CN117288418B - An intelligent identification method of aerodynamic forces in hypersonic wind tunnel based on wavelet decomposition - Google Patents

Abstract

The invention discloses a hypersonic wind tunnel aerodynamic force intelligent identification method based on wavelet decomposition, which comprises the following steps: obtaining a balance signal and an ideal step signal in a simulated wind tunnel test process, decomposing the balance signal and the ideal step signal by adopting a wavelet decomposition method to respectively obtain wavelet coefficients of the balance signal and the ideal step signal, and simultaneously inputting the wavelet coefficients into a deep learning model for training to obtain a trained deep learning model; obtaining a balance signal and carrying out wavelet decomposition on the balance signal to obtain a wavelet coefficient of the balance signal, inputting a trained deep learning model to remove interference components, and reconstructing the wavelet coefficient of the balance signal with the interference components removed by adopting a wavelet reconstruction method to obtain a real aerodynamic signal; the hypersonic wind tunnel aerodynamic force intelligent identification method based on wavelet decomposition can effectively filter inertia force and other interference signals and accurately obtain real aerodynamic force signals.

Description

An intelligent identification method of aerodynamic forces in hypersonic wind tunnel based on wavelet decomposition

Technical Field

The invention relates to the technical field of shock wave wind tunnel force measurement tests, and in particular to a hypersonic wind tunnel aerodynamic force intelligent identification method based on wavelet decomposition.

Background technique

Wind tunnel test is a key technology for the development of air-breathing hypersonic aircraft, and high-precision aerodynamic measurement is an important part of it. During the wind tunnel startup process, high-speed transient airflow will produce transient impact on the aircraft model installed in the force measurement system, thereby generating transient vibration. The model inertial force caused by the vibration will be collected by the force measurement system together with the aerodynamic force. Due to the short effective time of only 50-200ms, the inertial force is difficult to decay completely, and the output signal of the force measurement system finally shows an oscillation attenuation characteristic. During the wind tunnel test, the signal transmission and acquisition of the force measurement system will inevitably generate noise, mainly high-frequency white noise. In addition, during the test, variable frequency impact will occur at the junction of the aircraft model and the connection part of the force measurement system, so that variable frequency signals will appear in the output signal of the force measurement system, affecting the accuracy of aerodynamic identification. At present, the commonly used load identification methods in engineering are mainly: mean method, frequency domain method, time domain method and traditional neural network method, but the above methods cannot effectively reduce the interference of variable frequency signals, and some methods can only identify an aerodynamic load constant, which cannot effectively reflect the size and change process of dynamic aerodynamic loads during wind tunnel tests.

With the continuous development of aerospace technology, hypersonic technology has received extensive attention and in-depth research from various aerospace powers, and its scientific issues have important strategic significance. For the aerodynamic shape layout design and optimization of hypersonic aircraft, high-precision aerodynamic measurement tests play a decisive role. Shock tunnel force test can provide reliable data for the study of high-temperature real gas effects, and at the same time provide key technical support for my country's hypersonic aircraft research. At present, there are still many unresolved key technical problems in shock tunnel force test, which make shock tunnel force test a challenging research topic. One of the most important problems is the structural inertial vibration caused by the aerodynamic impact of the transient flow field on the model force measurement system. During the force measurement test, the force measurement system composed of the model-balance-support is subjected to instantaneous impact and generates structural vibration. These vibration signals cannot decay quickly in a short time, resulting in the output signal of the force measurement system containing interference signals generated by inertial vibration, which seriously affects the accuracy of the force measurement test.

Summary of the invention

In view of the above-mentioned deficiencies in the prior art, the present invention provides a method for intelligent identification of aerodynamic forces in a hypersonic wind tunnel based on wavelet decomposition. Through a dynamic calibration test bench, deep learning training data based on a balance signal and an ideal step signal are obtained. The acquired training data is decomposed by a wavelet decomposition method and sent to a deep learning model for training. The trained deep learning model is used to remove the interference components of the balance signal to obtain a real aerodynamic signal.

