A FADS solution system and method based on dimensionless input-output neural network

Abstract

The invention discloses a FADS solution system and method based on a dimensionless input-output neural network. The method comprises: determining the distribution of each pressure measuring point on the surface of an aircraft, collecting the pressure values ​​of each pressure measuring point under the corresponding incoming flow condition; performing dimensionless processing on the pressure values ​​of each pressure measuring point to obtain an input parameter set; using a first neural network to solve each dimensionless atmospheric parameter and contained in the atmospheric data parameter; using a second neural network to solve the pressure coefficient Cpi of each pressure measuring point; solving the pressure coefficient of each pressure measuring point according to the second neural network; the system comprises: a pressure value collection unit, a dimensionless processing unit, a dimensionless atmospheric parameter solving unit, and a dimensional atmospheric parameter calculating unit. The invention makes full use of the similarity under different free incoming flow static pressures, so that the aerodynamic model for solving the atmospheric data decouples the free incoming flow static pressure, and the samples for training the neural network are greatly reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of aerodynamics, and in particular to a FADS solution system and method based on a dimensionless input-output neural network. Background Art

[0002] The atmospheric data system is an important airborne electronic system on the aircraft. The atmospheric parameters such as angle of attack, sideslip angle, total pressure, static pressure, and total temperature provided by it are crucial for the guidance, navigation, and control of the aircraft. With the continuous improvement of the requirements of modern aircraft for flight speed, maneuverability, and stealth, the probe measurement method used by the traditional atmospheric data system cannot meet the measurement needs of the new generation of aircraft. Seeking a new atmospheric data measurement method has become an urgent need for atmospheric data sensing of the new generation of aircraft.

[0003] The flush air data sensing (FADS) system relies on the pressure sensor array embedded in the surface of the aircraft to measure the pressure distribution at the head of the aircraft or other local positions, and indirectly measures the Mach number, angle of attack, sideslip angle, static pressure and dynamic pressure and other flow parameters during the flight through a specific solution algorithm model. Compared with the traditional air data measurement method, the FADS system has great advantages in reducing the aircraft radar reflection area and adapting to the hypersonic high-heat and high-pressure flight environment. Therefore, the FADS system has become the first choice for many high-performance aircraft air data systems.

[0004] The corresponding solution method of the FADS system is the core of the entire embedded atmospheric data sensing system, which directly determines the accuracy, real-time performance and reliability of the system. The traditional solution method is based on the semi-empirical pressure model under ideal assumptions, such as the three-point method and the least squares method. However, the above algorithm has low accuracy and poor real-time performance, and the pressure sensor array can only be arranged in the front of the blunt nose of the aircraft (generally hemispherical), which has strict restrictions on the shape of the aircraft, greatly limiting the application scenarios of the FADS system.

[0005] The neural network algorithm can simulate complex nonlinear problems with its powerful nonlinear fitting ability, and is suitable for solving atmospheric parameters from pressure distribution. The corresponding solution method is called the FADS neural network algorithm. Compared with traditional solution methods, the FADS neural network algorithm has low requirements on the shape of the aircraft and the distribution of pressure measuring holes, which can further expand the application scenarios of the FADS system. However, this type of method usually directly establishes the mapping relationship between pressure distribution and atmospheric parameters, cannot decouple the static pressure of the incoming flow, and ignores the similarities under different working conditions. In order to ensure the accuracy and robustness of the FADS neural network algorithm, a large number of training samples are usually required for the parameter learning process of the neural network. However, the acquisition of training samples generally has a very high cost, which has become the main obstacle to the application of neural network algorithms in embedded atmospheric data systems.

[0006] Therefore, how to use fewer samples to train the FADS neural network while ensuring that the neural network algorithm obtained through training achieves better performance is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the invention

[0007] In view of the problems existing in the prior art, the present invention proposes a FADS solution system and method based on a dimensionless input-output neural network, aiming to solve the problem that the prior art usually requires a large number of training samples for the parameter learning process of the neural network, while the acquisition of training samples generally has a very high cost, which becomes the main obstacle to the application of neural network algorithms in embedded atmospheric data systems.

[0008] The present invention adopts the following technical solutions to solve the technical problems:

[0009] A FADS solution method based on a dimensionless input-output neural network, characterized in that the method comprises the following steps:

[0010] Step 1: determine the distribution of each pressure measuring point on the aircraft surface, and collect the pressure values ​​of each pressure measuring point under the corresponding incoming flow condition;

[0011] Step 2: Dimensionally transform the pressure values ​​of each pressure measuring point to obtain an input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ]; The dimensionless processing includes: respectively calculating the pressure ratio of the pressure value of each pressure measuring point to the pressure value of each pressure measuring point except itself Thus, the input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ];

[0012] Step 3: Use the first neural network to solve the dimensionless atmospheric parameters α, β and Ma contained in the atmospheric data parameters: After dimensionless processing, the input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ] as the input parameter of the first neural network, and the dimensionless atmospheric parameters including the angle of attack α, the sideslip angle β and the Mach number Ma as the output parameters of the first neural network. ij It is the ratio of the pressure value of each pressure measuring point to the pressure of every pressure measuring point except itself;

[0013] Step 4: Use the second neural network to calculate the pressure coefficient Cp of each pressure measuring point i Solve; use the output parameters α, β, Ma of the first neural network as the input parameters of the second neural network, and the pressure coefficient Cp of each pressure measuring point i As the output parameter of the second neural network, the Cp i is the pressure coefficient at each pressure measuring hole [Cp1, Cp2, …, Cp n ];

[0014] Step 5: Solve the pressure coefficients of each pressure measuring point [Cp1, Cp2, ..., Cp n ], combined with the corresponding measured pressure distribution, the dynamic pressure q of each dimensioned atmospheric parameter contained in the atmospheric data parameters is calculated. c , static pressure p ∞ The solution is obtained by using the least square method based on the pressure coefficients and pressure values ​​of multiple pressure measuring holes.

