Helicopter flight dynamics model parameter error correction method

Abstract

The invention discloses a helicopter flight dynamics model parameter error correction method comprising the following steps: S1, carrying out parameterization processing on helicopter pneumatic data in a discrete form to obtain a plurality of pneumatic parameters, and forming a flight dynamics parameterization model; s2, performing balancing error analysis on the parameterized model, determining each main input parameter through sensitivity analysis, and representing a model balancing error by adopting a balancing error value matrix and a balancing error-forward flight speed fitting coefficient matrix; and S3, taking the balancing error value matrix and the balancing error fitting coefficient matrix as input of a neural network, and identifying each error. According to the method, parameterization processing is carried out on the discrete aerodynamic data to obtain a flight dynamics parameterization model, the parameters are clearer, a training sample is calculated and generated by adopting an input parameter error cross combination mode, the error of each main input parameter is identified by using the neural network, the calculation cost is effectively reduced, and the correction efficiency and accuracy are improved.

Description

Technical Field

[0001] The invention relates to the technical field of helicopter flight mechanics, and in particular to a method for correcting parameter errors of a helicopter flight dynamics model. Background Art

[0002] The helicopter flight dynamics model is the basis for flight characteristics analysis and flight quality evaluation in the helicopter design phase, and is the key to the overall layout design and flight control system design of the helicopter. Improving the accuracy of flight dynamics modeling is of great significance for accurately evaluating the flight quality of helicopters and designing advanced flight control systems.

[0003] At present, the mainstream method of helicopter dynamics modeling is mechanism modeling, which has the advantages of clear model parameters, structure and physical meaning. However, in actual modeling, mechanism modeling also faces many difficulties. First of all, the rotor aerodynamic model is the core part of the helicopter flight dynamics model. Its aerodynamic characteristics and aerodynamic state are complex and difficult to calculate accurately. Secondly, it is also difficult to accurately calculate the aerodynamic interference between the various components of the helicopter. In addition, the motion coupling, inertial coupling and operational coupling of the helicopter will bring challenges to the accurate establishment of the helicopter flight dynamics model. Under the combined influence of the above factors, the flight dynamics model established by the mechanism modeling method is not accurate. To further improve the accuracy of flight dynamics modeling, it is necessary to verify and correct the parameter errors in the model.

[0004] The parameter errors in the helicopter flight dynamics model can be mainly divided into three aspects: First, the assumptions and simplifications used in the modeling process cause certain errors in the model itself, that is, structural errors, such as quasi-steady assumptions, rigid blade assumptions, etc. Second, the errors caused by the errors between the model input parameters and the true values, that is, parameter errors, such as inaccurate or even unknown values ​​of some design parameters during modeling, and inaccurate aerodynamic data obtained from wind tunnel tests or CFD calculations. Third, there are measurement errors between the flight test data used to verify the accuracy of the model and the actual flight characteristics of the helicopter.

[0005] The main method currently used to correct parameter errors in helicopter flight dynamics models is to use test flight data as a benchmark, determine the source of errors through the equilibrium equations derived from modeling and prior experience in model correction, and then manually correct the flight dynamics model to improve model accuracy. However, this method is inefficient, and when the model error is large and the error manifestation is complex, it is difficult to achieve an ideal correction effect. Therefore, an efficient and accurate method is needed for correcting helicopter flight dynamics models. Summary of the invention

[0006] In view of the above problems, the purpose of the present invention is to propose a method for correcting parameter errors of a helicopter flight dynamics model, parameterize discrete aerodynamic data, obtain a flight dynamics parameterized model, and make the model parameters and their meanings clearer; at the same time, the main parameters affecting the model balancing error are determined by sensitivity analysis, which effectively reduces the difficulty of parameter error correction; the model balancing error is characterized by a balancing error value matrix and a balancing error fitting coefficient matrix, and the important information in the flight dynamics model balancing error is accurately extracted; in addition, a neural network algorithm is used to identify the model input parameter error, which can efficiently and accurately identify the input parameter error and improve the efficiency and accuracy of model correction.

