Identification method of large-scale supersonic inlet flow state

Abstract

The present invention provides a method for identifying flow states in large-scale supersonic inlets. This method is a hybrid algorithm combining F-KNN and I-DAGSVM, designed to address the problem of identifying flow states in large-scale supersonic inlets. The hybrid algorithm first uses F-KNN to quickly find k accurate nearest neighbors for each test sample. Then, based on these nearest neighbors, I-DAGSVM is used to train a local classifier for classification. This hybrid algorithm combines the efficient search speed of F-KNN with the excellent classification performance of I-DAGSVM. It also overcomes the shortcomings of the F-KNN algorithm, which often misclassifies when the number of nearest neighbors is equal, and the I-DAGSVM algorithm, which is time-consuming to train when processing large data sets. This ensures excellent classification results within a reasonable runtime.

Description

Technical Field

[0001] The present invention relates to the technical field of supersonic inlet flow state identification, and in particular to a method for identifying large-scale supersonic inlet flow state. Background Art

[0002] As a key component of an air-breathing supersonic propulsion system, the flow state of a supersonic inlet has a direct impact on the overall system performance. For example, inlet surge is an extremely unstable flow state, often accompanied by violent, periodic shock wave oscillations. This severely degrades the thrust characteristics of the entire system and poses a constant risk of flameout. Furthermore, the periodic thermal loads generated by shock wave oscillations can dramatically degrade the controllability of the entire system and may even directly lead to structural damage, posing a threat to the safety of the aircraft and personnel. From a protection and control perspective, since the flow characteristics of a supersonic inlet vary under different flow states, different control measures are implemented to ensure normal operation under various flow conditions. Therefore, flow state identification is crucial for supersonic inlets.

[0003] At present, the methods for identifying the flow state of the air inlet mainly include methods based on high-speed schlieren photography and data-driven methods. The method based on high-speed schlieren photography is intuitive and reliable, but requires a complex optical path system. It is only feasible in ground tests and is obviously unrealistic in actual flights. The data-driven method is to use CFD simulation technology or experiments to collect samples under different conditions and different flow states, and then introduce machine learning and other technologies to train these samples, and use the trained model to distinguish the flow state to which the samples to be classified belong. This method does not require a complex optical path system and is suitable for both ground tests and actual flights. It can also use new data continuously generated in ground tests or actual flights to correct existing models, fully tap the effective information in historical data and real-time data, and improve data utilization while making the model more adaptable to external conditions. The present invention adopts a data-driven method to solve the problem of flow state identification in supersonic air inlets.

[0004] Currently, research on data-driven methods is primarily based on CFD steady-state simulation data, typically consisting of only a few hundred samples, whereas actual measurement devices typically collect large amounts of dynamic test data. Consequently, previous research often overlooks the dynamic characteristics of actual tests, as well as the influence of factors such as noise and interference. On the other hand, a vast amount of dynamic data has been accumulated from previous experiments. Fully utilizing this data could yield valuable insights that could even guide related experiments. However, the previously used SVM and its variants are not suitable for processing large amounts of data due to their time-consuming training process. Summary of the Invention

[0005] The purpose of this invention is to provide an efficient hybrid algorithm for identifying inlet flow states directly based on a large amount of dynamic test data. The algorithm not only has excellent classification performance but also has a reasonable runtime when processing large amounts of data.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] This paper proposes a hybrid algorithm for large-scale supersonic inlet flow state identification. The algorithm first uses Fast K-Nearest-Neighbors (F-KNN) to screen training samples related to the test sample. It then uses an Improved Directed-Acyclic-Graph Support-Vector Machine (I-DAGSVM) to train a local classifier to identify the flow state to which the test sample belongs.

[0008] First, we briefly introduce the shortcomings of K-Nearest-Neighbors (KNN) and Directed-Acyclic-Graph Support-Vector-Machine (DAGSVM). Then, we improve these two methods to form F-KNN and I-DAGSVM. Finally, we combine the two improved methods to form the hybrid algorithm proposed in this paper.

[0009] When processing big data, KNN, on the one hand, needs to calculate the distance between each test sample and all training samples, and all training samples need to be sorted according to the size of the distance, which is time-consuming and has a relatively low search efficiency. On the other hand, when classifying based on k nearest neighbor samples, if the number of samples belonging to different categories is equal, KNN cannot make further judgments, which can easily lead to misclassification. Therefore, in order to address the problem of low efficiency of KNN in processing big data, the present invention introduces an efficient indexing strategy, which only requires calculating the distance between each test sample and a small number of training samples, and does not require sorting, thus saving a lot of time and improving search efficiency. In response to the misclassification of KNN, the present invention chooses to combine it with DAGSVM and optimizes the structure of DAGSVM.

