An effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms

Abstract

The present invention belongs to the technical field of buffeting detection for aircraft model support systems, and discloses an effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms. First, according to Euler-Bernoulli theory, a buffeting estimation model of a force balance is established; then, the separation of variables method is used to solve the vibration dynamics equation and modal analysis is carried out; finally, based on the solutions of the obtained dynamics equation and the results of modal analysis, vibration decoupling and noise removal are carried out based on deep learning algorithms. Different from the finite element analysis method and the analytical calculation method, this method is based on deep learning algorithms and solves the vibration dynamics equation through modal analysis. Compared with the above two methods, it has the advantages of fast calculation speed, less consumption of computing resources, high calculation accuracy, and relatively simple solution process. The effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms proposed by this method is a relatively simple and practical method for general wind tunnel test scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of buffeting detection of aircraft model support systems, and particularly relates to an effective method for removing multi-source mixed noise in wind tunnel experiments based on a deep learning algorithm. Background Art

[0002] For high-performance aircraft such as supersonic fighter jets, hypersonic aircraft, and rockets, which are national strategic equipment. Such aircraft often fly at high speeds and with high maneuverability in a strong flow field environment, where the aerodynamic loads are complex and variable, and their stability and service safety are severely threatened. To ensure the stability of the aircraft and the reliability of various technical indicators, it is necessary to accurately measure multi-dimensional aerodynamic forces in a high-speed wind tunnel, and then accurately obtain the aerodynamic data of the aircraft. Taking the dynamic derivative test of a certain hypersonic aircraft in a pulse combustion wind tunnel as an example, it is necessary to complete the aerodynamic force measurement under the conditions of continuous multi-directional excitation to simulate the high maneuverability and disturbed conditions of the aircraft. The test wind speed reaches 5 to 11 times the speed of sound (Mach / Ma), the test time is only 200 ms, and the measurement error is <2%. The allowable installation space for the internal excitation device and the measurement balance in the large-scale wind tunnel model is only 150 mm × 50 mm × 50 mm. Especially under the condition of a large angle of attack, the turbulent flow field is complex, and the model support system generates strong vibration interference, resulting in an overload rate of the measurement system as high as 300%. Among them, under extreme test conditions such as a large angle of attack, the action of various complex flow fields such as wing airflow separation and pulsating airflow causes the wind tunnel model-support system to generate forced strong vibrations, and the test space is limited, and the test has high requirements for measurement dimensions and frequency responses. As a result, there are problems of multi-source noise mixing in the aerodynamic force measurement data in actual wind tunnel experiments and low accuracy of aerodynamic force data. The aerodynamic force measurement data is often accompanied by multi-source mixed noises such as additional jitter noise, vibration suppression residual vibration noise, and driving oscillation noise, and the signal-to-noise frequency overlaps and the time series are different. To solve the above problems, it is urgent to propose an effective method for removing multi-source mixed noise in wind tunnel experiments.

[0003] At present, the research on methods for removing multi-source mixed noise in wind tunnel experiments by domestic and foreign scholars mainly focuses on the finite element analysis method and the analytical calculation method. Among them, the advantage of the finite element analysis method is high calculation accuracy, accurate vibration mode analysis, and simple operation, but there are problems of long calculation time and being not conducive to the real-time monitoring of the experimental system and the application of noise removal; the advantage of the analytical calculation method is short calculation time, less consumption of computing resources, and being convenient for the implementation of the noise removal algorithm during the experiment, but there are defects of poor calculation accuracy, narrow applicability, and inaccurate vibration mode analysis.

