Wind tunnel flow field state prediction method based on multi-channel fusion

Abstract

The present invention discloses a method for predicting the state of a wind tunnel flow field based on multi-channel fusion, which relates to the field of aerospace aerodynamic test research, including: S1, determining the embedding dimension of a time series based on a Cao-style discriminant method to pre-process the wind tunnel operation data; S2, designing a trend encoder based on the pre-processed wind tunnel operation data, and realizing the correlation modeling of the wind tunnel flow field within a short distance based on a multi-head self-attention mechanism to construct a neural network sub-model I; S3, modeling the correlation of the long period of wind tunnel operation data based on a periodic encoder and a multi-head self-attention module to construct a neural network sub-model II; S4, channel splicing the fixed parameter part modules in the neural network sub-model I and the neural network sub-model II, and inputting the spliced ​​data into a multi-layer perceptron and an input module to predict the wind tunnel flow field. The present invention provides a method for predicting the state of a wind tunnel flow field based on multi-channel fusion, which is used to improve the prediction accuracy of the flow field.

Description

Technical Field

[0001] The present invention relates to the field of aerospace aerodynamic test research, and more specifically, to a wind tunnel flow field state prediction method based on multi-channel fusion. Background Art

[0002] As a key equipment for aerodynamic research, wind tunnel has very complex aerodynamic characteristics, and the flow field in the test section is easily affected by multi-source coupling interference. With the rapid development of my country's aerospace industry, it is urgent to establish a high-precision wind tunnel flow field model to support the innovative design of flow field control methods, so as to further improve the flow field control accuracy and meet the needs of wind tunnels to carry out multi-type large-scale / real aircraft high-speed special aerodynamic tests. In the flow field state of wind tunnel tests, there is a strong correlation between the Mach number and the total pressure and static pressure. The flow field state is directly controlled by the main adjustment opening, etc. The strong coupling relationship between the flow field state and between the flow field and the control parameters is a key technology that affects the accuracy of wind tunnel flow field modeling. The existing wind tunnel flow field modeling methods based on neural networks mainly learn and mine the correlation between the states of the flow field at different time points. There is no direct calculation of the correlation between the flow field states and between the flow field and the control parameters. Only some methods are indirectly established through the state transfer relationship between time points. Modeling the coupling relationship of key parameters of the wind tunnel flow field is the key to improving the accuracy of the wind tunnel flow field model. Summary of the invention

[0003] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0004] In order to achieve these purposes and other advantages of the present invention, a method for predicting wind tunnel flow field state based on multi-channel fusion is provided, comprising:

[0005] S1. Determine the embedding dimension of the time series based on the pseudo-neighbor method FNN to preprocess the wind tunnel operation data;

[0006] S2. Design a trend encoder based on the preprocessed wind tunnel operation data, and implement wind tunnel flow field correlation modeling within a short distance based on the multi-head self-attention mechanism to construct the neural network sub-model I;

[0007] S3, based on the periodic encoder and multi-head self-attention module, the long-term correlation modeling of wind tunnel operation data is carried out to construct the neural network sub-model II;

[0008] S4. Connect the fixed parameter modules in the neural network sub-model I and the neural network sub-model II through channels, and input the connected data into the multi-layer perceptron and input module. The output of the wind tunnel flow field prediction is Y o It can be characterized by the following formula:

[0009]

[0010] In the above formula, concat Indicates channel splicing, MLP Neural network layers representing multilayer perceptrons and output modules, Y thend Indicates the trend encoder output, Y couping represents the wind tunnel flow field correlation modeling output within a short distance, Y couping_l Represents the correlation modeling output for a long period of wind tunnel run data.

