Knowledge-guided flow field reconstruction and shock wave train leading edge position detection method and system

Abstract

The present invention belongs to the cross-integration field of the supersonic flow combustion technology field and the artificial intelligence field, and specifically relates to a knowledge-guided flow field reconstruction and shock wave train leading edge position detection method and system. In the prediction of the dynamic change of the internal flow field of the isolation section of the scramjet combustion chamber, the present invention uses knowledge distillation technology to combine the flow field reconstruction with the physical mechanism model, uses time, space coordinates, and suction parameter data as input, and generates condensed knowledge to enhance the accuracy of flow field reconstruction. A deep learning network model for leading edge position identification is constructed to accurately identify the position of the leading edge of the shock wave train. This method forms a new multi-level information processing framework by integrating physical knowledge and data-driven methods, which improves the adaptability and prediction accuracy of flow characteristics in complex aerodynamic environments. The system realizes real-time monitoring function on the hardware platform, providing strong technical support for aerodynamic design optimization and flight safety.

Description

Technical Field

[0001] The present invention belongs to the cross-integration field of supersonic flow combustion technology and artificial intelligence, and specifically relates to a knowledge-guided flow field reconstruction and shock wave train leading edge position detection method and system. Background Art

[0002] As an ideal hypersonic propulsion system for use in military protection and aerospace, scramjet engines have become the focus of global research. The isolation section in the engine is a key component responsible for regulating and stabilizing the airflow, involving complex gas dynamics phenomena. The interaction between the shock wave and the boundary layer in the isolation section, as well as the multi-scale turbulence effects, pose a major challenge to the stable control of the airflow. At present, supersonic flow technology usually relies on control strategies with preset parameters, but these strategies are difficult to effectively adapt to adaptive adjustments under high altitude, multi-speed and different Mach number conditions. Therefore, accurate monitoring and prediction of relevant information in the fluid dynamics process is crucial for accurately reconstructing the isolation section flow field and accurately detecting and controlling the leading edge position of the shock wave train.

[0003] Knowledge distillation algorithms have been widely used in the field of fluid mechanics in recent years due to their advantages in compressing model size and improving reasoning efficiency. By combining data-driven and physical mechanism methods, existing knowledge distillation neural network models have made significant progress. However, in the absence of large-scale high-quality datasets, the performance of these models is still limited when dealing with isolation segment reconstruction with shock trains and complex flow fields. In order to improve the accuracy of isolation segment flow field reconstruction and the detection capability of shock train leading edge position, a new neural network model combining knowledge distillation algorithms with physical mechanisms has been developed. By introducing physical constraints, the model's ability to capture complex flow field structures is enhanced, and the generalization ability of the model is significantly improved. In addition, the application of knowledge distillation technology effectively reduces the computational complexity of the model, and realizes efficient and real-time flow field reconstruction and shock train leading edge position detection. Summary of the invention

[0004] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a knowledge-guided variable parameter boundary layer suction isolation section flow field reconstruction and shock wave train leading edge position detection method, so as to improve the network's adaptability to the complex flow field structure inside the isolation section of the scramjet engine.

[0005] To achieve the above-mentioned purpose of the invention, the technical solution of the present invention is as follows:

[0006] A knowledge-guided flow field reconstruction and shock wave train leading edge position detection method comprises the following steps:

[0007] S1. Set different suction parameters and Mach numbers. The suction parameters include suction flow rate, suction hole angle, suction hole data, back pressure, and suction hole length. Set the Mach number to 2-8, and use the software to generate multi-physics field data, including: velocity field, density field, and pressure field;

[0008] S2, using wavelet transform and feature enhancement methods to preprocess data to construct a data set;

[0009] S3. Divide the data set according to the Mach number, construct a training set and a test set, and use them for training and testing the variable parameter boundary layer suction isolation section flow field reconstruction model;

[0010] S4, construct a variable parameter boundary layer suction isolation section flow field reconstruction model for the data set obtained in S2; the model includes two parts, the first part is the teacher model, and the second part is the student model:

[0011] The teacher model consists of a series of multi-scale convolutional blocks, residual modules, attention mechanism modules, fully connected layers, and Transformer encoder modules;

[0012] Multi-scale convolution blocks are used to extract features of different scales;

[0013] The residual module is used to enhance the deep learning ability of the network to alleviate the gradient vanishing problem;

[0014] The attention mechanism module is used to capture global context information and process complex flow fields with multi-scale flow characteristics. By introducing the multi-head attention mechanism, the model can focus on high-speed flow areas with speed discontinuities and flow direction changes, eddy structures and boundary layer characteristics, dynamically adjust the importance of input features, and automatically focus on key areas in the flow field.

[0015] The fully connected layer is used to: integrate and map the extracted flow field information;

[0016] The Transformer encoder module is used to: extract the flow field wave structure, thereby enhancing the accuracy of flow field reconstruction and improving the understanding of flow patterns and dynamic behaviors;

[0017] The student model is composed of a series of convolutional layers, pooling layers, and fully connected layers;

[0018] Convolutional layers are used to: capture spatial patterns and local structures in the flow field, and enhance the ability to understand and reconstruct complex flow phenomena;

[0019] The pooling layer is used to: reduce the spatial dimension of the data and extract more representative features, thereby improving the overall performance and accuracy in the flow field reconstruction process;

[0020] The fully connected layer is used to: integrate and map the extracted flow field information;

[0021] S5, training the variable parameter boundary layer suction isolation section flow field reconstruction model based on the training set constructed in S3, and outputting the training weights;

[0022] S6, based on the test set constructed in S3, the variable parameter boundary layer suction isolation section flow field reconstruction model is tested to obtain the prediction results;

[0023] S7, substituting the prediction results obtained in S6 into the Navier-Stokes equation and combining the boundary conditions and numerical solution methods to construct a physical mechanism model, and through the knowledge-guided method, the neural network model is integrated with the mechanism model to verify and optimize the fluid dynamics characteristics;

[0024] S8. Construct a shock wave front edge position detection model, which is composed of a convolution layer, a deconvolution layer and a fully connected layer in series. The density field is converted into a density gradient field as the input of the network, from which the shock wave front edge position is extracted;

[0025] Convolutional layers are used to: capture spatial patterns and local structures in the flow field, and enhance the ability to understand and reconstruct complex flow phenomena;

[0026] The deconvolution layer is used to restore the low-dimensional feature map to a high-dimensional output and reconstruct the spatial structure and details of the original flow field;

[0027] The fully connected layer is used to integrate and map the extracted flow field information.

