CN108182328A - A kind of big angle of attack Nonlinear Aerodynamic reduced-order model suitable for stall flutter - Google Patents

Abstract

A kind of big angle of attack Nonlinear Aerodynamic reduced-order model suitable for stall flutter of the invention designs sinusoidal multilevel signal as Nonlinear Aerodynamic identification signal first, and the amplitude of modelled signal need to meet substantially movement needs.Nonlinear Aerodynamic identification is carried out secondly based on the Recognition with Recurrent Neural Network model of deep learning, training output signal is calculated using CFD approach, dynamic Time-Domain Nonlinear Aerodynamic Model when substantially being vibrated with the Recognition with Recurrent Neural Network model foundation wing based on deep learning, and the system calculated with CFD under test signal responds, compared with the identification result of network model, the performance of model is verified.The present invention saves Flight Vehicle Design cost, improves stall flutter design efficiency, has developed the reduced-order model of Nonlinear Aerodynamic, so as to fast prediction stall flutter.

Description

A Nonlinear Aerodynamic Reduced-Order Model for Large Angle of Attack Applicable to Stall Flutter

technical field

The invention belongs to the field of aircraft design and system identification, and is a non-linear reduced-order model developed for rapidly predicting aircraft stall flutter characteristics and improving calculation efficiency.

Background technique

Stall flutter is a self-excited vibration caused by nonlinear aerodynamic force and elastic structure coupling caused by airflow separation when the aircraft wing or rudder surface is at a large angle of attack. Stall flutter exhibits strong nonlinear characteristics. When an aircraft flies at a high angle of attack (for example, a fighter or missile has high maneuverability and agility and needs to fly within a range of a large angle of attack) or encounters a gust of wind, it may cause the lifting surface to stall. When the angle of attack reaches a certain critical value, it is easy to This will lead to aerodynamic stall, which may lead to stall flutter, which will affect the structural safety of the aircraft; especially for propeller blades and helicopter blades, the tip of the blade is more prone to stall and stall flutter.

At present, the research methods of wing stall flutter include experimental research, empirical model and CFD-CSD fluid-structure interaction simulation. The test method is time-consuming and labor-intensive, and the cost is high. The calculation of the empirical model depends on the model and test data, and the accuracy of the result is not high. CFD-CSD calculation can obtain high-precision results, but the calculation is huge and time-consuming, which increases the calculation cost and aircraft development cycle.

To realize the prediction of aircraft stall flutter, the key technical problem to be solved is the accurate identification of nonlinear aerodynamic forces at large angles of attack. The main problems are as follows:

1. Traditional signals are less effective in predicting non-linear aerodynamics at non-design points. In order to accurately capture the vibration characteristics of the limit cycle with large angle of attack, it is necessary to design a reasonable identification input signal. The designed identification input signal must accurately describe the law of limit cycle vibration, large motion amplitude and frequency requirements.

2. For a well-designed identification input signal, it is necessary to develop a robust generalized reduced-order model and an algorithm to accurately identify the dynamic mathematical model of nonlinear aerodynamic forces separated by large angles of attack.

Based on the above situation, it is necessary to propose a new model in the input signal design and identification algorithm.

Contents of the invention

In view of the above problems, in order to save the cost of aircraft design and improve the efficiency of stall flutter design, a nonlinear aerodynamic reduction model with large angle of attack is proposed for stall flutter, so as to quickly predict stall flutter.

The present invention is applicable to the large angle of attack nonlinear aerodynamic reduction model of stall flutter, obtained through the following steps:

Step 1: According to the amplitude, frequency and vibration law of the wing limit cycle vibration, design a left-right symmetrical multi-level sinusoidal signal as the input data of the deep learning model system.

Step 2: Input the sinusoidal signal designed in step 1 into the CFD software, and obtain the aerodynamic coefficients of the wing at different speeds under the excitation of the signal, as the output data of the deep learning model system.

Step 3: Perform nonlinear aerodynamic identification based on the cyclic neural network model of deep learning, and obtain the nonlinear aerodynamic reduction model for large angle of attack.

