A shock train front edge position accurate detection method based on YOLOv8

Abstract

The present application relates to the technical field of aerodynamic test measurement, and discloses a shock train front edge position accurate detection method based on YOLOv8, comprising: establishing a space mapping of a pressure sensor array and image pixel coordinates, synchronously collecting a flow field schlieren image and wall pressure time series data through hardware triggering; generating a region mask using a YOLOv8 model, constructing a space probability field according to pressure gradient characteristics to constrain image features; analyzing the frequency domain characteristics of pressure data to determine the stability of the flow field, adaptively selecting a steady-state differential operator or a dynamic integral fitting operator, solving the shock train front edge sub-pixel coordinates under the constraint of the probability field, and combining distortion correction to inverse the physical position and instantaneous velocity. Through the multi-modal field coupling mechanism of aerodynamics and vision, the present application effectively solves the problem of precision decline caused by background noise interference and dynamic motion blur in single visual detection, and realizes high-precision quantitative detection of shock train.

Description

Technical Field

[0001] This invention relates to the field of aerodynamic test and measurement technology, specifically to a method for accurate detection of the leading edge position of shock trains based on YOLOv8. Background Technology

[0002] Shock trains are a typical flow field structure within the isolation section of a scramjet engine, formed by the interaction of a series of complex shock waves with the boundary layer. During engine operation, the dynamic changes in the leading edge position of the shock train directly reflect the matching between the combustion chamber back pressure and the incoming flow state. Monitoring the leading edge position of the shock train is of great significance for preventing the engine intake from failing to start and for achieving active control of combustion modes.

[0003] Existing shock wave detection methods primarily rely on contact wall pressure measurement and non-contact optical observation. Wall pressure measurement involves deploying an array of sensors along the flow direction and determining the shock wave location based on the abrupt increase in pressure distribution along the flow path. Optical observation utilizes schlieren or shading techniques to record the changes in the second derivative of the flow field density, converting the flow field structure into a visualized grayscale image. Subsequently, image processing algorithms are used to extract the geometric edges of the shock wave. In recent years, convolutional neural networks have been introduced into flow field image analysis, enabling automated identification of shock wave structures by learning from a large number of sample features.

[0004] Simply relying on a specific measurement mode makes it difficult to balance accuracy and robustness in detection. The spatial resolution of wall pressure measurement is limited by the sensor mounting spacing, making it impossible to accurately capture the minute displacements at the shock wave leading edge. While optical observation offers a continuous field of view, background texture noise, window contamination, and optical distortion can interfere with edge extraction algorithms. Deep learning methods can locate approximate areas, but face challenges in achieving sub-pixel accuracy. Furthermore, shock waves exhibit high-frequency oscillations in high-speed flow, and this dynamic behavior leads to motion blur in imaging. This causes traditional edge detection operators or fitting algorithms based on steady-state assumptions to produce significant positioning errors when handling blurred boundaries, failing to meet the high-precision quantitative measurement requirements of dynamic flow fields in wind tunnel experiments.

[0005] Therefore, the purpose of this invention is to provide a method for accurate detection of the leading edge position of shock trains based on YOLOv8, so as to overcome the shortcomings of the prior art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for accurate detection of the leading edge position of shock trains based on YOLOv8. This method solves the problems of limited spatial resolution of single-mode measurement, susceptibility of optical observation to background noise interference, and decreased edge positioning accuracy due to imaging motion blur under high-frequency oscillation conditions in the existing shock train detection technology.

[0007] To achieve the above objectives, a method for accurate detection of the leading edge position of shock trains based on YOLOv8 is proposed, comprising the following steps: Establish a spatial mapping relationship between the physical coordinates of the pressure sensor array components and the image pixel coordinates; Simultaneously acquire flow field schlieren images and wall pressure time-series data; Input the flow field schlieren image into the YOLOv8 neural network model to generate a binarized computational mask for the region of interest of the shock train. Based on the path distribution characteristics of wall pressure time series data, a spatial probability field is constructed to constrain the weights for image feature extraction. The frequency domain characteristics of wall pressure time series data are analyzed to determine the flow field stability, and the steady-state differential operator or dynamic integral fitting operator is adaptively selected accordingly. Within the region defined by the binarized computation mask, the flow field schlieren image is subjected to multi-field coupling solution by combining the spatial probability field and using the selected steady-state differential operator or dynamic integral fitting operator to obtain the sub-pixel level image coordinates of the shock train leading edge. Based on spatial mapping relationships, subpixel-level image coordinates are inverted into physical spatial locations and output.

