A three-layer precision progressive aerodynamic data fusion method

Abstract

This invention relates to the field of artificial intelligence and data fusion technology, and provides a three-layer progressive accuracy aerodynamic data fusion method. The method includes: acquiring low, medium, and high-precision datasets for the same aerodynamic problem; standardizing input features and output responses; constructing a low-precision basic prediction model to learn the low-precision mapping; constructing a medium-precision difference correction model, concatenating the low-precision baseline prediction value and input features as input, and learning the first difference from low to medium precision; constructing a high-precision difference correction model, concatenating the medium-precision aerodynamic prediction value and input features as input, and learning the second difference from medium to high precision; optimizing the model through staged or end-to-end joint training; during inference, forward propagating the samples to be predicted sequentially, progressively stacking the differences, and outputting a high-precision result. This invention reduces the learning difficulty and improves the prediction stability and extrapolation ability when high-precision samples are scarce through three-layer progressive accuracy difference learning.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data fusion technology, specifically relating to a pneumatic data fusion method with three-layer precision progression. Background Technology

[0002] In aircraft aerodynamic characteristic prediction and aerodynamic shape optimization design, using multi-precision data to build predictive models is a common method to improve modeling efficiency and accuracy. High-precision data usually comes from detailed wind tunnel experiments or high-precision CFD numerical simulations, which yield accurate results but are costly to obtain and have scarce samples. On the other hand, while low-precision data can be obtained in large quantities through engineering estimation methods, its prediction results have significant biases. Therefore, how to efficiently integrate aerodynamic data of different precisions to obtain high-precision prediction results under limited high-precision sample conditions has become a key issue in current aerodynamic data-driven modeling.

[0003] In existing technologies, multi-precision data fusion mainly relies on second-order direct mapping relationships, that is, directly establishing a mapping model between low-precision and high-precision aerodynamic data. These methods typically use the output of low-precision data as input features, learning the mapping to the high-precision output through models such as neural networks or Gaussian processes. However, there are complex nonlinear systematic biases between low-precision and high-precision data (e.g., caused by differences in grid resolution and turbulence models). Directly learning the complete mapping from low-precision to high-precision requires the model to simultaneously capture both the fundamental trend and local corrections, making the learning process quite challenging. When wind tunnel experiments or high-precision CFD samples are severely insufficient, the model is prone to overfitting, its extrapolation ability significantly decreases, and its prediction errors are large and its stability is poor under operating conditions such as angles of attack and Mach numbers outside the training data range, making it difficult to meet the reliability and physical consistency requirements in aircraft aerodynamic design.

[0004] Furthermore, existing technologies generally overlook the bridging role of intermediate-precision data in the fusion of multi-precision aerodynamic data. Intermediate-precision aerodynamic data (such as CFD results using RANS models) falls between the two in terms of accuracy and is typically abundant, reflecting the transition from low to high precision. However, existing methods lack mechanisms to effectively utilize this level of data and cannot reduce learning difficulty through progressive correction. Overall, existing two-level fusion methods suffer from high model complexity and convergence difficulties when processing multi-precision data, especially when high-precision samples are scarce, making it difficult to balance accuracy and generalization ability, thus limiting their application in complex aerodynamic problems (such as high angles of attack, unsteady flows, and multibody separation). Therefore, there is an urgent need to propose an aerodynamic data modeling method that can fully utilize low-precision, intermediate-precision, and high-precision aerodynamic data and achieve efficient fusion through a progressive differential correction mechanism. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a three-layer progressive aerodynamic data fusion method. This method addresses the shortcomings of existing two-level direct mapping methods, such as high learning difficulty, overfitting, and poor generalization ability when high-precision samples are scarce, and the failure to effectively utilize medium-precision data for progressive correction. The method learns from low-precision data by constructing a low-fidelity base model; then, it constructs a differential model to learn the differences between low and medium precision data to obtain medium-precision aerodynamic predictions; finally, it constructs a differential model to learn the differences between high precision data, accumulating these differences to obtain high-precision aerodynamic predictions. The network parameters of the model are then trained and optimized using either segmented independent training or end-to-end joint training. During inference, the final high-precision aerodynamic prediction result is output through forward propagation in the three-layer model, fully utilizing low, medium, and high-precision aerodynamic data. Through a progressive differential correction mechanism, the learning difficulty of each mapping step is reduced, improving prediction accuracy, generalization ability, and engineering reliability even with limited high-precision samples.

