Methods for establishing a predictive model for pressure distribution data
By dividing the surface of an aircraft into grid regions and constructing a neural network model with adaptive boundary expansion, the problem of low efficiency in pressure distribution data prediction in aircraft aerodynamic design is solved, and fast and accurate pressure distribution prediction is achieved, which is applicable to the aerodynamic design of aircraft with complex shapes.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CALCULATION AERODYNAMICS INST CHINA AERODYNAMICS RES & DEV CENT
- Filing Date
- 2022-10-21
- Publication Date
- 2026-06-30
AI Technical Summary
Existing technologies for predicting pressure distribution data in aircraft aerodynamic design are inefficient, computationally intensive, and time-consuming. Furthermore, traditional methods struggle to quickly and accurately predict pressure distribution data on aircraft surfaces.
A deep learning-based neural network approach is adopted. By dividing the surface of the aircraft into multiple grid regions, a reference neural network model is constructed, and adaptive boundary expansion is performed to establish a pressure distribution data prediction model. The preset expansion region is updated using the validation set until the preset conditions are met.
It improves the prediction accuracy and speed of pressure distribution data on aircraft surfaces, shortens modeling time, enhances prediction accuracy at boundaries, and is suitable for rapid aerodynamic design of complex-shaped aircraft.
Smart Images
Figure CN115544917B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of aircraft aerodynamic design technology, and in particular to a method for establishing a pressure distribution data prediction model. Background Technology
[0002] Currently, "aircraft design begins with aerodynamics." Aircraft are complex systems involving multiple disciplines, and aerodynamic performance serves as the foundation and starting point for other subsystems, possessing fundamental, global, and leading characteristics. The quality of aerodynamic performance largely determines the overall performance of the aircraft. Therefore, aerodynamic modeling is an important fundamental research topic in the development of aircraft. Furthermore, the optimization calculation process requires extensive use of original high-precision calculation models, leading to low aerodynamic optimization efficiency and problems such as excessive computational load and long processing times.
[0003] By constructing an aerodynamic dataset for a certain type of aircraft and establishing an aerodynamic characteristic prediction model using a deep learning-based neural network method, the aerodynamic characteristics (especially pressure distribution data) of the aircraft can be predicted quickly, accurately, and at low cost. This can effectively accelerate the aerodynamic design process of the aircraft and provide a reliable surrogate model for the rapid design of the aerodynamic layout of the aircraft. However, the existing achievements at home and abroad mainly focus on the research of integral aerodynamic force / local aerothermal prediction based on classical machine learning algorithms or deep neural network models. Generally, the amount of aerodynamic characteristic parameter data (such as pressure distribution data) predicted is relatively small.
[0004] In summary, how to construct a pressure distribution data prediction model to quickly and accurately predict the pressure distribution data on the surface of aircraft is an urgent problem to be solved. Summary of the Invention
[0005] In view of this, the purpose of this invention is to provide a method for establishing a pressure distribution data prediction model, which can construct a pressure distribution data prediction model to quickly and accurately predict the pressure distribution data on the surface of an aircraft. The specific solution is as follows:
[0006] In a first aspect, this application discloses a method for establishing a pressure distribution data prediction model, including:
[0007] The reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions are divided into training set, validation set and test set, and each training set is further divided into training subsets that correspond one-to-one with grid regions; the grid region is the area obtained by dividing the surface of the aircraft into grids.
[0008] A reference neural network model is set for each of the grid regions;
[0009] Based on the preset expansion region corresponding to each training subset, adaptive boundary expansion is performed on each training subset to obtain the expanded training subset;
[0010] Within each grid region, a corresponding reference neural network model is trained using the corresponding extended training subset, so that the reference neural network model has a mapping relationship between the working conditions and the pressure distribution data;
[0011] The preset expansion region is updated using the validation set, and the updated expansion region is used as the preset expansion region. Then, the process jumps to the step of the preset expansion region corresponding to each training subset until the first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region.
[0012] The pressure distribution data prediction model is tested using the test set. If the test results meet the first preset condition, the pressure distribution data prediction model is the final prediction model.
[0013] Optionally, the step of adaptively expanding the boundaries of each training subset based on a preset expansion region corresponding to each training subset to obtain an expanded training subset includes:
[0014] The number of first grid points in the grid region corresponding to the training subset and the number of second grid points in the corresponding preset extended region are counted, and the sum of the number of first grid points and the number of second grid points is taken as the number of grid points after extension.
