Wake area intelligent detection method based on convolution and machine learning

Abstract

The invention discloses a wake region intelligent detection method based on convolution and machine learning, and belongs to the technical field of wake region boundary detection. The problems that in the prior art, a traditional wake area detection method is low in manual interpretation efficiency, poor in universality and insufficient in precision are solved. The method comprises the following steps: standardizing tail rake pressure measurement data to obtain a normalized pressure coefficient vector; extracting local key features of the pressure distribution curve based on a convolution method, constructing a multi-dimensional feature vector, and generating a feature matrix; selecting a training sample to form a label column vector; dividing the sample set into a training set and a verification set, and selecting an optimal model as a wake region detection reference model; performing intelligent prediction on the tail rake pressure data of the new test train number, and determining the boundary of a wake region; the generalization ability of the model is continuously improved through incremental learning, and continuous iteration of the performance of the wake region detection reference model is achieved. The method effectively improves the accuracy of a wake region boundary detection result, and can be applied to a wind tunnel test.

Description

TECHNICAL FIELD

[0001] The present application relates to a wake region intelligent detection method, in particular to a wake region intelligent detection method based on convolution and machine learning, and belongs to the technical field of wake region boundary detection. BACKGROUND

[0002] As a core means of aerodynamics research and aircraft aerodynamic layout design, wind tunnel test provides key data support for aircraft performance evaluation. In wing profile wind tunnel test, wing profile drag coefficient is the core parameter representing aerodynamic performance, and its accurate measurement depends on momentum method drag measurement technology. The static pressure and total pressure distribution in the wake region are collected by fixed or continuous scanning wake measurement device, and the drag coefficient is calculated based on the total pressure loss of the wake. In this process, the accurate determination of the wake region boundary directly determines the effectiveness of the drag integration region, which is the key prerequisite to ensure the accuracy of drag measurement.

[0003] In the prior art, the wake region boundary detection method has the following significant limitations: first, the manual interpretation mode relies on the experience of test personnel to make subjective judgments on the wake region boundary, which has the problems of low efficiency and poor consistency of interpretation, and is difficult to adapt to the development needs of wind tunnel test efficiency and batch production; second, the traditional data-driven method, represented by the threshold method, has its core parameters (such as pressure threshold, gradient threshold) related to test wing chord length, installation position of wake rake (front-back distance), wake rake pressure point density, test Mach number and other working conditions, which leads to the non-universality of calibrated parameters for other test conditions. When the test object or working condition changes, parameter adjustment and optimization need to be re-performed, which has high adaptation cost; third, the wake region boundary is often accompanied by local mutation characteristics (such as gradient sudden change, curvature turning) of pressure distribution curve. Traditional wake region boundary detection methods are difficult to capture such subtle features, and can only give rough boundary range, which cannot meet the strict requirements of high-precision drag measurement on boundary determination. Since more accurate boundary values are needed, the number of parameters involved will increase significantly, and the solving method will also become more complex.

[0004] Therefore, there is an urgent need for a wake region intelligent detection method based on convolution and machine learning, which has high efficiency, universality and high precision. SUMMARY

[0005] A brief summary of the application is given in the following to provide a basic understanding of some aspects of the application. It should be understood that this summary is not an exhaustive overview of the application. It is not intended to identify key or important parts of the application, nor is it intended to limit the scope of the application. Its purpose is only to give some concepts in a simplified form as a prelude to the more detailed description discussed later.

[0006] In view of this, in order to solve the problems of low efficiency, poor universality and insufficient precision of the traditional wake region detection method in the prior art, the present application provides a wake region intelligent detection method based on convolution and machine learning.