[0006] This invention provides a three-layer precision-progressive aerodynamic data fusion method, comprising: Step 1: Obtain low-precision, medium-precision, and high-precision datasets for the same aerodynamic problem. Each sample contains input features and a corresponding output response. The input features include at least the following aerodynamic features: angle of attack, Mach number, and sideslip angle. Perform data preprocessing. Step 2: Construct a low-precision basic prediction model based on a fully connected deep neural network to learn the basic mapping relationship between input features and low-precision output responses in the low-precision dataset, and output a low-precision baseline prediction value. Step 3: Construct a medium-precision differential correction model based on a fully connected deep neural network to learn the first difference from the low-precision benchmark prediction value to the medium-precision output response, and output the first differential prediction value; add the low-precision benchmark prediction value and the first differential prediction value to obtain the medium-precision aerodynamic prediction value. Step 4: Construct a high-precision differential correction model based on a fully connected deep neural network to learn the second differential from the medium-precision aerodynamic prediction value to the high-precision output response, and output the second differential prediction value; sum the three terms of the low-precision benchmark prediction value, the first differential prediction value, and the second differential prediction value to obtain the final high-precision aerodynamic prediction value. Step 5: Using a phased independent training method or an end-to-end joint training method, train and optimize the network parameters of the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model respectively using the low-precision dataset, the medium-precision dataset, and the high-precision dataset. Step six: In the inference stage, after data preprocessing, the input features of the sample to be predicted are sequentially input into the low-precision basic prediction model, the medium-precision differential correction model and the high-precision differential correction model after network parameter optimization for forward propagation, so as to obtain the final high-precision aerodynamic prediction result of the sample to be predicted.

[0007] Compared with the prior art, the beneficial effects of the present invention include: (1) The invention decomposes the traditional two-level direct mapping into two differential correction subproblems: “low precision → medium precision” and “medium precision → high precision”. Each sub-network only needs to learn the systematic deviation between adjacent levels, which greatly reduces the learning difficulty of deep neural networks when modeling aerodynamics.

[0008] (2) Improved prediction accuracy: The model maintains a stable prediction trend and physical consistency in the extrapolation region (such as angle of attack and Mach number conditions outside the training range), effectively avoiding non-physical oscillations and abrupt changes common in traditional methods. In the case of aircraft aerodynamic data prediction, the differential model improves the prediction accuracy of aerodynamic coefficients. The relative error on the curve decreased from 10.74% for the traditional method to 2.97%, validating the effectiveness of the architecture in improving prediction accuracy.

[0009] (3) This invention introduces medium-precision data as a transition bridge, making full use of the transitional information contained in the medium-precision samples. This allows the model to obtain a stable correction foundation through sufficient training of the first two network levels, even under conditions where high-precision samples are extremely limited. Only a small number of high-precision samples are needed to complete the final correction, adapting to scenarios where high-precision samples are scarce. In the experimental case, under conditions of scarce high-precision data, the model... , , , , , The average relative errors of the six aerodynamic coefficients reached 0.30%, 0.24%, 0.33%, 0.75%, 2.97%, and 1.47%, respectively, demonstrating the practical value of this method in scenarios where high-precision samples are scarce.

[0010] (4) The solution of the present invention is designed for aerodynamic data characteristics, does not rely on specific network structure, and meets the data fusion requirements of different aerodynamic shapes and different aerodynamic working conditions, providing an efficient multi-precision data fusion solution for aerodynamic modeling. Attached Figure Description

