Aircraft unsteady aerodynamic modeling method and system based on neural network

Abstract

The invention relates to the technical field of aerodynamic simulation, in particular to an aircraft unsteady aerodynamic modeling method and system based on a neural network. The method comprises the following steps: establishing sample data about unsteady aerodynamic characteristics of an aircraft, and selecting characteristic parameters from the sample data to form a data set; carrying out normalization on the characteristic parameters; constructing a long-short-term memory neural network model to describe the relationship between the flight conditions and aerodynamic characteristics of aircrafts with different airfoils; adopting a quantum particle swarm optimization algorithm to optimize and adjust hyper-parameters of the long-short term memory neural network model; and based on the optimized hyper-parameters, training the long and short term memory neural network model to obtain an optimized aircraft unsteady aerodynamic model. According to the method, tedious aerodynamic effect decomposition and a traditional mathematical modeling process are avoided, the complete characteristics of the unsteady aerodynamic effect are reserved, the fidelity of the model is improved, and the simulation capability of the aerodynamic model on aerodynamic characteristics is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of aerodynamic simulation, and more specifically to a method and system for modeling aircraft unsteady aerodynamics based on neural network. Background Art

[0002] The aerodynamic model is one of the most important mathematical models of flight simulators. The accurate simulation of aerodynamic forces / torques involves the flight quality and fidelity of the simulator, and is an important evaluation basis for the simulator to pass technical appraisal. However, with the development of modern aircraft, the improvement of maneuverability has led to the continuous expansion of the flight envelope, and the worsening of flight conditions has made the aerodynamic model increasingly nonlinear and unsteady. How to accurately establish an unsteady aerodynamic model is of great significance in the field of aircraft design or simulator modeling.

[0003] Traditional unsteady aerodynamic modeling methods mainly rely on mathematical models, and establish the functional relationship between aerodynamic forces and flight state parameters through the aerodynamic physical mechanisms and statistical laws of the aircraft. For example, Wang Qing et al. established a high angle of attack unsteady aerodynamic model using the incremental method based on the flow separation and vortex structure rupture caused by high angles of attack. By comparing with wind tunnel tests, this method can reflect the unsteady aerodynamic characteristics under high angles of attack to a certain extent. Although these methods have clear physical meanings, for strong nonlinear and time-varying situations, it is usually necessary to couple complex physical mechanisms and polynomials, which makes modeling difficult and requires high data quality.

[0004] In recent years, with the development of artificial intelligence algorithms and hardware facilities, more and more studies have begun to use artificial intelligence methods to establish aerodynamic models. This method does not require understanding of complex aerodynamic mechanisms, but only uses artificial intelligence models to mine the laws hidden behind the data. Various neural network algorithms have been widely used in aerodynamic modeling. For example, Wang Peng et al. used traditional neural networks and back propagation algorithms to replace traditional aerodynamic models, established a three-layer neural network model with a single hidden layer and a four-layer neural network model with double hidden layers, and predicted the effects of free flow Mach number, angle of attack, sideslip angle and other parameters on the aerodynamic characteristics of a pointed wedge forebody aircraft. Compared with traditional methods, artificial intelligence methods can more efficiently handle complex nonlinear aerodynamic characteristics and have higher prediction accuracy.

[0005] However, the application of artificial intelligence methods, especially LSTM neural networks, in aerodynamic modeling also faces certain challenges. The prediction accuracy and generalization ability of LSTM neural networks are easily affected by the parameters of the network model. Generally, the training of LSTM neural networks relies on the experience of researchers to manually adjust, which makes it difficult to obtain the global optimal solution of model parameters.

[0006] In general, the traditional aerodynamic modeling method has a simple structure, but has limitations in modeling complex aerodynamic characteristics. For some strongly nonlinear and time-varying situations, polynomials of different physical mechanisms need to be coupled, which makes modeling difficult and time-consuming. The model is built based on some assumptions and simplifications, has relatively low fidelity, and has high requirements for the quality of aerodynamic data. Although artificial intelligence methods can overcome some of the shortcomings of traditional modeling methods, the selection and training process of model parameters still face challenges. Summary of the invention

[0007] The purpose of the present invention is to provide a method and system for modeling aircraft unsteady aerodynamics based on neural network, so as to solve the problems of poor accuracy and low efficiency of unsteady aerodynamic modeling in the prior art.

[0008] In order to achieve the above object, the present invention provides a method for modeling aircraft unsteady aerodynamics based on a neural network, comprising the following steps:

[0009] Establish sample data on the unsteady aerodynamic characteristics of the aircraft, and select characteristic parameters from the sample data to form a data set;

[0010] Normalize the feature parameters and divide the data set into training set and validation set;

[0011] Construct a long short-term memory neural network model to describe the relationship between flight conditions and aerodynamic characteristics of aircraft with different wing shapes;

[0012] The quantum particle swarm optimization algorithm is used to optimize and adjust the hyperparameters of the long short-term memory neural network model;

[0013] Based on the optimized hyperparameters, the long short-term memory neural network model is trained to obtain an optimized aircraft unsteady aerodynamic model, and the aircraft unsteady aerodynamic model is used to simulate aircraft aerodynamics.

[0014] In some embodiments, the step of establishing sample data on unsteady aerodynamic characteristics of the aircraft further comprises:

[0015] Using computational fluid dynamics simulation technology, the aerodynamic characteristics of the pitch reciprocating motion of the specified wing aircraft are numerically simulated to generate aircraft aerodynamic characteristics data under different flight conditions;

[0016] The aircraft aerodynamic characteristic data include aerodynamic response data under various different angle of attack mean values, angle of attack amplitudes, reduction frequencies, and Mach number factors.