[0010] DAGSVM can solve the misclassification of KNN very well, because it adopts a tree structure. When making judgments from top to bottom, it eliminates a category label each time, and finally produces a deterministic classification result. However, if the SVM with weaker distinguishing ability is placed at the top of the tree structure, there will be a high probability of misclassification from the beginning, thereby increasing the accumulation of errors. On the other hand, when DAGSVM is trained based on big data, it is very time-consuming because it adopts the idea of ​​dynamic programming. In response to the situation where the unreasonable tree structure of DAGSVM leads to misclassification, the present invention optimizes its structure by using an indicator to measure the separability between classes, and places the SVM with stronger distinguishing ability at the top of the tree structure, thereby reducing the accumulation of errors. In response to the shortcoming that DAGSVM is very time-consuming when training big data, the present invention combines it with KNN, and trains the layout classifier based on the found nearest neighbor samples, thereby greatly shortening the training time, and DAGSVM usually has higher classification ability than other algorithms when processing small amounts of data.

[0011] In summary, the hybrid algorithm proposed in the present invention mainly includes the following two steps:

[0012] In step 1, F-KNN is used to quickly find the k nearest neighbors of each test sample. Based on an efficient indexing strategy, F-KNN can quickly find the exact nearest neighbors of each test sample, thereby greatly improving the algorithm's search efficiency without sacrificing accuracy. The specific process is as follows:

[0013] First, a benchmark sample R is randomly determined, and an ordered queue and an index table are established based on the distance between each training sample and R. The index table records the first The position of each sample in the ordered queue and the distance from the corresponding training sample to R are calculated. Given a test sample x, the distance d between x and R is first calculated. Then, using a bisection method, the range of training samples q closest to R by distance d is determined from the index table. Next, q is found within this small range, and k training samples are selected from the ordered queue with q as the center as the initial nearest neighbors of x. By continuously searching for training samples before and after q that meet the requirements and replacing the initial nearest neighbors, the k precise nearest neighbors of x are eventually obtained.

[0014] It is worth noting that if the training sample s is a neighboring sample of x, then the following inequality will hold:

[0015] |d xR -d sR |≤(d k ) max (1)

[0016] Among them, d xR and dsR are the distances between x and s and R, (d k ) max is the farthest distance from x among the k nearest neighbors.

[0017] The proof of formula (1) is given below:

[0018] If s is a neighboring sample of x, then:

[0019] d sx ≤(d k ) max (2)

[0020] The following three cases are discussed:

[0021] (i) When 0≤d xR ≤d sR and 0≤d sx ≤d sR hour,

[0022] (a) If x, s, and R are on the same straight line, then d sx +d xR =d sR ;

[0023] (b) If x, s, and R are not on the same straight line, then d sR <d sx +d xR ;

[0024] Combining (a) and (b), we can see that

[0025] (ii) When 0≤d sR ≤d xR and 0≤d sx ≤d xR hour,

[0026] (a) If x, s, and R are on the same straight line, then d sx +d sR =d xR ;

[0027] (b) If x, s, and R are not on the same straight line, then d xR <d sx +d sR ;

[0028] Combining (a) and (b), we can see that

[0029] (iii) When 0≤d sR ≤d sx and 0≤d xR ≤dsx When |d xR -d sR |≤(d k ) max Obviously true.

[0030] The entire process of the F-KNN algorithm can be summarized as Algorithm 1.

[0031]

[0032]

[0033] In step 2, a local DAGSVM classifier is trained based on the k nearest neighbor samples found in step 1. A comprehensive metric that considers both the class center distance and the class sample distribution is used to measure the separability of the two classes. The SVMs corresponding to the two classes with greater separability are placed at the top of the DAGSVM tree structure, thereby reducing error accumulation and improving the reliability of the classification results. The specific process is as follows:

[0034] Assume N i is the number of samples in the i-th category, is the kth sample, then the mean vector and radius of the i-th class can be calculated by (3) and (4) respectively:

[0035]

[0036]

[0037] Similarly, the mean vector and radius of the jth class can be calculated. Then, the distance D between the i-th class and the j-th class is ij It can be obtained by subtracting the radius of each class from the distance between the average vectors of the two classes:

[0038] D ij =||m i -m j ||2-r i -r j (5)

[0039] The mean vector represents the center of the class, while the radius represents the distribution of samples within the class. A smaller radius indicates a denser distribution of samples. Therefore, a larger distance between class centers and a smaller radius indicate greater distinction between the two classes. The entire I-DAGSVM algorithm process can be summarized as Algorithm 2.