The above-mentioned nonlinear aerodynamic reduction model at large angle of attack is brought into the structural and aerodynamic coupling calculation equation, and the 4th-order Runge-Kutta method is used to perform time-domain propulsion calculations to predict the response process of each mode over time, so that To achieve the purpose of predicting stall flutter.

The advantages of the present invention are:

1. The present invention is applicable to the large angle of attack nonlinear aerodynamic reduction model of stall flutter. Through the analysis of the characteristics of stall flutter vibration, a multi-level sine identification signal is designed. This signal covers the amplitude and frequency of vibration range, it can well predict the nonlinear aerodynamic forces of the vibration process.

2. The present invention is applicable to a large angle of attack nonlinear aerodynamic reduction model for stall flutter, and is a nonlinear aerodynamic identification model based on a deep-learning cyclic neural network model, which has huge advantages in dynamic nonlinear flow field modeling Advantages and application prospects.

3. The present invention is suitable for a large angle of attack nonlinear aerodynamic reduction model of stall flutter, and has strong adaptability, high identification accuracy, relatively simple operation method, and can be realized by using MATLAB programming.

4. The present invention is applicable to the large angle of attack nonlinear aerodynamic reduction model of stall flutter. With the development of CFD technology, complicated experimental research is omitted, and the calculation efficiency can be increased by one to two orders of magnitude compared with CFD calculation. The identification method can also be used in other nonlinear aerodynamic identification systems, and has certain versatility.

Description of drawings

Fig. 1 is the flow chart of the design of the large angle of attack nonlinear aerodynamic order reduction model applicable to stall flutter of the present invention;

Fig. 2 is a schematic diagram of the identification signal of the large angle of attack nonlinear aerodynamic reduction model applicable to stall flutter in the present invention;

Fig. 3 is the structural diagram of the cyclic neural network model based on deep learning adopted in the large angle of attack nonlinear aerodynamic reduction model applicable to stall flutter in the present invention;

Fig. 4 is a structural diagram of neuron unit expansion calculation in the deep learning-based cyclic neural network model.

Detailed ways

The present invention will be described in further detail below in conjunction with the accompanying drawings.

The present invention is applicable to the large angle of attack nonlinear aerodynamic reduction model of stall flutter, as shown in Figure 1, obtained through the following steps:

Step 1: Design the input signal.

The input signal is designed to predict the stall flutter of the wing. When the wing has a large angle of attack limit cycle vibration, the vibration form of each cycle is sinusoidal. The design method of the present invention is, for a wing model, according to the amplitude, frequency and vibration law of its limit cycle vibration, design left and right symmetrical multi-level sinusoidal signals, these sinusoids include different amplitude and frequency information, and The required amplitude and frequency information are covered; the frequency is generally low-frequency vibration, so the signal frequency range only needs to cover the range of the stall flutter vibration frequency. as shown in picture 2. When targeting a specific model, it needs to be designed according to the needs of vibration amplitude and frequency.

Step 2: Develop a CFD solver to calculate the aerodynamic coefficients.

Input the sinusoidal signal designed in step 1 into the CFD software. Considering that under the condition of large angle of attack, different incoming flow velocities will have an impact on the calculated aerodynamic parameters, so the wings at different speeds are obtained through the CFD software The aerodynamic coefficient under the signal excitation includes lift coefficient, moment coefficient, drag coefficient, etc. The above-mentioned aerodynamic coefficient is the output data of the deep learning model system, and the designed sinusoidal signal is the input data of the deep learning model system.

Step 3: Nonlinear aerodynamic force identification.

In the present invention, the nonlinear aerodynamic identification is based on the deep learning cycle neural network model, that is, the deep learning model in step 2, as shown in FIG. 3 . Due to the complexity of the nonlinear flow field, the traditional model has high requirements for experience, and it is difficult to give a suitable explicit expression, while the neural network (Neural Network, NN) model based on deep learning does not need to give an identification system Advantages of displaying math expressions between input/output. The neural network model obtains the characteristics of the identification system by learning the characteristics of the input/output in the training model and the way the input affects the output, and uses the learned "experience" to predict the output under the new input.