[0015] If the number of grid points after expansion meets the second preset condition, then the training subset is directly expanded by adaptive boundary expansion based on the preset expansion region corresponding to the training subset to obtain the expanded training subset.
[0016] If the number of grid points after expansion does not meet the second preset condition, the preset expansion distance is adjusted based on the number of grid points after expansion to obtain the adjusted expansion distance, and the adjusted expansion area is determined based on the adjusted expansion distance.
[0017] The expanded region is determined based on the adjusted expanded region and the grid region, and the reference pressure distribution data corresponding to the expanded region is used as the expanded training subset. Then, the adjusted expanded region is used as the preset expanded region.
[0018] Optionally, the second preset condition is that the absolute value of the difference between the number of reference points and the number of expanded grid points does not exceed the target proportion of the number of reference points; the number of reference points is the product of the first number of grid points and the preset expansion multiple; the preset expansion multiple is an expansion multiple for the first number of grid points that is randomly set in advance according to the expansion parameter setting rules.
[0019] Optionally, updating the preset extended region using the verification set includes:
[0020] Based on the operating conditions corresponding to the validation set, and using each of the reference neural network models to collaboratively predict the first target pressure distribution data of the aircraft;
[0021] The preset expansion region is updated based on the first target pressure distribution data and the reference pressure distribution data in the verification set.
[0022] Optionally, updating the preset extended region based on the first target pressure distribution data and the reference pressure distribution data in the verification set includes:
[0023] Calculate the first mean absolute error between the first target pressure distribution data and the reference pressure distribution data in the verification set;
[0024] Acquire the second target pressure distribution data predicted without adaptive boundary expansion, and calculate the second mean absolute error between the reference pressure distribution data and the second target pressure distribution data;
[0025] The preset expansion factor is updated based on the first mean absolute error and the second mean absolute error to update the preset expansion distance and obtain the updated expansion distance. The updated expansion area is then determined based on the updated expansion distance.
[0026] Optionally, the step of predicting the first target pressure distribution data of the aircraft based on the operating conditions corresponding to the validation set and using each of the reference neural network models collaboratively includes:
[0027] Based on the working conditions corresponding to the validation set, and by using each of the reference neural network models to validate the local pressure distribution data corresponding to the grid region, the local pressure distribution data corresponding to each grid region is obtained.
[0028] The sum of all the local pressure distribution data is taken as the first target pressure distribution data of the aircraft.
[0029] Optionally, before performing adaptive boundary expansion on each training subset based on a preset expansion region corresponding to each training subset to obtain the expanded training subset, the method further includes:
[0030] The region boundary of the grid region corresponding to the training subset is determined, and the average of the maximum diagonal distance and the minimum diagonal distance of the boundary corresponding to the region boundary is used as the preset expansion distance. Then, the preset expansion region is determined based on the preset expansion distance to obtain the preset expansion region corresponding to each training subset.