[0007] The technical scheme is as follows: a wake region intelligent detection method based on convolution and machine learning, comprising the following steps:

[0008] S1. Standardizing the wake rake pressure measurement data to eliminate the influence of test working condition differences on data characteristics, and obtaining a normalized pressure coefficient vector;

[0009] S2. According to the normalized pressure coefficient vector, extracting local key features of the pressure distribution curve based on the convolution method, constructing a multi-dimensional feature vector, and generating a feature matrix;

[0010] S3. Selecting the wake rake pressure measurement data in the early test train as a training sample, assigning a wake region judgment label to the feature matrix to form a label column vector;

[0011] S4. Dividing the sample set constructed based on the feature matrix and the label column vector into a training set and a validation set, training the training set using a supervised classification algorithm, evaluating the model performance through the validation set, and selecting the optimal model as a wake region detection benchmark model;

[0012] S5. Based on the wake region detection benchmark model, intelligently predicting the wake rake pressure data of the new test train to determine the wake region boundary;

[0013] S6. Continuously improving the model generalization ability through incremental learning, realizing the continuous iteration of the performance of the wake region detection benchmark model, and further improving the accuracy of the wake region boundary.

[0014] Further, in S1, the following steps are included:

[0015] S11. Collecting the original data of the wake rake pressure measurement points, the wake rake comprising pressure measurement points, the spatial coordinates of the pressure measurement point being , with the unit being m, and the measured pressure being , with the unit being Pa;

[0016] S12. Based on the incoming flow parameters of step S11, the pressure coefficient of the pressure measurement point is calculated.

[0017] In S12, the pressure coefficient of the pressure measurement point is expressed as:

[0018]

[0019] wherein, P is the static pressure of the incoming flow, and has a unit of Pa, P is the dynamic pressure of the incoming flow, and has a unit of Pa;

[0020] S13. Normalizing the spatial coordinates of the pressure measuring points to obtain normalized coordinates with the chord length c of the airfoil as the reference , and normalizing the pressure coefficient of the i-th pressure measuring point to obtain a normalized pressure coefficient vector ;

[0021] In the S13, the normalized coordinate is expressed as:

[0022]

[0023] The normalized pressure coefficient vector is expressed as:

[0024]

[0025]

[0026]

[0027] wherein, Pmin is the minimum value of the pressure coefficient of the i-th pressure measuring point, Pmax is the maximum value of the pressure coefficient of the i-th pressure measuring point.

[0028] Further, in the S2, the following steps are included:

[0029] S21. For the normalized pressure coefficient vector , local feature extraction is performed using a difference convolution kernel to obtain a first-order differential feature and a second-order differential feature ;

[0030] In the S21, a first-order difference convolution kernel =[-1, 0, 1] is used for convolution calculation to obtain the first-order differential feature of each point, which is used to represent the gradient of the pressure curve at the i-th point, ;

[0031] The h-th element of the first-order differential feature vector is expressed as:

[0032] ​​​​

[0033] wherein, is the normalized pressure coefficient vector , th element of , , and is an integer;

[0034] The second-order differential convolution kernel =[1,−2,1] is used to calculate the second-order differential features of each point, which are used to represent the curvature of the pressure curve at the th point, ; ;

[0035] The hth element of the second-order differential feature vector is expressed as:

[0036]

[0037] S22. Generating a single measurement point feature vector according to the first-order differential feature and the second-order differential feature ;

[0038] In the S22, a twelve-dimensional feature row vector is constructed for the th pressure measurement point, which is divided into three categories. The first category is the basic feature: a one-dimensional feature row vector , is the incoming flow Mach number, a two-dimensional feature row vector , is the airfoil angle of attack, a three-dimensional feature row vector , and a four-dimensional feature row vector ; the second category is the first-order differential neighborhood feature: a five-dimensional feature row vector , a six-dimensional feature row vector , a seven-dimensional feature row vector , an eight-dimensional feature row vector , and a nine-dimensional feature row vector ; and the third category is the second-order differential neighborhood feature: , , ;

[0039] For edge pressure measurement points (without complete adjacent points), the missing first-order / second-order differential features are filled with zero values;

[0040] S23. Constructing a global feature matrix, repeating step S22 for all pressure measurement points to generate a dimensional feature matrix , To be a feature.

