Tilt rotor unsteady aerodynamic performance prediction method and system based on machine learning and storage medium
Through BiGRU timing neural network and physical constraint loss function optimization, the sample demand and calculation cost problems of non-stable aerodynamic performance prediction of tilt rotor aircraft are solved, and accurate and real-time aerodynamic performance prediction is achieved.
Patent Information
- Application Number
- CN202510178821.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-18
- Publication Date
- 2025-07-04
AI Technical Summary
The existing tilt rotor aerodynamic performance agent model fails to effectively consider its non-stableness during the tilt process, and the traditional method requires a large number of samples and high computing power consumption, making it difficult to accurately predict instantaneous aerodynamic characteristics.
Using a BiGRU-based timing neural network structure, combined with Latin hypercube sampling and overlapping grid CFD calculation, a tilt rotor non-stabilized aerodynamic performance prediction model is established, and through hierarchical sampling and physical constraint loss function optimization, the sample demand is reduced and the prediction accuracy is improved.
Real-time and accurate prediction of three-axis periodic instantaneous force, instantaneous moment, time-simultaneous moment, time-simultaneous moment, aerodynamic power and aerodynamic efficiency of the tilt rotor aircraft is realized, reducing calculation costs and possessing strong generalization capabilities.
Smart Images
Figure CN120257771A_ABST
Abstract
Description
Technical Field
[0001] The technical solution belongs to the field of aerodynamic design and analysis, and specifically relates to a method, system and storage medium for predicting the unsteady aerodynamic performance of a tilt-rotor based on machine learning. Background Art
[0002] As a new star in the aviation field, the tilt-rotor aircraft has both vertical take-off and landing capabilities and high-speed cruise capabilities. When the tilt-rotor aircraft performs operations such as take-off, hover, and cruise, due to its own tilt angle and complex oncoming flow conditions, the aerodynamic force of the tilt-rotor shows strong unsteadiness. During a stable rotation period, there are huge differences between the time-averaged value and the instantaneous extreme value. In the calculation of handling and stability performance and structural strength calculation, in addition to considering the time-averaged force and moment, the influence of the instantaneous aerodynamic force sometimes needs to be considered. However, at present, in the aerodynamic performance surrogate models of tilt-rotors, almost none of them consider the significant unsteadiness during the rotation period of the tilt-rotor, and only model the time-averaged aerodynamic performance, and cannot predict its unsteady aerodynamic characteristics.
[0003] In addition, traditional aerodynamic performance surrogate models require a large number of samples to make relatively accurate predictions. However, since the tilt-rotor often encounters the oncoming flow velocity component parallel to the rotor disk direction during the tilting process, only transient CFD (Computational Fluid Dynamics) simulations can be used, and a large number of samples and high-precision transient CFD simulations bring extremely high computing power consumption. Summary of the Invention
[0004] Aiming at the above deficiencies, the technical solution of the present invention is based on a time-series neural network structure with BiGRU (Bidirectional Gated Recurrent Unit) as the core, and proposes a method, system and storage medium for predicting the unsteady aerodynamic performance of a tilt-rotor based on machine learning. This method can accurately and real-time predict the three-axis periodic instantaneous force, instantaneous moment, time-averaged force, time-averaged moment, aerodynamic power and aerodynamic efficiency, and requires fewer sample numbers, and has strong generalization ability and scalability.
[0005] According to the first aspect of the technical solution of the present invention, a method for predicting the unsteady aerodynamic performance of a tilt-rotor based on machine learning is provided, wherein the prediction method includes the following steps:
[0006] Step 1: Determine the parameter space and obtain the original data. The spatial dimension of the parameter space has pitch angle, oncoming flow angle, oncoming flow magnitude, and rotational speed;
[0007] Step 2: Preprocess the original data, and divide the preprocessed original data into a training set, a validation set and a test set by using the method of stratified sampling;
[0008] Step 3: Define the neural network structure, and train the neural network using the training set and the validation set to obtain the trained neural network.
[0009] Step 4: Evaluate the performance of the trained neural network using the test set. When the accuracy requirement is met, output the surrogate model; otherwise, supplement new data samples and retrain until the surrogate model is output.
[0010] Step 5: Obtain the three-axis periodic instantaneous force, instantaneous torque, time-averaged force, time-averaged torque, aerodynamic power, and aerodynamic efficiency through the surrogate model.