[0031] As can be seen, this application divides the reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions into a training set, a validation set, and a test set, and divides each training set into a training subset corresponding to a grid region; the grid region is the region obtained by dividing the surface of the aircraft into grids; a reference neural network model is set for each grid region; based on the preset expansion region corresponding to each training subset, each training subset is adaptively expanded to obtain an expanded training subset; within each grid region, the corresponding reference neural network model is trained using the corresponding expanded training subset, so that the reference neural network model has a mapping relationship between the operating conditions and the pressure distribution data; the preset expansion region is updated using the validation set, the updated expansion region is used as the preset expansion region, and the process jumps to the step based on the preset expansion region corresponding to each training subset, until the first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region; the pressure distribution data prediction model is tested using the test set, and if the test result meets the first preset condition, the pressure distribution data prediction model is the final prediction model. Therefore, this application divides the aircraft surface into multiple grid regions and trains a pressure distribution data prediction model for each grid region. Compared with the single prediction model for the aircraft, this improves measurement accuracy and speeds up prediction. In this application, the model of each grid region is trained by adding data around the region boundary using an adaptive boundary extension method, thereby increasing the prediction accuracy of pressure distribution data at the grid region boundary. This application changes the preset extension region of each grid region, thereby finding a more suitable preset extension region for each grid region to improve prediction speed and accuracy. Attached Figure Description
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0033] Figure 1 A flowchart illustrating a method for establishing a pressure distribution data prediction model provided in this application;
[0034] Figure 2 A schematic diagram illustrating a grid region division and adaptive boundary expansion direction provided in this application;
[0035] Figure 3 A flowchart illustrating a specific method for establishing a pressure distribution data prediction model provided in this application;
[0036] Figure 4 A schematic diagram illustrating the process of establishing a pressure distribution data prediction model provided in this application. Detailed Implementation
[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0038] Currently, by constructing an aerodynamic dataset for a certain type of aircraft and using a neural network method based on deep learning to establish an aerodynamic characteristic prediction model, the aerodynamic characteristics (especially pressure distribution data) of the aircraft can be predicted quickly, accurately, and at low cost. This can effectively accelerate the aerodynamic design process of the aircraft and provide a reliable surrogate model for the rapid design of the aerodynamic layout of the aircraft. However, the existing achievements at home and abroad are mainly focused on the research of integral aerodynamic force / local aerothermal prediction based on classical machine learning algorithms or deep neural network models. Generally, the amount of aerodynamic characteristic parameter data (such as pressure distribution data) predicted is relatively small.
[0039] To overcome the above problems, this application provides a scheme for establishing a pressure distribution data prediction model, which can construct a pressure distribution data prediction model to quickly and accurately predict the pressure distribution data on the surface of an aircraft.
[0040] See Figure 1 As shown in the embodiment of this application, a method for establishing a pressure distribution data prediction model is disclosed, the method comprising:
[0041] Step S11: Divide the reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions into training set, validation set and test set, and divide each training set into a training subset corresponding to a grid region; the grid region is the area obtained by dividing the surface of the aircraft into grids.
[0042] In this embodiment, the National Numerical Wind Tunnel Engineering (NNW) series software is used to calculate the flow field data of a certain type of aircraft under different operating conditions. The surface pressure distribution data of the aircraft is obtained through post-processing using programs such as NNW-TopViz. A sample set of surface pressure distribution data for different grid regions is established to obtain a reference pressure distribution dataset for the aircraft under different operating conditions. It should be noted that the operating conditions include, but are not limited to, Mach number, roll angle, and angle of attack.
[0043] In this embodiment, the surface of the aircraft is divided into several grid regions. The pressure distribution data on the aircraft surface is enormous, making it difficult to quickly predict the pressure distribution of the entire aircraft by constructing a single aerodynamic model. Therefore, each training set is divided into training subsets corresponding to different grid regions to construct pressure distribution data prediction models for each grid region. Thus, the pressure distribution data sample set is divided according to different grid regions, and a neural network model for predicting pressure distribution is trained in each region. The construction of deep neural network models for predicting pressure distribution in different grid regions solves the problem of the enormous amount of aircraft pressure distribution data.
[0044] In one specific embodiment, a total of 799 sets of sample data were obtained under four different aerodynamic shapes and operating conditions; the number of training sets and validation sets for one different shape and operating condition were 183 and 12 sets, respectively; the number of training sets and validation sets for the four different shapes and operating conditions were 183*4=732 sets and 12*4=48 sets, respectively; and the number of test sets were 4 sets, 5 sets, 5 sets, and 5 sets, respectively. It should be noted that there should be no overlap between the test set, training set, and validation set; the surface of the aircraft contains 229 grid regions, and each training set is divided into 229 corresponding training set subsets according to the grid.
[0045] Step S12: Set a reference neural network model for each of the grid regions.
[0046] In this embodiment of the application, each grid region is provided with a reference neural network model so as to obtain a pressure distribution data prediction model through inverse training of the reference neural network model.
[0047] Step S13: Based on the preset expansion region corresponding to each training subset, perform adaptive boundary expansion on each training subset to obtain the expanded training subset.
[0048] In this embodiment, to enhance the prediction accuracy of the obtained pressure distribution data prediction model while ensuring the prediction effect of pressure distribution at the boundaries of different grid regions, the training set needs to be expanded outward by a certain range during the construction of the aerodynamic model for each grid region. That is, based on the preset expansion region corresponding to each training subset, adaptive boundary expansion is performed on each training subset to obtain the expanded training subset. It should be noted that adaptive boundary expansion, which increases the data around the boundary, can enhance the prediction accuracy of the grid region boundary data.
