Intelligent identification of shock wave structure and whistling mode analysis method based on time series images

Abstract

The present invention discloses a method for intelligent recognition of shock wave structure and whistling modal analysis based on time-series images, which relates to the field of flow field shock wave recognition, including: S1, preparing an image dataset with aircraft shock wave structure markers based on schlieren time-series images; S2, constructing a pixel-level semantic segmentation model for intelligent detection and recognition of shock wave structure; S3, using the pixel-level semantic segmentation model to perform intelligent detection and recognition of shock wave structure on the input schlieren time-series images; wherein, in the pixel-level semantic segmentation model, a composite loss function considering the target contour characteristics of the shock wave structure is introduced L CP The present invention provides a method for intelligent recognition of shock wave structure and whistling mode analysis based on time series images, by introducing a composite loss function that considers the target contour characteristics of the shock wave structure. L CP , which enables it to have more precise contour capture capability and more accurate complex shock wave structure recognition capability when dealing with shock wave structure targets with complex boundary shapes.

Description

Technical Field

[0001] The present invention relates to the field of flow field shock wave identification, and more specifically, to a method for intelligent shock wave structure identification and whistling mode analysis based on time-series images. Background Art

[0002] Noise research and noise reduction design are crucial for the development of aerospace vehicles. The nozzle, a crucial propulsion system structure, generates a strong noise source—the jet. Supersonic jets, characterized by their characteristic flow structures such as shear layers, shock grids, and turbulent vortices, have long garnered significant attention from researchers. Understanding the evolution of jet field structures and the physical mechanisms underlying their interactions holds significant theoretical value, providing theoretical support for a wide range of future engineering applications.

[0003] Among traditional research methods, CFD numerical calculation methods have limited simulation capabilities for complex unsteady working conditions; wind tunnel test methods mainly use pulsating pressure sensors to collect original wall pressure signals, and then analyze frequency domain characteristics. For example, an intelligent flow field shock wave identification and visualization method with patent application number 202411771394.0 adopts a layered progressive design through data preprocessing modules, feature extraction modules and topological association modules, so that the output of each step can be used as an efficient input for the next step, ensuring the logic and consistency of the entire method. However, as mentioned above, this method is to analyze the data in the hope of achieving efficient and accurate identification of shock wave characteristics under complex flow field conditions. However, the problem is that the measurement layout points of this wall sensor are limited, the wiring is difficult, and it is easy to interfere with the flow field. At the same time, more requirements are put forward for the model conditions of special complex configurations.

[0004] The flow field schlieren images acquired by high-speed photography contain rich non-steady information. Mature schlieren visualization technology provides a large number of high-quality images with rich flow characteristics, and has unique advantages such as wide field of view and non-contact acquisition. With the rapid development of computer vision and deep learning technology, this provides a new solution for exploring the evolution law of jet structure and realizing efficient, accurate and intelligent shock wave flow feature recognition and analysis. For example, a method for identifying and processing flow field features of schlieren images in impeller machinery blade cascade tests with patent application number: 202110697965.0 can effectively remove background noise in blade cascade schlieren images by adopting the L0 gradient minimization algorithm to retain flow field detail information such as weak compression waves / expansion waves and boundary layer fluid aggregation in the blade cascade flow field. However, this method mainly focuses on the preprocessing of images and cannot be directly applied to extract shock wave structures in images, that is, it cannot be applied to wind tunnel flow field tests;

