An intake passage flow state classification method based on improved support vector machine feature selection

Abstract

The application belongs to the field of flow field monitoring of hypersonic inlet, and proposes an inlet flow state classification method based on improved support vector machine feature selection. According to a typical hypersonic inlet model, an inlet flow state data set is constructed; based on the feature selection ReliefF method, the feature weight is calculated, and the feature is preliminarily selected; based on the granularity support vector machine GSVM-SNN of shared nearest neighbor similarity, two features are selected from the preliminarily selected features each time for the granularity support vector machine training of shared nearest neighbor similarity, the two features with the highest training accuracy are reserved as the final features, and the classification model obtained by training is used as the decision model; the decision model is used for the test set, and according to the test training effect, the flow state of the hypersonic inlet is finally obtained.

Description

An Inlet Flow State Classification Method Based on Improved Support Vector Machine Feature Selection Technical Field

[0001] This invention relates to the field of flow field monitoring in hypersonic inlets, and more particularly to an inlet flow state classification method based on improved support vector machine feature selection. Background Technology

[0002] The hypersonic inlet is a crucial component of a scramjet engine, and its inability to start is a major bottleneck in its development. The inlet's primary function is to capture air from the atmosphere during hypersonic flight, decelerating and pressurizing the low-pressure hypersonic inflow. When it is inactive, the inlet's performance in capturing airflow and restoring total pressure significantly decreases, additional drag increases sharply, affecting the normal operation of subsequent components such as the engine combustion chamber, and causing a sharp drop in the engine's net thrust. The scramjet engine is in an abnormal operating state, severely impacting the propulsion system's performance. Therefore, achieving flow monitoring and prediction of inlet inactivity is of significant importance and value.

[0003] In his paper "Computational Fluid Dynamics Analysis of the Central Institute of Aviation Motors / NASA Scramjet," Rodriguez conducted a quantitative study using numerical analysis. He performed CFD calculations under both high-speed and low-speed operating conditions and compared the results with actual flight data. The study showed that at low speeds (Ma 2.5–5.0), the inlet was not activated and significant flow separation occurred. Even at the designed inlet activation condition (Ma 6.4), due to hysteresis, significant flow separation still existed in the inlet, leading to a total pressure loss and causing the combustion chamber inlet velocity to be lower than expected, affecting engine performance and test results. In their paper "Multistability and Loops-Coupled Hysteresis: Flight-Test Analysis on Error Detection of Inlet Start / Unstart," Cui Tao et al., based on nonlinear interaction effects and dynamic system theory, discovered that hypersonic inlets exhibit multistability and hysteresis loop coupling effects. In their paper "Research Progress on the Problem of Hypersonic Inlet Start-up", Tan Huijun et al. summarized and analyzed the research progress on the problem of hypersonic inlet start-up. The inability to start up the scramjet engine inlet has attracted widespread attention from scholars at home and abroad, and it is necessary to make real-time and accurate judgments on the start-up / inability state of the hypersonic inlet. Summary of the Invention

[0004] To achieve accurate and real-time identification of the inactive state of a hypersonic inlet and overcome the impact of measurement noise and different feature selection combinations on classification results, this invention proposes an inlet flow state classification method based on the feature selection ReliefF method and granular support vector machine (GSVM-SNN) technology with shared nearest neighbor similarity. This method combines feature selection with inlet start / stop state classification, achieving rapid training and classification of the hypersonic inlet flow field state. It enables rapid identification and classification of different operating states of the hypersonic inlet, significantly improving the classification accuracy of the inactive state. This method has strong universal engineering application value.

[0005] The technical solution of this invention is as follows: A method for classifying the flow state of an inlet based on improved support vector machine feature selection, comprising three parts: inlet numerical simulation data processing, preliminary selection of classification features, and feature reselection based on the training results of an improved granular support vector machine. This method is for classifying the starting / stopping state of the inlet. Based on a typical hypersonic inlet model, an inlet flow state dataset is constructed; based on the ReliefF feature selection method, feature weights are calculated for preliminary feature selection; based on a granular support vector machine with shared nearest neighbor similarity (GSVM-SNN), two features are selected from the preliminary selection for training in each iteration, retaining the two features with the highest training accuracy as the final features, and using the trained classification model as the decision model; the decision model is used on a test set, and based on the test training effect, the flow state of the hypersonic inlet is finally obtained.