[0041] Further, in the S3, the following steps are included:

[0042] S31. For the first test, select the tail pressure data of the test train in the pre-test period as the training sample;

[0043] S32. Accurately determine the tail track area of each pressure measuring point in the sample by semi-intelligent screening combined with manual checking, and set the tail track area determination label, i.e. the tail track area measuring point label , non-tail track area measuring point label , forming a label column vector .

[0044] Further, in the S4, the following steps are included:

[0045] S41. Divide the data, and divide the sample set composed of the feature matrix and the label column vector into training set and validation set in the ratio of 7:3;

[0046] S42. Train the training set by using a supervised classification algorithm, and evaluate the model performance by using the validation set;

[0047] S43. Take the accuracy of the validation set as the model effectiveness threshold, select the optimal model, i.e. the model with the highest accuracy , as the tail track area detection benchmark model;

[0048]

[0049] wherein, is the number of samples actually in the tail track area and predicted as the tail track area, represents the number of samples actually in the non-tail track area and predicted as the non-tail track area, represents the number of samples actually in the tail track area and predicted as the non-tail track area, represents the number of samples actually in the non-tail track area and predicted as the tail track area.

[0050] Further, in the S5, the following steps are included:

[0051] S51. Repeat steps S1-S2 for the tail harrow pressure data of the new test train, and generate a new feature matrix ;

[0052] S52. Input the new feature matrix into the tail track area detection benchmark model determined in step S4, and output the tail track area determination label of each pressure measuring point;

[0053] S53. Determine the label of the wake region based on the distribution characteristics of each pressure measuring point The critical point of the transition from 0 to 1 is extracted as the boundary of the wake region.

[0054] Further, in the S6, the following steps are included:

[0055] S61. Perform result verification, manually review the wake region boundary output by step S5, and identify misjudgment labels;

[0056] S62. Correct the misjudgment label, incorporate the new sample containing the new feature matrix and the corrected label into the historical sample set, and obtain an updated sample set;

[0057] S63. Repeat step S4, train the wake region detection benchmark model based on the updated sample set, generate an optimized model, and realize continuous iteration of the model performance.

[0058] The beneficial effects of the present application are as follows: the present application realizes intelligent and high-precision determination of the wake region boundary, and improves the adaptability of the method to different test conditions. BRIEF DESCRIPTION OF DRAWINGS

[0059] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0060] Figure 1 It is a flowchart of an intelligent wake region detection method based on convolution and machine learning;

[0061] Figure 2 It is a comparison diagram of normalized pressure coefficient and wake region label. DETAILED DESCRIPTION

[0062] In order to make the technical solutions and advantages in the embodiments of the present application clearer, the following further describes the exemplary embodiments of the present application with reference to the drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, not an exhaustive enumeration of all embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0063] Reference Figure 1 and Figure 2 Detailed description of the embodiments, an intelligent wake region detection method based on convolution and machine learning, specifically including the following steps:

[0064] S1. Standardize the wake rake pressure measurement data to eliminate the influence of test condition differences on data characteristics, and obtain a normalized pressure coefficient vector;

[0065] S2. Extracting local key features of the pressure distribution curve based on the convolution method according to the normalized pressure coefficient vector, constructing a multi-dimensional feature vector, and generating a feature matrix;

[0066] S3. Selecting the tail rake pressure data in the early test train as the training sample, assigning the tail wake area judgment label to the feature matrix, and forming a label column vector;

[0067] S4. Dividing the sample set constructed based on the feature matrix and the label column vector into a training set and a validation set, training the training set using a supervised classification algorithm, evaluating the model performance through the validation set, and selecting the optimal model as the tail wake area detection benchmark model;

[0068] S5. Based on the tail wake area detection benchmark model, intelligently predicting the tail rake pressure data of the new test train and determining the tail wake area boundary;

[0069] S6. Continuously improving the model generalization ability through incremental learning, realizing the continuous iteration of the tail wake area detection benchmark model performance, and further improving the tail wake area boundary precision.