[0005] Furthermore, the existing art paper "Research on Multi-Scale Seismic Damage Identification and Assessment of Buildings Enhanced by Deep Learning with Geometric Constraints" proposes a Geometric Consistency Enhanced loss function (GCEloss) in its second chapter. This loss function comprises a Geometric Consistency Loss (GC loss) and a Cross-Entropy Loss (CE loss). GCloss explicitly utilizes the geometric features of image segmentation lines and regions to calculate the deviation between the segmentation line length, segmentation line curvature, and segmentation region area between the predicted image and the true labeled image. This is then combined with CE loss to guide model optimization. GCE loss leverages both the ability of GCloss to constrain the geometric features of segmented targets during model training and the stability and smoothness of CE loss during loss reduction. However, the geometric consistency enhancement loss function (GCE loss) is primarily used for the identification and assessment of multi-scale earthquake damage in buildings. This applies primarily to the multi-type (degree of damage) and multi-scale (large, medium, and small) earthquake structural damage of individual buildings from the perspectives of high-altitude satellites, low-altitude drones, and near-field drones. Specifically, high-altitude satellite imagery easily presents dense small targets with complex, fuzzy edges; low-altitude drone imagery is susceptible to weather interference and mutually occluded targets with varying scales and distributions; and near-field drone imagery presents component-level targets with multiple types of coupled and multiple damage states. Furthermore, multi-scale earthquake damage targets in buildings differ significantly from shock wave structures. Therefore, the GCE loss cannot be directly applied to the identification and detection of shock wave structures in wind tunnel flow field tests.

[0006] In summary, it is of great significance to conduct data mining on shock wave structure images of aircraft flow fields with rich flow characteristics, explore non-contact rapid prediction methods of pulsating pressure loads, and develop new whistling modal analysis methods. Summary of the Invention

[0007] An object of the present invention is to solve at least the above problems and / or disadvantages and to provide at least the advantages which will be described hereinafter.

[0008] In order to achieve these objectives and other advantages of the present invention, a method for intelligently identifying shock wave structures based on time-series images is provided, comprising:

[0009] S1. Prepare an image dataset with aircraft shock wave structure markers based on Schlieren time-series images;

[0010] S2. Build a pixel-level semantic segmentation model for intelligent detection and recognition of shock wave structures;

[0011] S3, using pixel-level semantic segmentation model to intelligently detect and identify shock wave structures on the input schlieren time series images;

[0012] Among them, in the pixel-level semantic segmentation model, a composite loss function considering the contour features of the shock wave structure target is introduced L CP , and the L CP It is characterized by the following formula:

[0013]

[0014] In the above formula, L CE is the cross entropy loss function, L DC is the Dice coefficient loss function, L CT is the contour loss function, L CL is the contour length loss function, L CC is the contour curvature loss function.

[0015] Preferably, in S1, the process of acquiring the image dataset includes:

[0016] S10, selecting 200 images at equal intervals from the 2000 collected schlieren time-series images;

[0017] S11, labeling the selected 200 images;

[0018] S12, perform 9 data augmentations on the 200 labeled images to obtain a dataset of 2000 images;

[0019] S13, divide the image dataset into training set, validation set and test set according to 7:2:1;

[0020] In S11, the labeling operation is to mark the shock wave structure in the image with a polygonal frame using a labeling tool to obtain the corresponding shock wave structure positioning information;

[0021] In S12, the enhancement refers to processing the selected 200 images by randomly combining noise addition, brightness, translation, and mirroring.

[0022] Preferably, in S2, the construction of the pixel-level semantic segmentation model includes:

[0023] S20. Establish a semantic segmentation network for intelligent detection and identification of shock wave structures;

[0024] S21. Design a composite loss function for semantic segmentation networks L CP ;

[0025] S22. Input the training set in the image dataset into the semantic segmentation network and use the composite loss function L CP Perform iterative training to obtain the trained model weight parameters;

[0026] During the iterative training process, after multiple hyperparameter adjustments and optimizations, and using the validation set in the image dataset to evaluate the model accuracy, the optimal settings for the pixel-level semantic segmentation model are:

[0027] The input image batch parameter is 6;

[0028] The initial learning rate of the network is 1e-3, and a step-down method is used, which is reduced to 0.92 times the previous step learning rate every 5 training rounds;

[0029] The total number of model training rounds is 100 rounds;

[0030] The network optimizer uses the Adam optimizer with parameters β 1 is 0.95, parameter β 2 is 0.999.