[0017] In some embodiments, the step of numerically simulating the aerodynamic characteristics of the pitch reciprocating motion of the specified wing-shaped aircraft further comprises:

[0018] The pitch reciprocating motion is represented as a sine function, and the frequency, mean value, and amplitude of the sine function are adjusted to optimize the simulation data.

[0019] In some embodiments, the step of numerically simulating the aerodynamic characteristics of the pitch reciprocating motion of the specified wing aircraft using computational fluid dynamics simulation technology further includes:

[0020] Generate the computational domain and perform meshing, set the distance between the far field inlet and the leading edge of the airfoil, the distance between the outlet and the trailing edge of the airfoil, and the distance between the lateral boundary and the upper and lower surfaces of the airfoil;

[0021] Use unstructured grid to generate fluid domain, set the mesh size of airfoil surface, set the first layer height, growth rate, maximum number of generated layers of boundary layer and total number of grids of boundary layer network;

[0022] The calculations were performed using a transient density-based solver, shear stress transport k-ω model, and ideal gas model;

[0023] Set the ambient temperature and the incoming flow direction, use the dynamic grid method to achieve the periodic motion of the airfoil in the pitch direction, and set the airfoil angle of attack and pitch angle under the initial conditions;

[0024] The implicit Roche single equation model-flux driven scheme format is used for discrete calculation of flux terms, the least squares lattice center format is used for gradient calculation, the second-order upwind format is used for spatial discrete calculation, and the double time step first-order implicit format is used for temporal discrete calculation.

[0025] In some embodiments, the step of selecting feature parameters from the sample data to form a data set further comprises:

[0026] Determining characteristic parameters of the long short-term memory neural network model according to the flight mission, the characteristic parameters including model input parameters for describing flight conditions and model output parameters for describing aerodynamic characteristics;

[0027] The model input parameters include flight altitude, flight Mach number, angle of attack and rate of change of angle of attack, and the model output parameters include lift coefficient, drag coefficient and pitch moment coefficient.

[0028] In some embodiments, the step of normalizing the characteristic parameters further comprises:

[0029] The min-max normalization method is used to normalize the characteristic parameter data.

[0030] In some embodiments, the constructed long short-term memory neural network model includes a cell state, an output signal, and a gating unit:

[0031] Capturing long-term dependencies between aircraft flight conditions and aerodynamic properties through cell states and gating units;

[0032] The input data of the long short-term memory neural network model is composed of model input parameters;

[0033] The output signal of the long short-term memory neural network model is composed of model output parameters.

[0034] In some embodiments, the gating unit includes an input gate, an output gate, and a forget gate:

[0035] The input gate is used to selectively memorize the information of the input data at the current moment and update the cell state;

[0036] The output gate is used to control the output signal of the long short-term memory neural network model at the current moment;

[0037] The forget gate is used to control the retention or discarding of the cell state at the previous moment.

[0038] In some embodiments, the hyperparameters of the long short-term memory neural network model include:

[0039] The number of network layers, the number of neurons in each network layer, the length of historical data, the length of predicted data, and the batch size of training data.

[0040] In some embodiments, the quantum particle swarm optimization algorithm optimizes the hyperparameters of the long short-term memory neural network by updating particle positions and evaluating fitness, wherein the particle fitness is calculated by mean square error.

[0041] In some embodiments, the quantum particle swarm optimization algorithm further comprises:

[0042] Initialize the relevant parameters of the quantum particle swarm optimization algorithm;

[0043] The particle position parameters are used to train the long short-term memory neural network and the particle fitness of each training is calculated;

[0044] According to the particle fitness value, determine the local optimal solution and global optimal solution of the individual particle, and update the position of the particle;

[0045] Determine whether the stopping condition is met. If so, output the optimal particle position. If not, enter the long short-term memory neural network training again using the particle position parameters and perform iterative calculation again.

[0046] In some embodiments, in the step of initializing relevant parameters of the quantum particle swarm optimization algorithm, the relevant parameters include particle dimension, particle population size, hyperparameters of the long short-term memory neural network to be optimized, and the maximum number of iterations.

[0047] In some embodiments, the stop condition includes:

[0048] The particle fitness value reaches a preset threshold, and / or

[0049] The number of iterations reached the maximum number of iterations.

[0050] In order to achieve the above object, the present invention provides an aircraft unsteady aerodynamic modeling system based on a neural network, comprising a memory and a processor:

[0051] The memory is used to store instructions executable by the processor;

[0052] The processor is used to execute the instructions to implement the method described above.

[0053] In order to achieve the above object, the present invention provides a computer storage medium on which computer instructions are stored, wherein when the computer instructions are executed by a processor, the above method is executed.

[0054] The neural network-based aircraft unsteady aerodynamic modeling method and system proposed in the present invention construct an aircraft unsteady aerodynamic model through an LSTM neural network, and use a quantum particle swarm optimization algorithm (QPSO) to automatically search for the optimal solution of the hyperparameters of the LSTM neural network, thereby avoiding cumbersome aerodynamic effect decomposition and traditional mathematical modeling processes, retaining the complete characteristics of the unsteady aerodynamic effect, thereby improving the realism of the model and enhancing the aerodynamic model's simulation capability of unsteady and nonlinear aerodynamic characteristics. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] The above and other features, properties and advantages of the present invention will become more apparent through the following description in conjunction with the accompanying drawings and embodiments, in which the same reference numerals always represent the same features, wherein:

[0056] Figure 1 A step diagram of a method for modeling aircraft unsteady aerodynamics based on a neural network according to an embodiment of the present invention is disclosed;

[0057] Figure 2 A flow chart of a method for modeling aircraft unsteady aerodynamics based on a neural network according to an embodiment of the present invention is disclosed;

[0058] Figure 3 A schematic diagram of a grid near an airfoil surface according to an embodiment of the present invention is disclosed;

[0059] Figure 4 A schematic diagram of an LSTM basic unit structure according to an embodiment of the present invention is disclosed. DETAILED DESCRIPTION

[0060] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the invention and are not used to limit the invention.