[0011] Preferably, in S1, the Cao-style discrimination criterion in the pseudo-nearest point method FFN method is as follows:

[0012]

[0013] In the above formula, L is the total number of state points of the wind tunnel running data sequence X, n is the embedding dimension of the time series, d min It is a status point X t The nearest neighbor distance, After increasing the embedding dimension by one X t The nearest neighbor distance, E ( n ) is the overall pseudo-neighbor statistic, J ´( n ) is the discriminant, E ( n+1 ) is the embedding dimension n+1 The overall pseudo-neighbor statistics of ;

[0014] The preprocessing refers to normalizing the wind tunnel operation data and labeling the data according to the embedding dimension.

[0015] Preferably, in S2, the design method of the trend encoder includes:

[0016] S210, constructing a trend encoder, if the scale of the processed wind tunnel operation data satisfies [wind tunnel operation time, embedding dimension, number of features], directly inputting it into the trend encoder;

[0017] S211, using the encoder part of Transformer, set 3 multi-head self-attention, the number of encoder blocks is 3, the embedding layer of the neural network is set to 512, the bit-by-bit feedforward network uses two layers of MLP, and the number of neurons in the middle layer is set to 256, then the trend encoder outputY trend for:

[0018]

[0019] In the above formula, X short Indicates the operation data of the wind tunnel in a short period. Encoder Network layers representing the trend encoder.

[0020] Preferably, in S2, the method for implementing wind tunnel correlation modeling within a short distance includes:

[0021] S221, performing dimension exchange on the processed wind tunnel operation data to obtain dimension exchanged wind tunnel short-period operation data X´ short ,and X´ short The scale satisfies [wind tunnel run time, number of features, embedding dimension], X´ short Contains short-distance wind tunnel operation trend information;

[0022] S222, will X´ short Input to the multi-head self-attention module MultiHead In this paper, the correlation of different features of the wind tunnel is calculated, and the wind tunnel correlation modeling output based on self-attention is performed. Y coupling Characterized by the following formula:

[0023]

[0024] Preferably, the output of the neural network sub-model I is Yo 1:

[0025]

[0026] In the above formula, represents the output layer of the neural network sub-model I, concat Indicates channel splicing.

[0027] Preferably, in S3, the long-period correlation modeling method of wind tunnel operation data is:

[0028] S300, construct a periodic encoder, and use the entire train wind tunnel operation data that meets the scale of [wind tunnel operation time, characteristic number] as the input of the periodic encoder, then the output of the periodic encoder is Y code for:

[0029]

[0030] In the formula, Encoder The network layers representing the periodic encoder, X long Represents the long-period operation data of the wind tunnel;

[0031] S310, performing dimension swapping on the output of the periodic encoder, and the data scale after the dimension swapping satisfies [wind tunnel running time, feature number, embedding dimension];

[0032] S320: If the wind tunnel operation data of the entire train contains periodic long-distance wind tunnel operation information, the dimension-swapped data is input into the multi-head self-attention module. MultiHead In this paper, the coupling calculation of different characteristics of the wind tunnel over a long distance is realized through the self-attention mechanism;

[0033] Calculated long-term correlation modeling output of wind tunnel operation data Y coupling_l for:

[0034]

[0035] In the above formula, represents the output of the periodic encoder after dimension transposition.

[0036] Preferably, the output of the neural network sub-model II is Yo 2 :

[0037]

[0038] In the above formula, represents the output layer of the neural network sub-model II, Decoder represents the decoder neural network layer for unsupervised learning, represents the output of the periodic encoder after dimension transposition.

[0039] Preferably, in S4, the method further includes:

[0040] S410, training the overall architecture neural network model based on the wind tunnel operation data in the data set, and only training the learnable parameters of the multilayer perceptron and the output module during the training process, the input scales of the overall architecture neural network model are the input scales of the neural network sub-model I and the neural network sub-model II, and the output or label of the overall architecture neural network model is consistent with the neural network sub-model I;

[0041] S420, performing flow field prediction based on multiple iterations to ensure prediction accuracy.