[0031] Preferably, the L CL It is characterized by the following formula:

[0032]

[0033] In the above formula, and Represents the target contour length of the predicted mask map and the true label mask map, 、 、 、 Indicates the predicted mask map and the true label mask map in H and W The length of the target contour line in the direction, , is the constraint coefficient.

[0034] Preferably, the L CC It is characterized by the following formula:

[0035]

[0036] In the above formula, and Represents the target contour curvature of the predicted mask map and the true label mask map, 、 Indicates that the predicted mask image is at high h He Kuan w The first-order gradient in the direction and, 、 、 Indicates that the predicted mask map is h - h 、 w - w 、 h - w The sum of the second-order gradients in the direction, 、 Indicates that the true label mask is at high h He Kuan w The first-order gradient in the direction and, 、 、 Indicates the true label mask map in h - h 、 w - w 、 h - w The sum of the second-order gradients in the direction.

[0037] A whistling modal analysis method further includes:

[0038] S4, pixel-level semantic segmentation model traverses all shock wave contours in each image one by one in the order of time-series schlieren images to obtain the centroid coordinates of each shock wave structure;

[0039] S5. Extract the vertical coordinates of the centroid of the shock wave structure, calculate the difference between adjacent time-series images, perform power spectrum analysis on them, and use the pulsation root mean square for dimensionless processing to obtain the corresponding pulsation frequency analysis spectrum for whistling mode analysis.

[0040] The present invention has at least the following beneficial effects:

[0041] First, compared with the prior art, the present invention introduces a composite loss function that considers the contour features of the shock wave structure target into the proposed pixel-level semantic segmentation model. L CP During network training, the model is constrained and optimized by utilizing the contour features of the shock wave structure target, so that it has a more refined contour capture capability and a more accurate complex shock wave structure recognition capability when dealing with shock wave structure targets with complex boundary shapes.

[0042] Secondly, the pixel-level semantic segmentation model in the method of the present invention makes it possible to extract the centroid coordinates of the shock wave structure in the later stage, and to implement whistling modal analysis by calculating and processing the pulsation frequency analysis spectrum. Compared with the existing technology, the analysis results of the present invention are basically consistent with the characteristic frequencies obtained by experimental measurement, meeting the needs of whistling modal analysis.

[0043] Other advantages, objectives and features of the present invention will be reflected in part from the following description and will be understood by those skilled in the art through study and practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the flow of the method for intelligent identification of shock wave structure and whistling mode analysis based on time series images of the present invention;

[0045] Figure 2 For the test of the collected Figure 1 Visualization diagram after labeling;

[0046] Figure 3 For the test of the collected Figure 2 Visualization diagram after labeling;

[0047] Figure 4 Schematic diagram of the processing of the pixel-level semantic segmentation model of the present invention;

[0048] Figure 5 The CP loss model and CE loss model are used to collect the experimental data. Figure 1 Schematic diagram of prediction visualization results;

[0049] Figure 6 for Figure 5 A magnified schematic diagram of part A;

[0050] Figure 7 for Figure 5 An enlarged schematic diagram of part B;

[0051] Figure 8 The CP loss model and CE loss model are used to collect the experimental data. Figure 2 Schematic diagram of prediction visualization results;

[0052] Figure 9 for Figure 8 A magnified schematic diagram of part C;

[0053] Figure 10 for Figure 8 An enlarged schematic diagram of part D in the middle;

[0054] Figure 11 This is the whistling mode analysis result diagram of the shock wave structure on the left;

[0055] Figure 12 This is the whistling mode analysis result diagram of the shock wave structure on the right;

[0056] Figure 13 This is the whistling modal analysis result after sensor testing. DETAILED DESCRIPTION

[0057] The present invention will be described in further detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0058] A method for intelligent identification of shock wave structure and whistling mode analysis based on time series images, Figure 1 The flow diagram of the present invention is shown. The specific steps include:

[0059] Step 1: Create a time-series image dataset of the aircraft shock wave structure.