In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

A hypersonic wind tunnel aerodynamic intelligent identification method based on wavelet decomposition comprises the following steps:

S1, obtaining a balance signal and an ideal step signal simulating a wind tunnel test process;

S2, decomposing the balance signal and the ideal step signal obtained in step S1 by using a wavelet decomposition method to obtain wavelet coefficients of the balance signal and the ideal step signal respectively;

S3, inputting the wavelet coefficients of the balance signal and the ideal step signal respectively obtained in step S2 into the deep learning model for training, to obtain a trained deep learning model;

S4, obtaining a balance signal, performing wavelet decomposition on the obtained balance signal to obtain the wavelet coefficients of the balance signal, and inputting the wavelet coefficients into the deep learning model trained in step S3 to remove interference components, and using the wavelet reconstruction method to reconstruct the wavelet coefficients of the balance signal with the interference components removed to obtain a true aerodynamic force signal.

Furthermore, step S1 specifically includes:

S11. The test bench realizes step loading in three directions of XYZ through the motor. The upper computer controls the motor loading according to the S-type force sensor connected to the loading head. When the expected load is reached, the pneumatic device of the loading head is released to realize unloading, and a balance signal including a signal before loading, a stable signal after loading, and an oscillation signal after unloading is obtained;

S12, using the average value of the stable signal after loading obtained in step S11 minus the average value of the signal before loading obtained in step S11 to obtain the step amplitude, and artificially generate an ideal step signal according to the jump time.

Furthermore, step S2 specifically includes:

According to the balance signal and the ideal step signal obtained in step S1, the frequency components of the balance signal and the ideal step signal are decomposed by wavelet decomposition method to obtain wavelet coefficients of the balance signal and the ideal step signal respectively.

Furthermore, the specific process of using wavelet decomposition method to decompose the frequency components of the balance signal and the ideal step signal is as follows:

Construct a decomposition high-pass filter and a decomposition low-pass filter and select bior3.3 wavelet to construct the parameters of the decomposition high-pass filter and the decomposition low-pass filter, convolve the balance signal with the ideal step signal and the constructed decomposition high-pass filter with the decomposition low-pass filter, and obtain the wavelet detail coefficients and wavelet approximation coefficients of the first-level decomposition balance signal and the ideal step signal respectively; convolve the wavelet approximation coefficients of the first-level decomposition balance signal and the ideal step signal respectively and the constructed decomposition high-pass filter with the decomposition low-pass filter, and obtain the wavelet detail coefficients and wavelet approximation coefficients of the secondary decomposition balance signal and the ideal step signal respectively.

Furthermore, step S3 specifically includes:

S31, using the wavelet coefficients of the balance signal obtained in step S2 as the original data, using the wavelet coefficients of the ideal step signal as the labels of the original data, and inputting the labeled original data as a data set into the bidirectional LSTM model;

S32, using a part of the data set in step S31 as a training set to train the bidirectional LSTM model, and using the remaining part of the data set as a test set to evaluate the performance of the trained bidirectional LSTM model;

S33. Select the mean square error as the loss function to predict the bidirectional LSTM model in training, and judge the difference between the predicted value of the bidirectional LSTM model and the wavelet coefficient of the ideal step signal, that is:

Among them, MSE represents mean square error, n represents the number of training times, _Yi represents the predicted value of the bidirectional LSTM model after the i-th training. Represents the wavelet coefficients of the ideal step signal for the i-th training;

S34. Select the Adam optimizer to optimize the parameters of the bidirectional LSTM model to generate a bidirectional LSTM model with smaller loss.

Furthermore, step S4 specifically includes:

S41, obtaining a balance signal, and decomposing the obtained balance signal using a wavelet decomposition method to obtain a wavelet coefficient of the balance signal;

S42, inputting the wavelet coefficients of the balance signal obtained in step S41 into the bidirectional LSTM model trained in step S3 to remove interference components;

S43, using the wavelet reconstruction method to reconstruct the wavelet coefficients of the balance signal from which the interference components are removed in step S42, to obtain a true aerodynamic force signal.

