Atmospheric parameter solving method of FADS system based on geometric feature fitting neural network

Abstract

This invention discloses a method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network, comprising: Step 1: Obtaining the surface pressure matrix p required for training the neural network using a wind tunnel and an aircraft scaled-down model. train and reference atmospheric data vector; Step 2: Construct the neural network training input matrix Γ train Step 3: Calculate the reference pressure coefficient matrix C p,reference Step 4: Construct a geometric feature fitting neural network; Step 5: Complete the training of the geometric feature fitting neural network; Step 6: Obtain the actual surface pressure vector at the current moment using the pressure measurement port at the front of the actual aircraft; Step 7: Construct the actual network input vector at the current moment; Step 8: Obtain the fitting angle of attack, sideslip angle, and fitting pressure coefficient vector at the current moment; Step 9: Calculate the dynamic pressure and static pressure at the current moment; Step 10: Determine the current flight state; Step 11: Calculate the current flight Mach number. This method can maintain a relatively low atmospheric data fitting error.

Description

Technical Field

[0001] This invention belongs to the field of atmospheric parameter calculation technology, specifically relating to a method for calculating atmospheric parameters in a FADS system based on a geometric feature fitting neural network. Background Technology

[0002] The Flush Air Data Sensing (FADS) system utilizes surface pressure information and a surface pressure distribution model to measure and sense atmospheric parameters such as angle of attack, sideslip angle, total static pressure, and Mach number. This method replaces invasive pitot tubes with pressure measurement ports embedded in the fuselage, resulting in an embedded air data system that offers advantages such as good stealth performance, strong environmental adaptability, and relatively high accuracy. Therefore, it has been widely used in various advanced aircraft, including stealth aircraft, hypersonic vehicles, and space shuttles.

[0003] The core of the FADS system atmospheric parameter calculation method is the surface pressure distribution model, which is strongly correlated with the geometric characteristics of the area where pressure gauges are located. Early FADS systems were mostly used on aircraft with nose cones. The surface pressure distribution model of the nose cone with pressure gauges can be established based on potential flow theory, and an analytical FADS system atmospheric parameter calculation method was constructed based on this.

[0004] With the development of aerospace technology and the needs of various application scenarios, modern advanced aircraft often adopt complex configurations such as blended wing-body designs and wedge-shaped forebody structures, making it difficult to establish surface pressure distribution models. Furthermore, the integration of various airborne sensor devices imposes more restrictions and constraints on the layout of pressure measurement orifices in FADS systems. Therefore, traditional analytical methods for calculating atmospheric parameters in FADS systems are no longer applicable.

[0005] Artificial neural networks, due to their strong nonlinear mapping capabilities, have been widely applied and researched in atmospheric parameter calculation methods for FADS systems. Rholff et al. from the University of California, in collaboration with NASA in 1998, constructed two neural networks based on F-18 SRA measurement data, respectively, to fit dynamic and static pressure while neglecting the effects of angle of attack and sideslip angle. With the development of sensor and neural network technologies, various types of neural networks have been used to design different methodologies to solve specific problems in different scenarios. In 2022, X. Na et al. used neural networks to construct a mapping relationship between flexible sensor array measurement information and angle of attack and airspeed. In 2021, Q. Jia et al. compared various FADS calculation algorithms, including neural network methods, and pointed out that although neural network-based FADS system algorithms can achieve relatively accurate results, they require a large amount of data for training, and network training is difficult to converge. As research deepens, some scholars have found that establishing multiple neural networks to fit different types of atmospheric parameters can effectively improve the fitting accuracy of atmospheric data. In 2022, L. Yang et al. from Northwestern Polytechnical University proposed a FADS sensing method based on a dimensionless input-output neural network to improve the fitting accuracy of atmospheric data with fewer training samples.

[0006] As a data-driven algorithm, the performance and fitting accuracy of neural networks largely depend on whether the training set can support the network's learned knowledge in covering all possible situations in real-world applications. Therefore, constructing a training dataset using real-world data is crucial to ensuring the consistency between the network's learned knowledge and actual application scenarios. In the aerospace field, wind tunnel tests and flight tests are the primary means of acquiring this type of data. However, due to cost and safety considerations, the data obtained from wind tunnel tests and flight tests is relatively limited, which may be insufficient for training the networks of most existing neural network-based FADS systems. Furthermore, due to the black-box nature of neural networks, the interpretability of neural network-based methods has always been a significant factor restricting their practical application. Summary of the Invention

[0007] The purpose of this invention is to provide a method for calculating atmospheric parameters of the FADS system based on a neural network fitting geometric features.