[0006] The specific steps are as follows:

[0007] S1: Numerical simulation data processing of the air intake duct;

[0008] S1.1: Conduct wind tunnel tests or fluid dynamics numerical simulations on typical hypersonic inlets to obtain inlet flow field data under different operating conditions;

[0009] S1.2: Based on the wind tunnel test or numerical simulation results and referring to the flow field cloud map, mark the flow state of the air intake and extract pressure point data to generate a dataset.

[0010] S1.3: Normalize the pressure data of all measuring points on the typical hypersonic inlet wall, with each measuring point corresponding to a feature;

[0011] S1.4: Divide the dataset samples into training and test sets according to the proportions for subsequent training of granular support vector machines that share nearest neighbor similarity;

[0012] S2: Preliminary feature selection is performed using the ReliefF algorithm;

[0013] S2.1: First, set the weight w of all features to 0, and set the number of samples n, the feature weight threshold δ, and the number of nearest neighbor samples t;

[0014] Extract training samples n times from the training set, and update the feature weights according to steps S2.2-S2.4:

[0015] S2.2: Randomly select a training sample S from the training set;

[0016] S2.3: Find the t nearest neighbors of the same class for this training sample, denoted as samples H1, H2, ..., Hn. t And the t nearest neighbors of different classes, denoted as samples M1, M2, ..., Mn. t ;

[0017] S2.4: The weight w[i] of the i-th feature is iteratively updated using the following formula.

[0018]

[0019] Where p(H) represents the proportion of samples of the same class, p(M) represents the proportion of samples of different classes, the function diff is used to represent the difference between different features; value(i,S1) represents the value of the i-th feature in sample S1;

[0020] S2.5: Sort the feature weights in descending order and discard features whose feature weights are lower than the feature weight threshold ε;

[0021] S3: Feature reselection based on granular support vector machine training results using shared nearest neighbor similarity;

[0022] S3.1: From the results of the initial feature selection, select two pressure measurement points on different walls as features, and train a granular support vector machine GSVM-SNN based on shared nearest neighbor similarity. The training steps are S3.2-S3.11. Iteratively select features based on the classification accuracy after training to obtain the final selected features and decision model.

[0023] S3.2: Set the threshold k for training the nearest neighbor KNN and the link threshold S in the shared nearest neighbor SNN;

[0024] S3.3: In step S1.1, the different working conditions are used as training set samples. Calculate the distance between the training set samples and construct the pairwise distance matrix D = (a i,j ) n×n Where: x i x jFor sample points,

[0025]

[0026] S3.4: Calculate the similarity between two samples and construct the similarity matrix SI;

[0027] Similarity(x i ,x j ) = size(nl(x i ),nl(x j ))

[0028] Wherein, nl(x i ) is x i The nearest neighbor list is used to calculate the size of the set; the similarity function is used to calculate the similarity between two samples.

[0029] S3.5: Construct a shared nearest neighbor graph SH using the similarity matrix SI as the adjacency matrix;

[0030] S3.6: Calculate the density of each sample. The density is defined as The nearest neighbor list with x i The number of similar points is shown in the following formula:

[0031]

[0032] in, Let be a sample point, and satisfy... The `count` function is used to count the number of similar points that are greater than the link threshold `s`; the `Density` function is used to calculate the sample density.

[0033] S3.7: Divide the points with sample density greater than the link threshold s into particles, and classify the points with density less than the link threshold s into noise points;

[0034] S3.8: For mixed particles {G} containing both starting and non-starting states mix The key information {G} is obtained by using the KNN algorithm for cropping. s};

[0035] S3.9: For a pure particle {G} containing only one state pure}, calculate its KNN connectivity, and extract edge points {G} b};

[0036] S3.10: Preserve noise points as noise particles {G} n};

[0037] S3.11: {G s}、{G b}、{G n After fusion, a granular support vector machine with shared nearest neighbor similarity is trained to obtain the decision model and the classification accuracy of the training set;

[0038] S4: Use the decision model to predict the test set data to obtain the classification results of the intake duct starting / not starting, and classify this typical intake duct starting / not starting state.

[0039] The specific steps for pruning the hybrid particles using the KNN algorithm are as follows;

[0040] S3.8.1: Calculate the nearest neighbor list for all samples in the mixed particles;

[0041] S3.8.2: When all nearest neighbor labels are the same, check whether the sample point has the same label. If they are the same, keep it; otherwise, delete it.

[0042] S3.8.3: When the labels of the nearest neighbors are inconsistent, determine whether they are the same as the category with more neighbor labels. If they are the same, keep them; otherwise, delete them.