[0015] The dimensionless input parameter Cp of the first neural network ij The functional relationship between the dimensionless output parameters α, β and Ma is: Assuming v1 = [α, β, Ma], we can inverse the function of formula (6) and obtain the functional relationship: v1 = Π1(Cp 12 ,Cp 21 ,...,Cp ij ) α, β, Ma are only dimensionless parameters [Cp 12 ,Cp 21 ,…,Cp ij ] is a function that is independent of the incoming flow static pressure, thereby decoupling the incoming flow static pressure.

[0016] The specific process of setting this function is as follows:

[0017] 1) Set a total of n pressure measuring holes;

[0018] 2) Each pressure measuring hole is compared with the remaining pressure measuring holes except itself;

[0019] 2) construct n·(n-1) equations of formula (6);

[0020] Where i≤n, j≤n, i≠j, when the pressure measuring hole distribution is determined, r i / l is a known parameter.

[0021] It is characterized in that: the derivation of formula (6) is as follows:

[0022] Assume that the pressure at a point i on the surface of the aircraft can be expressed as:

[0023] p i =q c Cp i +p ∞ (1)

[0024] where q c 、p ∞ is the dynamic pressure and static pressure of the incoming flow, Cp i is the pressure coefficient, which is a function of the relative position of the pressure hole, the angle of attack α, the sideslip angle β and the Mach number Ma. Assuming that the shape of the aircraft is determined, Cp i It can be written as

[0025]

[0026] where r i =[x i ,y i ,z i ] represents the spatial position of the i-th pressure measuring hole, and l is the reference length of the aircraft.

[0027] Assuming the pressure at the i-th and j-th pressure measuring holes, the pressure at the two points is compared to obtain the following formula:

[0028]

[0029] In the above formula, It is M ∞ The specific relationship is as follows:

[0030]

[0031] Substituting (4) into (3), we can obtain:

[0032]

[0033] Assuming there are n pressure measuring holes, each pressure measuring hole is ratioed with the remaining pressure measuring holes except itself, n·(n-1) equations can be constructed:

[0034]

[0035] The dimensionless input parameters α, β, Ma and the dimensionless output parameter Cp of the second neural network i The functional relationship is: v2=Π2(α,β,Ma), where v2=[Cp1,Cp2,...,Cp n ] (7)

[0036] The method of solving the pressure coefficient and pressure value based on multiple pressure measuring holes by using the least square method includes the following process:

[0037] 1) Construct the following superpositive definite equation:

[0038]

[0039] Solve α, β and Ma by formula (6). It is known from formula (2) that when the relative position of the pressure measuring hole is determined, Cp i As a function of α, β and Ma, the Cp at each pressure measuring hole under the corresponding flow condition can be determined. i , using the pressure coefficients at each pressure measuring hole and the known pressure distribution, the superpositive equation of formula (8) is constructed:

[0040] Formula (8) is solved using the least squares method, and is defined as:

[0041]

[0042] Then q c 、p ∞ The expression is:

[0043]

[0044] At this point, the angle of attack α, sideslip angle β, Mach number Ma, and dynamic pressure q of the outflow can be solved. c , static pressure p ∞ .

[0045] A FADS solution system of a dimensionless input-output neural network, characterized in that the system comprises:

[0046] A pressure value collection unit, used to determine the distribution of various pressure measuring points on the aircraft surface, and to collect the pressure values ​​of various pressure measuring points under the corresponding incoming flow conditions;

[0047] A dimensionless processing unit, used for performing dimensionless processing on the pressure values ​​of each pressure measuring point to obtain an input parameter set;

[0048] A dimensionless atmospheric parameter solving unit, used for solving each dimensionless atmospheric parameter included in the atmospheric data parameter by using the input parameter set and adopting a neural network;

[0049] The dimensioned atmospheric parameter calculation unit is used to use the dimensionless atmospheric parameters to obtain the pressure coefficient of each pressure measuring point using a neural network, and to calculate each dimensioned atmospheric parameter included in the atmospheric data parameter using the pressure coefficient of each pressure measuring point and the pressure value of each pressure measuring point.

[0050] The dimensionless parameter solving unit comprises:

[0051] A first calculation subunit is used to input the input parameter set as input data into a first neural network model, so that the first neural network model outputs each dimensionless atmospheric parameter included in the atmospheric data parameter;

[0052] The dimensioned parameter calculation unit comprises:

[0053] The second calculation subunit is used to input the dimensionless atmospheric parameters as input data into the second neural network model, so that the second neural network model outputs the pressure coefficient of each pressure measuring point.