Further, step S43 specifically includes:

S431, convolution of the wavelet detail coefficient of the balance signal with the reorganized high-pass filter plus convolution of the wavelet approximation coefficient of the balance signal with the reorganized low-pass filter, to obtain the wavelet approximation coefficient of the previous level of the balance signal;

S432, repeat the operation of step S431 with the wavelet approximation coefficient of the previous level of the balance signal obtained in step S431 to obtain the real aerodynamic force signal

The present invention has the following beneficial effects:

The present invention proposes a hypersonic wind tunnel aerodynamic intelligent identification method based on wavelet decomposition, which can accurately obtain the real aerodynamic signal, and can effectively filter out the inertial force and other interference signals even when the frequency overlaps with the aerodynamic signal; at the same time, the present invention adopts wavelet decomposition technology and deep learning technology to remove the interference frequency components of each level of wavelet coefficients, making the filtering process more explainable.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of a hypersonic wind tunnel aerodynamic intelligent identification method based on wavelet decomposition proposed by the present invention;

FIG2 is a schematic diagram of model training data and a power spectrum of model training data in an embodiment;

FIG3 is a schematic diagram of a decomposition low-pass filter and a decomposition high-pass filter constructed in the embodiment;

FIG4 is a schematic diagram of the training set and test set losses in an embodiment;

FIG5 is a schematic diagram of the model prediction results and the power spectrum of the model prediction results;

FIG6 is a schematic diagram of a reconstructed low-pass filter and a reconstructed high-pass filter;

FIG7 is a schematic diagram of a reconstructed signal of detail coefficients of the 9th level wavelet decomposition and a power spectrum of the reconstructed signal;

FIG8 is a schematic diagram of the reconstructed signal of the 13th level wavelet decomposition detail coefficients and the reconstructed signal power spectrum.

Detailed ways

The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

As shown in FIG1 , a hypersonic wind tunnel aerodynamic intelligent identification method based on wavelet decomposition includes the following steps S1-S4:

S1. Obtain the balance signal and ideal step signal of the simulated wind tunnel test process.

In an optional embodiment of the present invention, since deep learning requires a large amount of training data, the wind tunnel experiment is expensive and it is difficult to meet the required data volume. The installation method of the aircraft and the balance of the test bench is the same as that of the real wind tunnel test, and the dynamic characteristics obtained are similar to those of the real wind tunnel test. Therefore, a dynamic calibration test bench is used to simulate the wind tunnel experiment to obtain deep learning training data.

Specifically, step S1 specifically includes:

S11. The test bench realizes step loading in three directions of XYZ through the motor. The upper computer controls the motor loading according to the S-type force sensor connected to the loading head. When the expected load is reached, the pneumatic device of the loading head is released to realize unloading, and a balance signal including the pre-loading signal, the stable signal after loading and the oscillation signal after unloading is obtained.

S12, using the average value of the stable signal after loading obtained in step S11 minus the average value of the signal before loading obtained in step S11 to obtain the step amplitude, and artificially generate an ideal step signal according to the jump time.

The balance signal collected in this embodiment includes a pre-loading signal, a stable signal after loading, and an oscillating signal after unloading. The step load is obtained by subtracting the pre-loading signal from the stable signal after loading, and the oscillating signal after unloading contains the step load and the system inertial vibration, so the collected balance signal has the system inertial vibration and interference components. Therefore, the step amplitude is obtained by subtracting the mean value of the pre-loading signal from the mean value of the stable signal after loading, and an ideal step signal is artificially generated according to the jump time, and the generated ideal step signal is input as the label of the balance signal into the deep learning model for training.

As shown in Figure 2, Figure 2 is a schematic diagram of model training data and model training data power spectrum. Figure 2 (a) is a model training data diagram, the horizontal axis represents the training time, the vertical axis represents the normal force, the solid line represents the input balance signal, and the dotted line represents the ideal step signal as a label. Figure 2 (b) is a power spectrum diagram of model training data, the horizontal axis represents the frequency, the vertical axis represents the amplitude, the solid line represents the input balance signal, and the dotted line represents the ideal step signal as a label. From Figure 2 (b), it can be seen that the interference frequencies of the input balance signal are mainly 10.3Hz and 56.2Hz.

S2. Decompose the balance signal and the ideal step signal obtained in step S1 by using a wavelet decomposition method to obtain wavelet coefficients of the balance signal and the ideal step signal respectively.

Specifically, step S2 specifically includes:

According to the balance signal and the ideal step signal obtained in step S1, the frequency components of the balance signal and the ideal step signal are decomposed by wavelet decomposition method to obtain the wavelet coefficients of the balance signal and the ideal step signal respectively.

