the entire optimization process in a better direction. A tournament or championship approach is used to evaluate the performance of each group and rank them based on performance. Based on their performance in the competition, the policy function parameters and value function parameters of the agents are updated. For agents that perform well, their parameter values can be retained as the starting point for the next iteration. For agents that perform poorly, they can be adjusted or replaced according to the rules of genetic algorithms.

[0092] The process of local collaboration, overall competition, and parameter updating is repeated until a predetermined number of iterations is reached or a certain convergence condition is met. In each iteration, the performance of the agents is evaluated, and the parameters and strategies of the genetic algorithm are adjusted based on the evaluation results. Through continuous iteration and optimization, a set of excellent agent parameter values can be obtained, which can be used to solve similar problems or as a starting point for further research.

[0093] S3024, Calculation and Feedback of CFD Numerical Simulation Model

[0094] The updated initialization setting variable values after iteration are input into the CFD numerical simulation model for numerical calculation. The calculation results (such as aerodynamic performance parameters) are fed back to the agent as the current state of the CFD numerical simulation model, which is used by the agent to update its strategy and value function based on the new state and continue the optimization process. The simulation results fed back usually include key indicators such as aerodynamic performance, fluid flow characteristics, temperature field, etc. The agent adjusts the strategy according to these feedback to achieve more accurate simulation results

[0095] S3025, Repeat Iteration Until Reach Predetermined Goal

[0096] Steps S3022 and S3023 are repeated until the simulation results meet the predetermined aerodynamic performance requirements. This process is a dynamic adjustment process that gradually optimizes the initialization settings through multiple iterations, continuously improving the calculation accuracy and efficiency of the CFD model. When the model reaches the predetermined aerodynamic performance simulation goal (for example, the simulation error is below a certain threshold or the performance indicators meet the requirements), the algorithm terminates and outputs the optimized initialization settings.

[0097] Step four, aerodynamic performance simulation and knowledge graph iterative update

[0098] The historical flight data of the target large transport equipment under different flight conditions and weather conditions is input into the optimized adaptive CFD numerical simulation model for aerodynamic performance simulation. During the simulation, the adaptive CFD numerical simulation model will automatically adjust the grid division and calculation domain management according to the input parameters to ensure the accuracy and stability of the simulation results. These simulations will generate a large amount of simulation data, including lift, drag, pitch moment, etc. aerodynamic parameters. According to the comparison of simulation results and historical flight data, the aerodynamic force prediction knowledge graph is updated in real time. After each simulation, analyze the simulation data, extract key indicators such as aerodynamic coefficients, lift-drag ratio, lateral force coefficient, pitch moment, etc. Add the extracted key indicators as new nodes to the aerodynamic force prediction knowledge graph, extract new aerodynamic rules from the simulation data through machine learning algorithms (such as regression analysis, Kriging interpolation method, etc.), and update the network structure of the graph according to the relationship and action principle between the nodes. According to the relationship between the newly added nodes and the original nodes, the knowledge graph is iteratively learned and optimized to improve the prediction accuracy and accuracy.

[0099] By continuously updating the aerodynamic force prediction knowledge graph, the prediction model can have better adaptability and accuracy, providing more scientific and effective aerodynamic force prediction methods and technical support for the design, production and operation of large transport equipment.

[0100] Step five, deep reinforcement learning model training and aerodynamic performance prediction

[0101] A deep reinforcement learning (DRL) model is constructed to predict the aerodynamic performance of large transport equipment under various flight conditions. Deep reinforcement learning is trained from historical data and simulation results to optimize decision-making strategies to make prediction results more accurate. State space: flight conditions, weather conditions, structural parameters, etc. Action space: adjust flight parameters (such as speed, angle of attack, etc.). Reward function: evaluate the predicted aerodynamic performance through actual aerodynamic data and simulation results, and give rewards or punishments. The updated aerodynamic force prediction knowledge graph is an important input to the reinforcement learning model, so that the model can refer to known aerodynamic laws during training to quickly converge to accurate aerodynamic force prediction results. After training, the deep reinforcement learning model can predict the aerodynamic performance of the target large transport equipment under different flight conditions in real time, and provide flight control systems, design optimization, or pilot decision support for practical applications.

