A regression prediction method for flight attitude under complex wind fields
By constructing an unbalanced sample feature learning model and ARGAN network in complex wind fields, combined with the regression loss of distribution alignment, the quantitative prediction of the flight attitude of the wind field images is achieved, and the difficulty in evaluating the impact of the wind field on the flight attitude in the prior art is solved. The problem has improved the accuracy of flight safety warning.
Patent Information
- Application Number
- CN202311204487.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-09-18
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-09-18
AI Technical Summary
The prior art is difficult to quantitatively measure the impact of complex wind fields on flight attitudes, lacking effective flight attitude prediction algorithms, and it is difficult to provide pilots and aviation controllers with timely hazard assessment and early warning information.
A flight attitude regression prediction method under the action of complex wind fields was designed. By constructing an unbalanced sample feature learning model, using ARGAN network to generate simulated samples, combining the regression loss of distribution alignment, training discriminator and generator to realize wind field images quantitative prediction of flight attitude.
It improves the accuracy of flight attitude prediction, can better evaluate flight safety, provide auxiliary safety warnings, explore flight safety boundaries, and reminds pilots and aviation controllers to operate and command in advance to ensure flight safety.
Smart Images

Figure CN116994077B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of flight safety analysis, and in particular to a regression prediction method for flight attitude under the action of a complex wind field. Background Art
[0002] The wind field environment has an important impact on flight safety. When an aircraft encounters dangerous airflows such as shear wind during flight or landing, the original stable flight attitude will be rapidly changed, causing the aircraft to jolt, sway, or even crash, threatening flight safety. Therefore, studying and analyzing the impact of wind fields on flight attitude has important theoretical and practical value for flight safety.
[0003] Existing wind field analysis algorithms usually focus on the classification and identification of wind fields such as wind shear, turbulence and aircraft wake, and then provide early warnings of wind field types, but it is difficult to quantitatively measure their impact on flight attitude. At present, there is a lack of algorithms for predicting flight attitude based on wind field data, and related research and equipment are also difficult to provide pilots and air traffic controllers with timely hazard assessment and warning information. Summary of the invention
[0004] The purpose of the present invention is to overcome the deficiencies of the above-mentioned prior art and to provide a regression prediction method for flight attitude under the action of complex wind fields.
[0005] The technical solution provided by the present invention is as follows: a regression prediction method for flight attitude under the action of a complex wind field, characterized in that it specifically includes the following steps:
[0006] Step 1: Dataset construction
[0007] Use the collected wind field image real samples and the corresponding collected real pitch angles to form a data pair;
[0008] Step 2: Imbalanced Sample Feature Learning Model Construction
[0009] The model is trained using the set of real samples of wind field images and real pitch angles collected in step 1; an ARGAN network for regression tasks is designed to achieve regression prediction of pitch angles while generating simulated samples to augment image data; the network consists of two parts: a generator G and a discriminator D; the generator G module can generate simulated samples of wind field images based on the sampled simulated pitch angles, and the discriminator D can determine the authenticity of the input image while achieving regression prediction of the pitch angle parameter value.
[0010] Step 3: Regression loss design based on distribution alignment
[0011] The output of the pitch angle regression prediction branch of the discriminator D in step 2 is designed with an imbalanced regression loss based on distribution alignment; the real pitch angles collected by statistics and the sampled simulated pitch angles are used as prior distributions, and the generator G and the discriminator D are trained to align the distribution of the predicted pitch angles with the prior distributions to achieve attitude parameter regression;
[0012] Step 4: Model training
[0013] The models are trained respectively using the loss functions described in formulas (1), (2) and (3);
[0014] Step 5: Model testing
[0015] After completing the training of the model according to step 4, retain the discriminator D in the trained model of step 4, input the collected real samples into the discriminator D, and obtain the pitch angle prediction value output by the attitude parameter pitch angle regression prediction branch.