[0070] Further, in the S1, the following steps are included:

[0071] S11. Collecting tail rake pressure point original data, the tail rake containing pressure points, the spatial coordinates of the th pressure point being , with the unit being m, and the measured pressure being , with the unit being Pa;

[0072] S12. Based on the incoming flow parameters of step S11, the pressure coefficient of the th pressure point is calculated.

[0073] In the S12, the pressure coefficient of the th pressure point is represented as:

[0074]

[0075] wherein is the incoming flow static pressure, with the unit being Pa, is the incoming flow dynamic pressure, with the unit being Pa;

[0076] S13. Normalizing the pressure point spatial coordinates to obtain the normalized coordinates based on the airfoil chord length c, and normalizing the pressure coefficient of the th pressure point to obtain the normalized pressure coefficient vector ;

[0077] In S13, the normalized coordinates Represented as:

[0078]

[0079] Normalized pressure coefficient vector Represented as:

[0080]

[0081]

[0082]

[0083] in, For the first Pressure coefficient at each pressure measuring point The minimum value, For the first Pressure coefficient at each pressure measuring point The maximum value.

[0084] Specifically, the standardization process includes collecting raw data from the tail rake pressure measurement points, calculating the pressure coefficient, location coordinates, and normalizing the pressure coefficient data.

[0085] Furthermore, step S2 includes the following steps:

[0086] S21. For the normalized pressure coefficient vector Local feature extraction is performed using differential convolution kernels to obtain first-order differential features. and second-order differential characteristics ;

[0087] In step S21, a first-order difference convolution kernel is used. Convolution is performed on the interval [-1, 0, 1] to obtain the first-order differential features of each point. It is used to characterize the pressure curve in the first... gradient at a point ;

[0088] The h-th element of the first-order differential eigenvector is represented as:

[0089]

[0090] in, Normalized pressure coefficient vector No. item, , ,and It is an integer;

[0091] Using second-order difference convolution kernel Convolution is performed on the region [1,−2,1] to obtain the second-order differential features of each point. It is used to characterize the pressure curve in the first... The curvature of a point ;

[0092] Second-order differential eigenvector The h-th element is represented as:

[0093]

[0094] S22. Based on the characteristics of first-order differentials and second-order differential characteristics Generate single-point feature vectors;

[0095] In S22, the first... A twelve-dimensional feature row vector is constructed from each pressure measurement point. It is divided into three categories. The first category is the basic feature: one-dimensional feature row vector. , For the incoming Mach number, the two-dimensional feature row vector , For airfoil angle of attack, three-dimensional characteristic row vector Four-dimensional feature row vectors The second type is first-order differential neighborhood features: five-dimensional feature row vectors. Six-dimensional feature row vector Seven-dimensional feature row vector Eight-dimensional feature row vectors Nine-dimensional feature row vector (The second category covers the first-order differential features of the current point and its two left and right neighbors); the third category is the second-order differential neighborhood features: , , (The third category covers the second-order differential features of the current point and one neighboring point on each side).

[0096] For edge pressure measurement points (without complete neighboring points), their missing first / second-order differential features are filled with zero values;

[0097] S23. Construct a global feature matrix for all Repeat step S22 at each pressure measurement point to generate... 3D feature matrix , Features.

[0098] Specifically, the method extracts key local features of the pressure distribution curve based on convolution, including using first-order and second-order difference convolution kernels to extract local features for the normalized pressure coefficient vector.

[0099] Furthermore, step S3 includes the following steps:

[0100] S31. For the first test, the tail rake pressure measurement data from the test trains conducted in the early stage of the test were selected as training samples;

[0101] S32. A semi-intelligent screening method combined with manual verification is adopted to accurately determine the location of the wake region of each pressure measurement point in the sample, and a wake region determination label is set, that is, the pressure measurement point label of the wake region (including the boundary). Pressure measurement points in non-wake regions are labeled. This forms a label column vector. .

[0102] Furthermore, step S4 includes the following steps:

[0103] S41. Divide the data into feature matrices. With label column vector The sample set was divided into a training set and a validation set in a 7:3 ratio.