[0011] 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 some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0012] Figure 1 This is a flowchart illustrating the steps of a three-layer precision-progressive aerodynamic data fusion method in one embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the training of a low-precision basic deep neural network LF-DCNN in one embodiment of the present invention, wherein... This represents the standardized input feature components in a low-precision dataset. Components representing low-precision baseline predictions; Figure 3 This is a schematic diagram of a deep neural network MF-DCNN for accuracy difference correction during training, as described in one embodiment of the present invention. This represents the standardized components of the input features in a medium-precision dataset. This represents the component of the first difference prediction value. The components representing medium-precision aerodynamic prediction values; Figure 4 This is a schematic diagram of training a high-precision differential correction deep neural network HF-DCNN in one embodiment of the present invention, wherein, This represents the standardized input feature components in a high-precision dataset. This represents the component of the second difference prediction value. The component representing the final high-precision aerodynamic prediction value; Figure 5 This is a schematic diagram of a three-level precision progressive differential correction architecture in one embodiment of the present invention. Detailed Implementation

[0013] 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 a part of the embodiments of the present invention, and not all of them. 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.

[0014] In one embodiment, such as Figure 1 As shown, this invention provides a three-layer precision-progressive aerodynamic data fusion method, comprising: Step 1: Obtain low-precision, medium-precision, and high-precision datasets for the same aerodynamic problem. Each sample contains input features and a corresponding output response. The input features include at least the following aerodynamic features: angle of attack, Mach number, and sideslip angle. Perform data preprocessing. Step 2: Construct a low-precision basic prediction model based on a fully connected deep neural network to learn the basic mapping relationship between input features and low-precision output responses in the low-precision dataset, and output a low-precision baseline prediction value. Step 3: Construct a medium-precision differential correction model based on a fully connected deep neural network to learn the first difference from the low-precision benchmark prediction value to the medium-precision output response, and output the first differential prediction value; add the low-precision benchmark prediction value and the first differential prediction value to obtain the medium-precision aerodynamic prediction value. Step 4: Construct a high-precision differential correction model based on a fully connected deep neural network to learn the second differential from the medium-precision aerodynamic prediction value to the high-precision output response, and output the second differential prediction value; sum the three terms of the low-precision benchmark prediction value, the first differential prediction value, and the second differential prediction value to obtain the final high-precision aerodynamic prediction value. Step 5: Using a phased independent training method or an end-to-end joint training method, train and optimize the network parameters of the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model respectively using the low-precision dataset, the medium-precision dataset, and the high-precision dataset. Step six: In the inference stage, after data preprocessing, the input features of the sample to be predicted are sequentially input into the low-precision basic prediction model, the medium-precision differential correction model and the high-precision differential correction model after network parameter optimization for forward propagation, so as to obtain the final high-precision aerodynamic prediction result of the sample to be predicted.

[0015] Specifically, in step one, we first obtain datasets at three different levels of precision for the same aerodynamic problem: low-precision datasets. Medium-precision datasets and high-precision datasets .

[0016] Each sample in each dataset contains input features (vectors). (Different precision datasets correspond to different features, such as geometric parameters, boundary conditions, material properties, etc.) and their corresponding output responses (vectors). (e.g., force, temperature, displacement, etc.) Representing input features Dimensions Indicates the output response The dimension. Where: low precision dataset , It is the number of samples in a low-precision dataset, derived from simplified physical models or engineering methods. It has the most abundant number of samples, but the accuracy is relatively low. These represent the first and second elements in the low-precision dataset, respectively. The input features of each (data) sample and the corresponding output response (i.e., low-precision output response). Medium precision dataset , The number of samples in a medium-precision dataset is obtained using a medium-precision numerical model (such as CFD calculation of a RANS model). The number of samples is secondary, and the precision is between high and low. These represent the first and second halves of the medium-precision dataset, respectively. The input features of each (data) sample and the corresponding output response (i.e., medium-precision output response). High-precision dataset , It is the number of samples in a high-precision dataset, derived from high-fidelity CFD numerical simulations or wind tunnel experiments. It has the fewest samples but the highest accuracy. These represent the first and second elements in the high-precision dataset, respectively. The input features of each (data) sample and the corresponding output response (i.e., high-precision output response).

[0017] It should be noted that the input features of the three datasets with different precision mentioned above... , and They should come from the same feature space ( The feature space in which they reside, meaning they have the same physical meaning and dimension. However, the output responses of datasets with different precision... , and It is used only as a supervisory signal (label) during the training phase and does not directly affect the input of any neural network model. The input of each subsequent fully connected deep neural network is always composed of the standardized input features and (if any) the predictions of the previous stage model (network).