[0102] A deep reinforcement learning (DRL) method is used to further predict aerodynamic performance. Deep reinforcement learning learns the dynamic rules in simulation data and continuously optimizes the decision-making strategy in the prediction process. Aerodynamic features such as lift, drag, and side force are extracted from CFD simulation results as input features for the model. A reward function is designed based on target aerodynamic performance (such as lift-to-drag ratio, fuel efficiency, flight stability, etc.), and the reinforcement learning algorithm optimizes the control strategy through continuous iteration. In the deep reinforcement learning model, a graph neural network (GNN) is used to perform knowledge reasoning on the aerodynamic force prediction knowledge graph. The graph neural network can handle complex dependencies between nodes and can mine and model complex relationships between internal and external aerodynamic characteristics of large transport equipment. It can be used as a graph embedding module in the reinforcement learning model to reason about structural and flight condition parameters and improve the prediction ability of the model. Through GNN, deeper rules can be extracted from historical flight data and the aerodynamic force prediction knowledge graph, and the aerodynamic behavior of the equipment under different conditions can be mined, thereby improving the accuracy and reliability of the prediction model.

[0103] An aerodynamic force prediction device for large transport equipment, as shown in FIG. 1, includes: Figure 4

[0104] A data acquisition module is responsible for collecting structural parameters (such as size, shape, weight distribution, etc.) of the target large transport equipment, flight condition parameters (such as speed, altitude, Mach number, angle of attack, side slip angle, etc.), meteorological condition parameters (such as wind speed, wind direction, temperature, humidity, etc.), and historical flight data. Through sensors, data recording devices, or data interfaces, real-time or periodic data acquisition from the target large transport equipment and its related environment is performed, and the data is transmitted to the cloud server for processing and storage.

[0105] A knowledge graph construction module uses the cloud server to process data and construct an aerodynamic force prediction knowledge graph. Key information is extracted from structural parameters, flight conditions, meteorological conditions, and historical flight data as nodes of the aerodynamic force prediction knowledge graph. Aerodynamic knowledge, aerodynamic formulas, and physical models are integrated to form a multidimensional prediction framework. This may include fluid mechanics principles, aerodynamic equations, wind tunnel test data, etc. By associating nodes and integrating knowledge, a complete aerodynamic force prediction knowledge graph is constructed.

[0106] ​The numerical simulation module constructs an adaptive CFD numerical simulation model based on the aerodynamic force prediction knowledge graph and the structural parameters of the target large transport equipment for simulating aerodynamic performance. According to the structural parameters of the target large transport equipment and the aerodynamic force prediction knowledge graph, an adaptive CFD numerical simulation model is constructed. The aerodynamic performance of the target large transport equipment under different flight conditions and weather conditions is simulated using the CFD numerical simulation method. Based on the simulation process data, the aerodynamic force prediction knowledge graph is iteratively updated to improve its accuracy and reliability.

[0107] The aerodynamic force prediction module uses a deep reinforcement learning model to predict the aerodynamic performance of large transport equipment and trains it using the updated aerodynamic force prediction knowledge graph. The updated aerodynamic force prediction knowledge graph is used as training data to train the deep reinforcement learning model. The trained deep reinforcement learning model is used to predict the aerodynamic performance of the target large transport equipment under different conditions.

[0108] The user interaction module allows users to view the real-time state of the adaptive CFD numerical simulation model and the aerodynamic performance prediction results through a graphical interface, adjust flight conditions and weather conditions, and then automatically generate an aerodynamic performance report for the large transport equipment based on the simulation and prediction results. A user-friendly graphical interface is designed to display the real-time state of the adaptive CFD numerical simulation model and the aerodynamic performance prediction results. Users can adjust flight conditions and weather conditions through the graphical interface to observe their impact on aerodynamic performance. Based on the simulation and prediction results, an aerodynamic performance report for the large transport equipment is automatically generated, including key indicators such as aerodynamic drag, lift, and stability analysis.

[0109] In practical applications, these modules work together to form a complete aerodynamic force prediction system. Users can input the structural parameters and flight conditions of the target large transport equipment through the user interaction module, and the system will automatically perform data processing, knowledge graph construction, numerical simulation, aerodynamic force prediction, and report generation. This not only helps reduce the cost and time of aerodynamic force prediction, but also improves the accuracy and reliability of the prediction, providing strong support for the design and optimization of large transport equipment.

[0110] The output of the data acquisition module is connected to the input of the knowledge graph construction module, the knowledge graph construction module is bidirectionally connected to the numerical simulation module, the output of the numerical simulation module is connected to the input of the aerodynamic force prediction module, the output of the knowledge graph construction module is connected to the input of the aerodynamic force prediction module, the outputs of the data acquisition module and the aerodynamic force prediction module are connected to the input of the user interaction module, and the numerical simulation module is bidirectionally connected to the user interaction module.