[0008] Furthermore, establishing the spatial mapping relationship between the physical coordinates of the pressure sensor array components and the image pixel coordinates specifically includes: The homography matrix between the physical space coordinate system and the image pixel coordinate system is calculated using the feature points of the calibration plate; The homography matrix is ​​used to map the physical installation position of each sensor in the pressure sensor array assembly to a pixel position index in the image pixel coordinate system, and the pixel position index is used as the spatial reference for constructing the spatial probability field.

[0009] Furthermore, the steps for constructing the spatial probability field to constrain the image feature extraction weights specifically include: Calculate the pressure gradient along the wall pressure time series data and locate the sensor node corresponding to the maximum pressure gradient value; A one-dimensional Gaussian space weighting function is constructed with the mapping position of the sensor node in the image pixel coordinate system as the center. The one-dimensional Gaussian spatial weight function is extended into a spatial probability weight matrix with the same size as the flow field schlieren image, which serves as the spatial probability field.

[0010] Furthermore, the steps for analyzing the frequency domain characteristics of wall pressure time series data to determine flow field stability specifically include: Extract the local pressure time-series data of the sensor node corresponding to the maximum pressure gradient, and use this local pressure time-series data as the object of frequency domain analysis; perform fast Fourier transform on the local pressure time-series data to generate the corresponding power spectral density distribution; The ratio of characteristic band energy to full band energy is calculated based on the power spectral density distribution, and the oscillation index characterizing the dynamic stability of the flow field is obtained. Operator selection is performed based on the comparison result between the oscillation index and the preset threshold: When the oscillation index is less than or equal to the preset threshold, the flow field is determined to be in steady state and the steady-state differential operator is invoked. When the oscillation index is greater than the preset threshold, the flow field is determined to be dynamic and the dynamic integral fitting operator is invoked.

[0011] Furthermore, the steps for performing multi-field coupled solution of the flow field schlieren image using the selected steady-state differential operator specifically include: An edge detection operator is applied to the flow field schlieren image to perform a convolution operation, generating a horizontal grayscale gradient map; The binarized calculation mask and the spatial probability field are used as weight matrices to weight the gray-level gradient image, resulting in a weighted gradient energy distribution. Centroid calculation is performed within the peak neighborhood of the weighted gradient energy distribution, and the calculated weighted centroid is used as the subpixel-level image coordinates of the shock train leading edge.

[0012] Furthermore, the steps for performing multi-field coupled solution of the flow field schlieren image using the selected dynamic integral fitting operator specifically include: The flow field schlieren image is multiplied element-wise with the binarized computational mask and the spatial probability field to generate weighted image feature data used as the fitting object. A nonlinear least squares fitting objective function based on the error function distribution model is constructed, which is used to solve for the distribution parameters of the weighted image feature data; The weighted image feature data is substituted into the nonlinear least squares fitting objective function for iterative optimization to obtain the optimal fitting peak center parameters. The optimally fitted peak center parameters are used as subpixel-level image coordinates for the shock train leading edge.

[0013] Furthermore, the specific steps for inputting the flow field schlieren image into the YOLOv8 neural network model include: The flow field schlieren image is preprocessed and its dimensions are transformed to generate tensor data that meets the input requirements of the YOLOv8 neural network model; Tensor data is input into the YOLOv8 neural network model for inference, and the output is a prediction result containing multiple candidate bounding boxes; The non-maximum suppression algorithm is applied to the prediction results to select the optimal bounding box, and the geometric coordinates of the optimal bounding box are analyzed. A binary computation mask is constructed based on geometric coordinates, retaining only the numerical values ​​of the region inside the optimal bounding box.

[0014] Furthermore, the steps for simultaneously acquiring flow field schlieren images and wall pressure time-series data specifically include: By using hardware-triggered pulse sequences to simultaneously drive the optical observation subsystem and the aerodynamic measurement subsystem, a shared time reference for flow field schlieren images and wall pressure time series data is established. Based on the ratio between the sampling frequency of the aerodynamic measurement subsystem and the imaging frame rate of the optical observation subsystem, a linear mapping model is constructed between the frame index of the associated flow field schlieren image and the data point index of the wall pressure time series data. Substitute the frame index of the flow field schlieren image into the linear mapping model to calculate the starting pressure data point index and the ending pressure data point index of the exposure window covering the flow field schlieren image. Based on the starting pressure data point index and the ending pressure data point index, the corresponding data segment is extracted from the wall pressure time series data, and this data segment is used as the pressure data segment time-aligned with the flow field schlieren image.