[0060] The obtained schlieren time-series images of the aircraft's shock wave structure are labeled based on domain expertise and experience. Labels contain pixel-level information representing the shock wave structure's location within the schlieren image, as well as background and shock wave classification. Shock wave structure location information consists of the image coordinates of a polygonal box. Labeling can be performed using the deep learning semantic segmentation annotation tool, labelme.

[0061] In this example, to improve the generalization ability of the deep learning model subsequently developed, 200 images were first selected at equal intervals from a set of 2,000 original schlieren images (size 512×512) collected from a wind tunnel test. These 200 images were then labeled. Figure 2-Figure 3 Two example images from the dataset and their corresponding label visualizations are shown. Each image contains two shock wave structures, with the corresponding labels representing the shock wave structure in red and the background in black. These 200 images and their corresponding labels were then augmented nine times using random combinations of noise, brightness, translation, and mirroring. This resulted in a dataset of 2,000 images. This dataset was then partitioned into training, validation, and test sets with a 7:2:1 ratio for the training, validation, and testing of the deep learning model to be developed.

[0062] Step 2: Establish a high-precision pixel-level intelligent detection and recognition semantic segmentation model for shock wave structures

[0063] Figure 4 Schematic diagram of the pixel-level semantic segmentation model established by the present invention for high-precision intelligent detection and identification of shock wave structures. The construction process includes:

[0064] First, a semantic segmentation network is established to intelligently detect and identify shock wave structures. For example, a U-Net semantic segmentation network with an encoder-decoder structure is used as the semantic segmentation network.

[0065] Then, we design a loss function for training and optimizing the semantic segmentation network for intelligent detection and recognition of shock wave structures. In the process of semantic segmentation model training and optimization, the cross-entropy loss function (CE loss) is widely used in semantic segmentation tasks in the field of computer vision. L CE While CE loss can accurately evaluate the difference between the predicted and true values ​​of each pixel in a segmented image, it does not consider or utilize the geometric features of the segmented target at specific locations, such as contours. DC loss, a loss function defined based on the Dice coefficient, can be used to measure the similarity between the predicted and true values ​​of an image. However, this loss function does not consider the spatial positional relationship of the segmented target pixels in the image.

[0066] In this paper, a compound loss function (CP loss) is designed to consider the target contour characteristics of the shock wave structure in the schlieren image obtained by wind tunnel test. L CP Characterized by ), the loss function consists of the cross entropy loss function CE loss, the Dice coefficient loss function DC loss and the contour loss function (Contour loss function, abbreviated as CT loss, L CT The contour loss function is composed of the contour length loss function (Contourlength loss function, abbreviated as CL loss, L CL Contour curvature loss function (CC loss, used to represent the L CC to characterize) composition.

[0067] During the model training process, we can utilize the stabilizing and smoothing capabilities of CE loss during model training, the similarity constraints of DC loss during model training, and the ability to capture the contour details of shock wave targets when processing shock wave structure targets with background interference using CTloss. We can give full play to the advantages of combining the three loss functions to jointly guide model training optimization, thereby achieving accurate pixel-level detection and recognition of shock wave structures. Therefore, the composite loss function CP loss of the present invention can effectively avoid the limitations of using a single loss function for special target tasks. Specifically, the expression of the composite loss function CPloss of the present invention is as follows:

[0068]

[0069] The expression of CE loss is as follows:

[0070]

[0071] Where, H 、 W are the height and width of the image respectively. h and w Indicates the specific location of the pixel in the image, and its value range is 0≤ h ≤ H , 0≤ w ≤ W . v and u Represents the predicted mask map and the true label mask map respectively 。

[0072] The expression of DC loss is as follows:

[0073]

[0074] In the formula, the meaning of each symbol is the same as before.