[0061] The present invention proposes a neural network-based aircraft unsteady aerodynamic modeling method, which uses artificial intelligence methods to establish an aerodynamic model. It does not rely on complex aerodynamic mechanisms, but instead uses artificial intelligence models to mine the laws hidden behind the data and directly maps the flight state into aerodynamic force.

[0062] Compared with traditional aerodynamic modeling methods, the neural network model in the artificial intelligence method adopts a data-driven approach, avoiding the in-depth understanding of complex aerodynamic mechanisms, thereby lowering the technical threshold. In addition, the neural network model can fit any nonlinear function, so it has a stronger ability to express unsteady and nonlinear aerodynamic characteristics, thereby improving the modeling accuracy.

[0063] The present invention further proposes a neural network-based aircraft unsteady aerodynamics modeling method, which combines quantum particle swarm optimization (QPSO) technology to automatically search for the optimal solution of the long short-term memory neural network (LSTM) model hyperparameters, avoiding the arduous work of manually adjusting the LSTM hyperparameters and significantly improving the modeling efficiency.

[0064] Figure 1 A step diagram of a method for modeling an unsteady aerodynamics of an aircraft based on a neural network according to an embodiment of the present invention is disclosed, such as Figure 1 As shown, the present invention proposes a method for modeling aircraft unsteady aerodynamics based on a neural network, comprising the following steps:

[0065] Step S1, establishing sample data on unsteady aerodynamic characteristics of an aircraft, and selecting characteristic parameters from the sample data to form a data set;

[0066] Step S2, normalizing the feature parameters and dividing the data set into a training set and a validation set;

[0067] Step S3, constructing a long short-term memory neural network model to describe the relationship between flight conditions and aerodynamic characteristics of aircraft with different wing shapes;

[0068] Step S4, using quantum particle swarm optimization algorithm to optimize and adjust the hyperparameters of the long short-term memory neural network model;

[0069] Step S5, based on the optimized hyperparameters, the long short-term memory neural network model is trained to obtain an optimized aircraft unsteady aerodynamic model, wherein the aircraft unsteady aerodynamic model is used to simulate aircraft aerodynamics.

[0070] The present invention proposes a neural network-based aircraft unsteady aerodynamic modeling method, which uses artificial intelligence technology to establish an aircraft unsteady aerodynamic model. It does not require in-depth analysis of complex aerodynamic mechanisms, but instead uses artificial intelligence models to mine the laws hidden behind the data and directly maps relevant parameters such as flight status and configuration changes into aerodynamic forces, thereby reducing the difficulty of modeling.

[0071] By adopting the long short-term memory neural network (LSTM) model, the present invention can learn data from flight tests, wind tunnel experiments and numerical simulations, extract aircraft aerodynamic characteristics, such as the wing-body lift coefficient, the dynamic effect of landing gear retraction and extension, and the aerodynamic effect caused by the three-axis angular velocity, and establish an unsteady aerodynamic model of the aircraft for the full envelope, avoiding the tedious aerodynamic effect decomposition and traditional mathematical modeling process.

[0072] The present invention utilizes the nonlinear fitting ability of neural networks to learn and mine strong nonlinear, coupled and time-correlated aerodynamic effects from training data. By establishing a mapping relationship of high nonlinearity, strong coupling and time correlation between aircraft aerodynamic effects and their influencing factors, the present invention avoids the simplification and assumption of aircraft aerodynamic models in traditional methods, retains the complete characteristics of unsteady aerodynamic effects, thereby improving the realism of the model and enhancing the simulation ability of the aerodynamic model for unsteady and nonlinear aerodynamic characteristics.

[0073] These steps will be described in detail below. It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other and related to each other to form a preferred technical solution.

[0074] Figure 2 A flowchart of a method for modeling aircraft unsteady aerodynamics based on a neural network according to an embodiment of the present invention is disclosed. Figure 1 Hehe Figure 2 As shown, the present invention proposes a neural network-based aircraft unsteady aerodynamic modeling method, and the construction steps are as follows.

[0075] Step S1, establishing sample data on the unsteady aerodynamic characteristics of the aircraft, and selecting characteristic parameters from the sample data to form a data set.

[0076] Sample data of aircraft unsteady aerodynamic characteristics may include data from flight tests, wind tunnel experiments, and numerical simulations.

[0077] In step S1 of this embodiment, a method for constructing an unsteady aerodynamic sample data set that matches an artificial intelligence-long short-term memory neural network (LSTM) model is proposed, which can take into account both data quality and data generation efficiency.

[0078] The dataset construction method uses computational fluid dynamics (CFD) simulation technology to numerically simulate the aerodynamic characteristics of the pitch reciprocating motion of the wing-shaped aircraft to be simulated, calculate the aerodynamic coefficients, and generate aircraft aerodynamic characteristic data under different flight conditions.

[0079] Furthermore, the pitch reciprocating motion is represented as a sine function. By adjusting the frequency, mean and amplitude of the sine function, the quality and generation efficiency of the simulation data are optimized, thereby providing multi-dimensional, high-quality training data for the LSTM model, helping the model to better capture and learn complex unsteady aerodynamic characteristics.