[0028] As a preferred embodiment, step S1 is specifically as follows:

[0029] The entire data configuration used includes grid points, and the density distribution of the grid varies according to the accuracy requirements of different regions; in the shock wave reflection area, the data does not conform to the uniform Cartesian grid format; determine the length range of the selected study area in the x-axis direction and the range in the y-axis direction. The input data includes points on the boundary and internal residual points. These points have time t and spatial coordinates (x, y), and the output data covers the pressure field, velocity field and density field in the entire boundary and the selected area;

[0030] In order to study the flow field characteristics under different conditions, the initial inflow conditions at the inlet of the isolation section were set as: total temperature of 950K, total pressure of 0.82MPa, Mach numbers of 2.0, 3.0, 4.0 and 5.0 respectively, the number and angle of the suction holes were adjusted, and software was used for unsteady numerical simulation calculations to simulate two-dimensional or three-dimensional flows, and to handle stable and unstable flow states, as well as complete gas and chemical non-equilibrium flow conditions, so as to obtain multi-physics field data of velocity field, density field and pressure field.

[0031] As a preferred embodiment, step S2 is specifically as follows:

[0032] In order to construct a data set to support the knowledge-guided reconstruction of the flow field in the variable parameter boundary layer suction isolation section and the detection of the leading edge position of the shock train, wavelet transform and feature enhancement methods are used to preprocess the data: firstly, the flow field data is decomposed at multiple scales by wavelet transform to extract the local features and their information in the flow field; secondly, the extracted features are enhanced by feature enhancement methods;

[0033] After data preprocessing, three data forms are finally obtained: cutting target area, boundary data and internal residual points; among them, the cutting target area contains the spatial coordinates of the target flow field and its corresponding flow field data, which are used as label data of the neural network model for supervised learning; the boundary data are the spatial coordinates on the boundary of the target flow field area, which are used for flow field prediction; the internal residual points refer to the spatial coordinates inside the target flow field area, which are used to constrain the information of the physical mechanism part, thereby improving the physical rationality of the model.

[0034] As a preferred embodiment, step S3 is specifically as follows:

[0035] Multiple flow field snapshots were obtained, each snapshot contained the x-component (u) of the velocity field, the y-component (v) of the velocity field, the pressure field (p) and the density field (r). The data were stored in .dat file format. These flow field data showed significant differences in shock wave structures under different Mach number conditions. A local-global data partitioning method was adopted: first, the flow field data at different time points under each Mach number condition were mixed; then, the data under each Mach number condition were randomly divided according to the proportion to form their own training set and test set; finally, the training and test data sets under each Mach number were merged to construct the training and test data sets.

[0036] As a preferred embodiment, step S4 is specifically as follows:

[0037] The model is constructed by the knowledge distillation method, which is divided into two parts: the teacher model and the student model. When constructing the teacher model, time, spatial coordinates and suction parameters are first used as input variables. Then, the input data is processed using multi-scale convolution blocks to extract features of different scales. This process enhances the deep learning ability of the network by designing residual blocks. The multi-head attention mechanism is introduced to focus on the high-speed flow area and boundary layer features with discontinuities and flow direction changes in the input data, capturing global context information to enhance the richness of feature representation. Next, the Transformer encoder module is used to process the extracted features, combined with the fully connected layer and activation function, and finally the prediction results are output.

[0038] In the construction of the student model, based on the convolution layer and the maximum pooling layer, the local features of the flow field are extracted and downsampled; through the fully connected layer, the student model maps the extracted features to the output space to achieve target prediction; during the training process, the student model not only relies on the real labels for learning, but also uses the soft labels generated by the teacher model for guidance; by combining the cross entropy loss and the Kullback-Leibler divergence loss, the student model can learn the knowledge contained in the teacher model and try to get close to the performance of the teacher model;

[0039] In the process of using knowledge distillation, the loss function of the teacher model and the student model consists of two parts: one is the loss between the output of the student model and the true label, and the other is the loss between the student model and the output of the teacher model. These two parts of loss are combined by weight;

[0040] The multi-scale convolution layer extracts multi-scale features of the input data through convolution kernels of different sizes, allowing the model to capture a variety of information from local to global;

[0041] The multi-head attention mechanism is used to capture the global context information of the input features and process them in parallel through multiple attention heads;

[0042] The Transformer encoding layer combines a feed-forward neural network with a multi-head attention mechanism.

[0043] As a preferred embodiment, step S5 is specifically as follows:

[0044] A PyTorch-based environment framework is created to support the model training process, and the Adam optimizer with adaptive moment estimation is used. First, the constructed training set data is read, preprocessed and converted into a format suitable for model input. Then, the processed data is input into the model for training. During training, the model calculates the output through forward propagation, compares the predicted results with the true labels, and uses the loss function to evaluate the model performance. Subsequently, based on the calculated loss value, the error is back-propagated and the model parameters are updated. After multiple iterations, the model continuously learns and optimizes its internal weights to capture the complex dynamic features in flow field reconstruction. After training, the trained weights are saved.

[0045] As a preferred embodiment, step S6 is specifically as follows:

[0046] After the model training is completed, the variable parameter boundary layer suction isolation section flow field reconstruction model is systematically tested based on the constructed test set. First, the model weights trained in S5 are loaded. Then, the isolation section flow field reconstruction model is run using the data in the test set. The input data is passed to the model for forward reasoning to generate the prediction results of the flow field.

[0047] After the prediction is completed, a series of evaluation indicators are used to quantify the predictive ability and applicability of the model. The peak signal-to-noise ratio PSNR, structural similarity index SSIM and correlation coefficient CORR between the predicted results of the calculation model and the real data in the test set are evaluated. PSNR measures the reconstruction quality of the image, SSIM (Structural Similarity Index, SSIM) evaluates the image quality from the perspective of structural information, and CORR (Correlation Coefficient, CORR) reflects the overall similarity of the image.

[0048] As a preferred embodiment, step S7 is specifically as follows:

[0049] Using the Navier-Stokes (NS) equations, combined with boundary conditions and numerical solution methods;

[0050] First, the pressure field, density field, and velocity field predicted by the machine learning model are used as the initial conditions and boundary conditions of the Navier-Stokes equations. These conditions will provide a starting point for the numerical solution. Next, the fluid motion is simulated by the numerical solution method based on the control equations of the NS equations.

[0051] A physical loss term is introduced to constrain the deviation of the model output from the Navier-Stokes equations. The loss function consists of two parts: data-driven loss and physical mechanism loss. The data-driven loss focuses on measuring the difference between the model output and the real observation data, while the physical mechanism loss evaluates the extent to which the model output satisfies the NS equations and its boundary conditions.