[0016] Furthermore, in the step 1, the data set is constructed using real samples x of wind field images collected by meteorological radar and airborne sensors. r and the corresponding acquired true pitch angle θ r , and form a data pair <X, Θ>; where X is the wind field image real sample x in the XY plane where the aircraft passes r A set of, Θ is the actual pitch angle θ r A collection of .
[0017] Furthermore, in the step 2 of constructing the unbalanced sample feature learning model:
[0018] The input of the generator G contains randomly generated Gaussian distributed random noise z G and the simulated pitch angle θ obtained by sampling from the Weibull distribution f The two parts are spliced into the input vector z, and then the generator G generates the wind field image simulation sample x f ;
[0019] The discriminator D comprises a true or false label discrimination branch and a posture parameter pitch angle regression prediction branch. The true or false label discrimination branch outputs the authenticity discrimination result of each image contained in the input data block as a real sample or a simulated sample. The posture parameter pitch angle regression prediction branch outputs the pitch angle value θ predicted for the input image. p ; The collected wind field image real sample x r And the generated wind field image simulation sample x f Obtain data block x after confusing the order mix And input the discriminator D to obtain the prediction results of the true and false label discrimination branch and the attitude parameter pitch angle regression prediction branch respectively;
[0020] For the true and false label discrimination branch, the discriminator D is calculated based on the binary cross entropy to identify the real sample x of the wind field image. r and wind field image simulation sample x f The classification loss L D , the formula is:
[0021]
[0022] Where m is the batch size of the input image, i = {1,...,m};
[0023] For the attitude parameter pitch angle regression prediction branch, the MSE-based regression loss is used to calculate the true pitch angle θ of the generator G and the discriminator D. r and the simulated pitch angle θ f The prediction loss is:
[0024]
[0025] Where m is the batch size of the input image, θ r , is the actual pitch angle θ corresponding to the real sample among the m samples r And the pitch angle prediction value of the attitude parameter pitch angle regression prediction branch for the real sample, θ f , is the pitch angle prediction value of the aircraft pitch angle and attitude parameter pitch angle regression prediction branch corresponding to the simulation sample among the m samples.
[0026] Furthermore, in the step 3, the regression loss design based on distribution alignment is:
[0027] The real sample x r and simulation sample x f After mixing the order, we get data block x mix , then sample m samples, use KL divergence to measure the difference between the distribution of the pitch angle predicted by the attitude parameter pitch angle regression prediction branch for the input image and the prior distribution, and then feedback and adjust the distribution and model parameters of the simulation samples; the designed distribution alignment loss L KL It can be calculated as follows:
[0028]
[0029] In formula (3), C is a constant, represents the i-th image among m input images; For images The probability of the corresponding pitch angle, For the model The probability of the predicted pitch angle; the former is obtained by the pitch angle θ statistics of the model input, and the latter is the pitch angle θ predicted by the attitude parameter pitch angle regression prediction branch p Statistics are obtained; the calculation method is as follows:
[0030] ① First, let’s compare θ and θ p Divide the same K intervals; divide the intervals where the pitch angles are located with intervals of t; assume that θ and θ p In the entire set of values of the two, the maximum and minimum values are θ max and θ min Indicates that, with t as the interval, [θ min ,θ max ] is divided into K intervals, and K is calculated as: That is, t divides θ and θ p Take the upper integer after the range;
[0031] ② Secondly, calculate θ and θ respectively p The frequency of appearing in the K intervals divided by ①; for the above K intervals A=[A1,...,A K ], the pitch angle value range of the jth (j∈{1,...,K}) interval can be expressed as A j =[θ min +(j-1)·t,θ min +j·t); according to each interval A j The pitch angle value range is calculated, and the number of samples whose pitch angle values fall in each interval is counted to obtain the number of times the pitch angle appears in each interval F = [f1, ..., f K ];
[0032] ③Third, respectively for θ and θ p Calculating Probability and
[0033] Calculate the probability for θ For the i-th sample Assume that its pitch angle θ i falls in the jth interval, i.e. θ i ∈A j , then in formula (3) It can be defined as:
[0034]
[0035] in, is a step function, which is defined as is defined as follows:
[0036]
[0037] Among them, f j Represents the pitch angle θ in interval A j The number of occurrences, f n represents the number of times the pitch angle θ appears in the nth interval, n = {1, ..., K}, and ρ is a scale hyperparameter;
[0038] For θ p Calculating Probability For the i-th sample Assume that the predicted pitch angle θ pi falls in the jth interval, i.e. θ pi ∈A j , then in formula (3) It can be defined as:
[0039]
[0040] in, is a step function, which is defined as is defined as follows:
[0041]
[0042] Among them, f j Represents the predicted pitch angle θ pi In interval A j The number of occurrences, f n Represents the predicted pitch angle θ pi The number of occurrences in the nth interval, n = {1, ..., K}, ρ is the scale hyperparameter.