[0080] More specifically, the aircraft aerodynamic characteristic data generated by computational fluid dynamics simulation technology includes unsteady aerodynamic characteristic data, which includes a variety of different flight conditions, involving aerodynamic force responses under different angle of attack mean values, angle of attack amplitudes, reduction frequencies, Mach numbers, and other factors.

[0081] For example, by numerically simulating the aerodynamic characteristics of the pitch reciprocating motion of the NACA0012 airfoil, its aerodynamic characteristic data under different flight conditions are calculated.

[0082] The chord length of the NACA0012 airfoil is 1m, and the change of its angle of attack around the 1 / 4 chord length point is as follows:

[0083]

[0084] In the formula, For the airfoil Attack angle at all times, is the mean angle of attack, The angle of attack amplitude, k is the reduction frequency, is dimensionless time, t is time, is the free flow velocity, is the chord length of the airfoil, and C is the speed of sound.

[0085] The above expressions can be used to generate aerodynamic characteristic data based on data under different flight conditions.

[0086] Specifically, the reference axis of the pitch moment is at 1 / 4 chord length, and the mean angle of attack is The value range is [0º, 5º], the interval is 1º, and the angle of attack amplitude The value range of is [2.5º, 10.0º], the interval is 2.5º, the values ​​of the reduction frequency k are 0.05, 0.08, 0.12, and the values ​​of the Mach number Ma are 0.50, 0.60, 0.70, 0.75, 0.80, 0.85, 0.90, 0.95 respectively.

[0087] According to these different flight conditions, a total of 576 sets of unsteady aerodynamic sample data were generated. The total collection time of each set of aerodynamic sample data is 2 pitch motion cycles, and the collection interval of each data is 1 / 240 pitch motion cycle. These data provide rich unsteady aerodynamic characteristic data for training LSTM neural network, which helps to improve the accuracy and generalization ability of the model.

[0088] The calculation domain is generated and meshed using simulation software. The distance between the far-field inlet and the leading edge of the airfoil, the distance between the outlet and the trailing edge of the airfoil, and the distance between the lateral boundary and the upper and lower surfaces of the airfoil are all set to 25 times the chord length.

[0089] The fluid domain is generated using unstructured grids. The mesh size of the airfoil surface is set in the range of 1 to 8 mm. The first layer height of the boundary layer grid is set to 0.05 mm, the growth rate is 1.2, the maximum number of boundary layers generated is 40, and the total number of grids is 47,000. The grid distribution around the airfoil surface is as follows: Figure 3 shown.

[0090] During the calculation process, the transient density-based solver, the SST k-ω (shear stress transport k-ω model) model and the ideal gas model were used.

[0091] The ambient temperature is set to 288.15K, the incoming flow direction is horizontal, and the pitch direction periodic motion of the airfoil is realized by the moving mesh method. The airfoil is initially in the zero angle of attack state, and the pitch angle is 0º.

[0092] Among them, the transient density-based solver is a solver used for fluid dynamics calculations, which can handle density fields that change with time; the SST kw model is a turbulence model used to simulate turbulence phenomena in fluids and is suitable for complex flow conditions; the ideal gas model is a physical model that assumes that the gas follows the ideal gas law and is suitable for simulating gas behavior in engineering calculations; the dynamic mesh method is a numerical simulation technique that allows the mesh structure to change dynamically with the movement of the object during the calculation process.

[0093] The implicit Roe-FDS (Roe's One-Equation Model - Flux-Driven Scheme) format is used for discretization of convective flux terms, the least squares lattice center format is used for gradient calculation, the second-order upwind format is used for spatial discretization, and the double-time-step first-order implicit format is used for temporal discretization.

[0094] Among them, the implicit Roe-FDS format is a numerical method for processing convection terms in fluid flow. It can provide stable calculation results by solving it implicitly; the least squares grid center format is a method for calculating gradients, which uses the method of minimizing errors to obtain the gradient value of the grid center point. The second-order upwind format is a spatial discretization method that can better capture the characteristics of the flow direction and is suitable for numerical simulation of fluid dynamics problems. The dual-time-step first-order implicit format is a time integration method that improves the stability of the numerical solution through a virtual dual time step and is suitable for the simulation of unsteady flow problems.

[0095] The real-time iteration step is set to 1 / 240 of a pitch motion cycle. After the iteration is completed until the airfoil aerodynamic cycle is stable, it is iterated for two more cycles and the time history of the pitch angle and aerodynamic force is recorded.

[0096] After the unsteady aerodynamic characteristic data is generated, the characteristic parameters of the LSTM model are determined according to the flight mission requirements. The characteristic parameters include model input parameters for describing flight conditions and model output parameters for describing aerodynamic characteristics.

[0097] The model input parameters include flight conditions such as flight altitude, flight Mach number, angle of attack, rate of change of angle of attack, etc., while the model output parameters include aerodynamic characteristics such as lift coefficient, drag coefficient, pitch moment coefficient, etc.

[0098] The method for constructing an unsteady aerodynamic sample data set proposed in the present invention significantly improves the quality of simulated aerodynamic data and the efficiency of data generation, and provides efficient and accurate training data for subsequent model training.

[0099] Step S2, normalize the feature parameters and divide the data set into a training set and a validation set.

[0100] The purpose of normalization is to eliminate the impact of magnitude differences between different feature data on model convergence.

[0101] In this embodiment, the min-max normalization method is used to normalize the characteristic parameter data, and the aerodynamic characteristic data of different characteristics are converted to the [0, 1] interval.

[0102] The formula for min-max normalization is as follows:

[0103]

[0104] In the formula, is the normalized data of each column feature, x is the original data of each column feature, is the minimum value of each column of feature data, is the maximum value of each column of feature data.