[0011] Further, in the said Step 1, for the original data, if there is existing data, use the existing data; if there is no existing data, or the existing data does not meet the training requirements, sample and supplement new data samples.
[0012] Further, sample and supplement new data samples in the following way:
[0013] After determining the parameter space, establish the value dimension and range, and use the Latin hypercube sampling method to sample the parameter space to obtain sample points.
[0014] Perform CFD calculations for the said sample points, apply a CFD solver, and based on the overlapping grid method, calculate the parameter conditions of the sample points through transient solving until convergence to obtain new data samples.
[0015] Further, in the said Step 2, the data preprocessing specifically includes:
[0016] For the original data, take the periodic unsteady aerodynamic data after stabilization, and use the sample space parameters and the time series of the instantaneous force and instantaneous torque in the three spatial directions in the corresponding Cartesian coordinate system as a sample of the data set, and perform data normalization to achieve data preprocessing.
[0017] Further, in the said Step 2, the training set accounts for 60%, the validation set accounts for 20%, and the test set accounts for 20%.
[0018] Further, in the said Step 3, the neural network structure is built based on PyTorch, including 1 input layer, 2 BiGRU layers, 4 fully connected layers, and 1 output layer.
[0019] Further, in the neural network structure, use the Adam optimizer and adopt a training strategy with variable learning rates.
[0020] Furthermore, in the neural network structure, the output layer outputs equally-spaced time series of instantaneous forces and instantaneous torques in three spatial directions in the Cartesian coordinate system corresponding to the parameters. The average values of the six time series within the rotation period are calculated to obtain the time-averaged forces and time-averaged torques. Combining parameters such as the tilt-rotor rotational speed, oncoming flow direction, and oncoming flow velocity, the aerodynamic power and aerodynamic efficiency are calculated.
[0021] Furthermore, in the neural network structure, the loss function L total is expressed as:
[0022]
[0023] In the formula, n is the number of samples, T is the time series length, w i is the weight coefficient of the i-th sample; y it is the CFD result of the i-th sample at the t-th time step, is the predicted value of the i-th sample at the t-th time step, y i1 is the CFD result of the i-th sample at the 1st time step; y iT is the CFD result of the i-th sample at the last time step (the T-th time step) in one rotation period; λ1 and λ2 are the hyperparameters of L1 and L1 respectively, and L1 and L1 are physical constraints.
[0024] Here, the number of samples n is the total number of samples in the training set, which is an exact value; the time series length T depends on the number of samples within one rotation period of the tilt-rotor. For example, if a set of parameters in each sample corresponds to an instantaneous force and torque sequence of 6×1000, then T = 1000; the value of the weight coefficient w i can be adjusted according to the nonlinear degree in the nonlinear region. According to experience, values between 0 and 10 are often sufficient to describe the nonlinear degrees of different sample regions; there is no fixed value range for the hyperparameters. Usually, they need to be adjusted according to the specific data characteristics. The initial value usually starts from 1.0, and the value range is usually between 0.01 and 100. If the scale difference between L1 and L1 is too large due to the characteristics of the data itself, then both are first normalized to a similar range, and then λ1 and λ2 are set.
[0025] Furthermore, w i The weight coefficient is adjusted according to the physical prior knowledge of the distribution of the strong nonlinear region of the unsteady aerodynamic performance of the tilt-rotor.
[0026] Furthermore, in step 4, the absolute error δ is used to measure the error results of the time-averaged forces and time-averaged torques, and the root mean square error RMSE (Root Mean Square Error) is used to measure the error results of the instantaneous forces and torques.
[0027] Further, in step 4, when the absolute error δ of all samples in the test set does not exceed 5% and the RMSE does not exceed 0.1, it is considered that the accuracy requirement is met, and then the surrogate model is output.
[0028] According to the second aspect of the technical solution of the present invention, a tilt-rotor unsteady aerodynamic performance prediction system based on machine learning is provided. The system includes: a processor and a memory for storing executable instructions; wherein, the processor is configured to execute the executable instructions to perform the tilt-rotor unsteady aerodynamic performance prediction method based on machine learning as described in any of the above aspects.