[0052] As a preferred embodiment, step S8 is specifically as follows:

[0053] A shock front position detection model is constructed. The architecture of the model consists of a series of convolutional layers, deconvolutional layers and fully connected layers, and the flow field predicted by S6 is used as the input of the model: first, through multi-layer convolution operations, the model can identify potential patterns and features in the flow field; then, the feature map processed by the convolutional layer will be passed to the deconvolutional layer, and the deconvolutional layer will gradually expand the dimension of the feature map to the size of the original input, thereby ensuring that the shock front position detection can be performed on the same spatial scale as the input; finally, the feature map output by the deconvolutional layer is input to the fully connected layer to integrate the information from the convolutional layer and the deconvolutional layer to generate the final shock front position detection result; the fully connected layer forms a global understanding of the shock front position by comprehensively considering various features in the flow field, and adopts the LeakyReLU activation function to alleviate the problem of neuron inactivation that may be caused by traditional ReLU when processing negative input values.

[0054] As a preferred embodiment, step S9 is specifically as follows:

[0055] The Gradient Weighted Class Activation Mapping (Grad-CAM) method is used to visualize the key features in the flow field. The physical consistency of the model during feature extraction is verified by systematically comparing the generated activation maps with the physical properties of the flow field.

[0056] Combined with the decision boundary visualization method UMAP, the data distribution in the feature space is reduced in dimensionality; UMAP retains the local structure between high-dimensional data, and by mapping high-dimensional features to two-dimensional or three-dimensional space, the decision basis of the network in shock front detection can be intuitively displayed; by visually analyzing the feature points after dimensionality reduction, the decision boundaries and their distribution characteristics between different categories are observed, thereby deeply understanding the judgment basis of the network in the classification process.

[0057] As a preferred embodiment, step S10 is specifically as follows:

[0058] The isolation section flow field reconstruction and shock wave front edge position detection model obtained through S5 training are embedded into the FPGA platform. First, the trained weight parameters are loaded into the FPGA. Then, an LED visualization module is designed to display the flow field reconstruction results, shock wave front edge position detection results and related performance indicators in real time, which is specifically manifested in the graphical display of the flow field reconstruction results, the obvious identification of the shock wave front edge position and the numerical listing of performance indicators; the pulse code modulation (PCM) method is used to pre-define the encoding of the flow field reconstruction results, shock wave front edge position detection results and related performance indicators, and the synchronous transmission of data is realized through the four serial ports of the FPGA, so that the above information is seamlessly transmitted to the flow control equipment at the inlet of the scramjet engine combustion chamber;

[0059] Using pulse code modulated pulse signals, a specific binary sequence represents a specific data output; after being encoded, these data are synchronously transmitted to the flow control device at the entrance of the scramjet combustion chamber to adjust and optimize the engine operating state in real time. The control device dynamically adjusts the airflow distribution and fuel injection in the combustion chamber based on the real-time data received to optimize the performance of the scramjet engine.

[0060] The second object of the present invention is to provide a knowledge-guided flow field reconstruction and shock wave train leading edge position detection system, including a data acquisition module, a data processing module, an algorithm model module, and an FPGA platform;

[0061] Data acquisition module: Perform numerical simulation through software to generate multi-physics field data;

[0062] Data processing module: In order to establish the training data set required for the model, the acquired raw data is pre-processed using wavelet transform and feature enhancement methods, and then input into the algorithm model module after processing;

[0063] Algorithm model module: uses the processed data as input features of the network model;

[0064] FPGA platform: By integrating algorithms and sensors, the system can reconstruct the flow field in an aerodynamic environment and timely identify and locate the leading edge of the shock wave train.

[0065] As a preferred embodiment, the knowledge-guided flow field reconstruction and shock wave train leading edge position detection system, the FPGA platform includes an LED display module, a control module, and a pulse encoding module.

[0066] LED display module: used to output flow field reconstruction results, shock wave front position detection results and related performance indicators in real time.

[0067] Control module: including synchronous transmission of the detection results, relevant performance indicators and control feedback signals to the flow control equipment at the inlet of the scramjet combustion chamber to achieve dynamic flow regulation;

[0068] Pulse coding module: The pulse code modulation (PCM) method is used to pre-code the flow field reconstruction results, shock front position detection results and related performance indicators, and the synchronous transmission of data is realized through the four serial ports of the FPGA. The above information is seamlessly transmitted to the flow control equipment at the inlet of the scramjet engine combustion chamber to achieve dynamic flow regulation.

[0069] The beneficial effects of the present invention are as follows: in the prediction of the dynamic changes of the flow field inside the isolation section of the scramjet combustion chamber, the knowledge distillation technology is used to combine the flow field reconstruction with the physical mechanism model, and the time, space coordinates, and suction parameter data are used as inputs to enhance the accuracy of the flow field reconstruction through the generated condensed knowledge. At the same time, a deep learning network model for leading edge position identification is constructed to accurately identify the position of the leading edge of the shock wave train. This method forms a new multi-level information processing framework by integrating physical knowledge and data-driven methods, which improves the adaptability and prediction accuracy of flow characteristics in complex aerodynamic environments. Finally, the system realizes real-time monitoring functions on the hardware platform and realizes the automation of main flow control, providing strong technical support for aerodynamic design optimization and flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A schematic flow chart of a flow field reconstruction and shock wave train leading edge position detection method guided by the knowledge of the present invention;

[0071] Figure 2The overall structure diagram of the flow field reconstruction and shock wave train leading edge position detection method guided by the knowledge of the present invention;

[0072] Figure 3 Schematic diagram of the flow field reconstruction and shock wave train leading edge position detection system guided by the knowledge of the present invention. DETAILED DESCRIPTION

[0073] The following describes the embodiments of the present invention through specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention.

[0074] Example 1

[0075] like Figure 1 As shown, a knowledge-guided flow field reconstruction and shock wave train leading edge position detection method includes the following steps:

[0076] S1. Set different suction parameters and Mach numbers. The suction parameters include suction flow rate, suction hole angle, suction hole data, back pressure, and suction hole length. Set the Mach number to 2-8, and use the software to generate multi-physics field data, including: velocity field, density field, and pressure field;

[0077] S2, using wavelet transform and feature enhancement methods to preprocess data to construct a data set;

[0078] S3. Divide the data set according to the Mach number, construct a training set and a test set, and use them for training and testing the variable parameter boundary layer suction isolation section flow field reconstruction model;

[0079] S4, construct a variable parameter boundary layer suction isolation section flow field reconstruction model for the data set obtained in S2; the model includes two parts, the first part is the teacher model, and the second part is the student model:

[0080] The teacher model consists of a series of multi-scale convolutional blocks, residual modules, attention mechanism modules, fully connected layers, and Transformer encoder modules;

[0081] Multi-scale convolution blocks are used to extract features of different scales;

[0082] The residual module is used to enhance the deep learning ability of the network to alleviate the gradient vanishing problem;

[0083] The attention mechanism module is used to capture global context information, especially to process complex flow fields with multi-scale flow characteristics. By introducing the multi-head attention mechanism, the model can focus on high-speed flow areas with discontinuities and flow direction changes, eddy structures and boundary layer characteristics, dynamically adjust the importance of input features, and automatically focus on key areas in the flow field.