[0008] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0009] A method for calculating atmospheric parameters in a FADS system based on geometric feature fitting neural networks includes the following steps:

[0010] Step 1: Obtain the surface pressure matrix required for training the neural network using a scaled-down model of a wind tunnel and an aircraft. and a reference atmospheric data vector, which includes a reference angle of attack vector. Reference sideslip angle vector Reference static pressure vector and reference dynamic pressure vector ;

[0011] Step 2: Based on the surface pressure matrix Constructing the neural network training input matrix ;

[0012] Step 3: Based on the surface pressure matrix Reference static pressure vector and reference dynamic pressure vector Calculate the reference pressure coefficient matrix ;

[0013] Step 4: Construct a neural network for fitting geometric features;

[0014] Step 5: Train the input matrix using a neural network As the input for training the geometric feature fitting neural network, with reference angle of attack vector Reference sideslip angle vector and reference pressure coefficient matrix The geometric feature fitting neural network training output is used to complete the training of the geometric feature fitting neural network.

[0015] Step 6: Obtain the actual surface pressure vector at the current moment using the pressure measurement port at the front of the actual aircraft. ;

[0016] Step 7: Construct the actual network input vector at the current time step ;

[0017] Step 8: Utilize the actual network input vector at the current time. The geometric feature fitting neural network trained in step 5 is used to obtain the fitted angle of attack at the current time. Sideslip angle and fitted pressure coefficient vector ;

[0018] Step 9: Based on the actual surface pressure vector at the current moment Fit the pressure coefficient vector to the current time output of the neural network. Calculate the dynamic pressure at the current moment and static pressure ;

[0019] Step 10: Based on the calculated dynamic pressure at the current moment and static pressure Determine the current flight status;

[0020] Step 11: Calculate the current flight Mach number based on the determined flight status.

[0021] Compared with the prior art, the technical effects achieved by the present invention are as follows:

[0022] 1. The atmospheric parameter calculation method for the FADS system based on geometric feature fitting neural network proposed in this invention, after theoretical analysis, shows that the geometric feature fitting neural network used suppresses the influence of other atmospheric data such as Mach number and total static pressure parameter to the greatest extent when fitting the airflow angle, so that it is only related to the geometric features of the aircraft under certain conditions, thereby enhancing the extrapolation capability of the neural network.

[0023] 2. The geometric feature fitting neural network proposed in this invention classifies hidden layer neurons according to the actual physical characteristics of different network outputs. It adopts a non-fully connected architecture to form two relatively independent sub-networks, which reduces the complexity of the fitting target of each sub-network and can still meet the expected training target under limited training sample conditions.

[0024] 3. By decoupling the airflow angle parameter and total static pressure parameter in the surface pressure distribution model, the influence of the black box effect of the neural network on the atmospheric parameter calculation algorithm is reduced. On the other hand, based on the analysis of the physical characteristics of the input of the neural network that fits the geometric features, the network fitting target is linked to the actual physical object, which improves the interpretability of the neural network and thus enhances the overall interpretability of the method.

[0025] 4. Wind tunnel test data verification results show that, compared with the traditional neural network-based FADS system atmospheric parameter calculation method, the method proposed in this invention has higher accuracy and better stability of atmospheric parameters under limited training data conditions, and exhibits better calculation accuracy and algorithm stability under unfavorable conditions such as sparse feature points and network extrapolation. Attached Figure Description

[0026] Figure 1 A schematic diagram of a neural network architecture for fitting geometric features.

[0027] Figure 2 This is a schematic diagram of the distribution of pressure measuring holes in a scale model of a wind tunnel test.

[0028] Figure 3 This is a diagram showing the error distribution of the angle of attack fitting under normal conditions.

[0029] Figure 4 This is a diagram showing the error distribution of the sideslip angle fitting under normal conditions.

[0030] Figure 5 This is a diagram showing the error distribution of dynamic pressure calculation under normal conditions.

[0031] Figure 6 This is a diagram showing the error distribution of static pressure calculation under normal conditions.