Specifically, the specific process of using wavelet decomposition method to decompose the frequency components of the balance signal and the ideal step signal is as follows:

Construct a decomposition high-pass filter and a decomposition low-pass filter and select bior3.3 wavelet to construct the parameters of the decomposition high-pass filter and the decomposition low-pass filter, convolve the balance signal with the ideal step signal and the constructed decomposition high-pass filter with the decomposition low-pass filter, and obtain the wavelet detail coefficients and wavelet approximation coefficients of the first-level decomposition balance signal and the ideal step signal respectively; convolve the wavelet approximation coefficients of the first-level decomposition balance signal and the ideal step signal respectively and the constructed decomposition high-pass filter with the decomposition low-pass filter, and obtain the wavelet detail coefficients and wavelet approximation coefficients of the secondary decomposition balance signal and the ideal step signal respectively.

As shown in Figure 3, Figure 3 is a schematic diagram of the constructed decomposition low-pass filter and decomposition high-pass filter. In this embodiment, the frequency components of the acquired balance signal and the ideal step signal are decomposed by wavelet decomposition method, and are decomposed into 14 levels of wavelet coefficients, each level of wavelet coefficients contains the frequency information and phase information of the original signal.

S3. Input the wavelet coefficients of the balance signal and the ideal step signal respectively obtained in step S2 into the deep learning model for training to obtain a trained deep learning model.

In this embodiment, the deep learning model adopts a bidirectional LSTM model, and the wavelet coefficients of the balance signal and the ideal step signal obtained respectively are input into the deep learning model for training, and the number of hidden layers, hidden layer dimensions and dropout values are adjusted according to the effect of the model training in the bidirectional LSTM model training. Among them, the number of hidden layers is selected as 1, 2, and 4, the hidden layer dimensions are selected as 5 and 10, the dropout values are selected as 0, 0.2, and 0.6, the learning rate is selected as 0.01, 0.001, and 0.0001, the batch size is selected as 5 and 10, and the number of training rounds is selected as 1000, 5000, and 10000. By comprehensively considering the loss size and training time, it is finally determined that the hidden layer dimension is 10, including two hidden layers, and the dropout value is 0.2. And the mean square error MSE is selected as the loss function, the optimizer selects the Adam optimizer, the learning rate is 0.001, the batch size is 10, and the number of training rounds is 10000. In this embodiment, the loss function is an operation function used to measure the difference between the predicted value of the bidirectional LSTM model and the label value, i.e., the wavelet coefficient of the ideal step signal. It is a non-negative real-valued function. The smaller the loss function, the better the robustness of the model. The loss function is mainly used in the training phase of the bidirectional LSTM model. After each batch of training data is sent to the bidirectional LSTM model, the predicted value is output through forward propagation, and then the loss function calculates the difference between the predicted value and the label value, i.e., the wavelet coefficient of the ideal step signal, which is also the loss value. After obtaining the loss value, the bidirectional LSTM model updates each parameter through back propagation to reduce the loss between the label value and the predicted value, so that the predicted value of the bidirectional LSTM model moves closer to the label value, and the Adam optimizer uses the gradient descent method to optimize the parameters of the bidirectional LSTM model to generate a bidirectional LSTM model with less loss, thereby achieving the purpose of learning.

Specifically, step S3 specifically includes S31-S34:

S31, using the wavelet coefficients of the balance signal obtained in step S2 as the original data, using the wavelet coefficients of the ideal step signal as the labels of the original data, and inputting the labeled original data as a data set into the bidirectional LSTM model.

S32: Use a part of the data set in step S31 as a training set to train the bidirectional LSTM model, and use the remaining part of the data set as a test set to evaluate the performance of the trained bidirectional LSTM model.

S33. Select the mean square error as the loss function to predict the bidirectional LSTM model in training, and judge the difference between the predicted value of the bidirectional LSTM model and the wavelet coefficient of the ideal step signal, that is:

Among them, MSE represents mean square error, n represents the number of training times, _Yi represents the predicted value of the bidirectional LSTM model after the i-th training. Represents the wavelet coefficients of the ideal step signal for the i-th training.

S34. Select the Adam optimizer to optimize the parameters of the bidirectional LSTM model to generate a bidirectional LSTM model with smaller loss.