[0084] The fully connected layer is used to: integrate and map the extracted flow field information;

[0085] The Transformer encoder module is used to: extract the flow field wave structure, thereby enhancing the accuracy of flow field reconstruction and improving the understanding of flow patterns and dynamic behaviors;

[0086] The student model is composed of a series of convolutional layers, pooling layers, and fully connected layers;

[0087] Convolutional layers are used to: capture spatial patterns and local structures in the flow field, and enhance the ability to understand and reconstruct complex flow phenomena;

[0088] The pooling layer is used to: reduce the spatial dimension of the data and extract more representative features, thereby improving the overall performance and accuracy in the flow field reconstruction process;

[0089] The fully connected layer is used to: integrate and map the extracted flow field information;

[0090] S5, training the variable parameter boundary layer suction isolation section flow field reconstruction model based on the training set constructed in S3, and outputting the training weights;

[0091] S6, based on the test set constructed in S3, the variable parameter boundary layer suction isolation section flow field reconstruction model is tested to obtain the prediction results;

[0092] S7, substituting the prediction results obtained in S6 into the Navier-Stokes equation and combining the boundary conditions and numerical solution methods to construct a physical mechanism model, and through the knowledge-guided method, the neural network model is integrated with the mechanism model to verify and optimize the fluid dynamics characteristics;

[0093] S8. Construct a shock wave front edge position detection model, which is composed of a convolution layer, a deconvolution layer and a fully connected layer in series. The density field is converted into a density gradient field as the input of the network, from which the shock wave front edge position is extracted;

[0094] Convolutional layers are used to: capture spatial patterns and local structures in the flow field, and enhance the ability to understand and reconstruct complex flow phenomena;

[0095] The deconvolution layer is used to restore the low-dimensional feature map to a high-dimensional output and reconstruct the spatial structure and details of the original flow field;

[0096] The fully connected layer is used to integrate and map the extracted flow field information.

[0097] Example 2

[0098] like Figure 3 As shown, a knowledge-guided flow field reconstruction and shock wave train leading edge position detection system includes a data acquisition module, a data processing module, an algorithm model module, and an FPGA platform;

[0099] Data acquisition module: Perform numerical simulation through software to generate multi-physics field data;

[0100] Data processing module: In order to establish the training data set required for the model, the acquired raw data is pre-processed using wavelet transform and feature enhancement methods, and then input into the algorithm model module after processing;

[0101] Algorithm model module: uses the processed data as input features of the network model;

[0102] FPGA platform: By integrating algorithms and sensors, the system can reconstruct the flow field in an aerodynamic environment and timely identify and locate the leading edge of the shock wave train.

[0103] Preferably,

[0104] The FPGA platform includes LED display module, control module, pulse encoding module,

[0105] LED display module: used to output flow field reconstruction results, shock wave front position detection results and related performance indicators in real time.

[0106] Control module: including synchronous transmission of the detection results, relevant performance indicators and control feedback signals to the flow control equipment at the inlet of the scramjet combustion chamber to achieve dynamic flow regulation;

[0107] Pulse coding module: The pulse code modulation (PCM) method is used to pre-code the flow field reconstruction results, shock front position detection results and related performance indicators, and the synchronous transmission of data is realized through the four serial ports of the FPGA. The above information is seamlessly transmitted to the flow control equipment at the inlet of the scramjet engine combustion chamber to achieve dynamic flow regulation.

[0108] Example 3

[0109] like Figure 2 As shown, this embodiment provides a knowledge-guided flow field reconstruction and shock wave train leading edge position detection method, comprising the following steps:

[0110] S1. Set different suction parameters (suction flow, suction hole angle, suction hole data, pressure, suction hole length) and Mach number (flight Mach number 2-8) to generate multi-physics field data (velocity field, density field, pressure field) using internal and external flow coupling software;

[0111] As a preferred embodiment, step S1 is specifically as follows:

[0112] The combustion chamber consists of an isolation section and a combustion section. The cross-sectional area of ​​the isolation section inlet is 30*150mm 2 , with a length of 0.3m. The isolation section consists of two parts: a straight section with a length of 0.22m and an expansion section with a length of 0.08m, and the upper wall divergence angle is 1.4°. The combustion section length ranges from 0.30m to 1.07m, including a groove with a length-to-depth ratio of 11 and four expansion sections. The groove is 11mm deep and has a bottom divergence angle of 1.4°. The four expansion sections are as follows: the first expansion section is 105mm long and has a divergence angle of 1.4°; the second expansion section is 155mm long and has a divergence angle of 2°; the third expansion section is 142mm long and has a divergence angle of 8°; the fourth expansion section is 103mm long and has a divergence angle of 15°.

[0113] The initial inflow conditions at the inlet of the isolation section are set as follows: static temperature 555.3 K, static pressure P 104136.2 Pa, Reynolds number 20796.36, grid number 92, 342, Mach number Ma 2.0, 3.0, 4.0, 5.0. The numerical simulation was carried out using the large-scale parallel software for internal and external flow coupling developed by the China Aerodynamics Research and Development Center. The software can simulate two-dimensional or three-dimensional, stable or unstable, and completely gas or chemical non-equilibrium flows. The grid configuration of the scramjet engine used in this numerical simulation experiment has been verified by a large number of numerical simulation experiments.

[0114] S2, using wavelet transform and feature enhancement techniques to preprocess data to build a high-quality data set;

[0115] As a preferred embodiment, step S2 is specifically as follows:

[0116] In order to construct a high-quality data set to support the knowledge-guided reconstruction of the flow field in the variable parameter boundary layer suction isolation section and the detection of the leading edge position of the shock wave train, wavelet transform and feature enhancement methods were used for data preprocessing. First, the original flow field data was decomposed at multiple scales by wavelet transform, which can effectively extract local features and important information in the flow field, thereby reducing the influence of noise. In addition, the feature enhancement method was used to improve the extracted features and enhance the expression ability of key physical quantities in the data, so as to improve the effectiveness of subsequent model training.