[0111] The technical scheme of the present application has been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to the related technical features without departing from the principles of the present application, and the technical schemes after the changes or replacements will all fall within the protection scope of the present application.

Claims

1. A method of predicting aerodynamic forces of a large transport vehicle, characterized by, The method comprises the following steps: Step one, obtaining the structure parameters of the target large transport equipment, and historical flight data of large transport equipment with different structure parameters under different flight condition parameters and meteorological condition parameters; Step two, constructing an aerodynamic force prediction knowledge graph using a cloud server, wherein the cloud server extracts the structure parameters, flight condition parameters, meteorological condition parameters and historical flight data of multiple large transport equipment from the data obtained in step one as nodes of the aerodynamic force prediction knowledge graph, and integrates aerodynamic force prediction discipline knowledge and aerodynamic force formulas to form the network structure of the aerodynamic force prediction knowledge graph; Step three, constructing an adaptive CFD numerical simulation model based on the aerodynamic force prediction knowledge graph and the structure parameters of the target large transport equipment, and optimizing the adaptive CFD numerical simulation model using an adaptive optimization algorithm; In the step three, the construction of the adaptive CFD numerical simulation model comprises the following steps: S301, receiving the structure parameters of the target large transport equipment, and extracting the structure parameters, flight condition parameters and meteorological condition parameters matched with the target large transport equipment from the aerodynamic force prediction knowledge graph, and initializing the CFD numerical simulation model based on the extracted parameters, wherein the initialization setting at least includes geometry model setting, mesh division and boundary condition setting; S302, setting the adaptive optimization algorithm according to the structure parameters of the target large transport equipment and the simulation target, and optimizing the initialization setting of the CFD numerical simulation model using the adaptive optimization algorithm, wherein the adaptive optimization algorithm sets the aerodynamic performance simulation as the optimization target, and sets the constraint condition based on the aerodynamic force prediction discipline knowledge and the aerodynamic force formula; S303, selecting a physical model according to the characteristics of the target large transport equipment and the expected flight environment, so as to reflect the aerodynamic performance of the target large transport equipment in high-speed operation; S304, integrating the adaptive optimization algorithm and the CFD numerical simulation model to form the adaptive CFD numerical simulation model, and numerically simulating the aerodynamic performance of the target large transport equipment under different flight conditions and meteorological conditions according to the selected physical model and the adaptive CFD numerical simulation model; S305, comparing the historical flight data with the numerical simulation, and calibrating the adaptive CFD numerical simulation model according to the verification comparison result; The adaptive optimization algorithm optimizes the initialization setting of the CFD numerical simulation model based on the multi-agent proximal policy model and the local alternating genetic model, and the working method of the adaptive optimization algorithm comprises the following steps: S3021, regarding the initialization setting variable in the CFD numerical simulation model as an agent, and regarding the running environment of the agent as the CFD numerical simulation model, defining the strategy function and the value function of the agent using the aerodynamic force prediction discipline knowledge and the aerodynamic force formula, wherein the strategy function adjusts the value of the current initialization setting variable according to the current state of the CFD numerical simulation model and the historical experience of the initialization setting variable, and the value function evaluates the value of the current initialization setting variable according to the current state of the CFD numerical simulation model and the strategy function. S3022, the agent executes the adjustment action in the running environment according to the policy function, and updates the policy function parameters and the value function parameters of the agent based on the proximal policy optimization method, the proximal policy optimization method updates the policy and the value function through the policy loss function and the value loss function to constrain the update step of the policy and the value function; The expression of the policy loss function is: In formula (1), L P (θ) represents the update loss function of the policy function parameter θ, which is used to measure the difference between the new policy and the old policy, E t represents the expected value of all states and actions at time step t, r t (θ) represents the policy function parameter at time step t in state s t takes action a t The probability ratio of the policy, represents the advantage estimate value at time step t, which measures the advantage of taking action in state relative to the average action selection, ∈ represents the hyperparameter, which is used to limit the probability ratio r t (θ) range, clip(r t (θ),1-∈,1+∈) represents limiting the probability ratio r t (θ) between 1-∈ and 1+∈, and the expression of the value loss function is: In Equation (2), denotes the update loss function of the value function parameters denotes the value function of the agent in state s t at time step t, denotes the reward value obtained at time step t, E t denotes the expected value of all states and actions at time step t; S3023, in the process of updating the policy function parameters and the value function parameters, the agent is divided into multiple groups, and a local alternating genetic model is used for local cooperation and overall competition among multiple agents; S3024, input the initialized setting variable value after updating iteration into the CFD numerical simulation model for numerical calculation, and feed back the calculation result to the agent as the current state of the CFD numerical simulation model; S3025, repeat step S3022 and step S3023 until the predetermined aerodynamic performance simulation performance is reached; Step four, using the optimized adaptive CFD numerical simulation model to simulate the aerodynamic performance of the target large transport equipment under different flight conditions and weather conditions when running at high speed, and based on the simulation process data, the aerodynamic force prediction knowledge graph is iteratively updated; Step five, using the deep reinforcement learning model to predict the aerodynamic performance of the large transport equipment, and using the updated aerodynamic force prediction knowledge graph to train the deep reinforcement learning model.