[0105] In order to improve the generalization ability of the model and evaluate the accuracy of the model, the dataset needs to be divided into a training set and a validation set.

[0106] The training set is used for model training, while the validation set is used to evaluate the performance of the model and prevent overfitting.

[0107] Typically, the dataset is divided into 70% to 80% as a training set and the remaining 20% ​​to 30% as a validation set.

[0108] Step S3, constructing a long short-term memory neural network model to describe the relationship between flight conditions and aerodynamic characteristics of aircraft with different wing shapes;

[0109] Figure 4 A schematic diagram of the LSTM basic unit structure according to an embodiment of the present invention is disclosed, as shown in Figure 4 As shown, the LSTM neural network structure mainly includes the following core parameters: the input data at the current moment , the output signal at the previous moment , the cell state at the previous moment , the output signal at the current moment , the cell state at the current moment and three gated units.

[0110] Capturing long-term dependencies between aircraft flight conditions and aerodynamic properties through cell states and gating units;

[0111] The input data of the long short-term memory neural network model is composed of model input parameters;

[0112] The output signal of the long short-term memory neural network model is composed of model output parameters.

[0113] In this embodiment, the three gating units include a forget gate, an input gate, and an output gate:

[0114] The forget gate is used to control the retention or discarding of the cell state at the previous moment;

[0115] The input gate is used to selectively memorize the information of the input data at the current moment and update the cell state;

[0116] The output gate is used to control the output signal of the long short-term memory neural network model at the current moment.

[0117] More specifically, the forget gate controls the cell state at the last moment. , decide which historical information needs to be retained or discarded.

[0118] This operation is mainly implemented by the sigmoid function in formula (5). When the output of the sigmoid function is 0, it means that the cell state at the previous moment is not retained, and when the output signal is 1, it means that the cell state at the previous moment is retained.

[0119] Control variable of forget gate It can be expressed as:

[0120]

[0121] Where σ is the sigmoid activation function, and are the weight coefficient matrix and bias matrix of the forget gate respectively, is the output signal at the last moment, Enter data for the current moment.

[0122] More specifically, the function of the input gate is to selectively memorize the current input data information and update the cell status .

[0123] Control variable of input gate Mainly using formula (6), the sigmoid function is used to input the current moment And the output of the previous moment Perform nonlinear changes to obtain;

[0124] The candidate cell state at the current moment Then we mainly use formula (7) to calculate the input data at the current moment using the tanh function. And the output of the previous moment Perform nonlinear transformation to obtain, and then use formula (8) to update the current cell state .

[0125]

[0126] Where σ is the sigmoid activation function, and are the weight coefficient and bias matrix of the input gate respectively, and They are the information weight coefficients of the cell units. and number and information bias matrix, is the output signal at the last moment, Enter data for the current moment.

[0127] More specifically, the output gate is mainly used to control the current output of the LSTM neural network. .

[0128] Control variable of output gate , mainly using formula (9), through the sigmoid function to the current moment input data And the output of the previous moment Perform nonlinear changes to obtain;

[0129] The current output Then we mainly use formula (10) to first calculate the cell state at the current moment Perform tanh function transformation and then compare it with the control variable of the output gate Perform the dot product to get .

[0130]

[0131] In the formula, and are the weight coefficient and bias matrix of the output gate respectively, is the output signal at the last moment, Enter data for the current moment, is the cell state at the current moment.

[0132] The present invention utilizes an artificial intelligence-long short-term memory neural network (LSTM) model to learn the aircraft aerodynamic characteristics contained in aerodynamic data, thereby avoiding the complex aerodynamic generation mechanism and reducing the difficulty of aircraft aerodynamic modeling.

[0133] The LSTM neural network has the characteristic of realistically fitting any nonlinear function, thereby improving the model's simulation capability and fidelity for unsteady and nonlinear aerodynamic characteristics.

[0134] In this embodiment, the proposed LSTM aerodynamic model is constructed based on the classification and characteristics of the aircraft's unsteady aerodynamic effects. It is universal and suitable for aerodynamic modeling of different fixed-wing aircraft, rather than for specific aircraft or models. It can effectively capture the characteristics of aerodynamic data and improve the accuracy and adaptability of unsteady aerodynamic modeling.

[0135] Step S4, using quantum particle swarm optimization algorithm to optimize and adjust the hyperparameters of the long short-term memory neural network model;

[0136] The present invention uses a long short-term memory neural network (LSTM) to establish an unsteady aerodynamic model. Since the LSTM network involves many hyperparameters, these hyperparameters have a great influence on the prediction accuracy, and the efficiency of manual tuning is low, and it is difficult to find the global optimal parameter combination.

[0137] The hyperparameters of the constructed LSTM neural network include the number of LSTM network layers, the number of neurons in each layer, the length of historical data, the length of predicted data, and the batch size of training data.

[0138] In order to improve the prediction performance of the LSTM neural network, these hyperparameters need to be optimized. The quantum particle swarm optimization (QPSO) algorithm is used for global search and tuning. By setting different hyperparameter combinations, the model prediction error (that is, the difference between the predicted value and the true value) is calculated, and the hyperparameters are continuously adjusted iteratively to minimize the prediction error and obtain the optimal hyperparameter solution to ensure the optimal model performance.

[0139] In addition, the QPSO algorithm can improve the generalization ability of the network and enhance the adaptability and stability of the model when dealing with complex aerodynamic data modeling.