[0049] In this embodiment, when initially determining the preset expansion region, during the aerodynamic model construction process of each grid region, the training dataset needs to be expanded outward by a certain range. Since different grid regions have different areas and varying grid point sparsity, it is necessary to first calculate the maximum and minimum diagonal distances of the boundaries, and use their average as the initial expansion distance L0. Then, using all boundary points of the grid region as the starting point of expansion and the corresponding L0 as the initial expansion distance, the number of grid points absorbed during the expansion process of all boundary points is counted to obtain the expanded training subset corresponding to each grid region. It should be noted that before adaptively expanding each training subset to obtain the expanded training subset based on the preset expansion region corresponding to each training subset, the process further includes: determining the region boundary of the grid region corresponding to the training subset, and using the average of the maximum and minimum diagonal distances of the boundary corresponding to the region boundary as the preset expansion distance (i.e., half the distance between the longest and shortest diagonal lines of the boundary). Then, based on the preset expansion distance, the preset expansion region is determined to obtain the preset expansion region corresponding to each training subset. It should be noted that the division of the grid region and the direction of the adaptive boundary expansion are as follows: Figure 2 As shown.
[0050] In this embodiment of the application, in order to shorten the modeling time, enhance the prediction accuracy, and facilitate the counting of grid points in each grid region, it is necessary to control the preset expansion region corresponding to each grid region to be neither too large nor too small, that is, to constrain the number of grid points absorbed in the expansion process, and further control the size of the training subset. In addition, since the preset expansion region is determined by the preset expansion distance, it is necessary to control the preset expansion distance. It should be noted that the step of adaptively expanding the boundary of each training subset to obtain an expanded training subset based on the preset expansion region corresponding to each training subset includes: counting the number of first grid points in the grid region corresponding to the training subset and the number of second grid points corresponding to the preset expansion region, and taking the sum of the number of first grid points and the number of second grid points as the number of expanded grid points; if the number of expanded grid points meets a second preset condition, then the training subset is directly adaptively expanded based on the preset expansion region corresponding to the training subset to obtain an expanded training subset; if the number of expanded grid points does not meet the second preset condition, then the preset expansion distance is adjusted based on the number of expanded grid points to obtain an adjusted expansion distance, and the adjusted expansion region is determined based on the adjusted expansion distance; the expanded region is determined based on the adjusted expansion region and the grid region, and the reference pressure distribution data corresponding to the expanded region is taken as the expanded training subset, and then the adjusted expansion region is taken as the preset expansion region.
[0051] It should be noted that the second preset condition is that the absolute value of the difference between the number of reference points and the number of expanded grid points does not exceed the target proportion of the number of reference points; the number of reference points is the product of the first number of grid points and the preset expansion factor; the preset expansion factor is an expansion factor for the first number of grid points that is randomly set in advance according to the expansion parameter setting rules. It should be noted that, in order to control the preset expansion distance, this application presets the expansion factor in advance. Ideally, the final number of all grid points corresponding to the grid area after expansion is N times the original number of grid points (i.e., the first number of grid points), but it is also acceptable as long as the absolute value of the difference between the number of reference points and the number of expanded grid points does not exceed the target proportion of the number of reference points; the target proportion is 5%. In summary, the size of the expansion range in the boundary adaptive expansion technology is determined by the number of grid points in the grid area and the expansion factor N, thereby realizing the interaction of information from different grid areas, enabling reasonable expansion of boundary data for grid areas with different numbers of grid points and shapes, while preventing problems such as excessively high model training costs caused by too large an expansion range. It should be noted that the rule for setting the expansion parameters is that the upper limit of the preset expansion multiple is [value missing]. 3 Because model training time is affected by the number of samples in the training subset (stress data), setting an upper limit can ensure the efficiency requirements of rapid modeling.
[0052] Step S14: Within each grid region, train the corresponding reference neural network model using the corresponding extended training subset, so that the reference neural network model has a mapping relationship between the operating conditions and the pressure distribution data.