[0043] The specific steps in step S3.9 of extracting edge points from pure grains are as follows:

[0044] S3.9.1: Calculate the k-nearest neighbor connectivity of all sample points;

[0045] str(x i ,X)=∑nl(x i )

[0046] ∑nl(x i ) is x i The number of times a sample appears in the k nearest neighbor list; X is the sample set;

[0047] S3.9.2: Retain k nearest neighbor connectivity values ​​greater than threshold D m The sample;

[0048]

[0049] D m This represents the average connectivity of the samples.

[0050] The beneficial effects of this invention are as follows: This invention proposes to quickly perform preliminary feature selection using the ReliefF method, overcoming the problem of excessive computational complexity caused by bringing all features into the training; it adopts granular support vector machine technology with shared nearest neighbor similarity, which, compared with traditional support vector machines, can remove some non-support vectors in the granular part, reducing the number of training samples, and at the same time, parallel granular operations significantly improve training efficiency. Attached Figure Description

[0051] Figure 1 is a basic flowchart of an inlet flow state classification method based on improved support vector machine feature selection.

[0052] Figure 2 is a schematic diagram of the structural dimensions of a typical hypersonic air intake.

[0053] Figure 3 shows the feature weights after iterative calculation using the Relief algorithm.

[0054] Figure 4 shows the feature weights after the initial feature selection.

[0055] Figure 5 shows the shared nearest neighbor graph constructed based on the similarity matrix.

[0056] Figure 6 shows the noise points and information particles obtained from the particle division.

[0057] Figure 7 shows the classification results of the training set after training.

[0058] Figure 8 shows the classification results on the test set. Detailed Implementation

[0059] S1: Numerical simulation data processing of the air intake duct

[0060] S1.1: Wind tunnel tests or fluid dynamics numerical simulations were conducted on typical hypersonic inlets. By setting different parameters such as incoming Mach number, total incoming pressure, and inlet back pressure, inlet flow field data under different operating conditions were obtained. A typical hypersonic inlet structure is shown in Figure 2. In the figure, L1, L2, L3, and L4 are the abscissa distances of the fourth-stage wedges, θ1, θ2, θ3, and θ4 are the angles between the fourth-stage wedges and the horizontal plane, θ5 is the angle between the outer lip and the horizontal plane, and h is the inlet roar height.

[0061] S1.2: Based on the experimental or numerical simulation results and referring to the flow field cloud map, mark the flow state of the intake duct. The starting / stopping state is marked as 1 / 0 respectively. At the same time, extract pressure point data along the DEF and HG walls in Figure 2 to generate a dataset.

[0062] S1.3: Normalize the pressure data of all measuring points on the wall using the following formula;

[0063] Where, p *i p is the normalized dimensionless pressure at a certain pressure measuring point under the current operating conditions. i p represents the pressure data at a pressure measuring point under current operating conditions. min p max These represent the minimum and maximum pressure values ​​under all data acquisition conditions at a certain pressure measuring point;

[0064] S1.4: Divide the dataset samples into training and test sets according to a certain ratio for subsequent training of classification models.

[0065] S2: Preliminary feature selection using the ReliefF algorithm.

[0066] S2.1: First, set the weight w of all features to 0, set the number of sample extractions n, the feature weight threshold δ, and the number of nearest neighbor samples t; extract training samples n times from the training set, and update the weights according to steps S2.2-S2.4.

[0067] S2.2: Randomly select a training sample S from the training set;

[0068] S2.3: Find the t nearest neighbors of the same class for this training sample, denoted as samples H1, H2, ..., Hn. t And the t nearest neighbors of different classes, denoted as samples M1, M2, ..., Mn. t ;

[0069] S2.4: The weight w[i] of the i-th feature is iteratively updated using the following formula.

[0070]

[0071] Where p(H) represents the proportion of samples of the same class, p(M) represents the proportion of samples of different classes, and the function diff is used to represent the differences between different features; value(i,S1) represents the value of the i-th feature in sample S1. The results of iterative calculation of feature weights on the test data are shown in Figure 3;

[0072] S2.5: Sort the feature weights in descending order, perform preliminary feature selection according to the feature weights, and discard features with weights lower than ε. The test data results are shown in Figure 4.