[0042] Preferably, in S420, the specific process of performing flow field prediction based on multiple iterations is:

[0043] S421. Input the wind tunnel operation data in the test set into the neural network sub-model I to obtain the output of the neural network sub-model I. Yo 1, and the initial flow field state prediction value ;

[0044] S422, will Input into the neural network sub-model II, and based on the overall architecture of S4, the neural network sub-model I and sub-model II are channel-fused to obtain the flow field state predicted based on the overall architecture. , i Indicates the number of iterations;

[0045] S423, determine the flow field prediction accuracy, if the accuracy meets the requirements, output it, otherwise the flow field state Input into the neural network sub-model II for i+1 Iterate the prediction until the prediction accuracy meets the requirements.

[0046] The present invention has at least the following beneficial effects:

[0047] First, by exchanging the dimensions of the wind tunnel operation data input into the neural network, multivariable coupling and correlation modeling of the flow field is realized based on the long / short cycle information of the wind tunnel operation time data.

[0048] Second, the multi-channel fusion of flow field characteristics is realized based on the trend information and cycle information of wind tunnel operation data, and the long-distance evolution prediction of wind tunnel flow field is realized based on transfer learning and cyclic prediction, which effectively improves the prediction accuracy of flow field.

[0049] Other advantages, objectives and features of the present invention will be embodied in part through the following description, and in part will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 It is a schematic flow chart of a wind tunnel flow field state prediction method based on multi-channel fusion according to the present invention;

[0051] Figure 2 It is a schematic diagram of the specific structure of the trend encoder of the present invention;

[0052] Figure 3 It is a schematic diagram of the specific implementation function and principle of the dimension exchange module in the present invention;

[0053] Figure 4 A schematic diagram showing the comparison between the predicted value of the Mach number and the true value after the prediction method of the present invention is used to perform long-term prediction on the wind tunnel flow field model;

[0054] Figure 5 A schematic diagram showing the comparison between the predicted value and the true value of the air source pressure after long-term prediction of the wind tunnel flow field model using the prediction method of the present invention;

[0055] Figure 6 A schematic diagram showing the comparison between the predicted value of total pressure and the true value after long-term prediction of the wind tunnel flow field model using the prediction method of the present invention;

[0056] Figure 7 A schematic diagram showing the comparison between the predicted value and the true value of the nozzle static pressure after long-term prediction of the wind tunnel flow field model using the prediction method of the present invention;

[0057] Figure 8 The figure is a schematic diagram showing the comparison between the predicted value and the true value of the static pressure of the test chamber after long-term prediction of the wind tunnel flow field model using the prediction method of the present invention. DETAILED DESCRIPTION

[0058] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0059] In order to improve the accuracy of wind tunnel flow field models and provide theoretical guidance for advanced flow field control technology, this paper proposes a wind tunnel flow field state prediction method based on multi-channel fusion, which realizes direct modeling of the wind tunnel flow field state, the coupling relationship between the flow field and control parameters, and effectively improves the accuracy of flow field modeling. Specifically, it includes:

[0060] (1) The time series embedding dimension is determined based on the False Nearest Neighbors (FFN) method, and the wind tunnel operation data is processed.

[0061] S100, counting the overall pseudo-neighboring points of the wind tunnel operation data, and determining the embedding dimension of the time series based on the Cao-style discriminant criterion of the FFN method.

[0062] S110, normalizing the wind tunnel operation data, and labeling the data according to the embedding dimension.

[0063] (2) Based on the wind tunnel flow field correlation modeling and the flow field prediction of the trend encoder, the neural network sub-model I of this design scheme is constructed.

[0064] S200, the scale of the labeled wind tunnel operation data satisfies [wind tunnel operation time, embedding dimension, number of features], and this part of data is directly input into the trend encoder.

[0065] S210. Design a trend encoder and construct a neural network sub-model I.