[0043] Furthermore, in the step 4 model training:
[0044] The stochastic gradient descent method SGD is used as the optimizer, and the model is trained respectively using the loss functions described in formulas (1), (2) and (3). The specific steps are as follows:
[0045] ① According to step 2, randomly generated Gaussian distribution random noise z G and the simulated pitch angle θ obtained by sampling from the Weibull distribution f , obtain the input vector z, and use the generator G to generate a wind field image simulation sample x f ;
[0046] ② The real sample x r and simulation sample x f Obtain x after confusion mix , x mix Input the discriminator D, train the authenticity discrimination branch, calculate the authenticity discrimination loss according to formula (1), and update the discriminator D and generator G;
[0047] ③ x mix Input the discriminator D, train the attitude parameter pitch angle regression prediction branch, calculate the regression loss according to formula (2), and update the discriminator D and generator G;
[0048] ④ x mix After inputting into the discriminator D, the distribution alignment loss is calculated according to step 3 and formula (3) and then fed back to update the discriminator D and generator G.
[0049] The beneficial effects of the present invention are as follows: the present invention mainly studies the regression prediction of flight attitude under the action of complex wind fields. Since the wind force obeys the Weibull distribution, the greater the wind force, the smaller the sample size. Accordingly, since the wind force is the main factor affecting the attitude parameters, the present invention assumes that the change of attitude parameters also obeys the Weibull distribution. This research mainly faces two challenges: first, the sample size of complex wind fields is small, and the corresponding flight parameter data is small, which makes it difficult to meet the needs of safe flight research under boundary conditions; second, the number of samples in different wind fields is unbalanced, and the regression prediction model trained on unbalanced data will have learning bias, which affects the algorithm effect.
[0050] To address the problem of small number of complex wind field samples, the present invention designs a generative adversarial network ARGAN (Auxiliary Regression GAN) for regression tasks. While generating simulated samples to amplify complex wind field images, attitude parameter regression is achieved. The accuracy of flight attitude prediction is improved by simultaneously utilizing the collected real samples of wind field images and the generated simulated samples of wind field images. To address the problem of unbalanced wind field samples, the present invention combines the distribution characteristics of wind field data and designs an unbalanced regression loss based on distribution alignment to correct the distribution of generated simulated samples and enable the model to be more adaptable to unbalanced sample distributions. Ultimately, the goal of using wind field images to quantitatively predict flight attitude is achieved, thereby better assessing flight safety. This method can also be used in other image-oriented regression tasks.
[0051] Through step 2, an ARGAN model is constructed that can generate a specified wind field image simulation sample according to the pitch angle in the attitude parameter. This model can not only generate complex wind field images by sampling at the edge of the pitch angle distribution to improve the overfitting problem caused by the lack of real complex wind field image samples, but also predict the pitch angle parameters through the discriminator regression prediction branch training, so as to better learn the characteristics of wind fields of different levels from unbalanced training samples.