[0104] S42. Use supervised classification algorithms (including but not limited to decision trees, logistic regression, and support vector machines) to train the training set and evaluate the model performance using the validation set;

[0105] S43. Determine accuracy using the validation set. As a threshold for model effectiveness, the optimal model is selected, i.e., accuracy. The highest-performing model is used as the benchmark model for wake region detection;

[0106]

[0107] in, This represents the number of samples that are both actually wheezes and predicted to be wheezes. This represents the number of samples that are actually non-wheeze regions and are predicted to be non-wheeze regions. This represents the number of samples that are actually wake regions but are predicted to be non-wake regions. This represents the number of samples that are actually non-whet regions but are predicted to be whet regions.

[0108] Furthermore, step S5 includes the following steps:

[0109] S51. Repeat steps S1-S2 for the tail rake pressure data of the new test run to generate a new feature matrix. ;

[0110] S52. Transform the new feature matrix Input the wake region detection benchmark model determined in step S4, and output the wake region judgment label for each pressure measurement point. ;

[0111] S53. Determine the label based on the wake region of each pressure measurement point. Based on the distribution characteristics, the critical point where it transitions from 0 to 1 is extracted as the boundary of the wake region.

[0112] Furthermore, step S6 includes the following steps:

[0113] S61. Perform result verification, manually review the boundary of the wake region output in step S5, and identify misjudged labels;

[0114] S62. Correct the misclassified labels, and merge the new samples containing the new feature matrix and the corrected labels into the historical sample set to obtain the updated sample set;

[0115] S63. Repeat step S4 to train the baseline model for wake region detection based on the updated sample set, generate an optimized model, and achieve continuous iteration of model performance.

[0116] Although the invention has been described with reference to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of interpreting or limiting the subject matter of the invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. A method for intelligent detection of trail regions based on convolution and machine learning, characterized in that, Includes the following steps: S1. Standardize the tail rake pressure measurement data to eliminate the influence of differences in test conditions on data characteristics and obtain a normalized pressure coefficient vector. S2. Based on the normalized pressure coefficient vector, extract the local key features of the pressure distribution curve using the convolution method, construct a multi-dimensional feature vector, and generate a feature matrix; S3. Select tail rake pressure data from previous test runs as training samples, assign wake region judgment labels to the feature matrix, and form a label column vector; S4. Divide the sample set constructed based on the feature matrix and label column vector into a training set and a validation set. Use a supervised classification algorithm to train the training set, evaluate the model performance through the validation set, and select the optimal model as the benchmark model for wake region detection. S5. Based on the wake zone detection benchmark model, intelligently predict the tail rake pressure data of the new test train and determine the wake zone boundary; S6. By incrementally learning, the generalization ability of the model is continuously improved, and the performance of the baseline model for wake region detection is continuously iterated, thereby further improving the accuracy of the wake region boundary.

2. The intelligent tail region detection method based on convolution and machine learning according to claim 1, characterized in that, S1 includes the following steps: S11. Collect raw data from the tail rake pressure measurement points. The tail rake includes... The pressure measurement point, the first The spatial coordinates of the pressure measurement points are: Its unit is m, and the measured pressure is Its unit is Pa; S12. Based on the incoming flow parameters from step S11, calculate the first... Pressure coefficient at each pressure measuring point ; In S12, the first Pressure coefficient at each pressure measuring point Represented as: in, The static pressure of the incoming flow is expressed in Pa. The pressure is the pressure of the incoming flow, and its unit is Pa; S13. Normalize the spatial coordinates of the pressure measurement points to obtain normalized coordinates based on the airfoil chord length c. and for the first Pressure coefficient at each pressure measuring point Normalization is performed to obtain the normalized pressure coefficient vector. ; In S13, the normalized coordinates Represented as: Normalized pressure coefficient vector Represented as: in, For the first Pressure coefficient at each pressure measuring point The minimum value, For the first Pressure coefficient at each pressure measuring point The maximum value.