[0053] In this embodiment, after designing an appropriate neural network model framework, the factors affecting the prediction accuracy of the neural network are determined through debugging, including the number of epochs, learning rate, loss function, and activation function. The set of neural network parameters with the highest prediction accuracy is found, and the corresponding reference neural network model is trained accordingly, so that the reference neural network model has a mapping relationship between the operating conditions and the pressure distribution data. It should be noted that during the pressure prediction process, each neural network model only predicts the pressure data of the original region (corresponding grid region), and finally, the prediction results of all grid regions constitute the prediction result of the entire aircraft pressure distribution.
[0054] Step S15: Update the preset extended region using the validation set, use the updated extended region as the preset extended region, and jump to the step based on the preset extended region corresponding to each training subset until the first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region.
[0055] In this embodiment, the validation set can be used to adaptively expand the training set by a preset expansion factor N, thereby updating the preset expansion distance and further updating the preset expansion region. The updated expansion region is then used as the preset expansion region, and the process jumps to the step of adaptively expanding the boundaries of each training subset based on the preset expansion region corresponding to each training subset to obtain the expanded training subset, until a first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region. It should be noted that if the first preset condition cannot be met, the expansion is performed according to the maximum preset expansion distance to ensure model efficiency.
[0056] It should be noted that, in summary, the types of input and output data for the pressure distribution prediction model are as follows: In order to achieve the goal of rapid prediction of the surface pressure distribution of a certain type of aircraft based on boundary adaptive extension, the operating conditions of the aircraft (including Mach number, roll angle and angle of attack) are fed into the neural network model as one of the inputs for training and prediction. The output variable of the model is the surface pressure distribution data of interest to the designers of a certain type of aircraft.
[0057] Step S16: Test the pressure distribution data prediction model using the test set. If the test result meets the first preset condition, the pressure distribution data prediction model is the final prediction model.
[0058] In this embodiment of the application, the pressure distribution data prediction model is tested using the test set to ensure its practicality and applicability. If the test result meets the first preset condition, the pressure distribution data prediction model is the final prediction model. If the test result does not meet the first preset condition, it needs to be retrained to obtain a new pressure distribution data prediction model.
[0059] It should be noted that if the test result differs from the first preset condition by no more than the preset threshold, it indicates that the training algorithm is incorrect and the training algorithm should be modified. If the test result differs from the first preset condition by more than the preset threshold, it indicates that the reference pressure distribution dataset is incorrect, which may be due to insufficient dataset quantity or insufficient regularity of the dataset. It is necessary to increase the dataset quantity or replace the dataset.
[0060] As can be seen, this application divides the reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions into a training set, a validation set, and a test set, and divides each training set into a training subset corresponding to a grid region; the grid region is the region obtained by dividing the surface of the aircraft into grids; a reference neural network model is set for each grid region; based on the preset expansion region corresponding to each training subset, each training subset is adaptively expanded to obtain an expanded training subset; within each grid region, the corresponding reference neural network model is trained using the corresponding expanded training subset, so that the reference neural network model has a mapping relationship between the operating conditions and the pressure distribution data; the preset expansion region is updated using the validation set, the updated expansion region is used as the preset expansion region, and the process jumps to the step based on the preset expansion region corresponding to each training subset, until the first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region; the pressure distribution data prediction model is tested using the test set, and if the test result meets the first preset condition, the pressure distribution data prediction model is the final prediction model. Therefore, this application divides the aircraft surface into multiple grid regions and trains a pressure distribution data prediction model for each grid region. Compared with the single prediction model for the aircraft, this improves measurement accuracy and speeds up prediction. In this application, the model of each grid region is trained by adding data around the region boundary using an adaptive boundary extension method, thereby increasing the prediction accuracy of pressure distribution data at the grid region boundary. This application changes the preset extension region of each grid region, thereby finding a more suitable preset extension region for each grid region to improve prediction speed and accuracy.
[0061] See Figure 3 As shown in the figure, this application discloses a specific method for establishing a pressure distribution data prediction model, the method including:
[0062] Step S21: Divide the reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions into training set, validation set and test set, and divide each training set into a training subset corresponding to a grid region; the grid region is the area obtained by dividing the surface of the aircraft into grids.
[0063] For a more detailed description of the process of step S21, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0064] Step S22: Set a reference neural network model for each of the grid regions.