[0004] The patent "A Finite Element Simulation Method for the External Field Distribution of Transformer Core Vibration Noise" by Li Yong et al., with the patent number CN202110819734.2, introduces a finite element simulation method for the external field distribution of transformer core vibration noise. This method is based on the Helmholtz wave equation and the Fourier transform method of electromagnetic fields. Using the COMSOL finite element analysis software, a multi-field coupling finite element analysis model of the magnetic field and solid field of the vibration noise of an oil-immersed transformer core is established and calculated, and the vibration simulation image of the object is obtained. This method provides a simple and effective vibration simulation modeling method, but its limitation is that it only analyzes the vibration of the target object and does not decouple the vibration modes, so it cannot meet the engineering requirements of analyzing and removing external mixed noise.

[0005] Ren Zhuoyu, in the article "Research on Load Identification Based on Wilson- Method and Trend Term Removal Technology under High-Level Noise Conditions", based on the Wilson- method, aiming at the problem of removing high-level Gaussian noise generated under dynamic loads, introduces the five-point cubic smoothing method, mathematically fits the load spectrum and the noise term, and decouples the loading response according to the kinematic equation, so as to realize the identification and removal of the noise term. This method has the advantages of accurate modeling for specific dynamic loads, high mathematical calculation accuracy, and shorter calculation time compared with the finite element method, but it has the limitation of a narrow applicable load spectrum and cannot adapt to complex load conditions in wind tunnel experiments.

[0006] Based on the problems existing in the above technologies, it is necessary to propose an effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms. While shortening the calculation time of the traditional finite element method, by virtue of the high adaptability and high precision of deep learning algorithms in the field of vibration noise detection and analysis, appropriate physical image modeling is carried out to solve the defects of poor calculation accuracy, narrow applicability, and inaccurate mode shape analysis of the analytical calculation method. Summary of the Invention

[0007] The purpose of the present invention is to provide an effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms, overcoming the deficiencies of the prior art to solve various problems existing in the existing methods.

[0008] Technical Solution of the Present Invention:

[0009] An effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms. First, according to Euler-Bernoulli theory, a buffeting estimation model of the force balance is established; secondly, the vibration dynamics equation is solved by the method of separation of variables and modal analysis is carried out; finally, based on the solution of the obtained dynamics equation and the modal analysis results, vibration decoupling and noise removal are carried out based on deep learning algorithms;

[0010] The specific steps are as follows:

[0011] Step 1: Establish a buffeting estimation model for the force balance.

[0012] Consider an experimental system composed of a force balance and an aircraft model. Its configuration is such that the force balance and the aircraft model are fixedly connected to both ends of the force balance strut. Take the center of the circle corresponding to the tail cross-section on the side where the force balance strut is fixedly connected to the force balance as the coordinate origin, and take its axis as the axis, with the direction pointing from the origin to the opposite direction of the force balance strut.

[0013] According to the theory of elasticity, the force balance strut is equivalent to an elastic body beam with a variable cross-section. The differential equation of the forced vibration motion of the elastic body beam with a variable cross-section is:

[0014]

[0015] In the formula: is the bending moment function, is the time term, is the density of the force balance strut, is the cross-sectional area function of the force balance, is the exciting force function, is the deflection function;

[0016] According to the Euler-Bernoulli theory, the differential equation of the deflection vibration response of the beam with a variable cross-section is obtained, which is the buffeting estimation model of the force balance:

[0017]

[0018] In the formula: is the Young's modulus of the force balance material, is the moment of inertia of the cross-section of the force balance;

[0019] So far, the buffeting estimation model of the force balance has been established;

[0020] Step 2: Solve the vibration dynamics equation and conduct modal analysis;

[0021] According to the separation of variables method for solving partial differential equations, assume

[0022]

[0023] In the formula: is the position-related function term in the deflection function, called the vibration mode function; is the time-related function term in the deflection function;

[0024] It is called is the nth-order natural vibration mode function, which satisfies the characteristic equation:

[0025]

[0026] By using the orthogonality of the mode shape functions, we obtain:

[0027]

[0028] In the formula: is the generalized force corresponding to the th mode shape,

[0029] By using the Duhamel integral, we get:

[0030]