[0007] This is achieved through the following technical solutions: A helicopter flight dynamics model parameter error correction method is provided, wherein the method corrects the model error caused by multiple input parameters with errors, wherein the multiple input parameters include existing parameters of the model and multiple aerodynamic parameters obtained by parameterizing aerodynamic data, and the method comprises the following steps: S1. Parameterizing aerodynamic data of a plurality of data points in a discrete form to obtain a plurality of aerodynamic parameters and form a helicopter flight dynamics parameterized model; wherein the aerodynamic data includes aerodynamic force data and aerodynamic interference data; S2, performing a trim error analysis on the helicopter flight dynamics parameterized model in step S1, firstly determining each main input parameter that affects the model trim error by a sensitivity analysis method, and then characterizing the model trim error by using a trim error value matrix and a trim error fitting coefficient matrix with respect to forward flight speed; S3, taking the trim error value matrix and the trim error fitting coefficient matrix about the forward flight speed in step S2 as the input of the neural network, training the neural network to identify the error of each main input parameter, and then correcting each main input parameter with error according to each error.

[0008] A large number of discrete data points are parameterized so that the helicopter flight dynamics model can better capture the actual physical phenomena and make the subsequent analysis more intuitive and efficient. At the same time, the trim error of the model is characterized by the trim error value matrix and the trim error fitting coefficient matrix about the forward flight speed, which can effectively obtain the error of the helicopter flight dynamics model at any forward flight speed, making the subsequent correction effect more comprehensive. In addition, neural network training can accurately identify the error of each main input parameter and make accurate corrections.

[0009] Preferably, in step S1, the aerodynamic data includes the airfoil aerodynamic coefficient and each aerodynamic coefficient and each moment coefficient of each component, and the aerodynamic interference data includes the aerodynamic interference factor of the rotor to each component, the aerodynamic interference factor of the fuselage to the tail surface, and the dynamic pressure loss coefficient of the tail surface and the tail rotor. By including the airfoil aerodynamic coefficient and each aerodynamic coefficient and each moment coefficient of each component, the aerodynamic characteristics of each component of the helicopter can be described more accurately, and the introduction of the aerodynamic interference data takes into account the mutual interference between different components.

[0010] Preferably, before performing parameterization processing, according to the variation law of the aerodynamic data with respect to the angle of attack or the sideslip angle, within a small angle range, the data of each relevant data point with respect to the angle of attack or the sideslip angle in the aerodynamic data is fitted using a polynomial function of the highest second order to obtain each corresponding aerodynamic parameter; for the data of each relevant data point with respect to the angle of attack or the sideslip angle in the aerodynamic data, within a small angle of attack or the sideslip angle, the formula Processing is performed, wherein n is a non-negative integer, x is the angle of attack or sideslip angle corresponding to the data of each relevant data point above, is the error between the function fitting value and the corresponding arbitrary aerodynamic data before fitting, They are the corresponding aerodynamic parameters obtained by parameterizing the aerodynamic data. The angle of attack and sideslip angle are the influencing factors that cannot be ignored in the actual operation of the helicopter flight dynamics model. At the same time, fitting processing within a small angle range can effectively ensure the efficiency and effectiveness of the model calculation.

[0011] Preferably, when performing parameterization processing, the data of multiple data points in the aerodynamic interference data corresponding to each component of the rotor are first parameterized, and the parameterization processing is the starting point, turning point and end point of the interference factor curve corresponding to each component, and the wake tilt angle and interference factor value corresponding to the end point. Each component includes a fuselage, a tail surface and a tail rotor, and then the above points are used to interpolate the wake tilt angle to obtain the aerodynamic interference factor of the rotor to each component. The method of capturing the key characteristic points of the interference factor curve is adopted, and the wake tilt angle and interference factor value at the three key positions of the curve starting point, turning point and end point are parameterized to ensure that the model can capture the main change trends and characteristics of the aerodynamic interference; at the same time, it provides convenience for the setting of aerodynamic parameter errors, and when the aerodynamic parameters corresponding to the above key characteristic points are set, the changes in the aerodynamic interference factor can be reasonably and accurately estimated and described.