[0117] The entire data configuration consists of 92342 grid points, and the density distribution of the grid varies due to the different accuracy requirements in different areas. In the area where shock wave reflection occurs, the grid is denser to meet higher accuracy requirements, so the data does not conform to a uniform Cartesian grid format. The length of the selected area ranges from 0.0 to 0.7160000205 in the x-axis direction and from 0 to 0.08540000767 in the y-axis direction. In this study, the specific range is selected from 0.1027680859 to 0.6019849181 in the x-axis direction and from 0 to 0.02108996175 in the y-axis direction, containing 52739 grid points. The input data includes 1203 points on the boundary and 500 internal residual points, which have coordinates in time t and space (x, y). The output includes the pressure field, velocity field, and density field in the entire boundary and the selected area.

[0118] Through data preprocessing, three data forms are obtained: cutting target area, boundary data and internal residual points. Among them, the cutting target area refers to the spatial coordinates and flow field data of the target flow field obtained in this study, which are used as label data for neural network model calculation and supervised neural network learning. Boundary data refers to the spatial coordinates on the boundary of the target flow field area, which is used for flow field prediction. Internal residual points refer to the spatial coordinates inside the target flow field area, which are mainly used for information constraints in the physical mechanism part.

[0119] S3, dividing the data set according to the Mach number for training and testing the variable parameter boundary layer suction isolation section flow field reconstruction model;

[0120] As a preferred embodiment, step S3 is specifically as follows:

[0121] The traditional method of data set division is usually to mix all the data and then divide them randomly in proportion. However, when dealing with complex flow field data, especially when involving flow field shock wave structures under different Mach number conditions, simple random division may not ensure the representativeness of the training and test data sets.

[0122] In the present invention, in order to better reflect the flow field characteristics under different Mach number conditions, a total of 100 flow field snapshots were obtained, each snapshot contains the x component (u) of the velocity field, the y component (v) of the velocity field, the pressure field (p) and the density field (r), and the data are stored in .dat file format. Since the shock wave structure of the flow field under different Mach number conditions is different, in order to ensure that the training and test data sets contain diverse Mach number conditions, a local-global data partitioning method is adopted. Specifically, the flow field data at different times under each Mach number condition are first mixed together, and then the data under each Mach number condition are randomly divided according to a ratio of 8:2. Finally, the training and test data sets of each Mach number (from 2 to 8) are combined to form a complete and balanced training and test data set. This partitioning method not only maintains the diversity of the data set, but also ensures that the flow field characteristics under different Mach number conditions are well represented in both training and testing.

[0123] S4, constructing a variable parameter boundary layer suction isolation section flow field reconstruction model for the data set obtained in S2;

[0124] As a preferred embodiment, step S4 is specifically as follows:

[0125] The model adopts the knowledge distillation method and consists of two parts. The first part is the teacher model and the second part is the student model. When constructing the teacher model, the time, space coordinates and pumping parameters are used as inputs, and multi-scale convolution blocks are used to extract features of different scales. The residual block is used to enhance the deep learning ability of the network to alleviate the gradient vanishing problem. In order to capture the global context information, the multi-head attention mechanism is introduced, so that the model can focus on the important parts of the input data and further improve the richness of the feature representation. Then, the Transformer encoder module is used to process the features, and the final prediction results are output in combination with the fully connected layer and the activation function. In addition, the student model is based on the convolution layer and the maximum pooling layer to extract local features and downsample them, and finally the features are mapped to the output space through the fully connected layer. In order to realize knowledge transfer during the training process, the student model is not only trained with the real labels, but also guided by the soft labels of the teacher model. Combined with the cross entropy loss and the Kullback-Leibler divergence loss, the student model can effectively learn the knowledge in the teacher model, so as to keep the efficiency while keeping close to the performance of the teacher model.

[0126] In the process of using knowledge distillation, the loss function of the teacher model and the student model usually consists of two parts: one is the loss between the output of the student model and the true label, and the other is the loss between the student model and the output of the teacher model. These two parts of loss can be combined by weight. The total loss function of knowledge distillation is L It can be expressed as:

[0127] (1)

[0128] in L hard It is the loss between the student model output and the true label, commonly expressed as cross entropy loss:

[0129] (2)

[0130] L soft It is the loss between the output of the student model and the output of the teacher model, usually measured using the Kullback-Leibler divergence (KL divergence):

[0131] (3)

[0132] In order to effectively extract multi-scale features from the input data, capture global context information, and improve the expressiveness of the model, multi-scale convolutional layers, multi-head attention mechanisms, and Transformer encoding layers are used.

[0133] The multi-scale convolution layer extracts multi-scale features of the input data through convolution kernels of different sizes, enabling the model to capture a variety of information from local to global.

[0134] (4)

[0135] in, W i It is i The weight of the convolution kernel, b i is the corresponding bias term, n is the number of convolution kernels, and * indicates the convolution operation. k Values ​​(such as 3x3, 5x5) extract information of different scales.

[0136] The multi-head attention mechanism is used to capture the global context information of the input features. By processing multiple attention heads in parallel, the model's ability to focus on different features is improved. Specifically, for the input features , through linear transformation to obtain query, key and value matrices :

[0137] (5)

[0138] in, W Q , W K , W Vare the weight matrices for query, key, and value, respectively.

[0139] The Transformer encoding layer combines a feedforward neural network with a multi-head attention mechanism to further capture global dependencies. The output of the Transformer encoding layer is generated through the following steps:

[0140] (6)

[0141] (7)

[0142] Among them, the feedforward neural network (FFN) usually consists of two fully connected layers:

[0143] (8)

[0144] S5, training the variable parameter boundary layer suction isolation section flow field reconstruction model based on the constructed training set, and outputting the training weights;

[0145] As a preferred embodiment, step S5 is specifically as follows:

[0146] All models in this experiment were conducted on the Windows 11 operating system, and the computer used was equipped with an Intel(R)Core(TM) i5-1035G1 CPU @ 1.00GHz 1.19 GHz processor and 16 GB memory. A PyTorch-based environment framework was created to support the model training process. The Adam optimizer with adaptive moment estimation was used, and the batch size was set to 256. First, the constructed training set data was read, preprocessed and converted into a format suitable for model input. Then, the processed data was input into the model for training. During training, the model calculated the output through forward propagation, compared the predicted results with the true labels, and used the loss function to evaluate the model performance. Subsequently, based on the calculated loss value, the error was back-propagated and the model parameters were updated. After multiple iterations, the model continuously learned and optimized its internal weights to capture the complex dynamic characteristics in the flow field reconstruction. After training, the trained weights were saved.