[0065] For a more detailed description of the process of step S22, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0066] Step S23: Based on the preset expansion region corresponding to each training subset, perform adaptive boundary expansion on each training subset to obtain the expanded training subset.
[0067] For a more detailed description of the process of step S23, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0068] Step S24: Within each grid region, train the corresponding reference neural network model using the corresponding extended training subset, so that the reference neural network model has a mapping relationship between the operating conditions and the pressure distribution data.
[0069] For a more detailed description of the process of step S24, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0070] Step S25: Based on the operating conditions corresponding to the validation set, and using each of the reference neural network models to collaboratively predict the first target pressure distribution data of the aircraft.
[0071] In this embodiment, during the pressure prediction process based on the operating conditions corresponding to the validation set, each neural network model only predicts the pressure data of the original region (corresponding grid region). Finally, the prediction results of all grid regions constitute the prediction result of the entire aircraft pressure distribution, which is the first target pressure distribution data of the aircraft. It should be noted that the step of predicting the first target pressure distribution data of the aircraft based on the operating conditions corresponding to the validation set and using each of the reference neural network models collaboratively includes: verifying the local pressure distribution data corresponding to the grid region based on the operating conditions corresponding to the validation set and using each of the reference neural network models to obtain the local pressure distribution data corresponding to each grid region; and using the sum of all the local pressure distribution data as the first target pressure distribution data of the aircraft.
[0072] Step S26: Based on the first target pressure distribution data and the reference pressure distribution data in the validation set, update the preset expansion region, use the updated expansion region as the preset expansion region, and jump to the step based on the preset expansion region corresponding to each training subset, until the first preset condition is met to obtain the pressure distribution data prediction model.
[0073] In this embodiment, the preset expansion region is updated based on the first target pressure distribution data and the reference pressure distribution data in the validation set. Specifically, the preset expansion region is updated based on the first target pressure distribution data, the reference pressure distribution data in the validation set, and the second target pressure distribution data predicted without adaptive boundary expansion. This includes: calculating the first mean absolute error between the first target pressure distribution data and the reference pressure distribution data in the validation set; obtaining the second target pressure distribution data predicted without adaptive boundary expansion, and calculating the second mean absolute error between the reference pressure distribution data and the second target pressure distribution data; updating the preset expansion factor based on the first mean absolute error and the second mean absolute error to update the preset expansion distance to obtain the updated expansion distance, and determining the updated expansion region based on the updated expansion distance. It should be noted that updating the preset expansion region directly updates the preset expansion factor. After updating the preset expansion factor, the preset expansion distance can be automatically updated, and the preset expansion region can be further updated. It should also be noted that the range of adaptive boundary expansion is adjusted by the MAE value (first mean absolute error) of the pressure distribution prediction result, forming a closed-loop feedback loop and automating the pressure distribution modeling process.
[0074] It should be noted that if the first preset condition is met, the adaptive expansion stops and a pressure distribution data prediction model is obtained; if the first preset condition is not met, the expansion is carried out according to the maximum preset expansion distance to ensure model efficiency; it should be noted that the maximum preset expansion distance can be obtained when the adaptive expansion multiple (preset expansion multiple) is a pre-defined upper limit value; the upper limit value set in this application can be 3, and no specific limitation is made here.
[0075] It should be noted that updating the preset expansion factor based on the first mean absolute error and the second mean absolute error specifically requires ensuring that the predicted MAE (first mean absolute error) with adaptive boundary expansion is at least 20% lower on average than the predicted MAE (second mean absolute error) without considering boundary expansion; and the prediction time cannot exceed the target time, which is set according to actual conditions, and in this application, it is generally set to no more than 3 hours. It should also be noted that the first preset condition is that the predicted MAE (first mean absolute error) with adaptive boundary expansion is at least 20% lower on average than the predicted MAE (second mean absolute error) without considering boundary expansion (at this point, the mean absolute error decreases, and the prediction accuracy improves), and the prediction time cannot exceed the target time.
[0076] It should be noted that this application introduces the mean absolute error (MAE) to measure the quality of the prediction results. The measurement result is used to optimize the boundary adaptive expansion factor N. That is, the mean absolute error (MAE) of the predicted data relative to the real data is calculated, and the boundary adaptive expansion factor N is adjusted according to the MAE. Finally, the prediction accuracy and the time cost of model training are weighed, and N is set to 2.0. The prediction results MAE of 19 test sets for the four shapes are shown in Table 1. The prediction MAE based on boundary adaptive expansion is reduced by an average of more than 20% compared with the prediction MAE without considering boundary expansion.