[0054] Advantages and effects of the present invention

[0055] 1. The embedded atmospheric data system solution method based on the dimensionless input-output neural network of the present invention transforms the pressure distribution measured by the pressure sensor array to obtain a set of dimensionless parameters as the input of the neural network, and uses the dimensionless atmospheric parameters in the atmospheric parameters as the output of the neural network, thereby decoupling the static pressure of the incoming flow, causing a qualitative change in the aerodynamic model approximated by the neural network, and being able to fully utilize the similarities under different free incoming flow static pressures, thereby greatly reducing the samples for training the neural network and reducing the construction cost of the FADS system.

[0056] 2. The present invention uses the first neural network model to obtain three dimensionless atmospheric parameters: Mach number Ma, angle of attack α, and sideslip angle β. Then, the three dimensionless atmospheric parameters are used to use the second neural network to obtain the pressure coefficient Cpi of each pressure measuring point. Finally, the pressure coefficient Cpi of each pressure measuring point is used to calculate the dimensioned atmospheric parameter dynamic pressure q by an appropriate method. c and static pressure p ∞ On the one hand, it greatly reduces the sample data used to train the FADS system, and on the other hand, it lays the foundation for improving the applicability and robustness of the FADS system, making it possible to apply the FADS neural network algorithm on a large scale, which has significant economic benefits and application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a flow chart of an embedded atmospheric data system solution method based on a dimensionless input-output neural network provided by an embodiment of the present invention;

[0058] Figure 2 is a structural schematic diagram of a first neural network model provided by the present invention;

[0059] Figure 3 is a schematic diagram of the structure of a second neural network model provided by the present invention;

[0060] Figure 4 is a schematic diagram of the distribution of the pressure sensor array provided by the present invention on the airfoil head;

[0061] Figure 5a Pressure values ​​at pressure measuring hole 1 under different working conditions

[0062] Figure 5b Pressure values ​​at pressure measuring hole 2 under different working conditions

[0063] Figure 5c An input parameter Cp12 of the infinite input-output neural network (referring to the first neural network)

[0064] FIG6( a ) is a schematic diagram of the absolute error of the dimensionless atmospheric parameter (Mach number) provided by the present invention;

[0065] FIG6( b ) is a schematic diagram of the absolute error of the dimensionless atmospheric parameter (angle of attack) provided by the present invention;

[0066] FIG. 7( a ) is a schematic diagram of the absolute error of a dimensioned atmospheric parameter (dynamic pressure) provided by the present invention;

[0067] FIG7( b ) is a schematic diagram of the absolute error of a dimensioned atmospheric parameter (static pressure) provided by the present invention;

[0068] Figure 8 It is a schematic diagram of an embedded atmospheric data system solution method based on a dimensionless input-output neural network provided by the present invention. DETAILED DESCRIPTION

[0069] The following will be combined with the accompanying drawings in the embodiments of the present invention, the technical solution in the embodiments of the present invention will be clearly and completely described.

[0070] Design principle of the present invention

[0071] 1. Solution method of flush air data sensing (FADS) system: This solution method can be essentially considered as a regression problem of mapping pressure distribution to atmospheric parameters. The neural network algorithm can fit any nonlinear problem with its excellent nonlinear fitting ability. Compared with the traditional FADS solution method, the FADS neural network algorithm has great advantages in terms of real-time performance, stability and applicable appearance.

[0072] 2. Problems with existing FADS neural networks: The input of the existing FADS neural network is the pressure distribution measured by the pressure sensor array, rather than the input parameter set after dimensionless processing of the present invention. The output includes both dimensionless atmospheric parameters and dimensional atmospheric parameters. In order to build a complete FADS system, it is necessary to establish a FADS sample database covering the entire world. The large number of training samples greatly increases the cost of building a FADS neural network system, which greatly limits the application of neural network algorithms in the FADS system. In addition, whether the FADS neural network system established by any means is effective needs to be tested through flight tests. Due to the expensive cost of flight tests, a large number of FADS neural network algorithms remain at the level of theoretical and simulation research.

[0073] 3. The training samples of the present invention are reduced from 4000 to 400: 1) In the function after dimensionless processing, the four independent variables α, β, Ma, and p∞ are changed to three independent variables α, β, and Ma. 2) From the formula (6) of the present invention, it can be seen that there are only three independent variables in the function, while the formula (6) of the traditional method also includes the fourth variable p∞; 3) The difference between four independent variables and three independent variables: the magnitude of the four independent variables is 104. When the magnitude is 104 in actual measurement, the minimum number of training samples required is about 4000 to meet the requirements; while the magnitude of the three independent variables is 103. When the magnitude is 103 in actual measurement, the minimum number of training samples required is about 400 to meet the requirements; 4000 is 10 times different from 400.

[0074] 4. Design principle of using dimensionless processing to reduce training samples:

[0075] 1) The dimensions of physical quantities are used to analyze or check the relationship between several physical quantities. This method is called dimensional analysis. Usually, the dimensions of a physical quantity are composed of basic physical dimensions such as mass, length, time, charge, and temperature. For example, the dimension of speed is length per unit time, and the unit of measurement is meters per second, miles per hour, or other units. The important principle on which dimensional analysis is based is that physical laws must be independent of or related to the units of the physical quantities they measure. If they are independent of the units of the physical quantities they measure, they are dimensionless physical laws. If they are related to the units of the physical quantities they measure, they are dimensional physical laws. For any meaningful equation, the dimensions on the left and right sides of the equation must be the same. Checking whether this rule is followed is the most basic step in doing dimensional analysis.