[0031]

[0032] In the formula: is the modal amplitude corresponding to the th natural mode shape, and are respectively is the angular frequency corresponding to the th natural mode shape, is the integration parameter,

[0033] The natural mode shape function is called the

[0034] th modal shape function, and each vibration mode corresponds to a mode shape;

[0035] Step 3: Perform vibration decoupling and noise elimination based on the deep learning algorithm

[0036] As known from Equation (6), the solution of the mode generalized coordinate function is expressed by the

[0037]

[0038] th vibration mode and the corresponding integration constants. Therefore, based on the deep learning algorithm, according to the fitting function data of the acceleration measurement points in the experiment, the modal order to be solved is set to determine the integration constants. The relationship between the measured point acceleration and the modal amplitude function is: and respectively represent the acceleration at point and point represents the second derivative function of the modal amplitude obtained by fitting. And since The data of the acceleration measurement points for the experiment are used to fit the function, so it is marked as an estimator.

[0039] A GRU deep learning model is established, where the number of hidden layers of GRU is 2, the number of neurons is [3, 6], the number of hidden layers of the fully connected layer is 3, the number of neurons is [12, 6, 3], and the number of neurons in the output layer is 1; its estimated maximum error is 6.82×10-3g, the average value of the absolute error is 9.20×10-4g, and the root mean square error is 1.2×10-3g, where g represents the gravitational acceleration; estimate the buffeting acceleration at the force measuring balance, achieve vibration decoupling, and determine the numerical solutions of each mode in Equation (6) through the fitting of the measurement point data.

[0040] Due to the inertial force and gravity of the aircraft model, a support reaction force will be generated on the support rod of the force measuring balance, thereby generating a shear force in the lift direction on the force measuring balance; due to the non-coincidence of the center of mass and the measurement center of the aircraft model, a support moment will be generated under the influence of gravity; and a pitching inertial moment will be generated during the oscillation of the aircraft model. Based on Newton's theorem and the parallel axis theorem, list its functional form:

[0041]

[0042]

[0043] In the formula: and are the buffeting noise values of force and moment, is the gravity received by the aircraft model, is the angle between the aircraft model and the horizontal plane at the support point of the support rod, is the inertial acceleration of the aircraft model, is the angular acceleration of the aircraft model, L1 is the distance from the center of gravity of the aircraft model to the coordinate origin, L2 is the distance from the center of gravity of the aircraft model to the sensor, is the mass of the aircraft model;

[0044] Combining Equations (6), (7), (8), (9) with the fitting results of the measurement point data obtained by the GRU deep learning model, the numerical solutions of the two vibration noises can be obtained, and then decoupled from the experimental data to obtain more accurate experimental results.

[0045] So far, the vibration decoupling and noise elimination based on the deep learning algorithm have been completed.

[0046] The beneficial effects of the present invention are to propose an effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms. Different from the finite element analysis method and the analytical calculation method, this method is based on deep learning algorithms and solves the vibration dynamics equation through modal analysis. Compared with the above two methods, it has the advantages of fast calculation speed, less consumption of computing resources, high calculation accuracy, and relatively simple solution process. It provides an efficient method for removing external noise in vibration detection of wind tunnel vibration tests. At the same time, using deep learning algorithms to solve modal analysis problems has a certain theoretical foresight. In short, the effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms proposed by this method is a relatively simple and practical method for general wind tunnel test scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is a schematic diagram of an experimental system composed of a force balance and an aircraft model;

[0048] Figure 2 is a schematic diagram of finite element method dynamic simulation;

[0049] Figure 3 is a result comparison diagram of the method of the present invention. Among them, (a) is the sampling data diagram before noise removal, and (b) is the sampling data diagram after noise removal;

[0050] Figure 4 is a result comparison diagram of the method of the present invention. Among them, (a) is a schematic diagram of the reference data of the sampling point, and (b) is a schematic diagram of the relative error fitted by the deep learning algorithm;

[0051] Figure 5 is a flow chart of an effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms;

[0052] In the figure: 1 - force balance strut, 2 - aircraft model. DETAILED DESCRIPTION OF THE INVENTION

[0053] The following further elaborates on the specific implementation manners of the present invention in combination with the drawings and technical solutions.