[0018] In one embodiment, the input features from the same feature space (different precision datasets) include at least the following aerodynamic features: angle of attack, Mach number, and sideslip angle.

[0019] To ensure that the data are fused at the same scale, the next step is to perform data preprocessing, including standardizing the input features in all datasets.

[0020] In one embodiment, the input features (including input features from low-precision, medium-precision, or high-precision datasets) are standardized using the Z-score normalization method, which involves subtracting the mean from each set of input features and dividing by the standard deviation: ; in, , These are the mean and standard deviation of the input features in each dimension, respectively. with standard deviation It is obtained by jointly calculating all input features from three datasets: low, medium, and high. After preprocessing, each feature dimension follows a distribution with a mean of 0 and a standard deviation of 1.

[0021] The output response can be normalized according to the actual physical quantity range (e.g., linearly scaled to the [0,1] range), or it can be normalized using the Z-score normalization method to eliminate the dimensional influence between outputs of different precision. The normalized / standardized output response is only used for loss function calculation (as a supervision label).

[0022] Furthermore, in step two, a low-precision basic prediction model (Low-Fidelity Deep Neural Network, LF-DNN) is constructed based on a fully connected deep neural network, such as... Figure 2 As shown, the underlying mapping between input features and (low-precision) output responses is learned through training on a low-precision dataset.

[0023] A fully connected deep neural network used to construct an LF-DNN consists of at least one input layer, multiple hidden layers, and one output layer. The number of neurons in the input layer equals the dimension of the input features (e.g., angle of attack, Mach number, sideslip angle in aerodynamic prediction). The number of hidden layers and the number of neurons in each hidden layer can be adjusted according to the model's data size and correction complexity. The number of neurons in the output layer equals the dimension of the output response (e.g., aerodynamic coefficients). , , , , , wait).

[0024] In one embodiment, the number of hidden layers is 4; each hidden layer contains 128 neurons and uses the ReLU activation function.

[0025] The input to the low-precision basic prediction model is the standardized input features. Output low-precision benchmark prediction values .

[0026] In step three, a middle-fidelity discrepancy correction neural network (MF-DCNN) model is constructed based on a fully connected deep neural network. For example... Figure 3 As shown, the core function of the constructed medium-precision differential correction model is to learn the systematic deviation between the output of the low-precision base prediction model and the medium-precision real data, thereby achieving the first correction. That is, learning the first difference from the low-precision baseline prediction value to the medium-precision output response, and outputting the first differential prediction value. The fully connected deep neural network used to build MF-DCNN has the same network structure as that in LF-DNN.

[0027] The input to the medium-precision differential correction model consists of two parts: first, the low-precision baseline prediction value of the low-precision basic prediction model for the current sample; and second, the standardized input features. After outputting the first differential prediction value, the low-precision benchmark prediction value is added to the first differential prediction value to obtain the medium-precision aerodynamic prediction value.

[0028] Then, in step four, a high-fidelity discrepancy correction model (High-Fidelity Discrepancy Correction Neural Network, HF-DCNN) is constructed based on a fully connected deep neural network. For example... Figure 4 As shown, the constructed high-precision differential correction model is used to learn the residual deviation between the intermediate-precision correction prediction result and the high-precision real data, thereby achieving a second correction. This involves learning the second difference between the intermediate-precision aerodynamic prediction value and the high-precision output response, and outputting the second differential prediction value. The fully connected deep neural network used to build HF-DCNN has the same network structure as that in LF-DNN.

[0029] The input to the High-Precision Differential Correction Model (HF-DCNN) consists of medium-precision aerodynamic predictions and standardized input features. After splicing and constructing the second differential prediction value, the three terms—the low-precision reference prediction value, the first differential prediction value, and the second differential prediction value—are summed to obtain the final high-precision aerodynamic prediction value.

[0030] Furthermore, in step five, a phased independent training method or an end-to-end joint training method is adopted to train and optimize the network parameters of the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model using the low-precision dataset, the medium-precision dataset, and the high-precision dataset, respectively.