[0075] In CT loss, the expression of CL loss is as follows:

[0076]

[0077] In the above formula, and Represents the target contour length of the predicted mask map and the true label mask map, 、 、 、 Indicates the predicted mask map and the true label mask map in H and W The length of the target contour line in the direction, , is the constraint coefficient. In this example, all targets are considered to be H Direction and W The lengths of the contour lines in the two directions are between 2 and 3, so the constraint coefficients proportional to their lengths are given to the two directions. In this example, , 2.5 and 1.

[0078] In CT loss, the expression of CC loss is as follows:

[0079]

[0080] In the above formula, and Represents the target contour curvature of the predicted mask map and the true label mask map, 、 Indicates that the predicted mask image is at high h He Kuan w The first-order gradient in the direction and, 、 、 Indicates that the predicted mask map is h - h 、 w - w 、 h - w The sum of the second-order gradients in the direction, 、 Indicates that the true label mask is at high h He Kuan w The first-order gradient in the direction and, 、 、 Indicates the true label mask map in h - h 、 w - w 、 h - w The sum of the second-order gradients in the direction.

[0081] It should be noted that this step mainly involves adjusting the CL loss and CC loss formulas, and combining them with the conventionally used CE loss and DC loss to form the CP loss described in this invention. The CP loss is used for network training optimization, and the effect achieved is: it can achieve higher-precision pixel-level recognition of shock wave structure, thereby effectively realizing the extraction of shock wave features and modal analysis.

[0082] Finally, the training set obtained in step 1 is input into the built shock wave structure pixel-level intelligent detection and recognition semantic segmentation model, and the CP loss function is used for iterative training to obtain the trained weight parameter model. During the training process, after multiple hyperparameter adjustments and optimizations and using the validation set to evaluate the model accuracy, the optimal settings were finally determined: the input image batch parameter is 6; the network initial learning rate is 1e-3, and the step-down method is used, and it is reduced to 0.92 times the previous step learning rate every 5 rounds of training; the total number of model training rounds is 100 rounds; the network optimizer uses the Adam optimizer, and the parameters β 1 is 0.95, β 2 is 0.999.

[0083] In this step, a compound loss function (CP loss) is designed based on the characteristics of shock wave structure targets in the image. It includes CE loss, Dice coefficient-based loss function (DC loss), and contour loss function (CT loss). Compared with the GCE loss introduced in the background technology, this step has the following characteristics:

[0084] ① The DC loss is introduced to consider the similarity between the predicted image value and the true value to alleviate the imbalance between the background and shock wave target categories, and to improve the segmentation accuracy in the later stage;

[0085] ② Since the area of ​​the shock wave structure target in the time series image remains almost unchanged, the loss function that considers the target area is removed. This is used to improve the efficiency of model training;

[0086] ③ In view of the special morphological characteristics of the shock wave structure contour in this example, a constraint coefficient is quantitatively introduced into the loss function term that considers the target contour length. The role of the constraint coefficient is to achieve a proportional constraint on the length of the target contour by constraining the length (or height) H direction and the width W direction. After giving the target contour this proportional constraint, when training the model later, every time the loss function decreases, it can be given a larger penalty. Through the penalty, the model will focus on the pixels in the H direction with a longer contour length.

[0087] Step 3: Predict the image and obtain the time series image recognition result containing the shock wave structure.

[0088] The optimal weight parameter model selected in step 2 is loaded, and the time series image containing the shock wave structure is input into the model network to obtain the shock wave structure detection and recognition results.

[0089] In this step, to illustrate the effectiveness of the composite loss function (CP loss) proposed in this paper, a comparison set of conditions was established. This condition used the same network structure and hyperparameter settings, but only used CE loss, the most commonly used loss function in image processing, as the loss function for network training. The final weighted models obtained from both conditions were loaded and validated using the validation set. They were evaluated using the common semantic segmentation evaluation metric, Intersection over Union (IoU). The comparison results are shown in Table 1. It can be seen that the model using CP loss has higher accuracy. Compared with using CE loss alone, the background category, shock wave structure category, and average accuracy increased by 0.01%, 1.65%, and 0.83%, respectively.