[0140] Furthermore, the step S4 further comprises:

[0141] Step S41, initializing relevant parameters of the quantum particle swarm optimization algorithm;

[0142] Step S42, using the particle position parameters to train the long short-term memory neural network, and calculating the particle fitness of each training;

[0143] Step S43, according to the size of the particle fitness value, determine the local optimal solution and the global optimal solution of the individual particle, and update the position of the particle;

[0144] Step S44, determine whether the stop condition is met, if it is met, output the optimal particle position, if not, enter the long short-term memory neural network training again using the particle position parameters, and re-iterate the calculation.

[0145] The QPSO algorithm is based on the classic particle swarm optimization (PSO) algorithm. PSO is an optimization algorithm based on group behavior. Its basic idea is to imitate the migration and foraging behavior of birds, find a balance between individual exploration and social group behavior, and use iterative search to make particles follow the optimal particles in the solution space for global search.

[0146] Compared with general optimization algorithms, the PSO algorithm is easy to implement, has fewer parameters and higher computational efficiency. However, as the complexity of the target problem increases, the algorithm is prone to problems such as slow convergence, local optimality and poor robustness.

[0147] In order to solve the above problems, Sun Jun et al. proposed the QPSO algorithm based on the theory of quantum mechanics. Compared with the classical PSO algorithm, the QPSO algorithm cancels the moving direction attribute of the particle, so that the update of the particle position has no direct correlation with the historical movement of the particle, making the particle position update more random. At the same time, the QPSO algorithm introduces the average optimal position parameter to improve the cooperation ability between particles.

[0148] Compared with the classic PSO algorithm, the QPSO algorithm has the characteristics of fewer parameters, faster convergence speed, better global search and local search capabilities, and is not easy to fall into the local optimal point.

[0149] In step S41, the update formula of the QPSO algorithm is as follows:

[0150]

[0151] Where p(t) is the potential center point of each particle, is the local optimal position of the ith particle, is a random value between [0,1], is the global optimal particle, is the average optimal position of the particle, M is the total number of particles, x(t+1) is the position of the current particle, is the contraction-expansion factor. Usually, <1.781.

[0152] More specifically, in step S41, the relevant parameters include particle dimension, particle population size, long short-term memory (LSTM) neural network hyperparameters to be optimized, and maximum number of iterations, etc.

[0153] Step S42, training the LSTM neural network using the particle position parameters, and calculating the particle fitness of each training;

[0154] The position of each particle corresponds to a set of hyperparameters of the LSTM neural network (such as learning rate, number of layers, number of neurons, etc.). The LSTM neural network is trained using the current position parameters of the particle, and the training data set comes from unsteady aerodynamic samples.

[0155] The fitness function f of the particle is set to the mean square error during the training process, and its expression is:

[0156]

[0157] In the formula, N is the number of samples in the training set. is the true value of the training sample, is the predicted value of the training sample.

[0158] The smaller the fitness value, the smaller the prediction error of the LSTM model and the better the hyperparameters of the particle.

[0159] Step S43, according to the size of the particle fitness value, determine the local optimal solution and the global optimal solution of the individual particle, and update the position of the particle according to formula (11).

[0160] The local optimal solution is the position where the particle has the lowest fitness in its own history, and the global optimal solution is the position where the fitness of the entire particle swarm is the lowest.

[0161] Step S44, determine whether the stop condition is met, if it is met, output the optimal particle position, if not, enter the long short-term memory neural network training again using the particle position parameters, and re-iterate the calculation.

[0162] In this embodiment, the stopping conditions include: the particle fitness value reaches a preset threshold (such as a mean square error is less than a certain minimum value) and the number of iterations reaches a maximum number of iteration steps.

[0163] Whether to stop the calculation is determined according to the particle fitness value and the maximum number of iterations. If the stopping condition is met, the optimal particle position, that is, the optimal hyperparameter of the LSTM neural network, is output, and step S5 is executed. If the stopping condition is not met, steps S42 to S44 are returned to reiterate.

[0164] The present invention adopts the quantum particle swarm optimization algorithm (QPSO) to automatically search for the optimal solution of the hyperparameters of the LSTM neural network, ensuring the global optimality of the hyperparameter optimization process, effectively improving the prediction accuracy and training efficiency of the LSTM neural network, and reducing the impact of human factors on model training.

[0165] The present invention uses the QPSO algorithm to optimize the hyperparameters of the LSTM neural network, which can not only effectively improve the prediction accuracy and training efficiency of the LSTM neural network, but also reduce the impact of human factors on model training.

[0166] Step S5, based on the optimized hyperparameters, the long short-term memory neural network model is trained to obtain an optimized aircraft unsteady aerodynamic model, wherein the aircraft unsteady aerodynamic model is used to simulate aircraft aerodynamics.

[0167] In step S5, the hyperparameters optimized by QPSO are input into the LSTM neural network for training, and an aircraft unsteady aerodynamic model based on the QPSO-LSTM neural network is constructed. The model can accurately predict the unsteady aerodynamic characteristics of the aircraft and provide high-precision aerodynamic simulation capabilities for aircraft design, simulation and control.

[0168] The present invention also provides an aircraft unsteady aerodynamic modeling system based on a neural network. The aircraft unsteady aerodynamic modeling system based on a neural network includes a memory and a processor, the memory can store a program, and the processor can execute the program. When the processor executes the program, the aircraft unsteady aerodynamic modeling method based on a neural network provided by the present invention is implemented.

[0169] The present invention also provides a computer-readable storage medium, which is a non-volatile storage medium or a non-transient storage medium, on which computer instructions are stored. When the computer instructions are executed, the steps corresponding to any of the above methods are executed, which will not be repeated here.