[0029] According to the third aspect of the technical solution of the present invention, a computer-readable storage medium is provided, on which a computer program is stored, and when the computer program is executed by a processor, it implements the tilt-rotor unsteady aerodynamic performance prediction method based on machine learning as described in any of the above aspects.
[0030] Beneficial effects of the present invention:
[0031] When establishing the surrogate model in the solution of the present invention, the time series characteristics are considered, and the strong unsteadiness of the aerodynamic characteristics under the working state of the tilt-rotor is fully explored. It can more real-time and accurately predict the unsteady aerodynamic performance and time-averaged aerodynamic characteristics of the tilt-rotor under complex inflow conditions, which provides a more comprehensive reference for aerodynamic analysis work, performance handling and stability work, and structural strength work.
[0032] In addition, the learning of instantaneous data reduces the sample demand compared to directly modeling steady-state data. Compared with ordinary time series neural networks, the BiGRU, and the introduction of a loss function with physical constraints and weight correction can also reduce the sample demand and the consumption of computing power.
[0033] Finally, this method has strong generalization ability and scalability. It can be extended to the establishment of other periodic unsteady aerodynamic characteristic surrogate models, and can also be combined with multi-objective optimization algorithms such as NSGA-II (Non-dominated Sorting Genetic Algorithm II) based on the surrogate model to achieve rapid optimization work. Description of the Drawings
[0034] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on the structures shown in these drawings.
[0035] Figure 1 The flowchart of a machine learning-based unsteady aerodynamic performance prediction method for a tiltrotor according to an embodiment of the present invention is shown.
[0036] The realization, functional features and advantages of the object of the present invention will be further described with reference to the embodiments and the accompanying drawings. Detailed implementation manners
[0037] Exemplary embodiments will be described in detail herein, and examples thereof are shown in the accompanying drawings. When the following description refers to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The implementation manners described in the following exemplary embodiments do not represent all implementation manners consistent with the present disclosure. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present disclosure as detailed in the appended claims.
[0038] The terms "first", "second", etc. in the specification and claims of the present disclosure are used to distinguish similar objects and do not necessarily have to be used to describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments of the present disclosure described herein can be implemented in an order other than those illustrated or described herein. The orientation or positional relationship indicated by the terms "center", "longitudinal", "transverse", "left", "right", "vertical", "horizontal", "upper", "lower", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the embodiments of the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus cannot be construed as limiting the protection scope of the present invention.
[0039] In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.
[0040] Multiple, including two or more than two.
[0041] And / or, it should be understood that for the term "and / or" used in the present disclosure, it is merely an association relationship describing associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist simultaneously, and B exists alone.
[0042] The technical solution of the present invention first proposes a method, system and storage medium for predicting the unsteady aerodynamic performance of a tilt-rotor based on machine learning. By learning the unsteady aerodynamic characteristic samples, an unsteady aerodynamic characteristic surrogate model based on the BiGRU time-series neural network structure is established, which can realize the real-time and accurate prediction of the three-axis periodic instantaneous force, instantaneous moment, time-averaged force, time-averaged moment, aerodynamic power and aerodynamic efficiency. By fully mining the time-series features, improving the neural network structure and introducing physical constraints, the demand for the number of samples of the surrogate model can be reduced, thus greatly saving the computational cost. The technical solution of the present invention can be used for predicting the unsteady aerodynamic performance under complex working conditions of a tilt-rotor, and can realize the functions of accurately and real-time predicting the three-axis periodic instantaneous force, instantaneous moment, time-averaged force, time-averaged moment, aerodynamic power and aerodynamic efficiency, with fewer required samples, and having strong generalization ability and scalability.
[0043] First, determine the parameter space and sample. The parameter space generally has dimensions such as pitch angle, inflow angle, inflow magnitude, rotational speed, etc. Determine the value dimensions and ranges, and use the Latin hypercube sampling method to sample the parameter space. This process can be implemented through MATLAB programming.
[0044] Here, as a hierarchical sampling method, Latin hypercube sampling has the advantages of uniform coverage and high efficiency when used for selecting sample points of machine learning. In the technical solution of the present invention, the sample points are obtained by CFD transient calculation, and the required computational cost is very high. Latin hypercube sampling is suitable for scenarios where the selection of such sample points with high computational cost is required. The remaining sampling methods can be used as sub-optimal selection schemes.