[0032] Figure 7 This is a graph showing the error distribution for Mach number calculation under normal conditions.

[0033] Figure 8 This is a diagram showing the error distribution of the angle of attack fitting in a sparse feature point state.

[0034] Figure 9 The figure shows the distribution of the sideslip angle fitting error in the sparse feature point state.

[0035] Figure 10 This is a diagram showing the error distribution of dynamic pressure calculation in the sparse state of feature points.

[0036] Figure 11 This is a diagram showing the error distribution of sparse hydrostatic solution for feature points.

[0037] Figure 12 This is a diagram showing the error distribution of Mach number calculation in the sparse state of feature points.

[0038] Figure 13 This is a graph showing the error distribution of the network extrapolation state angle of attack fitting.

[0039] Figure 14 This is a graph showing the error distribution of the sideslip angle fitting for the network extrapolation state.

[0040] Figure 15 This is a diagram showing the error distribution of the dynamic pressure solution for the network extrapolation state.

[0041] Figure 16 This is a diagram showing the error distribution of the static pressure solution under network extrapolation.

[0042] Figure 17 This is a graph showing the error distribution in the extrapolation of the Mach number of the network state. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0044] A method for calculating atmospheric parameters in a FADS system based on geometric feature fitting neural networks includes the following steps:

[0045] Step 1: Obtain the surface pressure matrix required for training the neural network using a scaled-down model of a wind tunnel and an aircraft. and a reference atmospheric data vector, which includes a reference angle of attack vector. Reference sideslip angle vector Reference static pressure vector and reference dynamic pressure vector ;

[0046] Step 2: Based on the surface pressure matrix Constructing the neural network training input matrix ;

[0047] Step 3: Based on the surface pressure matrix Reference static pressure vector and reference dynamic pressure vector Calculate the reference pressure coefficient matrix ;

[0048] Step 4: Construct a neural network for fitting geometric features;

[0049] Step 5: Train the input matrix using a neural network As the input for training the geometric feature fitting neural network, with reference angle of attack vector Reference sideslip angle vector and reference pressure coefficient matrix The geometric feature fitting neural network training output is used to complete the training of the geometric feature fitting neural network.

[0050] Step 6: Obtain the actual surface pressure vector at the current moment using the pressure measurement port at the front of the actual aircraft. ;

[0051] Step 7: Construct the actual network input vector at the current time step ;

[0052] Step 8: Utilize the actual network input vector at the current time. The geometric feature fitting neural network trained in step 5 is used to obtain the fitted angle of attack at the current time. Sideslip angle and fitted pressure coefficient vector ;

[0053] Step 9: Based on the actual surface pressure vector at the current moment Fit the pressure coefficient vector to the current time output of the neural network. Calculate the dynamic pressure at the current moment and static pressure ;

[0054] Step 10: Based on the calculated dynamic pressure at the current moment and static pressure Determine the current flight status;

[0055] Step 11: Calculate the current flight Mach number based on the determined flight status.

[0056] Specifically, in step 1, the surface pressure matrix for 3D matrix The number of wind tunnel test condition combinations (wind tunnel test conditions include Mach number, angle of attack, and sideslip angle; different combinations of these three conditions constitute different wind tunnel test condition combinations. For example, when the only wind tunnel test condition is "Mach number 0.3, angle of attack -9°, sideslip angle -6°", the number of combinations is...) Labeled as 1, when the wind tunnel test conditions include two different combinations: "Mach number 0.3, angle of attack -9°, sideslip angle -6°" and "Mach number 0.3, angle of attack -6°, sideslip angle -6°", It can be denoted as 2, and so on. (This refers to the number of different combinations of conditions in the wind tunnel test). This represents the number of FADS pressure gauge holes. Each row contains surface pressure information measured at the same time. Different rows represent surface pressure information under different combinations of wind tunnel test conditions. The reference angle of attack vector is also included. Reference sideslip angle matrix Reference static pressure matrix and reference dynamic pressure matrix All A 3D vector, where each element is a bitwise AND vector. Atmospheric data corresponding to surface pressure information with the same row number, wherein the atmospheric data includes angle of attack. Sideslip angle static pressure Dynamic pressure The elements in the reference atmospheric data vector were all obtained directly through wind tunnel testing.