[0090] Table 1

[0091]

[0092] Figure 5-10 These are the test results of two randomly selected shock wave structure images. Figure 5 、 Figure 8 Figures (a), (b), and (c) are the original image collected during the experiment, the predicted visualization result obtained using CP loss, and the predicted visualization result obtained using CE loss, respectively. In the predicted visualization result diagram, Figure 5 、 Figure 8 A, B, C, and D in the figure are the segmentation results. The centroids of the shock wave structures on the left and right sides of the figure (b) and (c) are marked with green and blue dots respectively. Figure 6-Figure 7 , Figure 9-10 From the locally enlarged schematic diagram, it can be found that the prediction results of the model using CE loss have poor boundary recognition of shock wave structure targets (i.e. Figure 7 、 Figure 10 ), and the model using CP loss can effectively improve this situation (i.e. Figure 6 、 Figure 9 ), achieving excellent pixel-level segmentation and recognition of shock wave structure targets.

[0093] The above results prove that the composite loss function CP loss has a more refined ability to capture contours and more accurate ability to identify complex shock wave structures when dealing with shock wave structure targets with complex boundary shapes.

[0094] Step 4: Extract the centroid coordinates of the shock wave structure in each time series image

[0095] Based on the segmentation results obtained in the previous step, all shock wave contours in each image are traversed one by one in the order of the time-series images. The cv2.moments function in the OpenCV library is used to obtain the "contour moment" feature of each shock wave. Then, the centroid coordinates of each shock wave structure are calculated using the zero-order moment representing the area (total number of pixels) of the contour in the "contour moment" and the first-order moment representing the contour on the x and y axes.

[0096] Will Figure 5-Figure 8 The centroid coordinates of the shock wave structures in the two original images are compared with the centroid coordinates of the shock wave structures in the images predicted by the models trained with CP loss and CE loss, and the data comparison structure is shown in Table 2. It can be seen from Table 2 that the centroid coordinate extraction results of the shock wave structure images obtained by segmentation using CP loss are closer to the true values.

[0097] Table 2

[0098]

[0099] Step 5: Whistling Mode Analysis

[0100] Firstly, the vertical coordinates of the centroids of the two shock wave structures in 2000 original schlieren images were extracted. Then, the vertical differences of the corresponding shock wave structure centroids in adjacent time-series images were calculated. The power spectrum was analyzed and dimensionless processing was performed using the pulsation root mean square to obtain the analytical spectrum of its pulsation frequency.

[0101] like Figure 11-12 As shown, it can be found that the shock wave structures on the left and right sides of this set of 2000 time-series images both have a peak at a pulsation frequency of 4944. Figure 13 The figure shows the pulsation frequency analysis spectrum obtained by processing and analyzing the pulsation pressure sensor during the test. The pulsation frequency has a peak at 4949. Comparison results show that the characteristic frequencies of the whistling modal extracted by the method of the present invention are substantially consistent with those obtained by experimental measurements. The whistling modal analysis method described in this invention utilizes non-contact time-series image data of shock wave structures, effectively overcoming the limitations of traditional pressure signal measurement using contact sensors. It can obtain richer unsteady flow characteristics and provide a new approach for supersonic flow oscillation mechanism and noise modal analysis.

[0102] The above solution is only an illustration of a preferred embodiment, but is not limited thereto. When implementing the present invention, appropriate replacements and / or modifications can be made according to user needs.

[0103] Although the embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and exemplary embodiments. They can be applied to a variety of fields suitable for the present invention. Further modifications will be readily apparent to those skilled in the art. Therefore, the present invention is not limited to the specific details and illustrations shown and described herein without departing from the general concept defined by the claims and their equivalents.