[0045] After sampling, perform CFD calculations on the sample data set. The acquisition of the data set depends on CFD transient calculation. Apply CFD solvers such as FLUENT or STAR CCM+ and, based on the overlapping grid method, calculate the parameter conditions of the sample points selected in the previous step through transient solution until convergence.
[0046] The above steps are not necessary. When existing data is available, give priority to using the existing data. When the existing data cannot meet the training requirements, supplement the samples through the above steps.
[0047] Subsequently, perform data preprocessing. Take the stable periodic unsteady aerodynamic data, and use the sample space parameters (n parameters for n dimensions) and the time series of forces and moments in three spatial directions in the corresponding Cartesian coordinate system (6×200) as a sample of the data set, and perform data normalization to avoid the influence of feature scale differences. The preprocessing can be implemented through MATLAB programming.
[0048] After data preprocessing, use the method of stratified sampling to divide the training set (60%), validation set (20%) and test set (20%).
[0049] Define the neural network structure. The neural network structure proposed in this solution is built based on PyTorch, with BiGRU as the core, and is specifically composed of 1 input layer, 2 BiGRU layers, 4 fully connected layers, and 1 output layer. The input layer should confirm the input from the parameter space. The BiGRU layer is used to fully explore the unsteady characteristics of the tilt-rotor. The connection of the 4 fully connected layers can achieve the non-linear mapping from features to output. The final output layer itself can output equally spaced time series (6×200) of forces and moments in three spatial directions in the Cartesian coordinate system under the corresponding parameters. In addition, the time-averaged force, time-averaged moment, aerodynamic power, and aerodynamic efficiency can be calculated to achieve synchronous output. Regarding the optimizer, the Adam optimizer is adopted, and a training strategy with variable learning rate is used to improve the stability and fineness of training.
[0050] Regarding the loss function of the neural network, an improved mean squared error loss function with weight correction (MSELoss, Mean Squared Error Loss) is used, which can increase the sample weight in the strong non-linear region to improve the prediction accuracy, and introduce some physical information as a constraint to reduce the demand for the number of samples.
[0051] So far, the neural network training can be carried out. Using the training set and the validation set, when the loss function is small enough, the neural network training is completed.
[0052] Finally, the model performance is evaluated. The absolute error δ can be used to measure the error results of the time-averaged force and moment, and the root mean square error RMSE (Root Mean Square Error) is used to measure the error results of the instantaneous force and moment. When the accuracy requirements are met, the surrogate model can be output. Otherwise, according to the performance of the test set, supplement the number of samples and retrain.
[0053] The tools used in the implementation process of this solution include MATLAB, CFD solvers (such as STAR CCM+ or ANSYS FLUENT), and PyTorch.
[0054] Embodiment
[0055] As Figure 1 given, and the general content of each step is given. Here, each step is described in more detail.
[0056] First, determine the parameter space and sample it. In terms of the space dimensions, there are generally dimensions such as pitch angle, inflow angle, inflow magnitude, and rotational speed. The value range of each dimension should be determined according to the physical conditions and application scenarios of the tiltrotor. After establishing the parameter space, sampling is carried out, and the Latin hypercube sampling method is used for parameter space sampling. This process can be implemented through MATLAB programming. The specific process of the Latin hypercube sampling method is as follows:
[0057] (1) Divide the parameter range. For each parameter p i (where i = 1, 2, …, n), divide its possible value range into N equally divided sub-intervals, where N is the number of samples.
[0058] (2) Random sampling. For each parameter p i , generate N random numbers r ij , where j = 1, 2, …, N, and r ij is uniformly distributed in the range [0, 1]. Divide the value range of each parameter into N equally divided sub-intervals. For each parameter p i , map the random number r ij to the corresponding sub-interval to determine the sampling point.
[0059] (3) Generate the sample matrix. Combine the sampling points of all parameters in a cross-arrangement manner to form an N×n matrix, where each row represents a sample and each column represents the value of a parameter.