[0012] Preferably, in step S2, the sensitivity analysis method is a Monte Carlo integral improved Sobol method. When the Monte Carlo integral improved Sobol method is adopted, multiple input parameters are sampled by quasi-Monte Carlo to obtain corresponding input sample matrices A and B. , , where N is the number of samples and P is the number of sampled input parameters; replace the i-th column of the sample matrix A with the i-th column of the sample matrix B to obtain the matrix ; Substitute the matrix ABi into the parameterized model to obtain f(A), f(B) and f(ABi); The sensitivity calculation formula for the i-th input parameter is , V is the total variance of the output f(A) and f(B) corresponding to the sample matrix A and the sample matrix B. Compared with the ordinary Monte Carlo sampling method, the Monte Carlo integration improved Sobal method can further improve the accuracy of the results and improve the accuracy of error analysis.

[0013] Preferably, when performing the balancing error analysis in step S2, the balancing result when each main input parameter is an initial value is used as a benchmark, and then the benchmark is subtracted from the balancing result when the input parameter has an error to obtain a model balancing error; the model balancing error includes multiple balancing quantities with errors, and the multiple balancing quantities include rotor collective pitch, lateral cyclic pitch change, longitudinal cyclic pitch change, tail rotor collective pitch, roll attitude angle, pitch attitude angle and rotor required power, and the error of each balancing quantity together constitutes a balancing error value matrix.

[0014] Preferably, when setting the error of each main input parameter, each error is set by using an equal percentage value relative to the reference value of each main input parameter change, and the expression for setting any error is: , is the value when there is an error in any main parameter, is the initial value of any main parameter, is the reference value of change, Each error is set as a percentage relative to the reference value of each major input parameter change, which facilitates active adjustment of the percentage The size of the error makes error analysis easier and the errors of the main input parameters of different magnitudes and units can be described under a common standard, which makes the errors of large and small parameters be understood and handled in the same way.

[0015] Preferably, when performing the fitting process, a cubic polynomial is used to fit the trim error values ​​in the trim error value matrix with respect to the forward flight speed to obtain a trim error with respect to the flight speed fitting coefficient matrix. Using a cubic polynomial for fitting can effectively obtain the error of the helicopter flight dynamics model at any speed, and at the same time, the cubic polynomial can effectively express the relationship between the trim error and the forward flight speed.

[0016] Preferably, the neural network used in step S3 is a deep belief network or a BP neural network. The deep belief network has good generalization ability, that is, it can improve the accuracy of recognition errors, and the BP neural network is convenient for adjusting the network structure according to the actual error situation, thereby improving the correction effect.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The technical solution of the present invention performs parameterized processing on aerodynamic data in discrete form to obtain a parametric model of flight dynamics, making the model parameters and their meanings clearer; at the same time, the main parameters affecting the model balancing error are determined by a sensitivity analysis method, effectively reducing the difficulty of correcting the parameter error; the model balancing error is characterized by a balancing error value matrix and a balancing error fitting coefficient matrix, and accurate extraction of important information in the flight dynamics model balancing error is achieved; in addition, a neural network algorithm is used to identify the model input parameter error, which can efficiently and accurately identify the input parameter error and improve the efficiency and accuracy of model correction. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of a helicopter flight dynamics model parameter error correction method; Figure 2 This is a comparison chart of the calculation results of the before and after parameterization of a helicopter flight dynamics model and the flight test data; Figure 3 This is a comparison chart of the trim calculation results and flight test data of a helicopter flight dynamics model before and after parameter error correction; Figure 4 is a schematic diagram of a balancing error value matrix; Figure 5 A schematic diagram of the trim error fitting coefficient matrix with respect to forward flight speed. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] like Figure 1 As shown, it is a flow chart of a helicopter flight dynamics model parameter error correction method, which corrects the model error caused by multiple input parameters with errors, and the multiple input parameters include existing parameters of the model and multiple aerodynamic parameters obtained by parameterizing aerodynamic data, and the existing parameters include data such as total mass and fuselage length. The method specifically includes the following steps: S1. Parameterize the aerodynamic data of multiple data points in discrete form to obtain multiple aerodynamic parameters to form a helicopter flight dynamics parameterized model; wherein the aerodynamic data includes aerodynamic force data and aerodynamic interference data. Multiple aerodynamic parameters are important parameters of the helicopter flight dynamics model during operation. Parameterizing multiple discrete data points not only reduces the amount of calculation, but also facilitates analysis and processing, and ensures the accuracy of the calculation.