[0066] S220, using multi-head self-attention mechanism to achieve wind tunnel correlation modeling within a short distance. The specific implementation method is:

[0067] S221. The labeled data is dimensionally exchanged, and the scale of the data after the exchange satisfies [wind tunnel running time, number of features, embedding dimension].

[0068] S222. The wind tunnel operation data within an embedded dimension contains short-distance wind tunnel operation trend information. The data after dimension exchange is input into the multi-head self-attention module. The self-attention mechanism realizes the correlation calculation of different characteristics of the wind tunnel and realizes the implicit modeling of the coupling relationship of key parameters of the wind tunnel flow field.

[0069] S230, training neural network sub-model I is used to realize flow field prediction based on wind tunnel flow field correlation modeling and trend encoder, and neural network sub-model I is embedded into the overall architecture of this solution through transfer learning. The specific implementation method is:

[0070] S231. The training set of the neural network model is all the operating data regardless of the working conditions, and supervised learning is performed on the neural network sub-model I.

[0071] S232. After the training is completed, all parameters of the neural network sub-model I except the output layer are fixed, and the neural network layers except the output layer constitute the input of the two channels of multi-channel fusion in the overall architecture of the present design scheme.

[0072] (3) Flow field correlation modeling based on periodic encoder.

[0073] S300, design a periodic encoder and decoder, and construct a neural network sub-model II. The input of the periodic encoder is the wind tunnel operation data of the entire train, and the data scale satisfies [wind tunnel operation time, feature number].

[0074] S310, swapping the dimensions of the output of the periodic encoder, so that the scale of the data after the swap satisfies [wind tunnel running time, number of features, embedding dimension].

[0075] S320. The wind tunnel operation data of the entire train contains periodic long-distance operation information of the wind tunnel. The data after dimension exchange is input into the multi-head self-attention module. The self-attention mechanism realizes the correlation calculation of different characteristics of the wind tunnel within a long distance.

[0076] S330, train the neural network sub-model II to implement long-distance correlation modeling based on the wind tunnel flow field, and perform transfer learning to embed the neural network sub-model II into the architecture of this design scheme. The specific implementation method is:

[0077] S331. The training set of the neural network model is all the operating data regardless of the working conditions. A periodic decoder and output module are constructed to perform unsupervised learning on the neural network sub-model II.

[0078] S332. After the training is completed, the learnable parameters of the periodic encoder and the multi-head self-attention module are fixed. The periodic encoder and the multi-head self-attention module in the neural network sub-model II constitute a channel input of the multi-channel fusion in the overall architecture of this design scheme.

[0079] (4) The neural network sub-model I and the neural network sub-model II are channel-merged, and the learnable parameters of the non-neural network sub-model I and the neural network sub-model II modules are trained based on the wind tunnel operation data in the data set (the data set here is divided according to specific wind tunnel test conditions, such as subsonic and supersonic conditions, and the data set is divided into training set and test set according to test needs), and high-precision flow field prediction is achieved based on multiple iterations.

[0080] S400, channel-joining the outputs of the fixed parameter modules of the neural network sub-model I and the neural network sub-model II.

[0081] S410, the channel-joined data is input into a multi-layer perceptron and an output module to obtain a flow field prediction output of the design scheme.

[0082] S420. Train a neural network model of the overall design scheme based on wind tunnel operation data under specific working conditions. Only train the learnable parameters of the multilayer perceptron and the output module. The input scales of the model are the input scales of the neural network sub-model I and the neural network sub-model II, respectively. The model output / label is consistent with the label of the neural network sub-model I.

[0083] S430, flow field prediction based on multiple iterations, the specific process is as follows:

[0084] S431. Input the wind tunnel operation data of the test set into the neural network sub-model I to obtain the output of model I. Yo 1 , and the initial flow field state prediction .

[0085] S432, Initial flow field state predicted by neural network sub-model I The wind tunnel operation data of the test set is input into the neural network sub-model II, and multi-channel fusion is performed based on the overall architecture to obtain the iterative predicted flow field state. .