[0073] S3: Feature reselection based on granular support vector machine training results using shared nearest neighbor similarity

[0074] S3.1: In the initial feature selection, pressure measurement points on two different walls are chosen as features for training a granular support vector machine (GSVM-SNN) with shared nearest neighbor similarity, as per training steps S3.2-S3.11. Then, based on the classification accuracy after training, iterative feature selection is performed to obtain the final selected features and the trained model.

[0075] S3.2: Set the threshold k for training the KNN and the connection threshold s in the SNN.

[0076] S3.3: Calculate the distance between training samples and construct the pairwise distance matrix D = (a i,j ) n×n Where: x i x j For sample points,

[0077]

[0078] S3.4: Calculate the similarity between two samples and construct the similarity matrix SI;

[0079] Similarity(x i ,x j ) = size(nl(x i ),nl(x j ))

[0080] Wherein, nl(x i ) is x i The nearest neighbor list is used to calculate the size of the set; the similarity function is used to calculate the similarity between two samples.

[0081] S3.5: Construct a shared nearest neighbor graph SH using the similarity matrix SI as the adjacency matrix. The shared nearest neighbor graph SH constructed from the test data is shown in Figure 5.

[0082] S3.6: Calculate the density of each sample. The density is defined as The nearest neighbor list with x i The number of similar points is shown in the following formula:

[0083]

[0084] in, Let be a sample point, and satisfy... The `count` function is used to count the number of similar points that are greater than the link threshold `s`; the `Density` function is used to calculate the sample density.

[0085] S3.7: Points with a sample density greater than the link threshold s are divided into particles, and points with a density less than the link threshold s are classified as noise points. The density division of sample points yields noise points and several information particles X→{G1,G2,…,G...} i The results of the test data partitioning are shown in Figure 6.

[0086] S3.8: For mixed particles {G} containing both starting and non-starting states mix The KNN algorithm is used for cropping to obtain key information, namely support vector points {G}. s};

[0087] S3.9: For a pure particle {G} containing only one state, either started or not started. pure}, calculate its KNN connectivity. The algorithm steps are shown in S5. For connections with a connectivity greater than D, ... m Deleting sample points and extracting edge points {G b};

[0088] S3.10: The noise points obtained from granulation are retained as a single particle, called the noise particle {G}. n};

[0089] S3.11: {G s}、{G b}、{G n The results were fused to train the support vector machine, resulting in the decision model and the classification accuracy of the training set. The classification results of the training set are shown in Figure 7.

[0090] The specific steps for pruning mixed particles using the KNN algorithm are as follows:

[0091] S3.8.1: Calculate the nearest neighbor list for all samples in the mixed particles;

[0092] S3.8.2: If all neighbor labels are the same, check if the sample point is the same as them. If they are the same, keep it; otherwise, delete it.

[0093] S3.8.3: If the neighbor's label is inconsistent, determine whether it is the same as the category with more neighbor labels. If they are the same, keep it; otherwise, delete it.

[0094] The specific steps for extracting edge points from pure grains are as follows:

[0095] S3.9.1: Calculate the nearest neighbor connectivity of all sample points;

[0096] str(x i ,X)=∑nl(x i )

[0097] The above formula is used to solve for the sample point x. iThe nearest neighbor connectivity, denoted by X in the sample set, is ∑nl(x i ) represents x i Number of times it appears in the k nearest neighbor list

[0098] S3.9.2: Retain nearest neighbor connectivity greater than threshold D m The sample, where D m Defined as the average connectivity of the samples;

[0099]

[0100] S4: Using the trained model to predict the test set data yields the classification results for whether the intake is started or not, as shown in Figure 8. Based on the wall pressure data of this typical intake model, the start / stop state of this typical intake can be classified.