[0021] In this embodiment, in step S1, the aerodynamic data includes the aerodynamic coefficient of the airfoil and each aerodynamic coefficient and each moment coefficient of each component, and the aerodynamic interference data includes the aerodynamic interference factor of the rotor on each component, the aerodynamic interference factor of the fuselage on the tail surface, and the dynamic pressure loss coefficient of the tail surface and the tail rotor. The aerodynamic coefficient of the airfoil and each aerodynamic coefficient and each moment coefficient of each component are important coefficients that affect the simulation calculation of the helicopter flight dynamics model. Therefore, parameterizing these coefficients and obtaining each corresponding aerodynamic parameter can more accurately describe the aerodynamic characteristics of each component of the helicopter. The introduction of aerodynamic interference data takes into account the mutual interference between different components, and is also an important coefficient that affects the simulation calculation of the helicopter flight dynamics model.

[0022] In this embodiment, before performing parameterization processing, according to the variation law of the aerodynamic data with respect to the angle of attack or the sideslip angle, the data of each relevant data point of the aerodynamic data with respect to the angle of attack or the sideslip angle can be fitted using a polynomial function of the highest second order within a small angle range to obtain each corresponding aerodynamic parameter; for the data of each relevant data point of the aerodynamic data with respect to the angle of attack or the sideslip angle, the corresponding aerodynamic parameter can be obtained according to the formula within a small angle of attack or the sideslip angle. Processing is performed, where f(x) is the fitting function expression, n is a non-negative integer, and x is the angle of attack or sideslip angle corresponding to the data of each relevant data point mentioned above. is the error between the function fitting value and the corresponding arbitrary aerodynamic data before fitting, They are the corresponding aerodynamic parameters obtained by parameterizing the aerodynamic data. The angle of attack and sideslip angle are the influencing factors that cannot be ignored in the actual operation of the helicopter flight dynamics model. At the same time, fitting processing within a small angle range can effectively ensure the efficiency and effectiveness of the model calculation.

[0023] It should be noted that the small angle range generally refers to the angle of attack or sideslip angle not exceeding ±10 degrees to ±15 degrees. Within this range, the aerodynamic characteristics usually exhibit relatively linear behavior, so a polynomial function of up to the second order can be used to approximate the actual data well.

[0024] In this embodiment, when performing parameterization processing, it is necessary to first perform parameterization processing on the data of multiple data points in the aerodynamic interference data corresponding to each component of the rotor, and the parameterization processing is the starting point, turning point and end point of the interference factor curve corresponding to each component, and the wake inclination angle and interference factor value corresponding to the end point. Each component includes the fuselage, tail surface and tail rotor, and then the above points are interpolated through the wake inclination angle, and the data between the above three points are linearly interpolated to form an interference factor curve, so that the aerodynamic interference factor of the rotor to each component can be obtained. At the same time, the parameterization method of the aerodynamic interference factor of the fuselage to the tail surface is the same as the parameterization method of the aerodynamic interference factor of the rotor to each component, the difference is that the name of the object is different; the parameterization method of the dynamic pressure loss coefficient of the tail surface and the tail rotor is the same as the parameterization method of the aerodynamic interference factor of the rotor to each component, the difference is that the name of the object is different. The key characteristic points of the interference factor curve are captured, and the wake inclination angle and interference factor value at the three key positions of the curve, namely the starting point, turning point and ending point, are parametrically described to ensure that the model can capture the main changing trends and characteristics of the aerodynamic interference. At the same time, it facilitates the setting of aerodynamic parameter errors, and when the errors of the aerodynamic parameters corresponding to the above-mentioned key characteristic points are set, the changes of the aerodynamic interference factor can be reasonably and accurately estimated and described.

[0025] It should be noted that the key characteristic points obtained after the aerodynamic interference data is parameterized are not as many as the original discrete data points, nor are they processed continuously. The main purpose is to reduce the number of parameters on the basis of retaining the main change trends and characteristics of the aerodynamic interference, so as to facilitate the error setting of related parameters; and to reasonably and accurately describe the changes in the aerodynamic interference factors when there are errors in the parameters.