[0076] 2) The input parameter set obtained by dimensionless processing is used as the input layer of the neural network. Dimensional analysis is often more conducive to analyzing and judging the quantitative relationship between things and the general laws followed, and can better reveal the essence of a certain phenomenon. Based on dimensional analysis, the input layer of the FADS neural network algorithm is converted from the previous pressure signal to a dimensionless parameter. The dimensionless parameter in this embodiment is the ratio of pressure pi to pressure pj. The dimensionless atmospheric parameters α, β and Ma in the atmospheric parameters are used as the output layer of the neural network, which can decouple the static pressure in similar working conditions. The decoupled static pressure is to reduce the four dimensionless atmospheric parameters α, β, Ma, p∞ to three dimensionless atmospheric parameters α, β, Ma in the function of formula (6). This can greatly reduce the sample points used to train the neural network.

[0077] 3) The solution of atmospheric parameters is divided into two parts: the solution of dimensionless atmospheric parameters and the solution of dimensional atmospheric parameters. The solution of dimensionless parameters mainly changes the input and output of the previous neural network, changing the input layer of the neural network from the previous pressure signal to a dimensionless pressure ratio, and the output layer is the dimensionless atmospheric parameters of the incoming flow. The neural network constructed in this way decouples the static pressure of the incoming flow, greatly reducing the samples used to train the neural network, thereby reducing the construction cost of the FADS system. The solution of dimensional atmospheric parameters mainly solves the dynamic pressure and static pressure, which is to combine the pressure coefficient output by the second neural network with the pressure distribution measured by the pressure sensor array, and use the least squares method to solve the dynamic pressure and static pressure.

[0078] 4) Specific steps for solving dimensionless and dimensional atmospheric parameters: a. First, determine the distribution of the pressure sensor array on the aircraft surface and collect the pressure signals under the corresponding incoming flow conditions. Assuming that there are n pressure sensors collecting pressure signals, the pressure signal collected by the i-th pressure sensor is dimensionless, specifically, the i-th pressure signal is compared with the pressure signals other than itself, and (n-1) dimensionless quantities are obtained. For n pressure sensors, a total of n·(n-1) dimensionless parameters [Cp 12 ,Cp 21 ,…,Cpi j ], these dimensionless parameters are used as the input of the subsequent first neural network model. b. Solve the Mach number, angle of attack, and sideslip angle. Construct Figure 2 The first neural network model, [x1,x2,…,x m ] is the input of the neural network, where m = n(n-1), and the input layer is the dimensionless parameter [Cp 12 ,Cp 21 ,…,Cpi j], [y1, y2, y3] is the output of the neural network, specifically [α, β, Ma]. Based on the training samples, learning is performed to infinitely approximate the Π function in formula (7). The type and structural parameter settings of the first neural network model have a high degree of flexibility. For example, it can be a BP neural network model, etc. The purpose is to achieve accurate approximation of the Π function in formula (7), so that α, β and Ma of the incoming flow condition can be accurately obtained. c. Knowing α, β and Ma of the incoming flow condition, the pressure coefficient Cp at each pressure measuring hole under the corresponding condition can be solved i . Cp i is a function of α, β and Ma. In order to accurately solve Cp under different flow parameters i , can be constructed to solve Cp i The second neural network model, such as Figure 3 The input of the second neural network model is α, β, Ma, and the output is the pressure coefficient at each pressure measuring hole [Cp1, Cp2,…, Cp n ], based on the training samples, a second neural network can be obtained to solve the pressure coefficient at each pressure measuring hole. d. The pressure coefficient at each pressure measuring hole under the current flow condition is known, and the dynamic pressure q c , static pressure p ∞ The solution can be obtained by using the least square method based on the pressure coefficients and pressure values ​​of multiple pressure measuring holes. The solution process refers to equations (9) and (10).

[0079] Based on the above principles, the present invention designs an embedded atmospheric data system solution method based on dimensionless input-output neural network.

[0080] A FADS solution method based on a dimensionless input-output neural network, characterized in that the method comprises the following steps:

[0081] Step 1: determine the distribution of each pressure measuring point on the aircraft surface, and collect the pressure values ​​of each pressure measuring point under the corresponding incoming flow condition;

[0082] Step 2: Dimensionally transform the pressure values ​​of each pressure measuring point to obtain an input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ]; The dimensionless processing includes: respectively calculating the pressure ratio of the pressure value of each pressure measuring point to the pressure value of each pressure measuring point except itself Thus, the input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ];

[0083] Step 3: Use the first neural network to solve the dimensionless atmospheric parameters α, β and Ma contained in the atmospheric data parameters: After dimensionless processing, the input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ] as the input parameter of the first neural network, and the dimensionless atmospheric parameters including the angle of attack α, the sideslip angle β and the Mach number Ma as the output parameters of the first neural network. ij It is the ratio of the pressure value of each pressure measuring point to the pressure of every pressure measuring point except itself;

[0084] Step 4: Use the second neural network to calculate the pressure coefficient Cp of each pressure measuring point i Solve; use the output parameters α, β, Ma of the first neural network as the input parameters of the second neural network, and the pressure coefficient Cp of each pressure measuring point i As the output parameter of the second neural network, the Cp i is the pressure coefficient at each pressure measuring hole [Cp1, Cp2, …, Cp n ];

[0085] Step 5: Solve the pressure coefficients of each pressure measuring point [Cp1, Cp2, ..., Cp n ], combined with the corresponding measured pressure distribution, the dynamic pressure q of each dimensioned atmospheric parameter contained in the atmospheric data parameters is calculated. c , static pressure p ∞ The solution is obtained by using the least square method based on the pressure coefficients and pressure values ​​of multiple pressure measuring holes.