2. The method of claim 1, wherein In the step one, the structure parameters at least include the fuselage length, the fuselage width, the fuselage height, the fuselage shape, the wing shape, the wing area, the vertical tail area, the gravity center position, the engine thrust, the Mach number and the attack angle along the curved surface, the weather condition parameters at least include the wind speed, the wind direction, the temperature, the humidity and the air pressure, the flight condition parameters at least include the flight height, the flight speed, the flight attitude and the maneuvering action, and the flight data at least include the lift, the drag, the lateral force and the aerodynamic moment.

3. The method of claim 1, wherein In the second step, the cloud server includes a data preprocessing unit, a domain classification unit, an entity recognition and linking unit, a relationship extraction unit, a knowledge organization representation unit, and an incremental updating unit. The data preprocessing unit uses a natural language processor to extract structural parameters, flight condition parameters, meteorological condition parameters, and historical flight data of large transport equipment, and cleans and organizes the extracted parameters. The domain classification unit classifies the extracted parameters according to subject domain distribution through a text domain mining model LDA. The entity recognition and linking unit automatically identifies aerodynamic entities in the extracted parameters through a named entity recognition task NER, and links the aerodynamic entities to corresponding entities in the aerodynamic force prediction knowledge graph through an entity linking task EL. The relationship extraction unit automatically extracts relationships between entities from aerodynamic force prediction subject knowledge and aerodynamic force formulas through a relationship extraction algorithm, and stores them in the aerodynamic force prediction knowledge graph to form a network structure. The relationships between entities at least include the relationship between structural parameters and aerodynamic performance, the relationship between flight condition parameters and aerodynamic changes, and the relationship between meteorological condition parameters and aerodynamic changes. The knowledge organization representation unit stores the extracted parameters, aerodynamic force prediction subject knowledge, and aerodynamic formulas in a graph database or triple storage format. The incremental updating unit updates the aerodynamic force prediction knowledge graph in real time through a time window mechanism.

4. The method of claim 1, wherein The deep reinforcement learning model extracts aerodynamic force prediction features from adaptive CFD numerical simulation results as aerodynamic force prediction input features, and uses a graph neural network for knowledge reasoning of the aerodynamic force prediction knowledge graph to mine the aerodynamic characteristics inside the large transport equipment.

5. A large-scale transportation equipment aerodynamic force prediction device for use in the large-scale transportation equipment aerodynamic force prediction method according to any one of claims 1 to 4, characterized by It comprises: a data acquisition module that acquires structural parameters, flight condition parameters, meteorological condition parameters, and historical flight data of a target large transport equipment, and transmits them to a cloud server for processing and storage; a knowledge graph construction module that uses a cloud server to construct an aerodynamic force prediction knowledge graph, which extracts information from structural parameters, flight conditions, meteorological conditions, and historical flight data as nodes of the aerodynamic force prediction knowledge graph, and integrates aerodynamic subject knowledge, aerodynamic formulas, and physical models to form a multi-dimensional prediction framework; a numerical simulation module that constructs an adaptive CFD numerical simulation model based on the aerodynamic force prediction knowledge graph and the structural parameters of the target large transport equipment, for simulating aerodynamic performance, and iteratively updates the aerodynamic force prediction knowledge graph based on simulation process data; an aerodynamic force prediction module that uses a deep reinforcement learning model to predict the aerodynamic performance of the large transport equipment, and trains it using the updated aerodynamic force prediction knowledge graph; a user interaction module that views the real-time state of the adaptive CFD numerical simulation model and the aerodynamic performance prediction results through a graphical interface, adjusts the flight conditions and meteorological conditions, and then automatically generates an aerodynamic performance report of the large transport equipment based on the simulation and prediction results.

Images