[0060] After sampling, perform CFD calculations on the sample dataset. The acquisition of the dataset depends on CFD transient calculations. Apply CFD solvers such as FLUENT or STAR CCM+ and, based on the overlapping grid method, divide the component grids of the tiltrotor and the background grids of the entire external flow field respectively. Apply rotational motion to the component grid area, and the magnitude of the rotational speed depends on the specific sampling points in the parameter space. Adopt a compressible model and an implicit unsteady model, and establish the time step as 1 / 200 of the time required for the tiltrotor to rotate one week. The boundary condition of the external flow field adopts a free stream boundary, and the inflow direction and Mach number of the free stream boundary are determined by the specific sampling points in the parameter space. Generally, when calculating the 3rd to 4th cycles, the time-averaged aerodynamic force and the peak value of the instantaneous aerodynamic force hardly change with the increase in the number of rotation cycles, and it can be considered that the calculation converges.
[0061] In the above steps, parameter space sampling and CFD calculations are not necessary. When existing data is available, give priority to using the existing data. When the existing data cannot meet the training requirements, supplement the samples through the above steps.
[0062] After that, data preprocessing is carried out. Take the periodic unsteady aerodynamic data after stabilization. Use the sample space parameters (N parameters for N - dimension) and the time series of forces and moments in three spatial directions in the corresponding Cartesian coordinate system (6×200) as a sample of the dataset for preprocessing. During the preprocessing, data normalization should be performed on the sample space parameters to avoid affecting the neural network training efficiency due to differences in feature scales. The calculation formula for data normalization is:
[0063]
[0064] In the formula, X n represents the result after normalization, X is the original data, X max and X min represent the maximum and minimum values of the value range respectively.
[0065] After data preprocessing, the method of stratified sampling is adopted to divide the training set (60%), validation set (20%) and test set (20%). This method can ensure that the proportion of each category is roughly the same as that in the original dataset. Data preprocessing can be implemented through MATLAB programming.
[0066] Next, the neural network structure is defined. The neural network structure proposed in this scheme is built based on PyTorch, with BiGRU as the core, and is specifically composed of 1 input layer, 2 BiGRU layers, 4 fully - connected layers and 1 output layer.
[0067] The input layer should confirm the input by the parameter space. Regarding the BiGRU layer, each BiGRU layer is composed of two unidirectional GRU layers, which process the time - series data along the forward and reverse directions respectively. Therefore, it can consider the information of the past and future at a certain moment at the same time, and can fully explore the unsteady characteristics of the tilt - rotor. The connection of the 4 - layer fully - connected layers, combined with the Leaky ReLU (Leaky Rectified Linear Unit) activation function, can realize the non - linear mapping from features to output.
[0068]
[0069] In the formula, α represents the negative slope, and the empirical parameter 1 / 5.5 is taken.
[0070] The number of nodes in each layer of the BiGRU layer and the fully connected layer needs to be adjusted and selected according to the dimensions and quantity of the calculation samples. The final output layer itself can output equally spaced time series (6×200) of forces and torques in three spatial directions in the Cartesian coordinate system under corresponding parameters. The average values of the six time series within the rotation period are calculated respectively to obtain the time-averaged forces and torques. Combining parameters such as the tilt-rotor rotational speed, oncoming flow direction, and oncoming flow velocity, the aerodynamic power and aerodynamic efficiency can be calculated. Therefore, this solution can output three-axis periodic instantaneous forces, instantaneous torques, time-averaged forces, time-averaged torques, as well as aerodynamic power and aerodynamic efficiency.
[0071] The above steps include sub-optimal solutions. BiGRU is a variant of GRU (Gated Recurrent Unit). In essence, RNN (Recurrent Neural Network), GRU, LSTM (Long Short-Term Memory), and their variants can be used to replace the BiGRU layer. However, it is found in practice that the BiGRU layer has better effects and requires fewer samples under the same prediction accuracy.
[0072] In addition, during the definition of the neural network, regarding the optimizer, the Adam optimizer is adopted. As an optimization algorithm with an adaptive learning rate, it has advantages such as fast convergence speed, high computational efficiency, and excellent sparse data processing ability. Different learning rates are adopted in the early, middle, and late stages of training. A learning rate of 0.01 is adopted in the initial stage of training to enable the model to converge quickly, and learning rates of 0.001 and 0.0001 are adopted in the middle and late stages respectively to improve the stability and fineness of training.