[0086] The dimensionless input parameter Cp of the first neural network ij The functional relationship between the dimensionless output parameters α, β and Ma is:

[0087] Assuming v1 = [α, β, Ma], we can inverse the function of formula (6) and obtain the functional relationship: v1 = Π1(Cp 12 ,Cp 21 ,…,Cp ij )

[0088] α, β, Ma are only dimensionless parameters [Cp 12 ,Cp 21 ,…,Cp ij ] is a function that is independent of the incoming flow static pressure, thereby decoupling the incoming flow static pressure.

[0089] The specific process of setting this function is as follows:

[0090] 1) Set a total of n pressure measuring holes;

[0091] 2) Each pressure measuring hole is compared with the remaining pressure measuring holes except itself;

[0092] 2) construct n·(n-1) equations of formula (6);

[0093] Where i≤n, j≤n, i≠j, when the pressure measuring hole distribution is determined, r i / l is a known parameter.

[0094] It is characterized in that: the derivation of formula (6) is as follows:

[0095] Assume that the pressure at a point i on the surface of the aircraft can be expressed as:

[0096] p i =q c Cp i +p ∞ (1)

[0097] where q c 、p ∞ is the dynamic pressure and static pressure of the incoming flow, Cp i is the pressure coefficient, which is a function of the relative position of the pressure hole, the angle of attack α, the sideslip angle β and the Mach number Ma. Assuming that the shape of the aircraft is determined, Cp i It can be written as

[0098]

[0099] where r i =[x i ,y i ,z i ] represents the spatial position of the i-th pressure measuring hole, and l is the reference length of the aircraft.

[0100] Assuming the pressure at the i-th and j-th pressure measuring holes, the pressure at the two points is compared to obtain the following formula:

[0101]

[0102] In the above formula, It is M ∞ The specific relationship is as follows:

[0103]

[0104] Substituting (4) into (3), we can obtain:

[0105]

[0106] Assuming there are n pressure measuring holes, each pressure measuring hole is ratioed with the remaining pressure measuring holes except itself, n·(n-1) equations can be constructed:

[0107]

[0108] The dimensionless input parameters α, β, Ma and the dimensionless output parameter Cp of the second neural network i The functional relationship is: v2=Π2(α,β,Ma), where v2=[Cp1,Cp2,...,Cp n ] (7)

[0109] The method of solving the pressure coefficient and pressure value based on multiple pressure measuring holes by using the least square method includes the following process:

[0110] 1) Construct the following superpositive definite equation:

[0111]

[0112] Solve α, β and Ma by formula (6). It is known from formula (2) that when the relative position of the pressure measuring hole is determined, Cp i As a function of α, β and Ma, the Cp at each pressure measuring hole under the corresponding flow condition can be determined. i , using the pressure coefficients at each pressure measuring hole and the known pressure distribution, the superpositive equation of formula (8) is constructed:

[0113] Formula (8) is solved using the least squares method, and is defined as:

[0114]

[0115] Then q c 、p ∞ The expression is:

[0116]

[0117] At this point, the angle of attack α, sideslip angle β, Mach number Ma, and dynamic pressure q of the outflow can be solved. c , static pressure p ∞ .

[0118] A FADS solution system of a dimensionless input-output neural network, characterized in that the system comprises:

[0119] A pressure value collection unit, used to determine the distribution of various pressure measuring points on the aircraft surface, and to collect the pressure values ​​of various pressure measuring points under the corresponding incoming flow conditions;

[0120] A dimensionless processing unit, used for performing dimensionless processing on the pressure values ​​of each pressure measuring point to obtain an input parameter set;

[0121] A dimensionless atmospheric parameter solving unit, used for solving each dimensionless atmospheric parameter included in the atmospheric data parameter by using the input parameter set and adopting a neural network;

[0122] The dimensioned atmospheric parameter calculation unit is used to use the dimensionless atmospheric parameters to obtain the pressure coefficient of each pressure measuring point using a neural network, and to calculate each dimensioned atmospheric parameter included in the atmospheric data parameter using the pressure coefficient of each pressure measuring point and the pressure value of each pressure measuring point.

[0123] The dimensionless parameter solving unit comprises:

[0124] A first calculation subunit is used to input the input parameter set as input data into a first neural network model, so that the first neural network model outputs each dimensionless atmospheric parameter included in the atmospheric data parameter;

[0125] The dimensioned parameter calculation unit comprises:

[0126] The second calculation subunit is used to input the dimensionless atmospheric parameters as input data into the second neural network model, so that the second neural network model outputs the pressure coefficient of each pressure measuring point.

[0127] Example 1

[0128] The FADS neural network algorithm proposed in the embodiment of the present application can decouple the static pressure of the incoming flow, thereby greatly reducing the samples for training the FADS neural network: Taking the NACA0012 airfoil as the research object, refer to Figure 4 , a pressure measuring point 1 is arranged at the apex of the wing head, pressure measuring holes 2 and 3 are arranged on the upper and lower surfaces of the wing near the 12.5mm section of the head, and pressure measuring holes 4 and 5 are arranged near the 25mm section. A total of 5 pressure measuring points are used as the FADS system to measure the surface pressure sensor array of the model. Table 1 gives the specific coordinates of the pressure sensor array distributed on the airfoil head.