[0040]

[0041] Beneficial effects: 1. F-KNN uses an efficient indexing strategy to quickly find the precise nearest neighbor samples for each test sample. Therefore, compared with KNN, F-KNN's classification accuracy remains unchanged, but the search speed is greatly improved;

[0042] 2. I-DAGSVM uses a comprehensive measure of inter-class separability and sequentially uses SVMs with greater discriminative power for discrimination, thereby reducing error accumulation and improving the reliability of classification results. Therefore, compared with DAGSVM, I-DAGSVM has improved classification accuracy.

[0043] 3. This paper proposes a hybrid algorithm combining F-KNN and I-DAGSVM to solve the problem of flow state identification in large-scale supersonic inlets. This hybrid algorithm first uses F-KNN to quickly find k accurate nearest neighbors for each test sample. Then, based on these nearest neighbors, I-DAGSVM is used to train a local classifier for classification. This algorithm combines the efficient search speed of F-KNN with the excellent classification performance of I-DAGSVM. It also overcomes the shortcomings of the F-KNN algorithm, which often misclassifies when the number of nearest neighbors is equal, and the time-consuming training process of the I-DAGSVM algorithm when processing large data sets. This ensures excellent classification results within a reasonable runtime. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The following is a test model and its schematic diagram;

[0045] Figure 2 It is the static pressure time history diagram during the throttling process;

[0046] Figure 3 This is the experimental result of the hybrid algorithm. DETAILED DESCRIPTION

[0047] In order to better reflect the advantages of the present invention,

[0048] In the case of identifying supersonic inlet flow states, sample data from a certain type of external pressure supersonic inlet is first collected and preprocessed. The specific process is as follows: all samples are divided into training and test sets. Then, the ReliefF algorithm is used to select the sensor positions that are most effective in identifying the current inlet flow state. Each sample is composed of these selected sensor signals.

[0049] Next, for each test sample x to be judged, quickly find its k accurate neighboring samples. The specific process is as follows: First, randomly determine a benchmark sample R, and establish an ordered queue and an index table based on the distance between each training sample and R. The index table records the first The position of each training sample in the ordered queue and its distance to R are calculated. Given a test sample x, the distance d between x and R is first calculated. Then, a bisection method is used to determine the range of training samples q whose distance to R is closest to d in the index table. Next, q is found within this range, and k training samples are selected from the ordered queue with q as the center as the initial nearest neighbors of x. By continuously searching for training samples before and after q that meet the requirements and replacing the initial nearest neighbors, the k precise nearest neighbors of x are eventually obtained.

[0050] If the training sample s is a neighboring sample of x, then the following inequality will hold:

[0051] |d xR -d sR |≤(d k ) max (1)

[0052] Among them, d xR and d sR are the distances between x and s and R, (d k ) max is the farthest distance from x among the k nearest neighbors. The above process of quickly finding the k nearest neighbors of the test sample x can be summarized as Algorithm 1.

[0053]

[0054]

[0055]

[0056] Based on the k nearest neighbors of the test sample x, a local I-DAGSVM classifier is trained and its parameters are optimized using the grid search method to determine the category to which x belongs. The specific process is as follows:

[0057] Assume N i is the number of samples of category i among the k nearest neighbor samples, is the kth sample, then the mean vector and radius of the i-th class can be calculated by (2) and (3) respectively:

[0058]

[0059]

[0060] Similarly, the mean vector and radius of the j-th class sample can be calculated. Then, the distance D between the i-th class and the j-th class is ij It can be obtained by subtracting the radius of each class from the distance between the average vectors of the two classes:

[0061] Dij =||m i -m j ||2-r i -r j (4)

[0062] The mean vector represents the center of the class, while the radius represents the distribution of samples within the class. A smaller radius means a denser distribution of samples. Therefore, the greater the distance between class centers and the smaller the radius, the easier it is to distinguish between the two classes.

[0063] When determining the flow state to which x belongs, the largest D ij The corresponding binary SVM is used for discrimination, and the flow state of x is finally obtained. The above process of establishing the local I-DAGSVM can be summarized as Algorithm 2.

[0064]

[0065]

[0066] In summary, the present invention first uses F-KNN to find k nearest neighbor samples for each test sample, and then establishes a local I-DAGSVM based on these nearest neighbor samples to determine the category to which the test sample belongs.