[0052] Through step 3, by introducing the distribution of the actual collected pitch angle values and the pitch angle values sampled from a given distribution as prior information to guide the model's learning of the overall data distribution, the distribution of the generated simulation samples predicted by the model and the collected real samples can be aligned with the distribution of the corresponding pitch angle labels, prompting the model to better fit the sample distribution, thereby better learning the decision boundary and reducing the impact of learning bias caused by sample imbalance on model learning.
[0053] Through the regression prediction method designed in steps 1-6, the present invention can realize the function of quantitatively predicting the change of flight attitude using the real samples of the collected wind field images, and then achieve the goal of providing auxiliary safety warning for flight. The present invention is of great significance for in-depth understanding of the impact of complex wind fields on flight attitude, exploring flight safety boundaries, and reminding pilots and air traffic controllers to operate and command in advance to ensure flight safety. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 It is the overall framework diagram of the present invention;
[0055] Figure 2 The figure is a corresponding relationship diagram between weather and wind field images; (a) is the downdraft weather from a side view, and (b) is the collected wind field image from a top view;
[0056] Figure 3 It is a schematic diagram of the pitch angle θ.
[0057] Specific implementation method: The specific implementation method of the present invention is described in detail below with reference to the accompanying drawings:
[0058] This paper proposes an auxiliary regression GAN method ARGAN, which realizes the flight attitude regression prediction based on wind field data while amplifying small samples in unbalanced data. Its overall framework is as follows Figure 1 As shown in the figure. This framework is based on InfoGAN. By embedding the pitch angle in the input of the generator and embedding the same pitch angle in the output of the discriminator, the wind field image simulation samples corresponding to the given pitch angle are generated. By designing the distribution alignment loss function in the attitude parameter regression output, the distribution quality of the wind field image simulation samples is further corrected and improved, and the model is made more adaptable to the unbalanced sample distribution, so as to better learn the characteristics of the wind field images collected under unbalanced conditions. This method can also be used in other image-oriented regression tasks.
[0059] A regression prediction method for flight attitude under complex wind fields specifically comprises the following steps:
[0060] Step 1: Dataset construction
[0061] Since vertical wind shear, especially downdraft, is usually very sudden, it is easy to cause the aircraft to sink suddenly, which is the most harmful. Therefore, a coordinate system is constructed with the vertical direction as the z-axis and the point where the center of mass of the aircraft is located as the origin. The present invention focuses on the influence of the vertical wind shear in the z-direction of the coordinate axis on the pitch angle θ of the aircraft. The attitude parameter studied is the pitch angle θ. Specifically, the real sample x of the wind field image collected by the meteorological radar and the airborne sensor is r and the corresponding acquired true pitch angle θ r , and form a data pair <X, Θ>. Where X is the wind field image real sample x in the XY plane where the aircraft passes r A set of, Θ is the actual pitch angle θ r A collection of .
[0062] Step 2: Imbalanced Sample Feature Learning Model Construction
[0063] This step uses the actual pitch angle θ collected in step 1 r The set Θ and the real sample x of wind field image r Specifically, in order to solve the problem of algorithm overfitting caused by the imbalanced distribution of wind farm data and the small number of samples, an ARGAN network for regression tasks is designed to achieve regression prediction of pitch angles while generating simulated samples to amplify image data. The network consists of two parts: the generator G and the discriminator D. In particular, in order to solve the problem of weak algorithm generalization ability caused by the small number of wind farm images under high sea conditions, a generator G module is designed, which can predict the pitch angle θ according to the sampled simulated pitch angle θ. f Generate wind field image simulation samples, whose structure is as follows Figure 1 As shown on the left. Considering that the wind field size conforms to the Weibull distribution, the wind field is the main factor affecting the attitude change in normal flight, and the present invention studies the influence of the wind field on the pitch angle, therefore, the above simulated pitch angle θ f is generated by sampling from a Weibull distribution.