[0031] In one embodiment, a phased independent training approach is used to train and optimize the network parameters of the low-precision base prediction model, the medium-precision differential correction model, and the high-precision differential correction model. This includes: (1) The low-precision basic prediction model is trained using the low-precision dataset, with the low-precision output response as the supervision label, and the following mean squared error (MSE) loss function is adopted: ; in, This represents the low-precision true value, which is the low-precision output response after preprocessing. Here, the square of the L2 norm of the vector is simply denoted as the square of the vector, and the same method is used in the definition of all subsequent loss functions.

[0032] The network weights and biases are updated using the backpropagation algorithm. After the network parameters of the low-precision basic prediction model are trained to convergence, the network parameters of the low-precision basic prediction model are fixed as the basis for subsequent difference correction. The training process is as follows: Figure 2 As shown, It represents low-precision input features, which are the components of the standardized input features (vectors) in the low-precision dataset; The component representing the low-precision baseline prediction value; the input layer of the network is fully connected to multiple hidden layers and the output layer.

[0033] The role of low-precision basic prediction models is to capture the main trends in low-precision data and provide initial predictions for subsequent correction processes.

[0034] (2) Train the medium-precision difference correction model using the medium-precision dataset: output the low-precision baseline prediction value from the low-precision basic prediction model with fixed network parameters, and calculate the first difference by combining the medium-precision output response; use the mean square error between the first difference prediction value and the first difference as the loss function, train the network parameters of the medium-precision difference correction model until convergence, and then fix the network parameters of the medium-precision difference correction model.

[0035] When training the mid-precision differential correction model, the input is the low-precision baseline prediction value concatenated with the standardized input features of the mid-precision dataset; the mid-precision output response is the mid-precision true value. (Preprocessed output response).

[0036] Using medium-precision true values Compared with low-precision benchmark predictions Calculate the first difference: ; With the first difference Used as a supervisory label for training; Perform the medium-precision aerodynamic prediction values It is given by the following formula: ; in, This is the first difference prediction value.

[0037] like Figure 3 As shown, The medium-precision input features are the standardized input features (vectors) in the medium-precision dataset. The component representing the first differential prediction value (medium-precision differential output response) is added to the original low-precision output response (corresponding components are added together) to obtain the medium-precision aerodynamic prediction value. (of the quantity): The traditional method is to directly output medium-precision aerodynamic prediction values.

[0038] The definition of the loss function in the above training process is: ; This is equivalent to the loss function defined using the mean square error between the current model output and the medium-precision real data (medium-precision output response), differing from the loss function by only a constant offset.

[0039] Through the above training process, the medium-precision differential correction model learns the differential correction from low-precision prediction to medium-precision data through a fully connected deep neural network. After training, the network parameters of the model are fixed. Its core function is to correct the system bias in the low-precision prediction for the first time through medium-precision data, and obtain a correction result with accuracy between low and medium, providing a better data starting point for subsequent high-precision correction.

[0040] (3) Train the high-precision differential correction model using the high-precision dataset: output the medium-precision aerodynamic prediction value from the medium-precision differential correction model with fixed network parameters, and calculate the second difference by combining the high-precision output response; use the mean square error between the second difference prediction value and the second difference as the loss function to train the network parameters of the high-precision differential correction model until convergence.

[0041] When training the high-precision differential correction model, the input is a combination of medium-precision aerodynamic prediction values ​​and standardized input features from a high-precision dataset; the high-precision output response is the high-precision true value. (Preprocessed output response); Through high-precision true values Compared with medium-precision aerodynamic prediction values Calculate the second difference: ; With the second difference Used as a supervisory label for training; The final high-precision aerodynamic prediction value It is given by the following formula: ; Will Substituting the formula into the equation, we finally get: ; in, This is the second difference prediction value.

[0042] like Figure 4 As shown, It represents high-precision input features, which are the components of the standardized input features (vectors) in a high-precision dataset; The component representing the second differential prediction value (high-precision differential output response) is added to the original medium-precision output response (corresponding components are added together) to obtain the final high-precision aerodynamic prediction value. (of the quantity): The traditional method is to directly output high-precision aerodynamic prediction values.