[0026] S2. Perform trim error analysis on the helicopter flight dynamics parameterized model in step S1. First, determine each main input parameter that affects the model trim error through sensitivity analysis, and then characterize the model trim error using a trim error value matrix and a trim error fitting coefficient matrix with respect to forward flight speed.

[0027] In this embodiment, in step S2, the sensitivity analysis method is the Monte Carlo integral improved Sobol method. When the Monte Carlo integral improved Sobol method is used, multiple input parameters are sampled by quasi-Monte Carlo to obtain corresponding input sample matrices A and B. , , where N is the number of samples and P is the number of sampled input parameters; replace the i-th column of the sample matrix A with the i-th column of the sample matrix B to obtain the matrix ; The matrix AB i Substitute into the parameterized model and obtain f(A), f(B) and f(AB i ); The sensitivity calculation formula for the i-th input parameter is , V is the total variance of the output f(A) and f(B) corresponding to the sample matrix A and the sample matrix B. Compared with the ordinary Monte Carlo sampling method, the Monte Carlo integration improved Sobol method can further improve the accuracy of the results and improve the accuracy of error analysis.

[0028] In this embodiment, when performing the balancing error analysis in step S2, the balancing result when each main input parameter is an initial value is used as a benchmark, and then the benchmark is subtracted from the balancing result when there is an error in the input parameter to obtain a model balancing error; the model balancing error includes multiple balancing quantities with errors, and the multiple balancing quantities include rotor collective pitch, lateral cyclic pitch change, longitudinal cyclic pitch change, tail rotor collective pitch, roll attitude angle, pitch attitude angle and rotor required power, and the error of each balancing quantity together constitutes a balancing error value matrix.

[0029] like Figure 2 The figure shows a comparison of the trim calculation results of the helicopter flight dynamics model before and after parameterization and the flight test data. The comparison is made from the rotor collective pitch, lateral cyclic pitch change, longitudinal cyclic pitch change, tail rotor collective pitch, roll angle, pitch angle and rotor required power, verifying the accuracy of the parameterization processing. The AEFA test flight data is the flight test data. Before and after parameterization, the trim calculation of the original model and the parameterized model needs to be kept consistent to ensure the accuracy of subsequent calculations.

[0030] S3, taking the trim error value matrix and the trim error fitting coefficient matrix about the forward flight speed in step S2 as the input of the neural network, training the neural network to identify the error of each main input parameter, and then correcting each main input parameter with error according to each error.

[0031] In this embodiment, when setting the error of each main input parameter, each error is set by using an equal percentage value relative to the reference value of each main input parameter change. The expression for setting any error is: , is the value when there is an error in any main parameter, is the initial value of any main parameter, is the reference value of change, Each error is set as a percentage relative to the reference value of each major input parameter change, which facilitates active adjustment of the percentage The size of the error makes error analysis easier and the errors of the main input parameters of different magnitudes and units can be described under a common standard, that is, they can be calculated on the same scale. This makes it possible for the errors of any main input parameters, whether large or small, to be understood and processed in the same way, and avoids calculation errors caused by different dimensions of the values.

[0032] In this embodiment, when performing fitting processing, a cubic polynomial is used to fit the trim error values ​​in the trim error value matrix with respect to the forward flight speed, and a trim error with respect to the flight speed fitting coefficient matrix is ​​obtained. Using a cubic polynomial for fitting can effectively obtain the error of the helicopter flight dynamics model at any speed, and at the same time, the cubic polynomial can effectively express the relationship between the trim error and the forward flight speed.