[0015] Furthermore, the steps for inverting sub-pixel-level image coordinates into physical spatial locations based on spatial mapping relationships specifically include: Call the camera intrinsic parameter data to calculate the normalized radial distance of the sub-pixel level image coordinates relative to the image optical center; Substitute the normalized radial distance into the polynomial distortion model to calculate the coordinates of the corrected image after eliminating optical distortion. The physical spatial location is obtained by performing a projection transformation on the coordinates of the corrected image using the inverse of the homography matrix.

[0016] Furthermore, after inverting the sub-pixel-level image coordinates into physical spatial locations based on spatial mapping relationships and outputting them, the process also includes a step of calculating the instantaneous moving velocity of the shock train, specifically including: The physical spatial locations output at multiple consecutive sampling times are cached in time sequence to construct the physical location time sequence to be calculated; Obtain the sampling interval time of the flow field schlieren image and use the sampling interval time as the time step parameter for the differential operation; The physical location time sequence and the sampling interval time are substituted into the five-point center difference calculation model for numerical differentiation. The calculation result of numerical differentiation is output as the instantaneous moving velocity of the shock train leading edge.

[0017] This invention provides a method for accurate detection of the leading edge position of shock trains based on YOLOv8. It has the following beneficial effects: 1. This invention can realize multimodal field coupling of aerodynamic data and visual data, solve the problem that single optical observation is easily interfered with by background noise. By extracting the region of interest through the YOLOv8 neural network model and constructing a spatial probability field based on the wall pressure gradient to weight and constrain the image features, it can shield the texture clutter in non-shock regions in complex flow field environments, thereby improving the robustness and signal-to-noise ratio of shock train leading edge identification.

[0018] 2. This invention can determine the flow field stability based on the frequency domain oscillation index of wall pressure time-series data and adaptively switch the image calculation operator. Under steady-state flow conditions, differential operators and centroid calculations are used to obtain high-resolution edges; under dynamic flow conditions, integral fitting operators and error function models are used to overcome imaging motion blur. This allows for the maintenance of sub-pixel-level detection accuracy under different flow velocity conditions, balancing the accuracy of static measurements with the adaptability of dynamic measurements.

[0019] 3. This invention constructs a rigorous spatiotemporal mapping and physical inversion system, enhancing the physical accuracy and application value of measurement results. A shared time reference is established through hardware synchronous triggering, and optical perspective and lens distortion errors are eliminated by combining homography matrix and radial distortion models. Combined with a five-point center difference algorithm, it can not only output the instantaneous geometric coordinates of the shock train but also accurately calculate its moving velocity, providing high-confidence state feedback parameters for the active closed-loop control of the wind tunnel flow field. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the multimodal field coupling shock train leading edge precise positioning system of the present invention; Figure 2 This is a flowchart of the method for precise positioning of the leading edge of a shock train using multimodal field coupling according to the present invention.

[0021] Among them: 10, a multimodal field-coupled shock train leading edge precision positioning system; 100, an optical observation subsystem; 110, a schlieren optical path assembly; 120, a high-speed imaging assembly; 200, a pneumatic measurement subsystem; 210, a pressure sensor array assembly; 220, a data acquisition assembly; 300, a data processing subsystem; 310, a coordinate mapping module; 320, a region generation module; 330, a field constraint module; 340, an operator modulation module; 350, a coupling solution module. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 The present invention provides a multimodal field-coupled shock train leading edge precision positioning system 10, comprising: an optical observation subsystem 100, an aerodynamic measurement subsystem 200, and a data processing subsystem 300.

[0024] The optical observation subsystem 100 includes a schlieren optical path assembly 110 and a high-speed imaging assembly 120. The pneumatic measurement subsystem 200 includes a pressure sensor array assembly 210 and a data acquisition assembly 220. The data processing subsystem 300 logically includes: a coordinate mapping module 310, a region generation module 320, a field constraint module 330, an operator modulation module 340, and a coupling solution module 350.

[0025] The schlieren optical path assembly 110 constructs a parallel light field and forms a flow field light and shadow image. The high-speed imaging assembly 120 acquires the light and shadow image and converts it into a digital image signal.

[0026] The pressure sensor array assembly 210 consists of high-frequency dynamic pressure sensors installed at equal intervals along the flow direction. The data acquisition assembly 220 converts the analog voltage signal into a digital pressure timing signal.