Claims

1. An inlet flow state classification method based on improved support vector machine feature selection, characterized in that, Based on a typical hypersonic inlet model, an inlet flow state dataset is constructed. Based on the ReliefF feature selection method, feature weights are calculated for initial feature selection. Then, based on the granular support vector machine GSVM-SNN with shared nearest neighbor similarity, two features are selected from the initially chosen features for training. The two features with the highest training accuracy are retained as the final features, and the trained classification model is used as the decision model. The decision model is used on the test set, and based on the test training effect, the flow state of this typical hypersonic inlet is finally obtained. The specific steps include: S1: Numerical simulation data processing of the inlet; S... 1.1: Conduct wind tunnel tests or numerical simulations of fluid dynamics on typical hypersonic inlets to obtain inlet flow field data under different operating conditions; S1.2: Based on the wind tunnel test or numerical simulation results and referring to the flow field cloud map, mark the flow state of the inlet and extract pressure point data to generate a dataset; S1.3: Normalize the pressure data of all measuring points on the wall of a typical hypersonic inlet, with each measuring point corresponding to a feature; S1.4: Divide the dataset samples into training and testing sets according to the proportion for subsequent training of a granular support vector machine sharing nearest neighbor similarity; S2: Preliminary feature selection is performed using the ReliefF algorithm; S2.1: First, set the weight w of all features to 0, and set the number of sample extractions n, the feature weight threshold δ, and the number of nearest neighbor samples t; Extract training samples n times from the training set, and update the feature weights according to steps S2.2-S2.4: S2.2: Randomly select a training sample S from the training set; S2.3: Find the t nearest neighbors of the same type of the training sample, denoted as samples H1, H2, ..., Hn. t And t nearest neighbors of different classes, denoted as samples M1, M2, ..., Mn. t S2.4: The weight w[i] of the i-th feature is iteratively updated using the following formula. Where p(H) represents the proportion of samples of the same class, p(M) represents the proportion of samples of different classes, and the function diff is used to represent the difference between different features; value(i, S1) represents the value of the i-th feature in sample S1; S2.5: Sort the feature weights in descending order and discard features whose feature weights are lower than the feature weight threshold ε; S3: Feature reselection based on the training results of granular support vector machine with shared nearest neighbor similarity; S3.1: Select two pressure measurement points of different walls as features from the preliminary feature selection results, and train granular support vector machine GSVM-SNN based on shared nearest neighbor similarity. The training steps are S3.2-S3.11; Iteratively select features based on the classification accuracy after training to obtain the final selected features and decision model; S3.2: Set the threshold k value of training nearest neighbor KNN and the link threshold S in shared nearest neighbor SNN; S3.3: Use the different working conditions in step S1.1 as training set samples, calculate the distance between training set samples, and construct the distance matrix D between each pair of samples = (a i，j ) n×n Where: x i x j For sample points, S3.4: Calculate the similarity between two samples and construct the similarity matrix SI; Similarity(x i x j ) = size(nl(x i ), nl(x j Where, nl(x) i ) is x i The nearest neighbor sample list is used, and `size` is used to calculate the set size; the `Similarity` function is used to calculate the similarity between two samples; S3.5: Construct a shared nearest neighbor graph `SH` using the similarity matrix `SI` as the adjacency matrix; S3.6: Calculate the density of each sample. The density is defined as The nearest neighbor list with x i The number of similar points is shown in the following formula: in, Let be a sample point, and satisfy... `count` is used to count the number of similar points greater than the link threshold `s`; the function `Density` is used to calculate the sample density; S3.7: Points with a sample density greater than the link threshold `s` are divided into particles, and points with a density less than the link threshold `s` are classified as noise points; S3.8: For mixed particles {G} containing both active and inactive states... mix The key information {G} is obtained by using the KNN algorithm for cropping. s };S3.9:For a pure particle {G} containing only one state pure }, calculate its KNN connectivity, and extract edge points {G} b };S3.10: Preserve noise points as noise particles {G n };S3.11:Transfer {G s }、{G b }、{G n After fusion, a granular support vector machine with shared nearest neighbor similarity is trained to obtain the decision model and the classification accuracy of the training set; S4: The decision model is used to predict the test set data to obtain the classification result of the air intake starting / not starting, and the typical air intake starting / not starting state is classified.

2. The inlet flow state classification method based on improved support vector machine feature selection according to claim 1, characterized in that, The specific steps for pruning the mixed particles using the KNN algorithm are as follows: S3.8.1: Calculate the nearest neighbor list for all samples in the mixed particles; S3.8.2: When all nearest neighbor labels are consistent, check whether the sample point is the same as the label. If they are the same, keep it; if they are different, delete it; S3.8.3: When the labels of the nearest neighbors are inconsistent, determine whether it is the same as the class with more neighbor labels. If they are the same, keep it; if they are different, delete it.

3. The inlet flow state classification method based on improved support vector machine feature selection according to claim 1 or 2, characterized in that, The specific steps in step S3.9 for extracting edge points from pure grains are as follows: S3.9.1: Calculate the nearest neighbor connectivity of all sample points; str(x i ,X)=∑nl(x i )∑nl(x i ) is x i The number of times a sample appears in the k nearest neighbor list; X is the sample set; S3.9.2: Keep the nearest neighbor connectivity greater than the threshold D. m The sample; D m This represents the average connectivity of the samples.

Images