[0129] Table 1 Distribution of pressure measuring holes

[0130] Serial number X(mm) Y(mm) Z(mm) 1 0 0 0 2 13.5 19.6 0 3 12.2 -18.4 0 4 27.1 27.1 0 5 25.1 26.2 0

[0131] The measurement range of the FADS system is determined according to the flight of the aircraft. As shown in Table 2, the measurement range of the Mach number of the FADS system is determined to be 0.3-0.8, the measurement range of the angle of attack is -4-7°, and the measurement range of the static pressure is 2550-101325Pa (equivalent to an altitude of 0-25km).

[0132] Table 2 Measurement range of the atmospheric data system

[0133] Atmospheric data Measuring range Static pressure(pa) 2550-101325 Mach number 0.3-0.8 Angle of attack (deg) -4-7°

[0134] Note:

[0135] 1. Table 2 is a 2D example rather than a 3D example. The 2D example has only one free inflow incident angle α, which is independent of the variable sideslip angle β. If it is a 3D example, there are two free inflow incident angles: the angle of attack α and the sideslip angle β.

[0136] 2. Generation of neural network training set: select the flight profile with an altitude of 15km (corresponding to a free flow static pressure of 12.1kPa) and sample in this plane. The values ​​of Mach number Ma are: 0.3, 0.4, 0.5, 0.6, 0.65, 0.7, 0.75, 0.8, a total of eight points; the values ​​of angle of attack α are: -4, -3, -2, 0, 2, 3, 4, 5, 6, 7, a total of ten points. Therefore, the training set contains a total of 80 points.

[0137] 3. The proposed system and method need to be applicable to the entire flight envelope, so the test set of the neural network needs to cover the entire flight envelope. In the measurement range of the atmospheric data system given in Table 2, Mach number 0.3-0.8, angle of attack -4-7deg, free flow static pressure 101325-2550Pa, 253 points were randomly selected in the 3D space composed of these three variables to form a test set for testing the neural network. It should be noted that the training set comes from a section of the above 3D space, and the training set is distributed in the entire 3D space.

[0138] 4. Many points of pressure information in this embodiment are not shown in Table 2, but are represented by FIG. 5 .

[0139] In order to verify that the FADS neural network algorithm proposed in the embodiment of the present application can decouple the static pressure of the incoming flow, the training samples of the method provided in the embodiment of the present application are samples generated under a single static pressure of the incoming flow, and the test samples cover the test range of the atmospheric data system. As shown in Table 3, the training samples are generated under the incoming flow state at an altitude of 15 km, and the test samples randomly generate 253 sample points within the measurement range of Table 2.

[0140] Computational Fluid Dynamics (CFD) numerical simulation methods can be used to solve Navier-Stokes to obtain the pressure distribution on the airfoil surface for subsequent establishment and testing of the FADS system.

[0141] According to Table 3, the pressure distribution on the airfoil surface under various working conditions is obtained by CFD. Through data extraction and organization, a standard database for training and testing the FADS neural network algorithm is established. Taking pressure measuring holes 1 and 2 as an example, Figure 5a The pressure values ​​at the pressure measuring hole 1 under different working conditions are given. Figure 5bThe pressure values ​​at the pressure measuring hole 2 under different working conditions are given. Figure 5c An output parameter Cp12 of the infinite input-output neural network is given, where the grid plane represents the training data and the scattered points represent the measured data. As shown in Table 4, the input of the traditional FADS neural network algorithm is a specific pressure value. Figure 5a and Figure 5b The pressure value of two specific points is given, and the output is the Mach number Ma , angle of attack α , dynamic pressure q ∞ , static pressure p ∞ The input of the dimensionless input-output neural network (referring to the first neural network) is a series of pressure ratios [Cp 12 ,Cp 21 ,…,Cp ij ], Figure 5a Figure 5b shows the pressure ratio Cp of two specific points 12 , the output is the Mach number Ma , angle of attack α .

[0142] Table 3 CFD calculation conditions

[0143] Working conditions Training sample set Test sample set (randomly selected) Static pressure(pa) 12111 2550-101325 Mach number 0.3、0.4、0.5、0.6、0.65、0.7、0.75 0.3-0.8 Angle of attack (deg) -4、-3、-2、0、2、3、4、5、6、7 -4-7

[0144] Table 4 Input and output comparison of different neural network algorithms

[0145]

[0146] The obtained training samples are used to train the first neural network model, and the test sample set is used to test the performance of the first neural network model. Figure 6(a) shows the absolute error of the Mach number on the test set. It can be seen that the absolute error of all samples is less than 0.02, and the absolute error of most samples is less than 0.15. Figure 6(b) shows the absolute error of the angle of attack on the test set. It can be seen that the absolute error of the angle of attack is less than 0.5 degrees.

[0147] After calculating the pressure coefficient of each pressure measuring hole through the second neural network model, the dynamic pressure and static pressure are solved by the least squares method in combination with the pressure distribution measured by the pressure measuring hole. Figure 7 (a, b) shows the absolute errors of the dynamic pressure and static pressure on the test set. It can be seen from Figure 7 (a, b) that the absolute errors of the dynamic pressure and static pressure are both less than 500Pa.