[0033] like Figure 4 As shown in Figure 1, it is a schematic diagram of a balancing error value matrix, which contains data from e11 to e713, a total of 7 rows and 13 columns, such as Figure 5 The figure is a schematic diagram of the trim error with respect to the forward flight speed fitting coefficient matrix. Figure 4 The result of fitting the trim error matrix with a cubic polynomial includes data from P11 to P78, which has 7 rows and 8 columns. Since a cubic polynomial is used for fitting, the order of each item in the trim error fitting coefficient matrix for forward flight speed ranges from the third-order item to the constant item. Figure 4 and Figure 5 As shown, in the trim error value matrix, each row is the trim error, which corresponds to the rotor collective pitch, lateral cyclic pitch, longitudinal cyclic pitch, tail rotor collective pitch, roll attitude angle, pitch attitude angle and rotor required power. The roll attitude angle is also called the roll angle, and the pitch attitude angle is also called the pitch angle. Each column corresponds to a different forward flight speed between 0m / s and 80m / s. After fitting each error data in the trim error value matrix with respect to the forward flight data, the fitted coefficient matrix can be obtained. In the coefficient matrix obtained after fitting, each row is respectively Figure 4 The trim error fitting coefficients corresponding to different trim errors in the . After forming the trim error matrix, the trim error value matrix and the forward flight speed are fitted using a cubic polynomial to obtain a coefficient matrix, which can effectively obtain the error of the helicopter flight dynamics model at any speed. At the same time, the cubic polynomial can effectively express the relationship between the trim error and the forward flight speed.

[0034] In this embodiment, the neural network used in step S3 has multiple options, such as a deep belief network or a BP neural network. The deep belief network has good generalization ability, that is, it can improve the accuracy of recognition errors, and the BP neural network is convenient for adjusting the network structure according to the actual error situation, thereby improving the correction effect. Taking the use of a deep belief network as an example, the deep belief network can use a restricted Boltzmann machine RBM as the network core, and use the gradient descent method to update the data in different layers of the network. The RBM network in the restricted Boltzmann machine RBM is a two-layer structure, the lower layer is the visible layer, and the upper layer is the hidden layer. Assume that the number of nodes in the visible layer of the RBM network is q, and the input is , the node offset is ; The number of hidden layer nodes is m, and the output is , the node offset is , the visible layer and the hidden layer are connected by a weight matrix The activation function of each layer of neurons in the RBM network uses the sigmoid function. First, define the energy function of the input and output of the RBM network as:

[0035] Then the joint probability distribution between the visible layer and the hidden layer is It can be expressed as: , in the above formula, is the partition function.

[0036] There is also pre-training in the RBM network. Pre-training is to use the output of the previous layer as the input of the next layer, and obtain a stable network structure through layer-by-layer training. The optimal parameters of the RBM network can be obtained by maximizing the log-likelihood function:

[0037] Then, by using the contrast divergence fast learning method for each training sample, the training speed is fast and the training efficiency is high. During training, the relevant parameters of the RBM network are updated as follows:

[0038] In the above formula, is the learning rate.

[0039] After training, the deep belief network is fine-tuned. The multi-layer feedforward network trained by error back propagation is used to continuously adjust the entire network through error signals to achieve the overall optimality of the network. The input is processed and calculated by each layer of RBM and back propagation layer to obtain the output. The error between the network output and the target output is back propagated layer by layer, and the parameters of each layer of the network are continuously fine-tuned and updated using the gradient descent method. The weight update is as follows: , in the above formula, is the gradient of the k-th layer neuron, is the output of the k-1th layer of neurons, and n is the number of training times.

[0040] Specific application example 1: In the process of error identification of the input parameters of a helicopter flight dynamics model, the main input parameters affecting the model balancing error determined through sensitivity analysis include: v1-longitudinal position of the center of gravity, v2-vertical tail installation angle, v3-flapping hinge offset, v4-advance control angle, v5-tail rotor induced speed correction coefficient, v6-rotor induced speed correction coefficient, v7-fuselage zero angle of attack drag coefficient, v8-horizontal tail lift line slope, v9-rotor airfoil zero lift drag coefficient, v10-rotor lift line slope and v11-rotor vertical interference with the horizontal tail.

[0041] The generation of training samples of the neural network is shown in Table 1 below: Table 1: Neural network training sample table

[0042] In Table 1, all parameters are less than 0%, corresponding to the case of negative error; all parameters are 0%, corresponding to the case of no error; all parameters are greater than 0%, corresponding to the case of positive error; when setting the percentage of change, it is necessary to comprehensively consider three different situations and conduct separate training for multiple groups, so that the final calculated results can correspond to more actual situations and improve accuracy.