[0057] Specifically, in step 2, for A 3D matrix, with row vectors in each row. elements All only utilize Corresponding row vector Calculations show that forward Dimensional elements The calculation method is shown in equation (1):

[0058] (1)

[0059] in, It is a row vector The surface pressure value measured by the i-th pressure measuring hole. row vector The highest surface pressure value, row vector The smallest surface pressure value, Last dimension element The calculation method is shown in equation (2):

[0060] (2).

[0061] Specifically, in step 3, the reference pressure coefficient matrix for 3D matrix Each row vector It is a pressure coefficient vector composed of the pressure coefficients of each pressure measuring hole under the corresponding wind tunnel test conditions. Elements The calculation method is shown in equation (3):

[0062] (3)

[0063] in, It is the reference hydrostatic matrix In the middle and surface pressure vector The corresponding static pressure, It is the reference dynamic pressure matrix In the middle and surface pressure vector The corresponding dynamic pressure.

[0064] Specifically, the geometric feature fitting neural network constructed in step 4 is a three-layer non-fully connected neural network. The hidden layer neurons are divided into two sub-networks based on their function: sub-network 1 and sub-network 2. Sub-network 1 is a single hidden layer structure with 30 nodes, and sub-network 2 is a double hidden layer structure with 35 nodes per layer. The activation function for the input layer neurons is the Sigmoid function, whose function form is:

[0065] (4)

[0066] in, As input to the activation function, The output of the activation function;

[0067] Number of input layer nodes and neural network input matrix The column numbers are consistent, which is The number of output layer nodes is Angle of attack and sideslip angle The corresponding output layer nodes are only connected to the nodes of subnetwork 1, outputting a pressure coefficient vector. The nodes in subnetwork 1 are only connected to the nodes in subnetwork 2. The output layer activation function is the purelin function, a linear identity function. The constructed network architecture is as follows: Figure 1 As shown.

[0068] Specifically, in step 5, the neural network training method employs gradient descent, with a learning rate of... The training objective is set to .

[0069] Specifically, in step 7, the actual network input vector at the current moment... for The dimensional vector, its calculation method and step 2 Each row vector The calculation method is the same, and it is denoted as:

[0070] (5)

[0071] (6)

[0072] in, The surface pressure vector at the current moment The surface pressure corresponding to the i-th pressure measuring hole in the diagram. yes The highest surface pressure value, yes The smallest surface pressure value, for forward Weizhong and The corresponding i-th element, yes The last element in the middle.

[0073] Specifically, in step 9, dynamic pressure and static pressure The calculation formula is shown in equation (7):

[0074] , (7)

[0075] in:

[0076] : The vector of fitted pressure coefficients output by the geometric feature fitting neural network;

[0077] :and Unit vectors of the same dimension.

[0078] Specifically, in step 10, the method for determining the current flight status is as follows:

[0079] (8)

[0080] when It was judged to be subsonic flight at that time. It was determined to be supersonic flight.

[0081] Specifically, in step 11, based on the determined flight state, the specific Mach number is obtained by solving equation (9):

[0082] (9).

[0083] To verify the effectiveness and advancement of the atmospheric parameter calculation method for the FADS system proposed in this invention, wind tunnel tests were first conducted using a scale model of the nose of a certain type of aircraft. A total of 168 sets of wind tunnel test data under different conditions were obtained, which were used to construct the neural network training and validation sets. The training set was used to train the neural network used in this invention, and the validation set was used to test the effectiveness of this invention in calculating atmospheric parameters. The pressure gauge layout on the scale model is as follows... Figure 2 As shown.

[0084] To simulate three different states that may occur in real-world neural network applications—normal, sparse feature point, and network extrapolation—three sets of "training-test" datasets were constructed under different states. The mean error and root mean square error (RMSE) of 10 repeated experiments were used to analyze the accuracy and stability of the calculated atmospheric parameters. First, atmospheric data was calculated under the normal state, and after 10 repeated verifications, the mean error and RMS error of the 10 atmospheric data calculations were recorded. Then, verification and result recording were performed under the sparse feature point state and the network extrapolation state. The comparative results are shown in the attached table and figures.

[0085] Under normal conditions, the atmospheric data calculation results obtained based on the method proposed in this invention are compared with the true values ​​in wind tunnel test data in 10 repeated experiments. Figures 3 to 7 As shown in the attached figure, the mean and root mean square error of the fitting errors for the angle of attack, sideslip angle, static pressure, dynamic pressure, and Mach number of the method proposed in this invention vary under this condition. The triangle symbol represents the mean fitting error, and the length of the error bar represents the magnitude of the root mean square error.