[0073] Regarding the loss function of the neural network, since there is a strong non-linear region in the aerodynamic performance of the tilt-rotor, an improved mean squared error loss function (MSE Loss, Mean Squared Error Loss) with weight correction is proposed. The prediction accuracy is improved by increasing the sample weights in the strong non-linear region. To reduce the required number of samples, some physical information is introduced as a constraint in the loss function. During a stable rotation period of the tilt-rotor, the physical values at the start and end moments should be equal and the first-order derivatives should be equal. Therefore, L1 and L2 are introduced as physical constraints. The final loss function L total is expressed as a linear combination of MSE, L1, and L2.
[0074]
[0075] In the formula, n is the number of samples, T is the time series length, w i is the weight coefficient of the i-th sample, and this weight coefficient is adjusted according to the physical prior knowledge of the distribution of the strong non-linear region of the unsteady aerodynamic performance of the tilt-rotor. yit is the CFD result of the i-th sample at the t-th time step, is the predicted value of the i-th sample at the t-th time step, and λ1 and λ2 are the hyperparameters of L1 and L1 respectively.
[0076] So far, neural network training can be carried out. The complete neural network training process is as follows:
[0077] (1) Forward propagation. Use the input data to predict the aerodynamic force coefficients and moment coefficients predicted by the neural network under the current parameters.
[0078] (2) Calculate the loss. Calculate the error loss between the predicted values of the current aerodynamic force coefficients and moment coefficients and the CFD results.
[0079] (3) Backward propagation and update the model parameters. Perform backward propagation, calculate the gradients through the defined Adam optimizer, and update the parameters of the neural network model.
[0080] (4) Repeat the above steps until the loss function is small enough and the neural network training is completed.
[0081] Finally, the model performance is evaluated. Two error measurement indicators are adopted to describe the error between the prediction results of the neural network on the test set and the CFD results. The absolute error δ can be used to measure the error results of the time-averaged force and moment. The root mean square error RMSE (Root Mean Square Error) can be used to measure the error results of the instantaneous force and moment. When the absolute error of all samples in the test set does not exceed 5%, and the RMSE does not exceed 0.1, it can be considered that the accuracy requirement is met, and then the surrogate model can be output. Otherwise, according to the performance of the test set, supplement the number of samples and retrain.
[0082] The key points and pre-protection points of the technical solution of the present invention are as follows:
[0083] (1) Considering the unsteady characteristics of the tilt-rotor aerodynamics, a tilt-rotor unsteady aerodynamic performance surrogate model is constructed based on the time-series neural network structure (with BiGRU as the core).
[0084] (2) Develop a method for defining the loss function with physical constraints and weight correction to reduce the sample requirements and enhance the fitting ability for the strong nonlinear region.
[0085] Through experimental verification, the technical solution of the present invention is feasible. The neural network structure proposed in this solution has been tested on the instantaneous aerodynamic performance dataset of 84 sample data in two dimensions. It can achieve the expected functions of this solution and meet the prediction accuracy requirements on the test set (the absolute error of all samples in the test set does not exceed 5%, and the RMSE does not exceed 0.1).
[0086] It should be noted that in this text, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, such that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device. Without further limitation, an element defined by the phrase "comprising an..." does not exclude the presence of additional identical elements in the process, method, article or device comprising such element.
[0087] The serial numbers of the embodiments of the present invention above are only for description and do not represent the superiority or inferiority of the embodiments.
[0088] Through the description of the above embodiments, those skilled in the art can clearly understand that the above implementation methods can be realized by means of software plus a necessary general hardware platform. Of course, it can also be realized by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disc) and includes several instructions for causing a terminal (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0089] The embodiments of the present invention have been described above in conjunction with the accompanying drawings. However, the present invention is not limited to the above specific embodiments. The above specific embodiments are merely illustrative and not restrictive. Under the inspiration of the present invention, those of ordinary skill in the art can also make many forms without departing from the spirit and scope protected by the claims of the present invention, and these all fall within the protection scope of the present invention.
Claims
1. A method for predicting the unsteady aerodynamic performance of a tilt-rotor based on machine learning, characterized in that, The prediction method includes the following steps: Step 1: Determine the parameter space and obtain the original data. The spatial dimension of the parameter space has pitch angle, inflow angle, inflow magnitude, and rotational speed; Step 2: Preprocess the original data, and divide the preprocessed original data into a training set, a validation set, and a test set by using the method of stratified sampling; Step 3: Define the neural network structure, and train the neural network with the training set and the validation set to obtain the trained neural network; Step 4: Use the test set to evaluate the performance of the trained neural network. When the accuracy requirement is met, output the surrogate model; otherwise, supplement new data samples and retrain until the surrogate model is output; Step 5: Obtain the three-axis periodic instantaneous force, instantaneous torque, time-averaged force, time-averaged torque, aerodynamic power, and aerodynamic efficiency through the surrogate model.