[0147] S6. Testing the variable parameter boundary layer suction isolation section flow field reconstruction model based on the constructed test set to obtain prediction results;

[0148] As a preferred embodiment, step S6 is specifically as follows:

[0149] After the model training is completed, the variable parameter boundary layer suction isolation section flow field reconstruction model is systematically tested based on the constructed test set. First, the model weights trained in S5 are loaded. Then, the isolation section flow field reconstruction model is run using the data in the test set. The input data is passed to the model for forward reasoning to generate the prediction results of the flow field.

[0150] After the prediction is completed, a series of evaluation indicators are used to quantify the predictive ability and applicability of the model. The peak signal-to-noise ratio PSNR, structural similarity index SSIM and correlation coefficient CORR between the predicted results of the calculation model and the real data in the test set are evaluated. PSNR measures the reconstruction quality of the image, SSIM (Structural Similarity Index, SSIM) evaluates the image quality from the perspective of structural information, and CORR (Correlation Coefficient, CORR) reflects the overall similarity of the image.

[0151] S7, substituting the prediction results obtained in S6 into the Navier-Stokes equation and combining appropriate boundary conditions and numerical solution methods to construct a physical mechanism model;

[0152] As a preferred embodiment, step S7 is specifically as follows:

[0153] The innovation of the present invention is to use a knowledge-guided approach to use the pressure field, density field and velocity field output by the machine learning model as the input of the physical mechanism model, thereby complementing the traditional data-driven approach. The physical mechanism model is constructed based on the Navier-Stokes (NS) equations and boundary conditions to ensure that the generated results conform to the basic physical laws of fluid mechanics. During the training process, the deviation of the model output from the NS equation is further constrained by introducing a physical loss term.

[0154] Specifically, the loss function consists of two parts: data-driven loss and physical mechanism loss, in the following form:

[0155] (9)

[0156] Among them, data-driven loss Usually defined as the true label and the model prediction value The mean square error (MSE) between

[0157] (10)

[0158] here, N is the number of samples, y i It is iThe true value of the samples, and is the predicted value output by the model.

[0159] On the other hand, the physical mechanism loss Based on the residuals of the Navier-Stokes equations, the deviation when solving the following equations is obtained:

[0160] (11)

[0161] in, represents the residual of the Navier-Stokes equations, μ j , p j and ρ j are the predicted values ​​of velocity field, pressure field and density field of the jth sample respectively, and M is the number of samples of physical constraints.

[0162] Through this design, the loss function not only takes advantage of the efficiency of the data-driven model, but also ensures that the output results have high physical consistency and interpretability through the guidance of the physical mechanism model. Ultimately, the optimization process minimizes the overall loss, allowing the model to follow the basic laws of fluid mechanics while ensuring accuracy.

[0163] S8. Build a shock train leading edge position detection model, which is composed of a convolutional layer, a deconvolutional layer and a fully connected layer in series;

[0164] As a preferred embodiment, step S8 is specifically as follows:

[0165] The flow field predicted by S6 is used as input, and the input features are extracted using the convolution layer, and then the spatial information is restored through the deconvolution layer. The outputs of the convolution layer and the deconvolution layer are integrated through the fully connected layer to generate accurate shock front position detection results. The LeakyReLU activation function is used in the model architecture because it can alleviate the problem of neuron inactivation caused by ReLU when the input value is negative, thereby enhancing the learning ability and robustness of the model.

[0166] S9, using the Gradient Weighted Class Activation Mapping (Grad-CAM) and the Decision Boundary Visualization (UMAP) display modules to interpretably display the calculation process of the isolation section flow field reconstruction and the shock wave train leading edge position detection network;

[0167] As a preferred embodiment, step S9 is specifically as follows:

[0168] In order to enhance the interpretability of the isolation section flow field reconstruction and shock wave front position detection network, the key features of the flow field are visualized using Gradient Weighted Class Activation Mapping (Grad-CAM). By comparing the generated activation map with the physical characteristics of the flow field, it not only verifies the physical consistency of the model in the feature extraction process, but also improves the transparency and interpretability of the detection results. In addition, combined with decision boundary visualization (UMAP), the classification decision process of the network in the high-dimensional feature space is demonstrated, thereby revealing the decision basis of the network in shock wave front detection, further enhancing the reliability and understandability of the network.

[0169] S10, embed the test code for isolation section flow field reconstruction and shock wave front edge position detection and the training weights obtained in S5 into the FPGA platform, output the flow field reconstruction results, shock wave front edge position detection results and related performance indicators in real time, and transmit them to the flow control device at the inlet of the scramjet engine combustion chamber;

[0170] As a preferred embodiment, step S10 is specifically as follows:

[0171] The isolation section flow field reconstruction and shock wave front position detection model obtained through S5 training are embedded into the FPGA platform, and the flow field reconstruction results, shock wave front position detection results and related performance indicators are displayed through the LED visualization module. The flow field reconstruction results, shock wave front position detection results, flow field reconstruction and shock wave position detection performance indicators and control feedback signals are transmitted through the predefined coding method of pulse code modulation, and are synchronously transmitted to the flow control device at the inlet of the scramjet combustion chamber through the four serial ports of the FPGA to achieve dynamic flow regulation.

[0172] This encoding method uses a specific binary sequence to represent different types of output data, and transmits it synchronously to the flow control device at the inlet of the combustion chamber through the four serial ports of the FPGA. The encoding is defined in formula (12):

[0173] (12)

[0174] Using pulse code modulation pulse signals, specific binary sequences represent specific data outputs: '0001' is used to transmit the velocity, pressure and temperature distribution data FFRR of the flow field, '0010' is used to transmit the position coordinate information SWF of the shock front, '0100' is used to transmit the performance indicators PM related to flow field reconstruction and shock wave detection, and '1000' is used to control the feedback signal encoding to represent the adjustment instructions CFS generated based on real-time data. After encoding, these data are synchronously transmitted to the flow control equipment at the inlet of the scramjet combustion chamber to adjust and optimize the engine operating status in real time. The advantage of using pulse code modulation is that it can accurately transmit discrete samples of data and reduce signal distortion and noise effects during transmission in the form of binary coding.