As shown in Figure 4, Figure 4 is a schematic diagram of the loss of the training set and the test set, the horizontal axis represents the number of training rounds, the vertical axis represents the mean square error, the solid line represents the loss of the training set, and the vertical axis represents the loss of the test set. In this embodiment, the wavelet coefficients of the balance signal obtained in step S2 are used as the original data, the wavelet coefficients of the ideal step signal are used as the labels of the original data, and the original data assigned with the labels are input into the bidirectional LSTM model as a data set, including a total of 180 groups of data sets, 150 of which are divided into training sets for training the bidirectional LSTM model, and the remaining 30 groups are divided into test sets and the performance of the trained bidirectional LSTM model is evaluated. As can be seen from Figure 4, the test set loss and the training set loss are the smallest after training to 10,000 rounds.

S4, obtaining a balance signal, performing wavelet decomposition on the obtained balance signal to obtain the wavelet coefficients of the balance signal, and inputting the wavelet coefficients into the deep learning model trained in step S3 to remove interference components, and using the wavelet reconstruction method to reconstruct the wavelet coefficients of the balance signal with the interference components removed to obtain a true aerodynamic force signal.

As shown in Figure 5, Figure 5 is a schematic diagram of the model prediction result and the power spectrum of the model prediction result. Figure 5 (a) represents the model prediction result diagram, the horizontal axis represents time, and the vertical axis represents normal force; Figure 5 (b) represents the power spectrum diagram of the model prediction result, the horizontal axis represents frequency, and the vertical axis represents amplitude; the model prediction signal input to the model is a balance signal. In this embodiment, the acquired balance signal is input into the trained bidirectional LSTM model for prediction, and the real aerodynamic force signal can be obtained from Figure 5 (a).

Specifically, step S4 specifically includes S41-S42:

S41, obtaining a balance signal, and decomposing the obtained balance signal using a wavelet decomposition method to obtain wavelet coefficients of the balance signal.

S42, input the wavelet coefficients of the balance signal obtained in step S41 into the bidirectional LSTM model trained in step S3 to remove interference components.

S43, using the wavelet reconstruction method to reconstruct the wavelet coefficients of the balance signal from which the interference components are removed in step S42, to obtain a true aerodynamic force signal.

Specifically, step S43 specifically includes S431-S432:

S431, the convolution of the wavelet detail coefficient of the balance signal and the reconstructed high-pass filter is added to the convolution of the wavelet approximation coefficient of the balance signal and the reconstructed low-pass filter to obtain the wavelet approximation coefficient of the previous level of the balance signal.

S432, repeat the operation of step S431 with the wavelet approximation coefficient of the previous level of the balance signal obtained in step S431 to obtain a true aerodynamic force signal.

As shown in Figure 6, Figure 6 is a schematic diagram of the constructed recombinant low-pass filter and recombinant high-pass filter. In this embodiment, the wavelet coefficients of the balance signal decomposed into 14 levels are reconstructed using the wavelet reconstruction method to obtain the real aerodynamic force signal.

As shown in Figure 7, Figure 7 is a schematic diagram of the reconstructed signal of the 9th level wavelet decomposition detail coefficient and the reconstructed signal power spectrum. As can be seen from Figure 7, the 9th level wavelet decomposition detail component decomposes the interference frequency of 56.2 Hz separately.

As shown in Figure 8, Figure 8 is a schematic diagram of the reconstructed signal and the power spectrum of the reconstructed signal after the 13th level wavelet decomposition detail coefficient. As can be seen from Figure 8, the 13th level wavelet decomposition detail component decomposes the interference frequency of 10.3 Hz separately.

The present invention uses specific embodiments to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present invention.

Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims (5)