[0086] S433, determine the flow field prediction accuracy, if the accuracy meets the requirements, output; if the accuracy does not meet the requirements, the flow field state predicted by the multi-channel fusion of the overall architecture is output. Input into the neural network sub-model II and perform iterative prediction until the accuracy requirements are met.

[0087] Example:

[0088] Figure 1 This is the specific implementation process and overall architecture of the wind tunnel correlation modeling based on the attention mechanism of this design scheme. The design scheme consists of two sub-neural network models to form the overall neural network model, which requires three trainings. The implementation process is divided into the following four steps:

[0089] (1) The time series embedding dimension is determined based on the FFN method, and the wind tunnel operation data is processed.

[0090] S100, divide the wind tunnel operation data sequence X into different state points according to the order n , the overall pseudo-neighboring points of the wind tunnel operation data are counted, and the time series embedding dimension is determined based on the Cao-style discriminant criterion. The discriminant criterion is shown in the following formula:

[0091]

[0092] Among them, in the above formula, L is the total number of state points of the wind tunnel running data sequence X, n is the embedding dimension of the time series, d min It is a status point X t The nearest neighbor distance, E ( n ) is the overall pseudo-neighbor statistic, J ´( n ) is the discriminant. J ´( n ) will tend to be stable when the model order reaches the best, and the order corresponding to this point is the best embedding order. According to the Cao-type criterion, the best embedding order of the wind tunnel data in this embodiment is determined to be 10.

[0093] S110, performing maximum and minimum normalization on the wind tunnel operation data, and labeling the data according to the embedding dimension.

[0094] (2) Based on the wind tunnel flow field correlation modeling and the flow field prediction of the trend encoder, the neural network sub-model I of this design scheme is constructed.

[0095] S200, the scale of the labeled wind tunnel operation data satisfies [wind tunnel operation time, embedding dimension, number of features], and this part of data is directly input into the trend encoder.

[0096] Figure 2 For the specific structure of the trend encoder, Figure 3 It is the specific implementation function of the dimension exchange module. Figure 3 The horizontal axis is time, the vertical axis is dimensionless, and the value range is 0-1.

[0097] S210, design a trend encoder neural network model. Use the encoder part of Transformer, set 3 multi-head self-attention, and the number of encoder blocks is 3. The embedding layer of the neural network is set to 512, and the bit-by-bit feedforward network uses two layers of MLP, and the number of neurons in the middle layer is set to 256. The output of the trend encoder is:

[0098]

[0099] In the formula, Indicates the trend encoder output, Indicates the operation data of the wind tunnel in a short period. Encoder Network layers representing the trend encoder.

[0100] S220, using multi-head self-attention mechanism to achieve wind tunnel correlation modeling within a short distance. The specific implementation method is:

[0101] S221. The labeled data is dimensionally exchanged to convert the time series of wind tunnel operation data, i.e., [main adjustment opening, target Mach number, Mach number, total pressure, static pressure] at each time point. t The data matrix is ​​transposed so that its input sequence in the neural network model is a single variable sequence within the time range of 0~t, that is, according to [main tuning opening| 0~t , target Mach number | 0~t , Mach number | 0~t , total pressure | 0~t , static pressure | 0~t ], and after the exchange, the data scale satisfies [wind tunnel running time, number of features, embedding dimension].

[0102] S222, the wind tunnel operation data in one embedding dimension contains short-distance wind tunnel operation trend information, and the data after dimension exchange is input into the multi-head self-attention module, such as Figure 3 As shown in the figure, the self-attention mechanism realizes the correlation calculation of different characteristics of the wind tunnel and the implicit modeling of the coupling relationship of key parameters of the wind tunnel flow field. The output of the multivariable coupling model of the flow field based on self-attention is:

[0103]

[0104] In the formula, represents the multivariable coupled model output, Multihead Indicates multi-head self-attention, which is set to 3 in this example. X´ short Represents the dimensionally exchanged wind tunnel short-cycle operating data.