[0175] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical ideas disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A knowledge-guided flow field reconstruction and shock wave train leading edge position detection method, characterized in that The steps include: S1. Set different suction parameters and Mach numbers. The suction parameters include suction flow rate, suction hole angle, suction hole data, back pressure, and suction hole length. Set the Mach number to 2-8, and use the software to generate multi-physics field data, including: velocity field, density field, and pressure field; S2, using wavelet transform and feature enhancement methods to preprocess data to construct a data set; S3. Divide the data set according to the Mach number, construct a training set and a test set, and use them for training and testing the variable parameter boundary layer suction isolation section flow field reconstruction model; S4, construct a variable parameter boundary layer suction isolation section flow field reconstruction model for the data set obtained in S2; the model includes two parts, the first part is the teacher model, and the second part is the student model: The teacher model consists of a series of multi-scale convolutional blocks, residual modules, attention mechanism modules, fully connected layers, and Transformer encoder modules; Multi-scale convolution blocks are used to extract features of different scales; The residual module is used to enhance the deep learning ability of the network to alleviate the gradient vanishing problem; The attention mechanism module is used to capture global context information and process complex flow fields with multi-scale flow characteristics. By introducing a multi-head attention mechanism, the model can focus on high-speed flow areas with discontinuities and flow direction changes, eddy structures, and boundary layer characteristics, dynamically adjust the importance of input features, and automatically focus on key areas in the flow field. The fully connected layer is used to: integrate and map the extracted flow field information; The Transformer encoder module is used to: extract the flow field wave structure, thereby enhancing the accuracy of flow field reconstruction and improving the understanding of flow patterns and dynamic behaviors; The student model is composed of a series of convolutional layers, pooling layers, and fully connected layers; Convolutional layers are used to: capture spatial patterns and local structures in the flow field, and enhance the ability to understand and reconstruct complex flow phenomena; The pooling layer is used to: reduce the spatial dimension of the data and extract more representative features, thereby improving the overall performance and accuracy in the flow field reconstruction process; The fully connected layer is used to: integrate and map the extracted flow field information; S5, training the variable parameter boundary layer suction isolation section flow field reconstruction model based on the training set constructed in S3, and outputting the training weights; S6, based on the test set constructed in S3, the variable parameter boundary layer suction isolation section flow field reconstruction model is tested to obtain the prediction results; S7, substituting the prediction results obtained in S6 into the Navier-Stokes equation and combining the boundary conditions and numerical solution methods to construct a physical mechanism model, and through the knowledge-guided method, the neural network model is integrated with the mechanism model to verify and optimize the fluid dynamics characteristics; S8. Construct a shock wave front edge position detection model, which is composed of a convolution layer, a deconvolution layer and a fully connected layer in series. The density field is converted into a density gradient field as the input of the network, from which the shock wave front edge position is extracted; Convolutional layers are used to: capture spatial patterns and local structures in the flow field, and enhance the ability to understand and reconstruct complex flow phenomena; The deconvolution layer is used to restore the low-dimensional feature map to a high-dimensional output and reconstruct the spatial structure and details of the original flow field; The fully connected layer is used to integrate and map the extracted flow field information.

2. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S1 is specifically as follows: The entire data configuration used includes grid points, and the density distribution of the grid varies according to the accuracy requirements of different regions; in the shock wave reflection area, the data does not conform to the uniform Cartesian grid format; determine the length range of the selected study area in the x-axis direction and the range in the y-axis direction. The input data includes points on the boundary and internal residual points. These points have time t and spatial coordinates (x, y), and the output data covers the pressure field, velocity field and density field in the entire boundary and the selected area; In order to study the flow field characteristics under different conditions, the initial inflow conditions at the inlet of the isolation section were set as: total temperature of 950K, total pressure of 0.82MPa, Mach numbers of 2.0, 3.0, 4.0 and 5.0 respectively, the number and angle of the suction holes were adjusted, and software was used for unsteady numerical simulation calculations to simulate two-dimensional or three-dimensional flows, and to handle stable and unstable flow states, as well as complete gas and chemical non-equilibrium flow conditions, so as to obtain multi-physics field data of velocity field, density field and pressure field.

3. The knowledge-guided flow field reconstruction and shock wave train leading edge position detection method according to claim 1, characterized in that: Step S2 is specifically as follows: In order to construct a data set to support the knowledge-guided reconstruction of the flow field in the variable parameter boundary layer suction isolation section and the detection of the leading edge position of the shock train, the wavelet transform and feature enhancement methods are used to preprocess the data: firstly, the wavelet transform is used to perform multi-scale decomposition on the flow field data to extract the local features and their information in the flow field; Secondly, the extracted features are improved through feature enhancement methods; After data preprocessing, three data forms are finally obtained: cutting target area, boundary data and internal residual points. Among them, the cutting target area contains the spatial coordinates of the target flow field and its corresponding flow field data, which are used as label data of the neural network model for supervised learning; the boundary data is the spatial coordinates on the boundary of the target flow field area, which is used for flow field prediction. Internal residual points refer to the spatial coordinates inside the target flow field area, which are used to impose information constraints on the physical mechanism part, thereby improving the physical rationality of the model.

4. The knowledge-guided flow field reconstruction and shock wave train leading edge position detection method according to claim 1, characterized in that: Step S3 is specifically as follows: Multiple flow field snapshots were obtained, each snapshot contained the x-component (u) of the velocity field, the y-component (v) of the velocity field, the pressure field (p) and the density field (r). The data were stored in .dat file format. These flow field data showed significant differences in shock wave structures under different Mach number conditions. A local-global data partitioning method was adopted: first, the flow field data at different time points under each Mach number condition were mixed; then, the data under each Mach number condition were randomly divided according to the proportion to form their own training set and test set; finally, the training and test data sets under each Mach number were merged to construct the training and test data sets.

5. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S4 is specifically as follows: The model is constructed by using the knowledge distillation method, which is divided into two parts: the teacher model and the student model. When constructing the teacher model, time, spatial coordinates and suction parameters are first used as input variables. Then, the input data is processed using multi-scale convolution blocks to extract features of different scales. This process enhances the deep learning ability of the network by designing residual blocks; introduces a multi-head attention mechanism to focus on the discontinuities in the input data, high-speed flow areas with flow direction changes, and boundary layer features, captures global context information to enhance the richness of feature representation; next, the Transformer encoder module is used to process the extracted features, combined with the fully connected layer and activation function, and finally outputs the prediction results; In the construction of the student model, based on the convolution layer and the maximum pooling layer, the local features of the flow field are extracted and downsampled; through the fully connected layer, the student model maps the extracted features to the output space to achieve target prediction; during the training process, the student model not only relies on the real labels for learning, but also uses the soft labels generated by the teacher model for guidance; by combining the cross entropy loss and the Kullback-Leibler divergence loss, the student model can learn the knowledge contained in the teacher model and try to get close to the performance of the teacher model; In the process of using knowledge distillation, the loss function of the teacher model and the student model consists of two parts: one is the loss between the output of the student model and the true label, and the other is the loss between the student model and the output of the teacher model. These two parts of loss are combined by weight; The multi-scale convolution layer extracts multi-scale features of the input data through convolution kernels of different sizes, allowing the model to capture a variety of information from local to global; The multi-head attention mechanism is used to capture the global context information of the input features and process them in parallel through multiple attention heads; The Transformer encoding layer combines a feed-forward neural network with a multi-head attention mechanism.

6. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S5 is specifically as follows: Create a PyTorch-based environment framework to support the model training process, using the Adam optimizer with adaptive moment estimation. First, read the constructed training set data, preprocess it and convert it into a format suitable for model input. Then, input the processed data into the model for training. During training, the model calculates the output through forward propagation and compares the predicted results with the true labels, using the loss function to evaluate the model performance; then, based on the calculated loss value, the error is back-propagated and the model parameters are updated; After multiple iterations, the model continuously learns and optimizes its internal weights to capture the complex dynamic features in flow field reconstruction; After training is completed, save the trained weights.

7. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S6 is specifically as follows: After the model training is completed, the variable parameter boundary layer suction isolation section flow field reconstruction model is systematically tested based on the constructed test set. First, the model weights trained in S5 are loaded. Then, the isolation section flow field reconstruction model is run using the data in the test set. The input data is passed to the model for forward reasoning to generate the prediction results of the flow field. After the prediction is completed, a series of evaluation indicators are used to quantify the predictive ability and applicability of the model. The peak signal-to-noise ratio PSNR, structural similarity index SSIM and correlation coefficient CORR between the predicted results of the calculation model and the real data in the test set are evaluated. PSNR measures the reconstruction quality of the image, the structural similarity index SSIM evaluates the image quality from the perspective of structural information, and the correlation coefficient CORR reflects the overall similarity of the image.

8. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S7 is specifically as follows: Using the Navier-Stokes equations, combined with boundary conditions and numerical solution methods; First, the pressure field, density field, and velocity field predicted by the machine learning model are used as the initial conditions and boundary conditions of the Navier-Stokes equations. These conditions will provide a starting point for the numerical solution. Next, based on the governing equations of the Navier-Stokes equations, the fluid motion is simulated through a numerical solution method. A physical loss term is introduced to constrain the deviation of the model output from the Navier-Stokes equations. The loss function consists of two parts: data-driven loss and physical mechanism loss. The data-driven loss focuses on measuring the difference between the model output and the real observation data, while the physical mechanism loss evaluates the extent to which the model output satisfies the Navier-Stokes equations and its boundary conditions.

9. The knowledge-guided flow field reconstruction and shock wave train leading edge position detection method according to claim 1, characterized in that: Step S8 is specifically as follows: A shock front position detection model is constructed. The architecture of the model consists of a series of convolutional layers, deconvolutional layers and fully connected layers, and the flow field predicted by S6 is used as the input of the model: first, through multi-layer convolution operations, the model can identify potential patterns and features in the flow field; then, the feature map processed by the convolutional layer will be passed to the deconvolutional layer, and the deconvolutional layer will gradually expand the dimension of the feature map to the size of the original input, thereby ensuring that the shock front position detection can be performed on the same spatial scale as the input; finally, the feature map output by the deconvolutional layer is input to the fully connected layer to integrate the information from the convolutional layer and the deconvolutional layer to generate the final shock front position detection result; the fully connected layer forms a global understanding of the shock front position by comprehensively considering various features in the flow field, and adopts the LeakyReLU activation function to alleviate the problem of neuron inactivation that may be caused by traditional ReLU when processing negative input values.

10. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S9 is specifically as follows: Gradient-weighted class activation mapping (Grad-CAM) is used to visualize key features in the flow field. The physical consistency of the model during feature extraction is verified by systematically comparing the generated activation maps with the physical properties of the flow field. Combined with the decision boundary visualization method UMAP, the data distribution in the feature space is reduced in dimensionality; UMAP retains the local structure between high-dimensional data, and by mapping high-dimensional features to two-dimensional or three-dimensional space, the decision basis of the network in shock front detection can be intuitively displayed; by visually analyzing the feature points after dimensionality reduction, the decision boundaries and their distribution characteristics between different categories are observed, thereby deeply understanding the judgment basis of the network in the classification process.

11. The knowledge-guided flow field reconstruction and shock train leading edge position detection method according to claim 1, characterized in that: Step S10 is specifically as follows: The isolation section flow field reconstruction and shock wave front edge position detection model obtained through S5 training are embedded into the FPGA platform. First, the trained weight parameters are loaded into the FPGA. Then, an LED visualization module is designed to display the flow field reconstruction results, shock wave front edge position detection results and related performance indicators in real time, which is specifically manifested in the graphical display of the flow field reconstruction results, the obvious identification of the shock wave front edge position and the numerical listing of performance indicators; the pulse code modulation PCM method is used to pre-define the encoding of the flow field reconstruction results, shock wave front edge position detection results and related performance indicators, and the synchronous transmission of data is realized through the four serial ports of the FPGA, so that the above information is seamlessly transmitted to the flow control equipment at the inlet of the scramjet engine combustion chamber; Utilizing pulse code modulated pulse signals, a specific binary sequence represents a specific data output; after being encoded, these data are synchronously transmitted to the flow control device at the entrance of the scramjet engine combustion chamber to adjust and optimize the engine operating state in real time; the control device dynamically adjusts the airflow distribution and fuel injection in the combustion chamber based on the real-time data received to optimize the operating performance of the scramjet engine.

12. A knowledge-guided flow field reconstruction and shock wave train leading edge position detection system, used in any one of the detection methods of claims 1 to 11, characterized in that: Including data acquisition module, data processing module, algorithm model module, and FPGA platform; Data acquisition module: Perform numerical simulation through software to generate multi-physics field data; Data processing module: In order to establish the training data set required for the model, the acquired raw data is pre-processed using wavelet transform and feature enhancement methods, and then input into the algorithm model module after processing; Algorithm model module: uses the processed data as input features of the network model; FPGA platform: By integrating algorithms and sensors, the system can reconstruct the flow field in an aerodynamic environment and timely identify and locate the leading edge of the shock wave train.

13. The knowledge-guided flow field reconstruction and shock wave train leading edge position detection system according to claim 12, characterized in that: The FPGA platform includes LED display module, control module, pulse encoding module, LED display module: used to output flow field reconstruction results, shock wave front position detection results and related performance indicators in real time. Control module: including synchronous transmission of the detection results, relevant performance indicators and control feedback signals to the flow control equipment at the inlet of the scramjet combustion chamber to achieve dynamic flow regulation; Pulse coding module: The pulse code modulation (PCM) method is used to pre-code the flow field reconstruction results, shock front position detection results and related performance indicators, and the synchronous transmission of data is realized through the four serial ports of the FPGA. The above information is seamlessly transmitted to the flow control device at the inlet of the scramjet engine combustion chamber to achieve dynamic flow regulation.

Images