[0064] Specifically, the input of the generator G contains randomly generated Gaussian distributed random noise z G (63×1) and the simulated pitch angle θ obtained by sampling from the Weibull distribution f The two parts are spliced into an input vector z with a length of 64, and then the generator G generates a wind field image simulation sample x with a size of 128×128×3 f The structure of the generator G is similar to that of DCGAN, consisting of a fully connected layer (FC, Full Connection) and 5 deconvolution layers (DeConv, convolution kernel size k is set to 3×3, step size s is set to 2) and corresponding activation layers (ReLU).
[0065] The function of the discriminator D is to distinguish the authenticity of the input image and realize the regression prediction of the pitch angle parameter value. The input image of D is the real sample x of the collected wind field image. r (obtained according to step 1, recorded as the real sample) and the generated wind field image simulation sample x f (Obtained using the generator G constructed in step 2, denoted as the simulation sample) The obfuscated data block x mix .
[0066] The discriminator D comprises a true or false label discrimination branch and a posture parameter pitch angle regression prediction branch. The true or false label discrimination branch outputs the authenticity discrimination result of each image contained in the input data block as a real sample or a simulated sample. The posture parameter pitch angle regression prediction branch outputs the pitch angle value θ predicted for the input image. p (where, for the real sample x r The predicted value of the pitch angle is recorded as For the simulation sample x f The predicted value of the pitch angle is recorded as ). The real sample x collected r With the generated simulation sample x f Obtain data block x after confusing the order mix And input the discriminator D to obtain the prediction results of the true and false label discrimination branch and the attitude parameter pitch angle regression prediction branch respectively.
[0067] For the true and false label discrimination branch, the discriminator D is calculated based on the binary cross entropy to identify the real sample x. r and simulation sample x f The classification loss L D , the formula is:
[0068]
[0069] Among them, m is the batch size of the input image, that is, the real sample x r With the simulation sample x f The number of images after confusion merging, i = {1,...,m}.
[0070] For the attitude parameter pitch angle regression prediction branch, the MSE-based regression loss is used to calculate the true pitch angle θ of the generator G and the discriminator D. r and the simulated pitch angle θ f The prediction loss can be formalized as:
[0071]
[0072] Where m is the batch size of the input image, θ r , is the actual pitch angle θ corresponding to the real sample among the m samplesr And the pitch angle prediction value of the attitude parameter pitch angle regression prediction branch for the real sample, θ f , is the pitch angle prediction value of the aircraft pitch angle and attitude parameter pitch angle regression prediction branch corresponding to the simulation sample among the m samples.
[0073] Step 3: Regression loss design based on distribution alignment
[0074] In order to solve the regression learning bias problem caused by the imbalance of sample quantity, combined with the distribution characteristics of the data, the output of the pitch angle regression prediction branch of the discriminator D attitude parameter in step 2 is designed based on the imbalanced regression loss (referred to as distribution alignment loss). r and the sampled simulated pitch angle θ f As a prior distribution, the generator G and the discriminator D are trained to align the distribution of the predicted pitch angle θ with the above prior distribution to improve the quality of the distribution learned by the algorithm model and prevent the algorithm from generating a single sample due to the distribution deviation of the learned pattern, thereby better realizing the regression of posture parameters.