[0054] Embodiment

[0055] Select a vibration system formed by combining a force balance and an aircraft model 2 for modeling and calculation. Among them, the force balance strut 1 is mainly solved and analyzed.

[0056] An effective method for removing multi-source mixed noise in wind tunnel experiments based on deep learning algorithms is as Figure 5 shown, and the specific calculation steps are as follows:

[0057] First step, establish a buffeting estimation model for the force balance

[0058] Set the strut 1 of the force measuring balance as a homogeneous non-uniform cross-section elastic rod, and take its density as , and take its Young's modulus as , and take its length as , and its physical properties conform to various assumptions of classical elasticity mechanics. Set the mass of the aircraft model 2 as , and the radius of the cross-sectional circle at the fixed connection position is First, the boundary conditions of the vibration differential equation can be obtained

[0059]

[0060]

[0061] Establish a buffeting estimation model of the force measuring balance according to Equation (2)

[0062]

[0063] Second, solve the vibration dynamics equation and conduct modal analysis

[0064] According to Equation (6), the solution result of the vibration dynamics equation is

[0065]

[0066] Select the modal order , to make theoretical preparations for the deep learning algorithm in the third step.

[0067] Third, perform vibration decoupling and noise elimination based on the deep learning algorithm

[0068] Establish a GRU deep learning model, where the number of GRU hidden layers is 2, the number of neurons is [3, 6], the hidden layer of the fully connected layer is 3 layers, the number of neurons is [12, 6, 3], and the number of neurons in the output layer is 1. Its estimated maximum error is 6.82×10-3g, the average absolute error is 9.20×10-4g, and the root mean square error is 1.2×10-3g. Through the fitting of the measured point data, determine the numerical solutions of each mode of the vibration dynamics equation in the second step, and the obtained results are as Figure 4 shown.

[0069] Substitute the solution result of the deep learning model into Equations (8)(9), the force noise and moment noise to be eliminated can be obtained. Subtract Equation (6) from Equations (8)(9) to obtain the sampled data after eliminating the vibration noise, and the obtained results are as Figure 3 shown.

[0070] So far, the calculation of this embodiment is completed.

[0071] The effective elimination method for multi-source mixed noise in wind tunnel experiments based on deep learning algorithms aims at the problems commonly existing in vibration wind tunnel tests, such as limited test space, high requirements for measurement dimension frequency response, multi-source noise mixing, and low accuracy of aerodynamic data. A solution method for effective elimination of multi-source mixed noise based on deep learning algorithms is proposed. Starting from the vibration dynamics equation, this method conducts vibration modal analysis and gives the modal solution of the vibration dynamics equation based on deep learning algorithms, thereby analyzing and eliminating the mixed noise. It has the advantages of fast calculation speed, high accuracy, simple operation, and clear vibration modal analysis.

[0072] In addition to solving the embodiments proposed in this patent specification, this method can also use the idea of proposing modal solutions based on deep learning algorithms using vibration dynamics equations in this method to solve other problems existing in wind tunnel experiments. It has a certain degree of flexibility compared with traditional methods and is a simple and efficient new method.