[0043] The definition of the loss function in the above training process is: ; This is equivalent to the loss function defined by the mean square error between the current model output and the high-precision real data (high-precision output response), differing from the loss function by only a constant offset.

[0044] Through the above training process, the high-precision differential correction model learns the differential correction from the intermediate-precision correction result to the high-precision true value through a fully connected deep neural network.

[0045] In one embodiment, an end-to-end joint training approach is used to train and optimize the network parameters of the low-precision base prediction model, the medium-precision differential correction model, and the high-precision differential correction model. This includes: The fully connected deep neural networks in the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model are sequentially connected in series into a whole; The mean square error between the final high-precision aerodynamic prediction and the high-precision true value is used as the loss function: ; Alternatively, an auxiliary loss term for medium-precision prediction can be added: ; The total loss function is obtained as follows: .

[0046] The training data used in the above training process consists of low-precision, medium-precision, and high-precision datasets. The final training dataset is usually obtained by sampling proportionally or by calculating the loss separately and then weighting them.

[0047] In one embodiment, when training and optimizing the network parameters of the low-precision base prediction model, the medium-precision differential correction model, and the high-precision differential correction model: a convergence criterion or iteration upper limit is used to control training termination, and cross-validation is used to monitor overfitting; an adaptive moment estimation optimizer is used, with an adaptive adjustment strategy for the learning rate, an initial learning rate of 0.001, and a batch size of 32 or 64 depending on the amount of data. After training convergence, all network structures and parameters are saved for subsequent inference and prediction.

[0048] Finally, in step six, the model is applied for inference. During the inference phase, the input features of the sample to be predicted, after data preprocessing, are sequentially input into the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model optimized by the network parameters for forward propagation, to obtain the final high-precision aerodynamic prediction result of the sample to be predicted.

[0049] In the actual inference stage, for any sample to be predicted, the standardized input features are first input, and then forward propagated sequentially through LF-DNN, MF-DCNN, and HF-DCNN, finally outputting a high-precision prediction result. This result integrates the basic trend of low-precision data, the transition correction of medium-precision data, and the fine correction of high-precision data, and can maintain high prediction accuracy and stable extrapolation ability even under the condition that high-precision samples are scarce.

[0050] The complete architecture of the aerodynamic data fusion method proposed in this invention, which is oriented towards three-layer progressive accuracy, is as follows: Figure 5 As shown, the core of the framework is that, for the first time, intermediate precision is explicitly introduced as an independent level in data fusion, and the complex "low precision → high precision" direct mapping is decomposed into two smoother sub-problems with smaller amplitudes, namely "low precision → intermediate precision deviation" and "intermediate precision → high precision deviation", through two differential corrections. This significantly reduces the learning difficulty of deep neural networks and improves the generalization performance and physical consistency of the model outside the training data range.

[0051] Furthermore, to verify the effectiveness of the three-layer precision-progressive aerodynamic data fusion method architecture provided by this invention in improving prediction accuracy, experiments were conducted on an aircraft aerodynamic data prediction case. Under conditions of scarce high-precision data, the model... , , , , , The average relative errors of the six aerodynamic coefficients reached 0.30%, 0.24%, 0.33%, 0.75%, 2.97%, and 1.47%, respectively, demonstrating the practical value of the method of this invention in scenarios where high-precision samples are scarce. The difference model designed by the method of this invention is effective in aerodynamic coefficient... The relative error on the extrapolation model is reduced from 10.74% in the traditional two-stage direct mapping method to 2.97%. Experimental results show that the difference model designed in this invention can maintain a stable prediction trend and physical consistency in the extrapolation region, effectively avoiding the non-physical oscillations and abrupt changes common in traditional methods. Furthermore, by introducing medium-precision data as a transition bridge, this invention fully utilizes the transitional information contained in medium-precision samples. Even under conditions where high-precision samples are extremely limited, the model can still obtain a stable correction foundation through sufficient training of the first two network stages, requiring only a small number of high-precision samples to complete the final correction. Moreover, the method architecture designed in this invention does not depend on a specific network structure and can flexibly adapt to the data characteristics of different engineering fields, providing a general and efficient multi-precision data fusion solution for complex system modeling.