[0075] Specifically, the real sample x r and simulation sample x f After mixing the order, we get data block x mix , then sample m samples, use KL divergence to measure the difference between the distribution of the pitch angle predicted by the attitude parameter pitch angle regression prediction branch for the input image and the prior distribution, and then feedback and adjust the distribution and model parameters of the simulation samples. The designed distribution alignment loss L KL It can be calculated as follows:
[0076]
[0077] In formula (3), C is a constant, which is set to 0.01 in this embodiment. represents the i-th image among the m input images. For images The probability of the corresponding pitch angle, For the model The probability of the predicted pitch angle. The former is the pitch angle θ input by the model (including the actual pitch angle θ collected r and the sampled simulated pitch angle θ f ) is obtained by statistics, the latter is the pitch angle θ predicted by the attitude parameter pitch angle regression prediction branch p (including the predicted value of the pitch angle of the real sample And the predicted pitch angle of the simulation sample ) is obtained by statistics. The calculation method is as follows:
[0078] ① First, let’s compare θ and θ p Divide the pitch angle into K intervals. Considering that the pitch angle is a series of discrete points, in order to measure its distribution, the interval where the pitch angle is located is divided into intervals of t, where t is 0.1°. Specifically, assuming that θ and θ p In the entire set of values of the two, the maximum and minimum values are θ max and θ min Indicated by, then with t as the interval, θ min ,θ max ] is divided into K intervals, K can be calculated as: That is, t divides θ and θ p The upper integer after the range is taken.
[0079] ② Secondly, calculate θ and θ respectively p The frequency of appearing in the K intervals divided by ①. Since θ and θ p The calculation method is the same, so here we take θ as an example. Specifically, for the above K intervals A=[A1,...,A K ], the pitch angle value range of the jth (j∈{1,...,K}) interval can be expressed as A j =[θ min +(j-1)·t,θ min +j·t). According to each interval A j The pitch angle value range is calculated, and the number of samples of the pitch angle value θ falling in each interval is counted to obtain the number of times the pitch angle θ appears in each interval F = [f1, ..., f K ].
[0080] ③Third, respectively for θ and θ p Calculating Probability and
[0081] The calculation method for both is the same, and the probability of θ is calculated Specifically, for the i-th sample Assume that its pitch angle θ i falls in the jth interval, i.e. θ i ∈A j , then the formula
[0082] (3) It can be defined as:
[0083]
[0084] in, is a step function, which is defined as is defined as follows:
[0085]
[0086] Among them, f j Represents the pitch angle θ in interval A j The number of occurrences, f n represents the number of times the pitch angle θ appears in the nth interval, n={1,...,K}, and ρ is a scale hyperparameter used to control the distribution to be aligned, and is set to 1 in this embodiment.
[0087] For θ p Calculating Probability
[0088] Specifically, for the i-th sample Assume that the predicted pitch angle θ pi falls in the jth interval, i.e. θ pi ∈A j , then in formula (3) It can be defined as:
[0089]
[0090] in, is a step function, which is defined as is defined as follows:
[0091]
[0092] Among them, f j Represents the predicted pitch angle θ pi In interval A j The number of occurrences, f n Represents the predicted pitch angle θ pi The number of occurrences in the nth interval, n={1,...,K}, ρ is a scale hyperparameter used to control the distribution to be aligned, and is set to 1 in this embodiment.
[0093] Step 4: Model training
[0094] The stochastic gradient descent method SGD is used as the optimizer, the SGD learning rate parameter is set to 0.0002, the batch size m is set to 128, and the model is trained using the loss functions described in formulas (1), (2), and (3). The specific steps are as follows:
[0095] ① According to step 2, randomly generated Gaussian distribution random noise z G and the simulated pitch angle θ obtained by sampling from the Weibull distribution f , obtain the input vector z, and use the generator G to generate a wind field image simulation sample x f ;
[0096] ② The real sample x r and simulation sample x f Obtain x after confusion mix , x mix Input the discriminator D, train the authenticity discrimination branch, calculate the authenticity discrimination loss according to formula (1), and update the discriminator D and generator G;
[0097] ③ x mix Input the discriminator D, train the attitude parameter pitch angle regression prediction branch, calculate the regression loss according to formula (2), and update the discriminator D and generator G;
[0098] ④ x mix After inputting into the discriminator D, the distribution alignment loss is calculated according to step 3 and formula (3) and then fed back to update the discriminator D and generator G.