[0077] Table 1
[0078]
[0079] It should be noted that adaptive expansion stops after the first preset condition is met, that is, the acquisition of pressure prediction results stops, and the optimization process of the adaptive expansion factor is further stopped. Therefore, the adaptive expansion factor is fixed at this time.
[0080] Step S27: Test the pressure distribution data prediction model using the test set. If the test result meets the first preset condition, the pressure distribution data prediction model is the final prediction model.
[0081] For a more detailed description of the process of step S27, please refer to the relevant content disclosed in the foregoing embodiments, which will not be repeated here.
[0082] As can be seen, this application divides the reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions into a training set, a validation set, and a test set, and divides each training set into a training subset corresponding to a grid region; the grid region is the region obtained by dividing the surface of the aircraft into grids; a reference neural network model is set for each grid region; based on the preset expansion region corresponding to each training subset, each training subset is adaptively expanded to obtain an expanded training subset; within each grid region, the corresponding reference neural network model is trained using the corresponding expanded training subset, so that the reference neural network model has a mapping relationship between the operating conditions and the pressure distribution data; the preset expansion region is updated using the validation set, the updated expansion region is used as the preset expansion region, and the process jumps to the step based on the preset expansion region corresponding to each training subset, until the first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region; the pressure distribution data prediction model is tested using the test set, and if the test result meets the first preset condition, the pressure distribution data prediction model is the final prediction model. Therefore, this application divides the aircraft surface into multiple grid regions and trains a pressure distribution data prediction model for each grid region. Compared with the single prediction model for the aircraft, this improves measurement accuracy and speeds up prediction. In this application, the model of each grid region is trained by adding data around the region boundary using an adaptive boundary extension method, thereby increasing the prediction accuracy of pressure distribution data at the grid region boundary. This application changes the preset extension region of each grid region, thereby finding a more suitable preset extension region for each grid region to improve prediction speed and accuracy.
[0083] See Figure 4The diagram illustrates the process of establishing a pressure distribution data prediction model. First, a dataset of aircraft surface pressure distribution under different operating conditions is established. Second, the sample dataset and corresponding operating conditions are divided into training, validation, and test sets according to the operating conditions. Third, the training set data is divided into training subsets based on several grid regions of the aircraft. Fourth, each training subset undergoes adaptive boundary expansion to obtain an expanded subset, thereby expanding the relevant data of the grid region boundaries. Fifth, the neural network model corresponding to the corresponding grid region is debugged using each training subset. Sixth, the validation set and the neural network prediction model corresponding to each grid region are used to collaboratively predict the predicted pressure data of the aircraft. Seventh, the first mean absolute error of the actual pressure data in the predicted pressure data set is calculated to measure the prediction accuracy. Eighth, the range of adaptive boundary expansion is optimized based on the prediction accuracy, and the process jumps back to step four to repeat the adaptive expansion until a pressure prediction model corresponding to each grid region is obtained. Ninth, the pressure distribution data prediction model is tested using the test set. If the test results meet the target accuracy, the pressure distribution data prediction model is the final prediction model.
[0084] It should be noted that this application measures prediction accuracy by the difference between the second absolute error and the first mean absolute error between the predicted data without adaptive expansion and the actual pressure data. It should also be noted that this application improves prediction accuracy by optimizing the range of adaptive expansion of the boundary based on prediction accuracy. Furthermore, this application solves the problem of massive data volume in pressure distribution prediction for complex-shaped aircraft by pre-fetching a neural network model for each grid, making neural network-based aerodynamic modeling more closely aligned with engineering application scenarios.
[0085] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the pressure distribution data prediction model establishment method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.