[0067] All experiments were performed on a laptop with an Intel® Core™ i7-7700HQ CPU, 2.80 GHz clock speed, 8 GB memory, Windows 10 system, and MATLAB 2018b version.

[0068] The research object of this invention is a type of dual external pressure supersonic inlet, such as Figure 1 As shown in the figure, the leading edge of the lip shield of the inlet is blunted to prevent large-scale flow separation inside the lip shield. The upper compression surface adopts a partial isentropic design, consisting of a first-stage inclined wedge and a first-stage isentropic surface. Downstream of the inlet is a single-sided expansion channel, followed by a section of constant-flow channel. Optical glass is embedded on both sides of the inlet flow channel for observing the internal flow field, and the viewing window covers the core area of ​​the internal flow channel. To simulate the downstream throttling effect in actual operation, a wedge-shaped throttle cone is installed at the inlet outlet, with its leading edge located at the center height of the channel outlet. Driven by a stepper motor, the throttle cone can be translated axially to establish a controllable blockage degree at the inlet outlet. To obtain dynamic pressure information of the inlet flow field during throttling, a total of 31 dynamic pressure measurement points are set on the centerlines of the upper and lower inner surfaces of the flow channel. The sensors are numbered C01 to C17 and R01 to R14 according to their location.

[0069] Figure 2The static pressure time histories of sensors C01 and R01 during throttling are presented. Before surge occurs, the intake duct is in a stable flow state. As the blockage increases, the intake duct experiences three distinct flow states: small surge, mixed surge, and large surge.

[0070] Based on the four flow states described above, the raw test data can be divided into four categories of samples. If the intake is in a stable state, the corresponding label is 1; if the intake is in a slight gasp state, the corresponding label is 2; if the intake is in a mixed gasp state, the corresponding label is 3; and if the intake is in a large gasp state, the corresponding label is 4. In practice, due to weight and structural limitations, it is not suitable to install too many sensors on the intake. Therefore, the ReliefF algorithm is first used to determine the optimal sensor locations: R01, R02, R03, C01, and C14. After this processing, each sample is composed of signals from these five sensors. Then, all 348,621 collected samples are divided into a training set (244,035 samples) and a test set (104,586 samples) according to the commonly used 7:3 ratio.

[0071] The effectiveness of F-KNN, I-DAGSVM and hybrid algorithms will be demonstrated below.

[0072] (1) F-KNN

[0073] To demonstrate the effectiveness of the F-KNN algorithm in training sets of varying sizes, the training set sizes were set to 30,000, 40,000, 50,000, and 244,035, respectively, while the test set remained unchanged. The value of k ranged from [1 to 10]. The results are shown in Table 1.

[0074] The data in the table shows that in every case, F-KNN achieves the same accuracy as KNN, but its runtime is significantly shorter than KNN. This accuracy is achieved because F-KNN finds the same k nearest neighbors for each test sample as KNN. However, KNN requires calculating the distance between each test sample and all training samples and sorting all training samples by distance, which is very time-consuming when the data volume is very large. F-KNN, on the other hand, first narrows the search for nearest neighbors, so it only needs to calculate the distance between the test sample and a small number of training samples. Furthermore, F-KNN eliminates the step of sorting all training samples. Therefore, F-KNN runs in much less time than KNN.

[0075] (2) I-DAGSVM

[0076] In order to show the excellent classification performance of the I-DAGSVM algorithm in training sets of different sizes, the training set sizes were set to 30,000, 40,000, and 50,000 respectively (the training of 244,035 samples was very time-consuming, so it was not calculated). The grid search method was used to select the best SVM parameters, and the parameter range was set to [2 -2 ,2 10 The training process for DAGSVM and I-DAGSVM is the same. The only difference is that when distinguishing the categories of test samples, the structure of I-DAGSVM is optimized using a metric that measures inter-class separability, thereby reducing error accumulation and improving the reliability of the classification results. The results are shown in Table 2.

[0077] Table 2 shows that, given the same parameters, I-DAGSVM achieves higher accuracy than DAGSVM, demonstrating the effectiveness of optimizing the DAGSVM architecture. However, the training time for both approaches is very long. Combined with Table 1, we can see that in each case, DAGSVM and I-DAGSVM achieve higher accuracy than the best F-KNN approach, but their runtimes are significantly longer than F-KNN.