[0043] In order to match the data format required for neural network training, the balanced error value matrix and the balanced error fitting coefficient matrix are transposed, and then connected in sequence to form a row matrix as the input of the neural network, and then the row matrix composed of the percentage change of each model input parameter is used as the network output. The dimensions of the two neural network input data are 7×13=91 and 7×8=56 respectively, and the dimensions of the network output data are both 11.

[0044] Secondly, in order to avoid affecting the prediction performance of the neural network, the sample data after structural transformation is dimensionally normalized so that the data of each dimension are in the same range. Dimensional normalization generally uses the maximum and minimum values ​​of each dimension of the sample data to normalize the original data to between [-1, 1] or [0, 1]. The relevant expressions are as follows:

[0045] The generated samples from group 1 to group 7 are used as training sets, and the trim error value and the trim error with respect to the speed fitting coefficient are used as the input of the neural network. The BP neural network and the deep belief network DBN are trained respectively. After the network training is completed, the 8th group of samples is used as the validation set. The maximum absolute error between the predicted value and the true value of the percentage change of each parameter is shown in Table 2 below: Table 2: Maximum absolute error between the predicted value and the true value of each parameter change percentage in the two networks

[0046] From the comparison of the results in Table 2 above, it can be seen that the maximum absolute error between the predicted value and the true value of the percentage change of each input parameter does not exceed 1% in general, indicating that the method of the present invention can accurately identify the error of the input parameter through the model balancing error.

[0047] Specific application example 2: Taking the poorly accurate BO105 helicopter flight dynamics model as an example, the above method is used to correct it. The specific correction process is as follows: First, it is assumed that the values ​​of a set of input parameters are accurate, and the balancing result of the model under this set of input parameters is used as the reference value. Then, the training sample generation method in Example 1 is adopted, and the percentage change of the input parameters is taken as ±10% and ±20% to calculate and generate 7 groups of 895 training samples for training the neural network. Finally, the error between the assumed reference value and the flight test data is input into the neural network, and the error of the input parameter is identified, thereby completing the correction of the flight dynamics model. The model correction parameters and the initial values ​​and corrected values ​​of each parameter are shown in Table 3 below: Table 3: Model correction parameters and initial and corrected values ​​of each parameter

[0048] like Figure 3 As shown, it is a comparison diagram of the trim calculation results of the helicopter flight dynamics model before and after the parameter error correction and the flight test data, that is, the comparison of the test flight data before and after the correction of the BO105 helicopter flight dynamics model in this embodiment based on the data in Table 3, which is used to verify the accuracy of the helicopter flight dynamics model after the error correction. As can be seen from the figure, the model trim result after the parameter error correction has a more consistent trend with the flight test data, and its accuracy has been significantly improved, which proves the feasibility of the helicopter flight dynamics model parameter error correction method of the present invention.

[0049] In summary, the present invention parameterizes a large number of discrete data points, so that the helicopter flight dynamics model can better capture the actual physical phenomena and make the subsequent analysis more intuitive and efficient; at the same time, the model balancing error is characterized by using a balancing error value matrix and a balancing error fitting coefficient matrix with respect to the forward flight speed, which can effectively obtain the error of the helicopter flight dynamics model at any forward flight speed, making the subsequent correction effect more comprehensive; in addition, a neural network is used for training, which can accurately identify the error of each main input parameter, and then make accurate corrections, which has significant progress.

[0050] The above embodiments are only for illustrating the technical idea of ​​the present invention, and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the present invention.

Claims

1. A helicopter flight dynamics model parameter error correction method, characterized in that: The method is to correct the model error caused by multiple input parameters with errors, the multiple input parameters include existing parameters of the model and multiple aerodynamic parameters obtained by parameterizing aerodynamic data, and the method includes the following steps: S1. Parameterizing aerodynamic data of a plurality of data points in a discrete form to obtain a plurality of aerodynamic parameters and form a helicopter flight dynamics parameterized model; wherein the aerodynamic data includes aerodynamic force data and aerodynamic interference data; S2, performing a trim error analysis on the helicopter flight dynamics parameterized model in step S1, firstly determining each main input parameter that affects the model trim error by a sensitivity analysis method, and then characterizing the model trim error by using a trim error value matrix and a trim error fitting coefficient matrix with respect to forward flight speed; S3, taking the trim error value matrix and the trim error fitting coefficient matrix about the forward flight speed in step S2 as the input of the neural network, training the neural network to identify the error of each main input parameter, and then correcting each main input parameter with error according to each error.