The recurrent neural network model based on deep learning is divided into input layer, hidden layer and output layer. Among them, the number of hidden layers is greater than 4 layers, and the specific number of layers can be adjusted according to the actual model. From the input layer to the output layer, the neuron units of each layer and the neuron units of the previous and subsequent layers are connected and transmit data. Each neuron unit in the hidden layer feeds back as its own input.

An expanded view of the internal structure of each of the above neuron units is shown in FIG. 4 .

The neuron unit is expanded in time series, and the expansion diagram is described as follows:

1) x ^(t) is the input of training samples at time t, and similarly, x ^(t-1) and x ^(t+1) are the inputs of training samples at time t-1 and time t+1 respectively.

2) h ^(t) is the hidden state of the model at time t, and h ^(t) is jointly determined by x ^(t) and h ^(t-1) .

3) o ^(t) represents the output of the model at time t, and o ^(t) is only determined by the current hidden state h ^(t) of the model.

4) L ^(t) represents the loss function of the model at time t.

5) y ^(t) represents the real output of the training sample at time t.

6) The three matrices U, W, and V are the linear relationship parameters of the model, which are shared in the entire recurrent network, reflecting the idea of "circular feedback" in the recurrent network model.

Applying the above-mentioned deep learning-based cyclic neural network model can approximate any dynamic nonlinear mapping system, and can well identify dynamic nonlinear unsteady aerodynamic models. The forward propagation algorithm of the entire cyclic neural network is the same as that of the ordinary feed-forward neural network. The activation function of the hidden layer is selected from the continuous hyperbolic tangent function, and the activation function of the output layer is selected from the softmax function. The gradient descent method in the error backpropagation algorithm is used in the training of the cyclic neural network parameters, and the network output is finally approached to the actual output by repeatedly adjusting the connection weights and thresholds inside the cyclic neural network.

Step 4: Use the trained network model to predict the stall flutter of the wing. The network model is brought into the structural and aerodynamic coupling calculation equation, and the 4th-order Runge-Kutta method is used for time-domain propulsion calculation to predict the response process of each mode over time, so as to achieve the purpose of predicting stall flutter.

Claims (5)

1. A large angle of attack nonlinear aerodynamic reduction model applicable to stall flutter, characterized in that: obtain by the following steps: Step 1: According to the amplitude, frequency and vibration law of the wing limit cycle vibration, design a left-right symmetrical multi-level sinusoidal signal as the input data of the deep learning model system; Step 2: Input the sinusoidal signal designed in step 1 into the CFD software, and obtain the aerodynamic coefficient of the wing at different speeds under the excitation of the signal, as the output data of the deep learning model system; Step 3: Perform nonlinear aerodynamic identification based on the cyclic neural network model of deep learning, and obtain the nonlinear aerodynamic reduction model for large angle of attack. 2. A large angle-of-attack nonlinear aerodynamic reduced-order model suitable for stall flutter as claimed in claim 1, characterized in that: in step 1, the multi-level sinusoidal signals contain different amplitude and frequency information. 3. A kind of large angle of attack nonlinear aerodynamic reduction model applicable to stall flutter as claimed in claim 1, characterized in that: in the cyclic neural network model based on deep learning, the hidden layer activation function selects continuous hyperbolic The tangent function, the output layer activation function selects the softmax function. 4. A kind of large angle-of-attack nonlinear aerodynamic reduction model applicable to stall flutter as claimed in claim 1, is characterized in that: when the cyclic neural network model based on deep learning is used for parameter training, the error backpropagation algorithm is selected The gradient descent method. 5. A large angle of attack nonlinear aerodynamic reduced order model suitable for stall flutter as claimed in claim 1, characterized in that: the large angle of attack nonlinear aerodynamic reduced order model is brought into the structure and aerodynamic coupling calculation In the equation, the 4th-order Runge-Kutta method is used to carry out time-domain advance calculations to predict the response process of each mode over time, so as to achieve the purpose of predicting stall flutter.