[0099] Step 5: Model testing
[0100] After completing the model training according to step 4, retain the discriminator D in the trained model in step 4 and convert the collected real samples x r It is input as a test sample into the discriminator D to obtain the pitch angle prediction value θ output by the attitude parameter pitch angle regression prediction branch r p ,The predicted value of the pitch angle is the effect of the given wind field image on the pitch angle of the aircraft under study.
[0101] It should be understood that the parts not elaborated in detail in this specification belong to the prior art. The above embodiments are only descriptions of the preferred implementation methods of the present invention, and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made by ordinary engineers and technicians in this field to the technical solution of the present invention should fall within the protection scope determined by the claims of the present invention.
Claims
1. A regression prediction method for flight attitude under complex wind fields, characterized in that The specific steps include: Step 1: Dataset construction Use the collected wind field image real samples and the corresponding collected real pitch angles to form a data pair; Step 2: Unbalanced sample feature learning model construction Use the set of real samples of wind field images and real pitch angles collected in step 1 to train the model; design an ARGAN network for regression tasks to achieve regression prediction of pitch angles while generating simulated samples to augment image data; The network consists of two parts: the generator G and the discriminator D. The generator G module can generate wind field image simulation samples according to the sampled simulated pitch angles, and the discriminator D can distinguish the authenticity of the input image and realize the regression prediction of the pitch angle parameter value. Step 3: Regression loss design based on distribution alignment The output of the pitch angle regression prediction branch of the discriminator D in step 2 is designed with an imbalanced regression loss based on distribution alignment; the real pitch angles collected by statistics and the sampled simulated pitch angles are used as prior distributions, and the generator G and the discriminator D are trained to align the distribution of the predicted pitch angles with the prior distributions to achieve attitude parameter regression; In the step 3, the regression loss design based on distribution alignment is: The real sample x r and simulation sample x f After mixing the order, we get data block x mix , then sample m samples, use KL divergence to measure the difference between the distribution of the pitch angle predicted by the attitude parameter pitch angle regression prediction branch for the input image and the prior distribution, and then feedback and adjust the distribution and model parameters of the simulation samples; the designed distribution alignment loss L KL It can be calculated as follows: In formula (3), C is a constant, represents the i-th image among m input images; For images The probability of the corresponding pitch angle, For the model The probability of the predicted pitch angle; the former is obtained by the pitch angle θ statistics of the model input, and the latter is the pitch angle θ predicted by the attitude parameter pitch angle regression prediction branch p Statistics are obtained; the calculation method is as follows: ① First, let’s compare θ and θ p Divide the same K intervals; divide the intervals where the pitch angles are located with intervals of t; assume that θ and θ p In the entire set of values of the two, the maximum and minimum values are θ max and θ min Indicates that, with t as the interval, [θ min ,θ max ] is divided into K intervals, and K is calculated as: That is, t divides θ and θ p Take the upper integer after the range; ② Secondly, calculate θ and θ respectively p The frequency of appearing in the K intervals divided by ①; for the above K intervals A=[A1,...,A K ], the pitch angle value range of the jth (j∈{1,...,K}) interval is represented by A j =[θ min +(j-1)·t,θ min +j·t); according to each interval A j The pitch angle value range is calculated, and the number of samples whose pitch angle values fall in each interval is counted to obtain the number of times the pitch angle appears in each interval F = [f1, ..., f K ]; ③Third, respectively for θ and θ p Calculating Probability and Calculate the probability for θ For the i-th sample Assume that its pitch angle θ i falls in the jth interval, i.e. θ i ∈A j , then in formula (3) Defined as: in, is a step function, which is defined as is defined as follows: Among them, f j Represents the pitch angle θ in interval A j The number of occurrences, f n represents the number of times the pitch angle θ appears in the nth interval, n = {1, ..., K}, and ρ is a scale hyperparameter; For θ p Calculating Probability For the i-th sample Assume that the predicted pitch angle θ pi falls in the jth interval, i.e. θ pi ∈A j , then in formula (3) Defined as: in, is a step function, which is defined as is defined as follows: Among them, f j Represents the predicted pitch angle θ pi In interval A j The number of occurrences, f n Represents the predicted pitch angle θ pi The number of occurrences in the nth interval, n = {1, ..., K}, ρ is the scale hyperparameter; Step 4: Model training Train the models separately; Step 5: Model testing After completing the training of the model according to step 4, retain the discriminator D in the trained model of step 4, input the collected real samples into the discriminator D, and obtain the pitch angle prediction value output by the attitude parameter pitch angle regression prediction branch.