[0148] It can be seen that by applying the method provided in the embodiment of the present application, the total number of sample points can be controlled at the order of hundreds (about 80, about 400 after considering the sideslip angle), the Mach number error is less than 0.1, the accuracy of the angle of attack and the sideslip angle is less than 0.3deg, and the absolute error of the static pressure and the dynamic pressure is less than 800Pa.

[0149] Comparative Example 1

[0150] Reference 1 "Air Data Sensing from Surface Pressure Measurements Using a Neural Network Method_rohloff1998" is an early application of the FADS neural network algorithm, which has been used on the US F-18B SRA verification aircraft. The system is used to solve dynamic pressure and static pressure. According to the document, the total number of sample points is about 40,000, the dynamic pressure accuracy is less than 575Pa (12psf), and the static pressure accuracy is less than 800Pa (17psf).

[0151] Comparative Example 2

[0152] Reference 2 "Real-time solution design and verification of FADS system for supersonic aircraft_Chen Guangqiang" consists of FADS solution algorithm and solver based on BP neural network technology. 4000 sample points are randomly generated within the measurement range. By extracting and sorting the pressure values ​​of the measuring point array from the flow field data, establishing a data table of the sample point flow state and the pressure value of the measuring point array and training to obtain the model, the FADS solution model based on BP neural network technology can be established. Reference 3 "Design of FADS system for air-breathing air-to-air missile_Chen Guangqiang" gives a case of FADS neural network algorithm applied to air-breathing air-to-air missiles. Through the nonlinear mapping relationship based on BP neural network, 4000 sample points are randomly generated within the measurement range for training neural network. According to the descriptions in References 2 and 3, the number of sample points is 4000, the Mach number accuracy is less than 0.1, the angle of attack and sideslip angle accuracy is less than 1deg, and the absolute error of static pressure measurement is less than 970Pa.

[0153] It can be seen from Example 1 and Comparative Examples 1-2 that, under the premise of ensuring that the measurement errors of all variables meet the design index requirements, the method provided in the embodiment of the present application only requires about 400 training samples during training, while the method in Comparative Example 1 requires about 40,000, and the method in Comparative Example 2 requires about 4,000. It can be seen that a set of dimensionless pressure parameters based on dimensional analysis provided in the embodiment of the present application is used to construct a FADS neural network, decouple the pressure information in the incoming flow, so that the FADS neural network algorithm can be constructed using fewer samples. Since the number of required training samples is reduced, the construction cost of the FADS system is greatly reduced, and the robustness of the FADS system is improved.

[0154] See also Figure 8 , which corresponds to an atmospheric data solving method for an embedded atmospheric data system provided in an embodiment of the present application, such as Figure 8As shown, the embodiment of the present application also provides an atmospheric data solution system for an embedded atmospheric data system, which may specifically include:

[0155] The pressure value collecting unit 201 is used to determine the distribution of each pressure measuring point on the aircraft surface and collect the pressure value of each pressure measuring point under the corresponding incoming flow condition;

[0156] A dimensionless processing unit 202, used for performing dimensionless processing on the pressure values ​​of each pressure measuring point to obtain an input parameter set;

[0157] A dimensionless parameter solving unit 203, used to solve each dimensionless atmospheric parameter included in the atmospheric data parameter using a neural network by using the input parameter set;

[0158] The dimensioned parameter calculation unit 204 is used to obtain the pressure coefficient of each pressure measuring point using the dimensionless atmospheric parameter, and calculate each dimensioned atmospheric parameter included in the atmospheric data parameter using the pressure coefficient of each pressure measuring point and the pressure value of each pressure measuring point.

[0159] Furthermore, the dimensionless parameter solving unit includes:

[0160] A first calculation subunit is used to input the input parameter set as input data into a first neural network model, so that the first neural network model outputs each dimensionless atmospheric parameter included in the atmospheric data parameter;

[0161] The dimensioned parameter calculation unit comprises:

[0162] The second calculation subunit is used to input the dimensionless atmospheric parameters as input data into the second neural network model, so that the second neural network model outputs the pressure coefficient of each pressure measuring point.

[0163] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0164] It can be known from the description of the above implementation methods that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general hardware platform. Based on such an understanding, the technical solution of the present application can be essentially or partly contributed to the prior art in the form of a software product, which can be stored in a storage medium such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments of the present application or certain parts of the embodiments.