3. The intelligent tail region detection method based on convolution and machine learning according to claim 2, characterized in that, S2 includes the following steps: S21. For the normalized pressure coefficient vector Local feature extraction is performed using differential convolution kernels to obtain first-order differential features. and second-order differential features ; In S21, a first-order difference convolution kernel is used. Convolution is performed on the interval [-1, 0, 1] to obtain the first-order differential features of each point. It is used to characterize the pressure curve in the first... gradient at a point ; The h-th element of the first-order differential eigenvector is represented as: in, Normalized pressure coefficient vector No. item, , ,and It is an integer; Using second-order difference convolution kernel Convolution is performed on the region [1,−2,1] to obtain the second-order differential features of each point. It is used to characterize the pressure curve in the first... The curvature of a point ; Second-order differential eigenvector The h-th element is represented as: S22. Based on the characteristics of first-order differentials and second-order differential features Generate single-point feature vectors; In S22, the first A twelve-dimensional feature row vector is constructed from each pressure measurement point. It is divided into three categories. The first category is the basic feature: one-dimensional feature row vector. , For the incoming Mach number, the two-dimensional feature row vector , For airfoil angle of attack, three-dimensional characteristic row vector Four-dimensional feature row vectors The second type is first-order differential neighborhood features: five-dimensional feature row vectors. Six-dimensional feature row vector Seven-dimensional feature row vector Eight-dimensional feature row vectors Nine-dimensional feature row vector The third category is the second-order differential neighborhood feature: , , ; For edge pressure measurement points (without complete neighboring points), their missing first / second-order differential features are filled with zero values; S23. Construct a global feature matrix for all Repeat step S22 at each pressure measurement point to generate... 3D feature matrix , Features.

4. The intelligent tail region detection method based on convolution and machine learning according to claim 3, characterized in that, S3 includes the following steps: S31. For the first test, the tail rake pressure measurement data from the test trains conducted in the early stage of the test were selected as training samples; S32. A semi-intelligent screening method combined with manual verification is adopted to accurately determine the location of the wake region of each pressure measurement point in the sample, and a wake region determination label is set, that is, a label for the pressure measurement point in the wake region. Pressure measurement points in non-wake regions are labeled. This forms a label column vector. .

5. The intelligent detection method for trail regions based on convolution and machine learning according to claim 4, characterized in that, S4 includes the following steps: S41. Divide the data into feature matrices. With label column vector The sample set was divided into a training set and a validation set in a 7:3 ratio. S42. A supervised classification algorithm is used to train the model on the training set, and the model performance is evaluated using the validation set; S43. Determine accuracy using the validation set. As a threshold for model effectiveness, the optimal model is selected, i.e., accuracy. The highest-performing model is used as the benchmark model for wake region detection; in, This represents the number of samples that are both actually wheezes and predicted to be wheezes. This represents the number of samples that are actually non-wheeze regions and are predicted to be non-wheeze regions. This represents the number of samples that are actually wake regions but are predicted to be non-wake regions. This represents the number of samples that are actually non-whet regions but are predicted to be whet regions.

6. The intelligent detection method for trailing regions based on convolution and machine learning according to claim 5, characterized in that, S5 includes the following steps: S51. Repeat steps S1-S2 for the tail rake pressure data of the new test run to generate a new feature matrix. ; S52. Transform the new feature matrix Input the wake region detection benchmark model determined in step S4, and output the wake region judgment label for each pressure measurement point. ; S53. Determine the label based on the wake region of each pressure measurement point. Based on the distribution characteristics, the critical point where it transitions from 0 to 1 is extracted as the boundary of the wake region.

7. The intelligent tail region detection method based on convolution and machine learning according to claim 6, characterized in that, S6 includes the following steps: S61. Perform result verification, manually review the boundary of the wake region output in step S5, and identify misjudged labels; S62. Correct the misclassified labels, and merge the new samples containing the new feature matrix and the corrected labels into the historical sample set to obtain the updated sample set; S63. Repeat step S4 to train the wake region detection benchmark model based on the updated sample set, generate an optimized model, and achieve continuous iteration of model performance.

Images