[0052] In one embodiment, the present invention also provides a pneumatic data fusion device for three-layer precision progression, comprising: The first module is used to acquire low-precision, medium-precision, and high-precision datasets for the same aerodynamic problem, where each sample contains input features and a corresponding output response; the input features include at least the following aerodynamic features: angle of attack, Mach number, and sideslip angle; and to perform data preprocessing. The second module is used to build a low-precision basic prediction model based on a fully connected deep neural network. It is used to learn the basic mapping relationship between the input features and the low-precision output response in the low-precision dataset and output a low-precision benchmark prediction value. The third module is used to construct a medium-precision differential correction model based on a fully connected deep neural network. It is used to learn the first difference from the low-precision benchmark prediction value to the medium-precision output response and output the first differential prediction value. The low-precision benchmark prediction value is added to the first differential prediction value to obtain the medium-precision aerodynamic prediction value. The fourth module is used to construct a high-precision differential correction model based on a fully connected deep neural network. It is used to learn the second differential from the medium-precision aerodynamic prediction value to the high-precision output response, and output the second differential prediction value. The low-precision benchmark prediction value, the first differential prediction value and the second differential prediction value are summed to obtain the final high-precision aerodynamic prediction value. The fifth module is used to train and optimize the network parameters of the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model using the low-precision dataset, the medium-precision dataset, and the high-precision dataset, respectively, by adopting a phased independent training method or an end-to-end joint training method. The sixth module is used in the inference stage. After the input features of the sample to be predicted are preprocessed, they are sequentially input into the low-precision basic prediction model, the medium-precision differential correction model and the high-precision differential correction model after network parameter optimization for forward propagation, so as to obtain the final high-precision aerodynamic prediction result of the sample to be predicted.

[0053] On the other hand, in one embodiment, the present invention provides a computer device including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the three-layer precision-progressive pneumatic data fusion method provided in any of the above embodiments. The computer device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected via a system bus. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and computer program in the non-volatile storage medium. The database of the computer device stores sample data. The network interface of the computer device is used for communication with external terminals via a network connection.

[0054] On the other hand, in one embodiment of the present invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the three-layer precision progressive aerodynamic data fusion method provided in any of the above embodiments.

[0055] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0056] Matters not covered in this invention are common knowledge.

[0057] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0058] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for aerodynamic data fusion with progressively higher accuracy across three layers, characterized in that, include: Step 1: Obtain low-precision, medium-precision, and high-precision datasets for the same aerodynamic problem. Each sample contains input features and a corresponding output response. The input features include at least the following aerodynamic features: angle of attack, Mach number, and sideslip angle. Perform data preprocessing. Step 2: Construct a low-precision basic prediction model based on a fully connected deep neural network to learn the basic mapping relationship between input features and low-precision output responses in the low-precision dataset, and output a low-precision baseline prediction value. Step 3: Construct a medium-precision differential correction model based on a fully connected deep neural network to learn the first difference from the low-precision benchmark prediction value to the medium-precision output response, and output the first differential prediction value; add the low-precision benchmark prediction value and the first differential prediction value to obtain the medium-precision aerodynamic prediction value. Step 4: Construct a high-precision differential correction model based on a fully connected deep neural network to learn the second differential from the medium-precision aerodynamic prediction value to the high-precision output response, and output the second differential prediction value; sum the three terms of the low-precision benchmark prediction value, the first differential prediction value, and the second differential prediction value to obtain the final high-precision aerodynamic prediction value. Step 5: Using a phased independent training method or an end-to-end joint training method, train and optimize the network parameters of the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model respectively using the low-precision dataset, the medium-precision dataset, and the high-precision dataset. Step six: In the inference stage, after data preprocessing, the input features of the sample to be predicted are sequentially input into the low-precision basic prediction model, the medium-precision differential correction model and the high-precision differential correction model after network parameter optimization for forward propagation, so as to obtain the final high-precision aerodynamic prediction result of the sample to be predicted.