[0105] S230, train the neural network sub-model I to realize the flow field prediction based on the wind tunnel flow field correlation modeling and trend encoder, and embed the neural network sub-model I into the overall architecture of this solution through transfer learning. The specific implementation method is:

[0106] S231, the neural network model training set is all the operating data regardless of the working condition. The neural network sub-model I is supervised for learning. The optimizer is set to the adaptive moment estimation method (Adam), and the learning rate (lr) is set to 0.005. The output of the neural network sub-model I is:

[0107]

[0108] In the formula, represents the output of the neural network sub-model I, represents the output layer of the neural network sub-model I, c oncat Indicates channel splicing.

[0109] S232. After the training is completed, all parameters of the neural network sub-model I except the output layer are fixed, and the neural network layers except the output layer constitute a channel input of the multi-channel fusion of the overall architecture of this design solution. Figure 1 As shown in the dotted line frame neural network sub-model Ⅰ, after the training is completed, the output layer is removed. Y thend Representation and Y couping Constitutes a channel input to the multi-channel fusion method.

[0110] (3) Flow field correlation modeling based on periodic encoder.

[0111] S300, design a periodic encoder and decoder, and construct a neural network sub-model II. The input of the periodic encoder is the wind tunnel operation data of the entire train, and the data scale satisfies [wind tunnel operation time, feature number]. The output of the periodic encoder is:

[0112]

[0113] In the formula, Y code represents the output of the periodic encoder, Encoer The network layers representing the periodic encoder, X long Represents the long-term operating data of the wind tunnel.

[0114] S310, swapping the dimensions of the output of the periodic encoder, so that the data scale after the dimension swap satisfies [wind tunnel running time, feature number, embedding dimension].

[0115] S320, the wind tunnel operation data of the entire train contains periodic long-distance wind tunnel operation information. The data after dimension exchange is input into the multi-head self-attention module. The self-attention mechanism realizes the coupling calculation of different characteristics of the wind tunnel over a long distance. The long-term correlation modeling output of the wind tunnel operation data is:

[0116]

[0117] In the formula, represents the long-period multivariate coupled modeling output, Multihead Indicates multi-head self-attention, which is set to 3 in this example.

[0118] represents the output of the periodic encoder after dimension transposition.

[0119] S330, train the neural network sub-model II to realize the long-distance correlation modeling based on the wind tunnel flow field, and embed the neural network sub-model II into the architecture of this design scheme through transfer learning. The specific implementation method is:

[0120] S331, the neural network model training set is all the operating data regardless of the working condition, the periodic decoder and output module are constructed, the neural network sub-model II is constructed, and the neural network sub-model II is subjected to unsupervised learning. The optimizer is set to the adaptive moment estimation method (Adam), and the learning rate (lr) is set to 0.0036. The output of the neural network sub-model II is:

[0121]

[0122] In the formula, represents the output of the neural network sub-model II, represents the output layer of the neural network sub-model II, Decoder represents the decoder neural network layer for unsupervised learning, represents the output of the periodic encoder after dimension transposition.

[0123] S332, after the training is completed, the learnable parameters of the periodic encoder and the multi-head self-attention module are fixed, and the periodic encoder and the multi-head self-attention module constitute a channel input of the multi-channel fusion method of the present design scheme. Figure 1 As shown in the dotted line frame neural network sub-model II, after the training is completed, the output layer and the periodic decoder are removed. Constitutes a channel input to the multi-channel fusion method.

[0124] (4) The neural network sub-model I and the neural network sub-model II are channel-merged, and the learnable parameters of the non-neural network sub-model I and the neural network sub-model II modules are trained based on the wind tunnel operation data under specific working conditions, that is, Figure 1The learnable parameters of the multi-layer perceptron module are used to achieve high-precision flow field prediction based on multiple iterations.