2. According to claim 1, a regression prediction method for flight attitude under complex wind field, characterized in that The real sample x of wind field images collected by weather radar and airborne sensors in step 1 of the dataset construction is r and the corresponding acquired true pitch angle θ r , and form data pairs Where X is the real sample x of the wind field image on the XY plane where the aircraft passes r A collection of is the actual pitch angle θ r A collection of .
3. According to claim 1, a regression prediction method for flight attitude under complex wind field, characterized in that In the step 2 of building the unbalanced sample feature learning model: The input of the generator G contains randomly generated Gaussian distributed random noise z G and the simulated pitch angle θ obtained by sampling from the Weibull distribution f The two parts are spliced into the input vector z, and then the generator G generates the wind field image simulation sample x f ; The discriminator D comprises a true or false label discrimination branch and a posture parameter pitch angle regression prediction branch. The true or false label discrimination branch outputs the authenticity discrimination result of each image contained in the input data block as a real sample or a simulated sample. The posture parameter pitch angle regression prediction branch outputs the pitch angle value θ predicted for the input image. p ; The collected wind field image real sample x r And the generated wind field image simulation sample x f Obtain data block x after confusing the order mix And input the discriminator D to obtain the prediction results of the true and false label discrimination branch and the attitude parameter pitch angle regression prediction branch respectively; For the true and false label discrimination branch, the discriminator D is calculated based on the binary cross entropy to identify the real sample x of the wind field image. r and wind field image simulation sample x f The classification loss L D , the formula is: Where m is the batch size of the input image, i = {1, ..., m}; For the attitude parameter pitch angle regression prediction branch, the MSE-based regression loss is used to calculate the true pitch angle θ of the generator G and the discriminator D. r and the simulated pitch angle θ f The prediction loss is: Where m is the batch size of the input image, θ r , is the actual pitch angle θ corresponding to the real sample among the m samples r And the pitch angle prediction value of the attitude parameter pitch angle regression prediction branch for the real sample, θ f , is the pitch angle prediction value of the aircraft pitch angle and attitude parameter pitch angle regression prediction branch corresponding to the simulation sample among the m samples.
4. According to claim 3, a regression prediction method for flight attitude under complex wind field, characterized in that In the step 4 model training described above: The stochastic gradient descent method SGD is used as the optimizer, and the model is trained respectively using the loss functions described in formulas (1), (2) and (3). The specific steps are as follows: ① According to step 2, randomly generated Gaussian distribution random noise z G and the simulated pitch angle θ obtained by sampling from the Weibull distribution f , obtain the input vector z, and use the generator G to generate a wind field image simulation sample x f ; ② The real sample x r and simulation sample x f Obtain x after confusion mix , x mix Input the discriminator D, train the authenticity discrimination branch, calculate the authenticity discrimination loss according to formula (1), and update the discriminator D and generator G; ③ x mix Input the discriminator D, train the attitude parameter pitch angle regression prediction branch, calculate the regression loss according to formula (2), and update the discriminator D and generator G; ④ x mix After inputting into the discriminator D, the distribution alignment loss is calculated according to step 3 and formula (3) and then fed back to update the discriminator D and generator G.