2. The aerodynamic data fusion method for three-layer precision progression according to claim 1, characterized in that, In step one, the data preprocessing includes: The input features in low-precision, medium-precision, or high-precision datasets are standardized using the Z-score normalization method. The output response can be normalized or standardized using the Z-score standardization method.

3. The aerodynamic data fusion method for three-layer precision progression according to claim 2, characterized in that, The fully connected deep neural network includes at least one input layer, multiple hidden layers, and one output layer; the number of neurons in the input layer is equal to the dimension of the input features; the number of hidden layers and the number of neurons in each hidden layer are adjusted according to the data scale and correction complexity in the low-precision basic prediction model, medium-precision differential correction model, or high-precision differential correction model; the number of neurons in the output layer is equal to the dimension of the corresponding output response. The low-precision baseline prediction model takes standardized input features as input and outputs a low-precision baseline prediction value. ; The input to the medium-precision differential correction model is the concatenation of the low-precision baseline prediction value and the standardized input features, and the output is the first differential prediction value. ; The high-precision differential correction model takes as input the concatenation of the medium-precision aerodynamic prediction value and the standardized input features, and outputs a second differential prediction value. .

4. The aerodynamic data fusion method for three-layer precision progression according to claim 3, characterized in that, The number of hidden layers is 4; each hidden layer contains 128 neurons and uses the ReLU activation function.

5. The aerodynamic data fusion method for three-layer precision progression according to claim 1, characterized in that, Step five, which involves adopting a phased independent training approach, includes: The low-precision basic prediction model is trained using a low-precision dataset. The low-precision output response is used as a label. The mean squared error loss function is used. The network weights and biases are updated through the backpropagation algorithm. After the network parameters of the low-precision basic prediction model are trained to convergence, the network parameters of the low-precision basic prediction model are fixed. The intermediate-precision difference correction model is trained using an intermediate-precision dataset: the low-precision base prediction model with fixed network parameters outputs a low-precision baseline prediction value, and the first difference is calculated by combining the intermediate-precision output response; the mean square error between the first difference prediction value and the first difference is used as the loss function to train the network parameters of the intermediate-precision difference correction model until convergence, and then the network parameters of the intermediate-precision difference correction model are fixed. The high-precision differential correction model is trained using a high-precision dataset: the medium-precision aerodynamic prediction value is output from the medium-precision differential correction model with fixed network parameters, and the second difference is calculated by combining the high-precision output response; the network parameters of the high-precision differential correction model are trained to convergence using the mean square error between the second difference prediction value and the second difference as the loss function.

6. The aerodynamic data fusion method for three-layer precision progression according to claim 5, characterized in that, When training the intermediate-precision differential correction model, the intermediate-precision output response is the intermediate-precision true value. ; Using medium-precision true values Compared with low-precision benchmark predictions Calculate the first difference: ； With the first difference Used as a supervisory label for training; The medium-precision aerodynamic prediction value It is given by the following formula: ； in, This is the first difference prediction value.

7. The aerodynamic data fusion method for three-layer precision progression according to claim 6, characterized in that, When training the high-precision differential correction model, the high-precision output response is the high-precision true value. ; Through high-precision true values Compared with medium-precision aerodynamic prediction values Calculate the second difference: ； With the second difference Used as a supervisory label for training; The final high-precision aerodynamic prediction value It is given by the following formula: ； in, This is the second difference prediction value.

8. The aerodynamic data fusion method for three-layer precision progression according to claim 1, characterized in that, The end-to-end joint training method includes: The fully connected deep neural networks in the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model are sequentially connected in series into a whole; The total loss function is obtained by using the mean square error between the final high-precision aerodynamic prediction and the high-precision true value, or by adding an auxiliary loss term from the medium-precision prediction.

9. The aerodynamic data fusion method for three-layer precision progression according to claim 1, characterized in that, In step five, when training and optimizing the network parameters of the low-precision basic prediction model, the medium-precision differential correction model, and the high-precision differential correction model: the training termination is controlled by convergence judgment or iteration upper limit, and the overfitting is monitored by cross-validation method; an adaptive moment estimation optimizer is used, the learning rate adopts an adaptive adjustment strategy, the initial learning rate is set to 0.001, and the batch size is set to 32 or 64 according to the amount of data.

Images