Claims

1. An effective method for eliminating multi-source mixed noise in wind tunnel experiments based on deep learning algorithm, characterized in that: Here are the steps: The first step is to establish a force balance buffeting estimation model; Consider an experimental system consisting of a force balance and an aircraft model. The two ends of the force balance support rod are fixedly connected to the force balance and the aircraft model respectively. The center of the circle corresponding to the tail cross section of the force balance support rod fixed to the force balance is taken as the origin of the coordinate system, and its axis is taken as Axis, the direction points from the origin to the opposite direction of the force balance support rod; According to the theory of elastic mechanics, the force balance support rod is equivalent to a variable cross-section elastic beam, and the differential equation of the forced vibration motion of the variable cross-section elastic beam is: Where: is the bending moment function, is the time term, is the density of the force balance rod, is the cross-sectional area function of the force balance, is the exciting force function, is the deflection function; According to the Euler-Bernoulli theory, the differential equation of the deflection vibration response of the variable cross-section beam is obtained, which is the buffeting estimation model of the force measuring balance: Where: is the Young's modulus of the force balance material, is the moment of inertia of the cross section of the force balance; At this point, the force balance buffeting estimation model has been established; The second step is to solve the vibration dynamics equation and perform modal analysis; According to the separation of variables method for solving partial differential equations, let Where: is the position-dependent function term in the deflection function, called the vibration mode function; is the time-dependent function term in the deflection function; say for The order natural vibration mode function satisfies the characteristic equation: Using the orthogonality of the vibration mode function, we get: Where: for The generalized force corresponding to the order vibration mode is, is the length of the variable cross-section elastic beam; Using Duhamel points, we get: Where: for The modal amplitude corresponding to the order natural vibration mode is, and They are The two constants under the order natural vibration mode are: for The angular frequency corresponding to the order natural vibration mode is, is the integral parameter, is a constant parameter; The natural vibration function called The modal shape function, each vibration mode corresponds to a vibration shape; At this point, the solution of the vibration dynamics equations and the modal analysis are completed; Step 3: Vibration decoupling and noise removal based on deep learning algorithm From formula (6), we know that the solution of the vibration mode generalized coordinate function is The order vibration mode and the corresponding integral constant are expressed. Therefore, based on the deep learning algorithm, according to the fitting function data of the acceleration measurement point in the experiment, the modal order to be determined is set to determine the integral constant. The relationship between the acceleration of the measurement point and the modal amplitude function is: Where: , Respectively Point and The acceleration at the point, represents the second-order derivative function of the fitted modal amplitude, and due to It is the fitting function of the experimental acceleration measurement point data, and is marked as an estimated quantity; A GRU deep learning model is established, where the number of hidden layers of GRU is 2, the number of neurons is [3, 6], the number of hidden layers of the fully connected layer is 3, the number of neurons is [12, 6, 3], and the number of neurons in the output layer is 1; its estimated maximum error is 6.82×10-3g, the average absolute error is 9.20×10-4g, and the root mean square error is 1.2×10-3g, where g represents the gravitational acceleration; the vibration acceleration at the force balance is estimated to achieve vibration decoupling, and the numerical solutions of each mode of equation (6) are determined by fitting the measurement point data; Due to the inertia and gravity of the aircraft model, a support reaction force will be generated on the force balance support rod, thereby generating a shear force in the lift direction on the force balance; since the center of mass of the aircraft model does not coincide with the measurement center, a support moment will be generated under the influence of gravity; and a pitch inertia moment will be generated during the oscillation of the aircraft model. Based on Newton's theorem and the parallel axis theorem, its function form is listed as follows: Where: , is the chattering noise value of force and torque, is the gravity acting on the aircraft model, is the angle between the aircraft model and the horizontal plane at the support point of the strut, is the inertial acceleration of the aircraft model, is the angular acceleration of the aircraft model, L1 is the distance from the center of gravity of the aircraft model to the origin of the coordinate system, L2 is the distance from the center of gravity of the aircraft model to the sensor, is the mass of the aircraft model; Combining equations (6), (7), (8), and (9) with the fitting results of the measurement point data obtained by the GRU deep learning model, the numerical solutions of the two vibration noises are obtained, which are then decoupled from the experimental data to obtain more accurate experimental results. At this point, vibration decoupling and noise removal based on deep learning algorithm are completed.

Images