2. A helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: In step S1, the aerodynamic data includes the aerodynamic coefficient of the airfoil and each aerodynamic coefficient and each moment coefficient of each component, and the aerodynamic interference data includes the aerodynamic interference factor of the rotor on each component, the aerodynamic interference factor of the fuselage on the tail surface, and the dynamic pressure loss coefficient of the tail surface and the tail rotor.

3. A helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: Before parameterization, according to the variation law of the aerodynamic data on the angle of attack or the sideslip angle, the data of each relevant data point on the angle of attack or the sideslip angle in the aerodynamic data is fitted using the highest second-order polynomial function within a small angle range to obtain each corresponding aerodynamic parameter; For each relevant data point of the angle of attack or sideslip angle in the aerodynamic data, in the range of small angle of attack or sideslip angle, according to the formula Processing is performed, wherein n is a non-negative integer, x is the angle of attack or sideslip angle corresponding to the data of each relevant data point, is the error between the function fitting value and the corresponding arbitrary aerodynamic data before fitting, They are the corresponding aerodynamic parameters obtained by parameterizing the aerodynamic data.

4. A helicopter flight dynamics model parameter error correction method according to claim 2, characterized in that: When performing parameterization processing, firstly, the data of multiple data points in the aerodynamic interference data corresponding to each component of the rotor are parameterized, and the parameterization processing is the starting point, turning point and end point of the interference factor curve corresponding to each component, and the wake inclination angle and interference factor value corresponding to the end point. Each component includes a fuselage, a tail surface and a tail rotor, and then the above points are used to interpolate the wake inclination angle to obtain the aerodynamic interference factor of each component of the rotor; When performing parameterization processing, the parameterization method of the aerodynamic interference factor of the fuselage on the tail surface is the same as the parameterization method of the aerodynamic data; the parameterization method of the dynamic pressure loss coefficient of the tail surface and tail rotor is the same as the parameterization method of the aerodynamic interference factor of the rotor on each component.

5. The helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: In step S2, the sensitivity analysis method is the Monte Carlo integral improved Sobol method. When the Monte Carlo integral improved Sobol method is used, multiple input parameters are sampled by quasi-Monte Carlo to obtain corresponding input sample matrices A and B. , , where N is the number of samples and P is the number of sampled input parameters; replace the i-th column of the sample matrix A with the i-th column of the sample matrix B to obtain the matrix ; The matrix AB i Substitute into the parameterized model and obtain f(A), f(B) and f(AB i ); The sensitivity calculation formula for the i-th input parameter is , V is the total variance of the outputs f(A) and f(B) corresponding to the sample matrix A and the sample matrix B.

6. A helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: When performing the trim error analysis in step S2, the trim result when each main input parameter is an initial value is used as a benchmark, and then the benchmark is subtracted from the trim result when the input parameter has an error to obtain a model trim error; the model trim error includes multiple trim quantities with errors, and the multiple trim quantities include rotor collective pitch, lateral cyclic pitch change, longitudinal cyclic pitch change, tail rotor collective pitch, roll attitude angle, pitch attitude angle and rotor required power, and the error of each trim quantity together constitutes a trim error value matrix.

7. A helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: When setting the error of each main input parameter, use an equal percentage value relative to the reference value of each main input parameter to set each error. The expression for setting any error is: , is the value when there is an error in any of the main parameters, is the initial value of any main parameter, is the reference value of change, As a percentage.

8. A helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: When performing the fitting process, a cubic polynomial is used to fit the trim error values ​​in the trim error value matrix with respect to the forward flight speed, and a fitting coefficient matrix of the trim error with respect to the flight speed is obtained.

9. A helicopter flight dynamics model parameter error correction method according to claim 1, characterized in that: The neural network used in step S3 is a deep belief network or a BP neural network.

Images