[0125] S400, channel-joining the outputs of the fixed parameter modules of the neural network sub-model I and the neural network sub-model II.

[0126] S410, the data after channel splicing is input into the multi-layer perceptron and the output module to obtain the flow field prediction output of the design scheme, and the flow field prediction output is:

[0127]

[0128] In the formula, Y o is the flow field prediction output, MLP represents the multi-layer perceptron and output module neural network layers, concat Indicates channel splicing.

[0129] S420. Based on the wind tunnel operation data at subsonic speeds in the training set, the overall architecture neural network model of the design scheme is trained. Only the learnable parameters of the multi-layer perceptron and the output module are trained. The input scales of the model are the input scales of the neural network sub-model I and the neural network sub-model II, respectively. The model output / label is consistent with the label of the neural network sub-model I.

[0130] S430, flow field prediction based on multiple iterations, the specific process is as follows:

[0131] S431. Input the wind tunnel operation data of the test set into the neural network sub-model I to obtain the flow field state predicted by sub-model I. Yo 1, that is, the initial flow field state .

[0132] S432, Neural Network Sub-Model I predicts the initial flow field state The wind tunnel operation data of the test set is input into the neural network sub-model II, and the flow field state is predicted based on the overall neural network model. , i Indicates the number of iterations.

[0133] S433, determine the flow field prediction accuracy, if the accuracy meets the requirements, output; if the accuracy does not meet the requirements, the flow field state predicted by the overall architecture neural network model is output. Input into the neural network sub-model II and perform i +1 iterative prediction until the accuracy requirement is met. Finally, the wind tunnel evolution prediction result of this embodiment is as follows Figure 4-Figure 8 As shown in the figure, the prediction accuracy of the wind tunnel flow field is still high after long-distance prediction.

[0134] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacement and / or modification can be performed according to user needs.

[0135] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and the embodiments. It can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily realized. Therefore, without departing from the general concept defined by the claims and equivalent scope, the present invention is not limited to the specific details and the illustrations shown and described here.

Claims

1. A wind tunnel flow field state prediction method based on multi-channel fusion, characterized in that: include: S1. Determine the embedding dimension of the time series based on the pseudo-neighbor method FNN to preprocess the wind tunnel operation data; S2. Design a trend encoder based on the preprocessed wind tunnel operation data, and implement wind tunnel flow field correlation modeling within a short distance based on the multi-head self-attention mechanism to construct the neural network sub-model I; S3, based on the periodic encoder and multi-head self-attention module, the long-term correlation modeling of wind tunnel operation data is carried out to construct the neural network sub-model II; S4. Connect the fixed parameter modules in the neural network sub-model I and the neural network sub-model II through channels, and input the connected data into the multi-layer perceptron and input module. The output of the wind tunnel flow field prediction is Y o It can be characterized by the following formula: In the above formula, concat Indicates channel splicing, MLP Neural network layers representing multilayer perceptrons and output modules, Y trend Indicates the trend encoder output, Y coupling represents the wind tunnel flow field correlation modeling output within a short distance, Y coupling_l Represents the correlation modeling output for a long period of wind tunnel run data.

2. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 1, characterized in that: In S1, the Cao-style discrimination criterion in the pseudo-nearest point method FFN method is as follows: In the above formula, L is the total number of state points of the wind tunnel running data sequence X, n is the embedding dimension of the time series, d min It is a status point X t The nearest neighbor distance, After increasing the embedding dimension by one X t The nearest neighbor distance, E ( n ) is the overall pseudo-neighbor statistic, E ( n+1 ) is the embedding dimension n+1 The overall pseudo-neighbor statistic of , J ´( n ) is the discriminant, E ( n+1 ) is the embedding dimension n+1 The overall pseudo-neighbor statistics of ; The preprocessing refers to normalizing the wind tunnel operation data and labeling the data according to the embedding dimension.

3. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 1, characterized in that: In S2, the design method of the trend encoder includes: S210, constructing a trend encoder, if the scale of the processed wind tunnel operation data satisfies [wind tunnel operation time, embedding dimension, number of features], directly inputting it into the trend encoder; S211, using the encoder part of Transformer, set 3 multi-head self-attention, the number of encoder blocks is 3, the embedding layer of the neural network is set to 512, the bit-by-bit feedforward network uses two layers of MLP, and the number of neurons in the middle layer is set to 256, then the trend encoder output Y trend for: In the above formula, X short Indicates the operation data of the wind tunnel in a short period. Encoder Network layers representing the trend encoder.

4. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 3 is characterized in that: In S2, the implementation of wind tunnel correlation modeling within short distances includes: S221, performing dimension exchange on the processed wind tunnel operation data to obtain dimension exchanged wind tunnel short-period operation data X´ short ,and X´ short The scale satisfies [wind tunnel run time, number of features, embedding dimension], X´ short Contains short-distance wind tunnel operation trend information; S222, will X´ short Input to the multi-head self-attention module MultiHead In this paper, the correlation of different features of the wind tunnel is calculated, and the wind tunnel correlation modeling output based on self-attention is performed. Y coupling Characterized by the following formula: 。 5. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 4, characterized in that: The output of the neural network sub-model I is Yo 1: In the above formula, represents the output layer of the neural network sub-model I, concat Indicates channel splicing.

6. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 1, characterized in that: In S3, the long-term correlation modeling method of wind tunnel operation data is: S300, construct a periodic encoder, and use the entire train wind tunnel operation data that meets the scale of [wind tunnel operation time, characteristic number] as the input of the periodic encoder, then the output of the periodic encoder is Y code for: In the formula, Encoder The network layers representing the periodic encoder, X long It represents the operation data of the wind tunnel for a long period; S310, performing dimension swapping on the output of the periodic encoder, and the data scale after the dimension swapping satisfies [wind tunnel running time, feature number, embedding dimension]; S320: If the wind tunnel operation data of the entire train contains periodic long-distance wind tunnel operation information, the dimension-swapped data is input into the multi-head self-attention module. MultiHead In this paper, the coupling calculation of different characteristics of the wind tunnel over a long distance is realized through the self-attention mechanism; Calculated long-term correlation modeling output of wind tunnel operation data Y coupling_l for: In the above formula, represents the output of the periodic encoder after dimension transposition.

7. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 6, characterized in that: The output of the neural network sub-model II is Yo 2: In the above formula, represents the output layer of the neural network sub-model II, Decoder represents the decoder neural network layer for unsupervised learning, represents the output of the periodic encoder after dimension transposition.

8. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 1, characterized in that: In S4, it also includes: S410, training the overall architecture neural network model based on the wind tunnel operation data in the data set, and only training the learnable parameters of the multilayer perceptron and the output module during the training process, the input scales of the overall architecture neural network model are the input scales of the neural network sub-model I and the neural network sub-model II, and the output or label of the overall architecture neural network model is consistent with the neural network sub-model I; S420, performing flow field prediction based on multiple iterations to ensure prediction accuracy.

9. The wind tunnel flow field state prediction method based on multi-channel fusion according to claim 8, characterized in that: In S420, the specific process of performing flow field prediction based on multiple iterations is as follows: S421. Input the wind tunnel operation data in the test set into the neural network sub-model I to obtain the output of the neural network sub-model I. Yo 1, and the initial flow field state prediction value ; S422, will Input into the neural network sub-model II, and based on the overall architecture of S4, the neural network sub-model I and sub-model II are channel-fused to obtain the flow field state predicted based on the overall architecture. , i Indicates the number of iterations; S423, determine the flow field prediction accuracy, if the accuracy meets the requirements, output it, otherwise the flow field state Input into the neural network sub-model II for i+1 Iterate the prediction until the prediction accuracy meets the requirements.

Images