[0086] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0087] The steps of the algorithm described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0088] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0089] The above provides a detailed description of the method for establishing a pressure distribution data prediction model provided by the present invention. Specific examples have been used to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A pressure distribution data prediction model establishment method, characterized in that, include: The reference pressure distribution dataset of the aircraft under different operating conditions and the corresponding operating conditions are divided into training set, validation set and test set, and each training set is further divided into training subsets that correspond one-to-one with grid regions. The grid region is the area obtained by dividing the surface of the aircraft into a grid. A reference neural network model is set for each of the grid regions; Based on the preset expansion region corresponding to each training subset, adaptive boundary expansion is performed on each training subset to obtain the expanded training subset; Within each grid region, a corresponding reference neural network model is trained using the corresponding extended training subset, so that the reference neural network model has a mapping relationship between the working conditions and the pressure distribution data; The preset expansion region is updated using the validation set, and the updated expansion region is used as the preset expansion region. Then, the process jumps to the step of the preset expansion region corresponding to each training subset until the first preset condition is met to obtain the pressure distribution data prediction model corresponding to each grid region. The pressure distribution data prediction model is tested using the test set. If the test results meet the first preset condition, the pressure distribution data prediction model is the final prediction model.
2. The pressure distribution data prediction model establishing method according to claim 1, characterized in that, The step of adaptively expanding the boundaries of each training subset based on a preset expansion region to obtain an expanded training subset includes: The number of first grid points in the grid region corresponding to the training subset and the number of second grid points in the corresponding preset extended region are counted, and the sum of the number of first grid points and the number of second grid points is taken as the number of grid points after extension. If the number of grid points after expansion meets the second preset condition, then the training subset is directly expanded by adaptive boundary expansion based on the preset expansion region corresponding to the training subset to obtain the expanded training subset. If the number of grid points after expansion does not meet the second preset condition, the preset expansion distance is adjusted based on the number of grid points after expansion to obtain the adjusted expansion distance, and the adjusted expansion area is determined based on the adjusted expansion distance. The expanded region is determined based on the adjusted expanded region and the grid region, and the reference pressure distribution data corresponding to the expanded region is used as the expanded training subset. Then, the adjusted expanded region is used as the preset expanded region.
3. The pressure distribution data prediction model establishing method according to claim 2, characterized in that, The second preset condition is that the absolute value of the difference between the number of reference points and the number of expanded grid points does not exceed the target proportion of the number of reference points; the number of reference points is the product of the first number of grid points and the preset expansion factor; The preset expansion factor is a random expansion factor for the number of points in the first grid, set in advance according to the expansion parameter setting rules.
4. The pressure distribution data prediction model establishing method according to claim 3, characterized in that, Updating the preset extended region using the verification set includes: Based on the operating conditions corresponding to the validation set, and using each of the reference neural network models to collaboratively predict the first target pressure distribution data of the aircraft; The preset expansion region is updated based on the first target pressure distribution data and the reference pressure distribution data in the verification set.
5. The pressure distribution data prediction model establishing method according to claim 4, characterized in that, The step of updating the preset expansion region based on the first target pressure distribution data and the reference pressure distribution data in the verification set includes: Calculate the first mean absolute error between the first target pressure distribution data and the reference pressure distribution data in the verification set; Acquire the second target pressure distribution data predicted without adaptive boundary expansion, and calculate the second mean absolute error between the reference pressure distribution data and the second target pressure distribution data; The preset expansion factor is updated based on the first mean absolute error and the second mean absolute error to update the preset expansion distance and obtain the updated expansion distance. The updated expansion area is then determined based on the updated expansion distance.
6. The method for establishing a pressure distribution data prediction model according to claim 4, characterized in that, The step of predicting the first target pressure distribution data of the aircraft based on the operating conditions corresponding to the validation set and using each of the reference neural network models collaboratively includes: Based on the working conditions corresponding to the validation set, and by using each of the reference neural network models to validate the local pressure distribution data corresponding to the grid region, the local pressure distribution data corresponding to each grid region is obtained. The sum of all the local pressure distribution data is taken as the first target pressure distribution data of the aircraft.
7. The method for establishing a pressure distribution data prediction model according to any one of claims 1 to 6, characterized in that, Before performing adaptive boundary expansion on each training subset based on the preset expansion region corresponding to each training subset to obtain the expanded training subset, the method further includes: The region boundary of the grid region corresponding to the training subset is determined, and the average of the maximum diagonal distance and the minimum diagonal distance of the boundary corresponding to the region boundary is used as the preset expansion distance. Then, the preset expansion region is determined based on the preset expansion distance to obtain the preset expansion region corresponding to each training subset.