1. A hypersonic wind tunnel aerodynamic intelligent identification method based on wavelet decomposition, characterized in that it comprises the following steps: S1. Obtain the balance signal and ideal step signal of the simulated wind tunnel test process. The specific process is as follows: S11. The test bench realizes step loading in three directions of XYZ through the motor. The upper computer controls the motor loading according to the S-type force sensor connected to the loading head. When the expected load is reached, the pneumatic device of the loading head is released to realize unloading, and a balance signal including a signal before loading, a stable signal after loading, and an oscillation signal after unloading is obtained; S12, using the average value of the stable signal after loading obtained in step S11 minus the average value of the signal before loading obtained in step S11 to obtain a step amplitude, and artificially generate an ideal step signal according to the jump time; S2, decomposing the balance signal and the ideal step signal obtained in step S1 by using a wavelet decomposition method to obtain wavelet coefficients of the balance signal and the ideal step signal respectively; S3, inputting the wavelet coefficients of the balance signal and the ideal step signal respectively obtained in step S2 into the deep learning model for training, and obtaining a trained deep learning model. The specific process is as follows: S31, using the wavelet coefficients of the balance signal obtained in step S2 as the original data, using the wavelet coefficients of the ideal step signal as the labels of the original data, and inputting the labeled original data as a data set into the bidirectional LSTM model; S32, using a part of the data set in step S31 as a training set to train the bidirectional LSTM model, and using the remaining part of the data set as a test set to evaluate the performance of the trained bidirectional LSTM model; S33. Select the mean square error as the loss function to predict the bidirectional LSTM model in training, and judge the difference between the predicted value of the bidirectional LSTM model and the wavelet coefficient of the ideal step signal, that is: Where MSE represents mean square error, n represents the number of training times, and _Yt represents the predicted value of the bidirectional LSTM model after the i-th training. Represents the wavelet coefficients of the ideal step signal for the i-th training; S34. Select the Adam optimizer to optimize the parameters of the bidirectional LSTM model to generate a bidirectional LSTM model with smaller loss. S4, obtaining a balance signal, performing wavelet decomposition on the obtained balance signal to obtain the wavelet coefficients of the balance signal, and inputting the wavelet coefficients into the deep learning model trained in step S3 to remove interference components, and using the wavelet reconstruction method to reconstruct the wavelet coefficients of the balance signal with the interference components removed to obtain a true aerodynamic force signal. 2. The method for intelligent identification of aerodynamic forces in a hypersonic wind tunnel based on wavelet decomposition according to claim 1, characterized in that step S2 specifically comprises: According to the balance signal and the ideal step signal obtained in step S1, the frequency components of the balance signal and the ideal step signal are decomposed by wavelet decomposition method to obtain wavelet coefficients of the balance signal and the ideal step signal respectively. 3. According to claim 2, a hypersonic wind tunnel aerodynamic intelligent identification method based on wavelet decomposition is characterized in that the specific process of using wavelet decomposition method to decompose the frequency components of the balance signal and the ideal step signal is: Construct a decomposition high-pass filter and a decomposition low-pass filter and select bior3.3 wavelet to construct the parameters of the decomposition high-pass filter and the decomposition low-pass filter, convolve the balance signal with the ideal step signal and the constructed decomposition high-pass filter with the decomposition low-pass filter, and obtain the wavelet detail coefficients and wavelet approximation coefficients of the first-level decomposition balance signal and the ideal step signal respectively; convolve the wavelet approximation coefficients of the first-level decomposition balance signal and the ideal step signal respectively and the constructed decomposition high-pass filter with the decomposition low-pass filter, and obtain the wavelet detail coefficients and wavelet approximation coefficients of the secondary decomposition balance signal and the ideal step signal respectively. 4. The method for intelligent identification of aerodynamic forces in a hypersonic wind tunnel based on wavelet decomposition according to claim 1, characterized in that step S4 specifically comprises: S41, obtaining a balance signal, and decomposing the obtained balance signal using a wavelet decomposition method to obtain a wavelet coefficient of the balance signal; S42, inputting the wavelet coefficients of the balance signal obtained in step S41 into the bidirectional LSTM model trained in step S3 to remove interference components; S43, using the wavelet reconstruction method to reconstruct the wavelet coefficients of the balance signal from which the interference components are removed in step S42, to obtain a true aerodynamic force signal. 5. The method for intelligent identification of aerodynamic forces in a hypersonic wind tunnel based on wavelet decomposition according to claim 4, characterized in that step S43 specifically comprises: S431, convolution of the wavelet detail coefficient of the balance signal with the reorganized high-pass filter plus convolution of the wavelet approximation coefficient of the balance signal with the reorganized low-pass filter, to obtain the wavelet approximation coefficient of the previous level of the balance signal; S432, repeat the operation of step S431 with the wavelet approximation coefficient of the previous level of the balance signal obtained in step S431 to obtain a true aerodynamic force signal.