[0165] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without creative work.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A FADS solution method based on dimensionless input-output neural network, characterized in that: The method comprises the following steps: Step 1: determine the distribution of each pressure measuring point on the aircraft surface, and collect the pressure values ​​of each pressure measuring point under the corresponding incoming flow condition; Step 2: Dimensionally transform the pressure values ​​of each pressure measuring point to obtain an input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ]; The dimensionless processing includes: respectively calculating the pressure ratio of the pressure value of each pressure measuring point to the pressure value of each pressure measuring point except itself , Thus, the input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ]; Step 3: Use the first neural network to solve the dimensionless atmospheric parameters α, β and Ma contained in the atmospheric data parameters: After dimensionless processing, the input parameter set [Cp 12 ,Cp 21 ,…,Cp ij ] as the input parameter of the first neural network, and the dimensionless atmospheric parameters including the angle of attack α, the sideslip angle β and the Mach number Ma as the output parameters of the first neural network. ij It is the ratio of the pressure value of each pressure measuring point to the pressure of every pressure measuring point except itself; Step 4: Use the second neural network to calculate the pressure coefficient Cp of each pressure measuring point i Solve; use the output parameters α, β, Ma of the first neural network as the input parameters of the second neural network, and the pressure coefficient Cp of each pressure measuring point i As the output parameter of the second neural network, the Cp i is the pressure coefficient at each pressure measuring hole [Cp1, Cp2, …, Cp n ]; Step 5: Solve the pressure coefficients of each pressure measuring point [Cp1, Cp2, ..., C..., Cp n ], combined with the corresponding measured pressure distribution, the dynamic pressure q of each dimensioned atmospheric parameter contained in the atmospheric data parameters is calculated. c , static pressure p ∞ The solution is obtained by using the least square method based on the pressure coefficient and the pressure value of multiple pressure measuring holes; The dimensionless input parameter Cp of the first neural network ij The functional relationship between the dimensionless output parameters α, β and Ma is: Assuming v1 = [α, β, Ma], we can inverse the function of formula (6) and obtain the functional relationship: v1 = Π1(Cp 12 ,Cp 21 ,…,Cp ij ) α, β, Ma are only dimensionless parameters [Cp 12 ,Cp 21 ,…,Cp ij ] is a function that is independent of the static pressure of the incoming flow, thereby decoupling the static pressure of the incoming flow; The specific process of setting this function is as follows: 1) Set a total of n pressure measuring holes; 2) Each pressure measuring hole is compared with the remaining pressure measuring holes except itself; 2) Construct n2(n-1) equations of formula (6); Where i≤n, j≤n, i≠j, when the pressure measuring hole distribution is determined, r i / l is a known parameter; The derivation of formula (6) is as follows: Assume that the pressure at a point i on the surface of the aircraft can be expressed as: p i =q c Cp i +p ∞ (1) where q c 、p ∞ is the dynamic pressure and static pressure of the incoming flow, Cp i is the pressure coefficient, which is a function of the relative position of the pressure measuring hole, the angle of attack α, the sideslip angle β and the Mach number Ma. Assuming that the shape of the aircraft is determined, Cp i It can be written as where r i =[x i ,y i ,z i ] represents the spatial position of the i-th pressure tap, l is the reference length of the aircraft; Assuming the pressure at the i-th and j-th pressure measuring holes, the pressure at the two points is compared to obtain the following formula: In the above formula, is a function of M∞, and the specific relationship is as follows: In the formula, we can get: Assuming there are n pressure measuring holes, each pressure measuring hole is ratioed with the remaining pressure measuring holes except itself, n·(n-1) equations can be constructed:

2. The FADS solution method based on dimensionless input-output neural network according to claim 1, characterized in that: The dimensionless input parameters α, β, Ma and the dimensionless output parameter Cp of the second neural network i The functional relationship is: v2=Π2(α,β,Ma), where v2=[Cp1,Cp2,...,Cp n ] (7).

3. The FADS solution method based on dimensionless input-output neural network according to claim 1, characterized in that: The method of solving the pressure coefficient and pressure value based on multiple pressure measuring holes by using the least square method includes the following process: 1) Construct the following superpositive definite equation: Solve α, β and Ma by formula (6). When the relative position of the pressure measuring hole is known by formula (2), Cp i As a function of α, β and Ma, the Cp at each pressure measuring hole under the corresponding flow condition can be determined. i , using the pressure coefficients at each pressure measuring hole and the known pressure distribution, the superpositive equation of formula (8) is constructed: Formula (8) is solved using the least squares method, and is defined as: Then q c 、p ∞ The expression is: At this point, the angle of attack α, sideslip angle β, Mach number Ma, and dynamic pressure q of the outflow can be solved. c , static pressure p ∞ .

4. A FADS solving system of a dimensionless input-output neural network used in a FADS solving method based on a dimensionless input-output neural network as claimed in any one of claims 1 to 3, characterized in that: The system comprises: A pressure value collection unit, used to determine the distribution of various pressure measuring points on the aircraft surface, and to collect the pressure values ​​of various pressure measuring points under the corresponding incoming flow conditions; A dimensionless processing unit, used for performing dimensionless processing on the pressure values ​​of each pressure measuring point to obtain an input parameter set; A dimensionless atmospheric parameter solving unit, used for solving each dimensionless atmospheric parameter included in the atmospheric data parameter by using the input parameter set and adopting a neural network; The dimensioned atmospheric parameter calculation unit is used to use the dimensionless atmospheric parameters to obtain the pressure coefficient of each pressure measuring point using a neural network, and to calculate each dimensioned atmospheric parameter included in the atmospheric data parameter using the pressure coefficient of each pressure measuring point and the pressure value of each pressure measuring point.

5. The FADS solution system of dimensionless input-output neural network according to claim 4, characterized in that: The dimensionless parameter solving unit comprises: A first calculation subunit is used to input the input parameter set as input data into a first neural network model, so that the first neural network model outputs each dimensionless atmospheric parameter included in the atmospheric data parameter; The dimensioned parameter calculation unit comprises: The second calculation subunit is used to input the dimensionless atmospheric parameters as input data into the second neural network model, so that the second neural network model outputs the pressure coefficient of each pressure measuring point.

Images