Claims

1. A method for intelligent identification of shock wave structure based on time series images, characterized in that: include: S1. Prepare an image dataset with aircraft shock wave structure markers based on Schlieren time-series images; S2. Build a pixel-level semantic segmentation model for intelligent detection and recognition of shock wave structures; S3, using pixel-level semantic segmentation model to intelligently detect and identify shock wave structures on the input schlieren time series images; Among them, in the pixel-level semantic segmentation model, a composite loss function considering the contour features of the shock wave structure target is introduced L CP , and the L CP It is characterized by the following formula: In the above formula, L CE is the cross entropy loss function, L DC is the Dice coefficient loss function, L CT is the shock wave structure profile loss function, L CL is the shock wave structure contour length loss function, L CC is the curvature loss function of the shock wave structure profile; described L CL It is characterized by the following formula: In the above formula, and Represents the target contour length of the predicted mask map and the true label mask map, 、 、 、 Indicates the predicted mask map and the true label mask map in H and W The length of the target contour line in the direction, , is the constraint coefficient; described L CC It is characterized by the following formula: In the above formula, and Represents the target contour curvature of the predicted mask map and the true label mask map, 、 Indicates that the predicted mask image is at high h He Kuan w The first-order gradient in the direction and, 、 、 Indicates that the predicted mask map is h - h 、 w - w 、 h - w The sum of the second-order gradients in the direction, 、 Indicates that the true label mask is at high h He Kuan w The first-order gradient in the direction and, 、 、 Indicates the true label mask map in h - h 、 w - w 、 h - w The sum of the second-order gradients in the direction.

2. The method for intelligent identification of shock wave structure based on time series images according to claim 1, characterized in that: In S1, the process of acquiring the image dataset includes: S10, of the 2000 collected schlieren time-series images, select 200 images at equal intervals; S11, labeling the selected 200 images; S12, perform 9 data augmentations on the 200 labeled images to obtain a dataset of 2000 images; S13, divide the image dataset into training set, validation set and test set according to 7:2:1; In S11, the labeling operation is to mark the shock wave structure in the image with a polygonal frame using a labeling tool to obtain the corresponding shock wave structure positioning information; In S12, the enhancement refers to processing the selected 200 images by randomly combining noise addition, brightness, translation, and mirroring.

3. The method for intelligent identification of shock wave structure based on time series images according to claim 1, characterized in that: In S2, the construction of the pixel-level semantic segmentation model includes: S20. Establish a semantic segmentation network for intelligent detection and identification of shock wave structures; S21. Design a composite loss function for semantic segmentation networks L CP ; S22. Input the training set in the image dataset into the semantic segmentation network and use the composite loss function L CP Perform iterative training to obtain the trained model weight parameters; During the iterative training process, after multiple hyperparameter adjustments and optimizations, and using the validation set in the image dataset to evaluate the model accuracy, the optimal settings for the pixel-level semantic segmentation model are: The input image batch parameter is 6; The initial learning rate of the network is 1e-3, and the step-down method is used, which is reduced to 0.92 times of the previous step learning rate every 5 rounds of training; The total number of model training rounds is 100 rounds; The network optimizer uses the Adam optimizer with parameters β 1 is 0.95, parameter β 2 is 0.999; L CP The weight coefficient in α 1, α 2The value is 1.

4. A whistling mode analysis method, which adopts the shock wave structure intelligent identification method based on time series images according to any one of claims 1 to 3, characterized in that: Also includes: S4, pixel-level semantic segmentation model traverses all shock wave contours in each image one by one in the order of time-series schlieren images to obtain the centroid coordinates of each shock wave structure; S5. Extract the vertical coordinates of the centroid of the shock wave structure, calculate the difference between adjacent time series images, and perform Fourier transform to obtain the corresponding pulsation frequency analysis spectrum for whistling mode analysis.

Images