[0078] In fact, when the amount of data is very large, even with a fast training method such as Sequential Minimization Optimization (SMO), training SVM is time-consuming, because the time complexity of SMO is about o(n 2 ), the number of operations is roughly proportional to the square of the number of samples. On the other hand, when both the amount of data and the number of categories are small, SVM classification performance generally outperforms other algorithms. Therefore, to overcome the time-consuming training limitations of SVM when processing large data sets while leveraging its excellent classification performance, we combined it with F-KNN.

[0079] (3) Hybrid algorithm

[0080] The hybrid algorithm proposed in this paper first uses F-KNN to find training samples related to the test samples, and then uses I-DAGSVM to train a local classifier for discrimination. The results are shown in Table 3 and Figure 3 shown.

[0081] As can be seen, in every case, the hybrid algorithm achieves higher accuracy than F-KNN. Furthermore, as shown in Table 2, the hybrid algorithm's runtime is significantly shorter than that of I-DAGSVM. Therefore, the hybrid algorithm combines the efficient search speed of F-KNN with the excellent classification performance of I-DAGSVM. It also overcomes the shortcomings of F-KNN, which suffers from misclassification when the number of neighbor samples belonging to different classes is equal, and the time-consuming training process of I-DAGSVM when processing large data sets. This ensures that the hybrid algorithm can achieve excellent classification results within a reasonable runtime.

[0082] Table 1 Comparison results between KNN and F-KNN

[0083]

[0084] (I) Number of training samples = 30,000

[0085]

[0086] (II) Number of training samples = 40,000

[0087]

[0088] (III) Number of training samples = 50,000

[0089]

[0090] (IV) Number of training samples = 244035

[0091] Table 2 Comparison between I-DAGSVM and DAGSVM

[0092]

[0093] Table 3 Experimental results of hybrid algorithm

[0094]

[0095] (I) Number of training samples = 30,000

[0096]

[0097] (II) Number of training samples = 40,000

[0098]

[0099] (III) Number of training samples = 50,000

[0100]

[0101] (IV) Number of training samples = 244035

[0102] It should be noted that the above embodiments are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent replacements or substitutions made on the basis of the above embodiments all fall within the scope of protection of the present invention.

Claims

1. A method for identifying large-scale supersonic inlet flow conditions, characterized in that: The steps include: Step 1: Collect the dynamic pressure time history signal of each pressure measuring point on the inlet wall during the throttling process of the supersonic inlet at rated state; Step 2: Using the collected signals as samples and the intake flow state categories corresponding to the collected signals as sample labels, a sample set is established and divided into a training set and a test set. Step 3: Use the F-KNN algorithm to find the k nearest neighbor samples of each test sample x in the training set; Step 4: Build DAGSVMs between the two categories of k nearest neighbor samples and calculate the separability between categories. Organize each DAGSVM into a tree structure from top to bottom in the order of separability between categories from large to small. Step 5: When judging the test sample, it is input into the sub-classifiers in the tree structure in sequence. After each judgment, the categories with smaller probabilities are eliminated until only one category remains. The test sample is then judged to belong to the remaining category. In step 1, the ReliefF algorithm is used to select the most effective sensor position for identifying the current flow state of the inlet duct, and the sample consists of the selected sensor signals; The specific steps of step 3 include: Step 3.1: Randomly determine a benchmark sample R in the test sample set, and establish an ordered queue and an index table based on the distance between each training sample and R; the index table records the first The position of the training sample in the ordered queue and its distance to R; Step 3.2: Calculate the distance d between the test sample x and R, and use the bisection method to determine the range of training samples q with the closest distance d to R in the index table; Step 3.3: Find q within the range and take k training samples from the ordered queue with q as the center as the initial nearest neighbor samples of x. By continuously searching for training samples before and after q that meet the conditions and replacing the initial nearest neighbor samples, k accurate nearest neighbor samples of x are finally obtained. The separability between classes in step 4 is represented by the distance between the class centers of the two classes of samples; The steps for calculating the distance between cluster centers are as follows: Assume N i is the number of samples in the i-th category, is the kth sample, then the mean vector and radius of the i-th class are calculated by (1) and (2) respectively: Calculate the average vector and radius of the jth class; then the distance D between the i-th class and the j-th class ij The distance between the mean vectors of the two classes minus their respective radii is obtained: D ij =||m i -m j ||2-r i -r j (3) The mean vector represents the center of the class, and the radius represents the distribution of samples in the class. The smaller the radius, the denser the sample distribution. Therefore, the greater the distance between class centers and the smaller the radius, the easier it is to distinguish between the two classes.

Images