[0086] The mean error of the fitting of 10 repeated trials is shown in Table 1.

[0087] Table 1. Mean values ​​of atmospheric parameter calculation errors in repeated validation experiments under normal conditions

[0088]

[0089] Verification results show that the calculation errors of various atmospheric parameters obtained by the method proposed in this invention are all maintained at a low level. Specifically, the average errors of angle of attack and sideslip angle calculation are all less than 0.1°, the average errors of dynamic pressure, static pressure and Mach number calculation are all less than 0.1%, and the average error of static pressure equivalent height is less than 20m.

[0090] The atmospheric data solution obtained based on the method proposed in this invention under sparse feature points is compared with the true values ​​in wind tunnel test data in 10 repeated experiments. Figures 8 to 12 As shown in the figure, the accompanying figure illustrates the variations in the mean and root mean square error of the fitting errors for the angle of attack, sideslip angle, static pressure, dynamic pressure, and Mach number of the method proposed in this invention under this condition.

[0091] The mean error of the fitting of 10 repeated trials is shown in Table 2.

[0092] Table 2. Mean error of atmospheric parameter calculation in repeated verification experiments under sparse feature point conditions.

[0093]

[0094] The atmospheric parameters calculated under sparse feature points have a certain degree of error compared with those under normal conditions, but the average error of angle of attack and sideslip angle is still less than 0.1°, the average error of dynamic pressure and static pressure is less than 0.3%, and the average error of Mach number is less than 2%.

[0095] The atmospheric data solution obtained by the method proposed in this invention under network extrapolation conditions is compared with the true values ​​in wind tunnel test data in 10 repeated experiments. Figures 13 to 17 As shown in the figure, the mean and root mean square error of the fitting errors for the angle of attack, sideslip angle, static pressure, dynamic pressure, and Mach number of the method proposed in this invention vary under this condition.

[0096] The mean error of the fitting of 10 repeated trials is shown in Table 3.

[0097] The verification results under network extrapolation conditions show that the average error of atmospheric parameter calculation obtained by the method proposed in this invention can still be maintained at a relatively low level, but the average root mean square error is significantly increased compared with the previous two conditions. Specifically, the average error of angle of attack and sideslip angle calculation is still less than 0.1°, the average error of dynamic pressure and static pressure calculation is less than 0.7%, and the average error of Mach number calculation is less than 5%.

[0098] Table 3. Mean error of atmospheric parameter calculation in repeated verification experiments of network extrapolation state

[0099]

[0100] The verification results show that the method proposed in this invention can maintain a relatively low atmospheric data fitting error under various conditions. Among them, the fitting errors of angle of attack and sideslip angle are relatively stable, while the fitting errors of dynamic pressure, static pressure and Mach number increase as the overlap between the validation set and the training set decreases, but overall they are maintained at a relatively low level.

[0101] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for calculating atmospheric parameters in a FADS system based on a geometric feature fitting neural network, characterized in that, Includes the following steps: Step 1: Obtain the surface pressure matrix p required for training the neural network using a scaled-down model of a wind tunnel and an aircraft. train and a reference atmospheric data vector, which includes a reference angle of attack vector. Reference sideslip angle vector Reference static pressure vector and reference dynamic pressure vector Step 2: Based on the surface pressure matrix p train Constructing the neural network training input matrix Γ train ; Step 3: Based on the surface pressure matrix p train Reference static pressure vector and reference dynamic pressure vector Calculate the reference pressure coefficient matrix C p,reference ; Step 4: Construct a neural network for fitting geometric features; Step 5: Train the input matrix Γ using a neural network train As the training input for the geometric feature fitting neural network, with reference angle of attack vector Reference sideslip angle vector and reference pressure coefficient matrix C p,reference The training output of the geometric feature fitting neural network is used to complete the training of the geometric feature fitting neural network. Step 6: Obtain the actual surface pressure vector at the current moment using the pressure measurement port at the front of the actual aircraft. Step 7: Construct the actual network input vector at the current time step Step 8: Utilize the actual network input vector at the current time. The fitted angle of attack at the current moment is obtained by combining the geometric feature fitting neural network trained in step 5. Sideslip angle and fitted pressure coefficient vector Step 9: Based on the actual surface pressure vector at the current moment Fit the pressure coefficient vector to the current time output of the neural network. Calculate the dynamic pressure at the current moment and static pressure Step 10: Based on the calculated dynamic pressure at the current moment and static pressure Determine the current flight status; Step 11: Calculate the current flight Mach number based on the determined flight status.

2. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 1, characterized in that, In step 1, the surface pressure matrix p train This is an m×n dimensional matrix, where m is the number of wind tunnel test condition combinations and n is the number of FADS pressure gauge holes. Each row contains surface pressure information measured at the same time, and different rows represent surface pressure information under different wind tunnel test condition combinations. The reference angle of attack vector is used. Reference sideslip angle matrix Reference static pressure matrix and reference dynamic pressure matrix Both are m-dimensional vectors, and each element is related to p. train The atmospheric data corresponding to the surface pressure information of the same row number, wherein the atmospheric data includes angle of attack α, sideslip angle β, static pressure P0, and dynamic pressure q. c The elements in the reference atmospheric data vector were all obtained directly through wind tunnel testing.

3. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 2, characterized in that, In step 2, Γ train Given an m×(n+1) dimensional matrix, each row contains row vectors. element Γ mi All only use p train Corresponding row vector Calculations show that The first n-dimensional element Γ mi The calculation method is shown in equation (1): Where, p mi It is a row vector The surface pressure value p measured by the i-th pressure measuring hole m,max row vector The highest surface pressure value, p m,min row vector The smallest surface pressure value, The last element Γ m,end The calculation method is shown in equation (2):

4. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 3, characterized in that, In step 3, the reference pressure coefficient matrix C p,reference For an m×n dimensional matrix, C p,reference Each row vector It is a pressure coefficient vector composed of the pressure coefficients of each pressure measuring hole under the corresponding wind tunnel test conditions. Elements The calculation method is shown in equation (3): in, It is the reference hydrostatic matrix In the middle and surface pressure vector The corresponding static pressure, It is the reference dynamic pressure matrix In the middle and surface pressure vector The corresponding dynamic pressure.

5. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 4, characterized in that, The geometric feature fitting neural network constructed in step 4 is a three-layer non-fully connected neural network. The hidden layer neurons are divided into two sub-networks based on their function: sub-network 1 and sub-network 2. Sub-network 1 is a single hidden layer structure with 30 nodes, and sub-network 2 is a double hidden layer structure with 35 nodes per layer. The activation function for the input layer neurons is the Sigmoid function, whose function form is: Where x is the input of the activation function and σ(x) is the output of the activation function; Number of input layer nodes and neural network input matrix Γ train The number of columns is consistent, n+1, the number of output layer nodes is n+2, the output layer nodes corresponding to the angle of attack α and sideslip angle β are only connected to the nodes of subnetwork 1, and the output pressure coefficient vector is... The nodes of the first network are only connected to the nodes of the second network, and the activation function of the output layer is the purelin function, a linear identity function.

6. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 5, characterized in that, In step 5, the neural network training method uses gradient descent with a learning rate of 1×10⁻⁶. -3 The training objective is set to 1×10. -6 .

7. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 6, characterized in that, In step 7, the actual network input vector at the current moment For an n+1 dimensional vector, its calculation method and step 2 Γ train Each row vector The calculation method is the same, and it is denoted as: Where, p i The surface pressure vector at the current moment The surface pressure corresponding to the i-th pressure measuring hole, p max yes The highest surface pressure value, p min yes The minimum surface pressure value, Γ i for In the first n dimensions and p i The corresponding i-th element, Γ end yes The last element in the middle.

8. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 7, characterized in that, In step 9, dynamic pressure and static pressure The calculation formula is shown in equation (7): in: The fitting pressure coefficient vector output by the geometric feature fitting neural network; and Unit vectors of the same dimension.

9. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 8, characterized in that, In step 10, the method for determining the current flight status is as follows: When Ma≤1, it is judged as subsonic flight; when Ma>1, it is judged as supersonic flight.

10. The method for calculating atmospheric parameters of a FADS system based on a geometric feature fitting neural network according to claim 9, characterized in that, In step 11, based on the determined flight state, the specific Mach number is obtained by solving equation (9):

Images