[0170] When the implementation process file of the neural network-based aircraft unsteady aerodynamic modeling method is a computer program, it can also be stored in a computer-readable storage medium as a product. For example, a computer-readable storage medium may include, but is not limited to, a magnetic storage device (e.g., a hard disk, a floppy disk, a magnetic strip), an optical disk (e.g., a compact disk (CD), a digital versatile disk (DVD)), a smart card, and a flash memory device (e.g., an electrically erasable programmable read-only memory (EPROM), a card, a stick, a key drive). In addition, the various storage media described herein can represent one or more devices and / or other machine-readable media for storing information. The term "machine-readable medium" may include, but is not limited to, wireless channels and various other media (and / or storage media) that can store, contain and / or carry code and / or instructions and / or data.

[0171] The aircraft unsteady aerodynamic modeling method and system based on neural network proposed in the present invention have the following beneficial effects:

[0172] 1) It avoids the tedious work of aircraft aerodynamic effect analysis, decomposition and mathematical modeling, eliminates the need to analyze complex aerodynamic generation mechanisms, and reduces the difficulty of aircraft aerodynamic modeling;

[0173] 2) Avoid various simplifications and assumptions used in the aircraft aerodynamic modeling process in existing technologies, enhance the aerodynamic model's ability to simulate various aerodynamic characteristics of the aircraft, and improve the fidelity of the aerodynamic model;

[0174] 3) By using the quantum particle swarm optimization algorithm (QPSO) to automatically search for the optimal solution of the hyperparameters of the LSTM neural network, the prediction accuracy and training efficiency of the LSTM neural network are effectively improved, while reducing the impact of human factors on the model training process;

[0175] 4) Avoid the binding relationship between traditional modeling methods and modeling objects (aircraft), and enhance the versatility of aircraft aerodynamic models. In traditional methods, the parameters and expressions of aerodynamic mathematical models correspond to the aerodynamic effects of a specific aircraft, which results in weak versatility of the model. The structure of the artificial intelligence-long short-term memory neural network (LSTM) model is based on the unsteady aerodynamic effects of fixed-wing aircraft, rather than a specific model, so it has strong versatility;

[0176] 5) A method for constructing an unsteady aerodynamic sample dataset that matches the artificial intelligence-long short-term memory neural network (LSTM) model is proposed, which not only ensures data quality but also takes into account the efficiency of data generation.

[0177] Although the above methods are illustrated and described as a series of actions for simplicity of explanation, it should be understood and appreciated that these methods are not limited by the order of the actions, because according to one or more embodiments, some actions may occur in a different order and / or concurrently with other actions from those illustrated and described herein or not illustrated and described herein but understandable to those skilled in the art.

[0178] As shown in this application and claims, unless the context clearly indicates an exception, the words "a", "an", "an" and / or "the" do not refer to the singular and may also include the plural. Generally speaking, the terms "include" and "comprise" only indicate the inclusion of the steps and elements that have been clearly identified, and these steps and elements do not constitute an exclusive list. The method or device may also include other steps or elements.

[0179] Those skilled in the art will appreciate that information, signals, and data may be represented using any of a variety of different technologies and techniques. For example, data, instructions, commands, information, signals, bits, symbols, and chips cited throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or optical particles, or any combination thereof.

[0180] Those skilled in the art will further appreciate that the various illustrative logic blocks, modules, circuits, and algorithm steps described in conjunction with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or a combination of the two. To clearly illustrate this interchangeability of hardware and software, various illustrative components, blocks, modules, circuits, and steps are generally described above in terms of their functionality. Whether such functionality is implemented as hardware or software depends on the specific application and the design constraints imposed on the overall system. The technician may implement the described functionality in different ways for each specific application, but such implementation decisions should not be interpreted as resulting in a departure from the scope of the present invention.

[0181] The various illustrative logic modules and circuits described in conjunction with the embodiments disclosed herein may be implemented or executed with a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA) or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to perform the functions described herein. A general purpose processor may be a microprocessor, but in an alternative, the processor may be any conventional processor, controller, microcontroller, or state machine. The processor may also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in cooperation with a DSP core, or any other such configuration.

[0182] The steps of the method or algorithm described in conjunction with the embodiments disclosed herein may be embodied directly in hardware, in a software module executed by a processor, or in a combination of the two. The software module may reside in a RAM memory, a flash memory, a ROM memory, an EPROM memory, an EEPROM memory, a register, a hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to a processor so that the processor can read and write information from / to the storage medium. In an alternative, a storage medium may be integrated into a processor. The processor and the storage medium may reside in an ASIC. The ASIC may reside in a user terminal. In an alternative, the processor and the storage medium may reside in a user terminal as discrete components.

[0183] The above embodiments are provided for persons familiar with the art to implement or use the present invention. Personnel familiar with the art can make various modifications or changes to the above embodiments without departing from the inventive concept of the present invention. Therefore, the protection scope of the present invention is not limited to the above embodiments, but should be the maximum scope of the innovative features mentioned in the claims.

Claims

1. A neural network-based aircraft unsteady aerodynamic modeling method, characterized in that: The following steps are involved: Establish sample data on the unsteady aerodynamic characteristics of the aircraft, and select characteristic parameters from the sample data to form a data set; Normalize the feature parameters and divide the data set into training set and validation set; Construct a long short-term memory neural network model to describe the relationship between flight conditions and aerodynamic characteristics of aircraft with different wing shapes; The quantum particle swarm optimization algorithm is used to optimize and adjust the hyperparameters of the long short-term memory neural network model; Based on the optimized hyperparameters, the long short-term memory neural network model is trained to obtain an optimized aircraft unsteady aerodynamic model, and the aircraft unsteady aerodynamic model is used to simulate aircraft aerodynamics.