2. The tilt-rotor unsteady aerodynamic performance prediction method according to claim 1, wherein In the said Step 1, for the original data, if there is existing data, use the existing data; if there is no existing data, or the existing data does not meet the training requirements, sample and supplement new data samples, wherein, the following method is used to sample and supplement new data samples: After determining the parameter space, establish the value dimension and range, and use the Latin hypercube sampling method to sample the parameter space to obtain sample points; Perform CFD calculations on the sample points, apply a CFD solver, and based on the overlapping grid method, calculate the parameter conditions of the sample points through transient solution until convergence to obtain new data samples.
3. The tilt-rotor unsteady aerodynamic performance prediction method according to claim 1, characterized in that In the said Step 2, the data preprocessing specifically includes: For the original data, take the steady-state periodic unsteady aerodynamic data, and use the sample space parameters and the time series of instantaneous forces and instantaneous torques in three spatial directions in the corresponding Cartesian coordinate system as a sample of the data set, and perform data normalization to achieve data preprocessing.
4. The method for predicting the unsteady aerodynamic performance of a tilt-rotor according to claim 1, characterized in that In the said Step 3, the neural network structure is built based on PyTorch, and includes 1 input layer, 2 BiGRU layers, 4 fully connected layers, and 1 output layer, wherein, in the neural network structure, an Adam optimizer is used, and a training strategy with variable learning rate is used.
5. The tilt-rotor unsteady aerodynamic performance prediction method according to claim 4, wherein In the neural network structure, the output layer outputs the equally spaced time series of instantaneous forces and instantaneous torques in three spatial directions in the Cartesian coordinate system corresponding to the parameters. The average values of the six time series within the rotation period are obtained to get the time-averaged force and time-averaged torque, and combined with parameters such as the tilt-rotor rotational speed, inflow direction, and inflow velocity, the aerodynamic power and aerodynamic efficiency are calculated.
6. The method for predicting the unsteady aerodynamic performance of a tilt-rotor according to claim 1, wherein In the neural network structure, the loss function L total is expressed as: L1 = (y i1 - y iT ) 2 L total = MSE + λ1L1 + λ2L2 where n is the number of samples, T is the length of the time series, and w i is the weight coefficient of the i-th sample; y it is the CFD result of the i-th sample at the t-th time step, is the predicted value of the i-th sample at the t-th time step, y i1 is the CFD result of the i-th sample at the 1st time step; y iT is the CFD result of the i-th sample at the last time step in a rotation period, that is, the CFD result at the T-th time step; λ1 and λ2 are the hyperparameters of L1 and L1 respectively, and L1 and L1 are physical constraints, Among them, w i The weight coefficient is adjusted according to the physical prior knowledge of the distribution in the strong nonlinear region of the unsteady aerodynamic performance of the tilt-rotor.
7. The tilt-rotor unsteady aerodynamic performance prediction method according to claim 1, characterized in that In the said Step 4, the absolute error δ is used to measure the error results of the time-averaged force and time-averaged torque, and the root mean square error is used to measure the error results of the instantaneous force and torque.
8. The method for predicting the unsteady aerodynamic performance of a tiltrotor according to claim 1, wherein In the said Step 4, when the absolute error δ of all samples in the test set does not exceed 5%, and the RMSE does not exceed 0.1, it is considered that the accuracy requirement is met, and the surrogate model is output.
9. A prediction system for unsteady aerodynamic performance of a tiltrotor based on machine learning, the system comprising: A processor and a memory for storing executable instructions; characterized in that the processor is configured to execute the executable instructions to execute the machine learning-based tilt-rotor unsteady aerodynamic performance prediction method according to any one of claims 1 to 8.
10. A computer-readable storage medium, characterized in that, A computer program is stored thereon, and when the computer program is executed by a processor, it implements the machine learning-based unsteady aerodynamic performance prediction method for a tiltrotor as described in any one of claims 1 to 8.