2. The aircraft unsteady aerodynamic modeling method based on neural network according to claim 1, characterized in that: The step of establishing sample data on the unsteady aerodynamic characteristics of the aircraft further comprises: Using computational fluid dynamics simulation technology, the aerodynamic characteristics of the pitch reciprocating motion of the specified wing aircraft are numerically simulated to generate aircraft aerodynamic characteristics data under different flight conditions; The aircraft aerodynamic characteristic data include aerodynamic response data under various different angle of attack mean values, angle of attack amplitudes, reduction frequencies, and Mach number factors.

3. The aircraft unsteady aerodynamic modeling method based on neural network according to claim 2, characterized in that: The step of numerically simulating the aerodynamic characteristics of the pitch reciprocating motion of the specified wing-shaped aircraft further comprises: The pitch reciprocating motion is represented as a sine function, and the frequency, mean value, and amplitude of the sine function are adjusted to optimize the simulation data.

4. The method for modeling aircraft unsteady aerodynamics based on neural network according to claim 3, characterized in that: The step of numerically simulating the aerodynamic characteristics of the pitch reciprocating motion of the specified wing aircraft by using computational fluid dynamics simulation technology further includes: Generate the computational domain and perform meshing, set the distance between the far field inlet and the leading edge of the airfoil, the distance between the outlet and the trailing edge of the airfoil, and the distance between the lateral boundary and the upper and lower surfaces of the airfoil; Use unstructured grid to generate fluid domain, set the mesh size of airfoil surface, set the first layer height, growth rate, maximum number of generated layers of boundary layer and total number of grids of boundary layer network; The calculations were performed using a transient density-based solver, shear stress transport k-ω model, and ideal gas model; Set the ambient temperature and the incoming flow direction, use the dynamic grid method to achieve the periodic motion of the airfoil in the pitch direction, and set the airfoil angle of attack and pitch angle under the initial conditions; The implicit Roche single equation model-flux driven scheme format is used for discrete calculation of the flux term, the least squares lattice center format is used for gradient calculation, the second-order upwind format is used for spatial discrete calculation, and the double time step first-order implicit format is used for temporal discrete calculation.

5. The aircraft unsteady aerodynamic modeling method based on neural network according to claim 1, characterized in that: The step of selecting feature parameters from the sample data to form a data set further includes: Determining characteristic parameters of the long short-term memory neural network model according to the flight mission, wherein the characteristic parameters include model input parameters for describing flight conditions and model output parameters for describing aerodynamic characteristics; The model input parameters include flight altitude, flight Mach number, angle of attack and rate of change of angle of attack, and the model output parameters include lift coefficient, drag coefficient and pitch moment coefficient.

6. The aircraft unsteady aerodynamic modeling method based on neural network according to claim 1, characterized in that: The step of normalizing the characteristic parameters further comprises: The min-max normalization method is used to normalize the characteristic parameter data.

7. The method for modeling aircraft unsteady aerodynamics based on neural network according to claim 5, characterized in that: The constructed long short-term memory neural network model includes cell states, output signals, and gating units: Capturing long-term dependencies between aircraft flight conditions and aerodynamic properties through cell states and gating units; The input data of the long short-term memory neural network model is composed of model input parameters; The output signal of the long short-term memory neural network model is composed of model output parameters.

8. The method for modeling aircraft unsteady aerodynamics based on neural network according to claim 7, characterized in that: The gate control unit includes an input gate, an output gate and a forget gate: The input gate is used to selectively memorize the information of the input data at the current moment and update the cell state; The output gate is used to control the output signal of the long short-term memory neural network model at the current moment; The forget gate is used to control the retention or discarding of the cell state at the previous moment.

9. The method for modeling aircraft unsteady aerodynamics based on neural network according to claim 1, characterized in that: The hyperparameters of the long short-term memory neural network model include: The number of network layers, the number of neurons in each network layer, the length of historical data, the length of predicted data, and the batch size of training data.

10. The aircraft unsteady aerodynamic modeling method based on neural network according to claim 1, characterized in that: The quantum particle swarm optimization algorithm optimizes the hyperparameters of the long short-term memory neural network by updating particle positions and evaluating fitness, wherein the particle fitness is calculated by mean square error.

11. The aircraft unsteady aerodynamic modeling method based on neural network according to claim 10, characterized in that: The quantum particle swarm optimization algorithm further comprises: Initialize the relevant parameters of the quantum particle swarm optimization algorithm; The particle position parameters are used to train the long short-term memory neural network and the particle fitness of each training is calculated; According to the particle fitness value, determine the local optimal solution and global optimal solution of the individual particle, and update the position of the particle; Determine whether the stopping condition is met. If so, output the optimal particle position. If not, enter the long short-term memory neural network training again using the particle position parameters and perform iterative calculation again.

12. The neural network-based aircraft unsteady aerodynamic modeling method according to claim 11, characterized in that: In the step of initializing relevant parameters of the quantum particle swarm optimization algorithm, the relevant parameters include particle dimension, particle population size, hyperparameters of the long short-term memory neural network to be optimized, and the maximum number of iterations.

13. The neural network-based aircraft unsteady aerodynamic modeling method according to claim 11, characterized in that: The stop conditions include: The particle fitness value reaches a preset threshold, and / or The number of iterations reached the maximum number of iterations.

14. An aircraft unsteady aerodynamic modeling system based on neural network, characterized in that: Including memory and processor: The memory is used to store instructions executable by the processor; The processor is used to execute the instructions to implement the method according to any one of claims 1-13.

15. A computer storage medium having computer instructions stored thereon, wherein when the computer instructions are executed by a processor, the method according to any one of claims 1 to 13 is executed.

Images