[0027] The data processing subsystem 300 controls the high-speed imaging component 120 and the data acquisition component 220 through a synchronous trigger signal generator to ensure that the digital image signal and the digital pressure timing signal have the same time reference.

[0028] The coordinate mapping module 310 stores homography matrix parameters, maps the physical installation coordinates of the pressure sensor array assembly 210 to the image pixel coordinate system, and inverts the image pixel coordinates back to the physical space coordinates at the end of the calculation.

[0029] The region generation module 320 runs a pre-trained YOLOv8 neural network model, receives digital image signals, outputs the coordinates of the shock wave region bounding box, and generates a binarized computational mask.

[0030] The field constraint module 330 processes digital pressure time-series signals, calculates the pressure gradient distribution along the flow direction, identifies the location of pressure surges, and constructs a one-dimensional Gaussian space weighting function.

[0031] The operator modulation module 340 executes frequency domain analysis logic, performs a fast Fourier transform on the local pressure time-series signal corresponding to the pressure surge location, and calculates the oscillation index. Based on the comparison result between the oscillation index and a preset threshold, the operator modulation module 340 selects either a steady-state differential operator or a dynamic integral fitting operator.

[0032] The coupled solution module 350 receives a binarized computation mask, a one-dimensional Gaussian spatial weighting function, and a selected image processing operator. Within the region defined by the computation mask, the coupled solution module 350 performs weighted feature extraction and curve fitting on the image data using the one-dimensional Gaussian spatial weighting function, and calculates the sub-pixel level coordinates of the shock train leading edge.

[0033] See attached document Figure 2 The structural design of the shock train leading edge precise positioning system 10 based on multimodal field coupling, and the corresponding implementation process of shock train leading edge precise positioning are as follows: Figure 2 As shown, it includes the following steps: Perform spatial coordinate mapping calibration, solve the homography matrix between the physical spatial coordinate system and the image pixel coordinate system, and convert the physical position of each sensor in the pressure sensor array component 210 into the pixel position index in the image coordinate system. Multimodal data synchronous acquisition is performed. The data processing subsystem 300 sends a synchronous trigger signal to control the high-speed imaging component 120 to acquire the schlieren image frame at the acquisition time, and at the same time controls the data acquisition component 220 to acquire the exposure window pressure time sequence data segment at the coverage time. The YOLOv8 region of interest extraction is performed. The region generation module 320 inputs the schlieren image frame into the YOLOv8 model and outputs the shock string target bounding box, defining the bounding box region as the effective computation domain. The pressure field spatiotemporal features are deconstructed. The field constraint module 330 calculates the pressure gradient along the path based on the pressure time series data segment, locates the sensor node corresponding to the pressure surge center, and generates a one-dimensional Gaussian space weight function to limit the image search range based on the sensor node position. Operator adaptive modulation is performed. The operator modulation module 340 performs spectral analysis on the pressure data of the pressure surge center node and calculates the oscillation index. When the oscillation index is less than or equal to the preset threshold, the Sobel differential operator is invoked. When the oscillation index is greater than the preset threshold, the dynamic Gaussian fitting operator is invoked. Multi-field coupled subpixel localization is performed. Within the effective computational domain, the coupling solution module 350 processes the image grayscale data using selected operators and multiplies it by a one-dimensional Gaussian space weighting function to obtain a weighted feature distribution curve. The weighted feature distribution curve is then fitted to obtain the subpixel coordinates in the image coordinate system. Perform physical coordinate inversion by using the inverse transformation of the homography matrix to convert the sub-pixel coordinates into actual position values ​​in physical space and output them.

[0034] For spatial coordinate mapping calibration, the specific implementation process includes the following: Establish the physical coordinate system and image coordinate system definitions: Set the physical coordinate system as , The axis is defined as the direction of airflow along the wind tunnel channel. The axis is defined as the direction normal to the wall (i.e., the height direction perpendicular to the flow channel wall). The pressure sensor array assembly 210 is mounted on... Wall location, number The physical coordinates of the pressure sensor are as follows: ,in , This represents the total number of sensors. The image coordinate system is set as follows: , The axis is defined as the horizontal pixel direction of the image. An axis is defined as the direction perpendicular to the pixels of an image.

[0035] Acquire calibration images and extract feature points: Before ventilation in the wind tunnel test section, a calibration plate with known geometrically sized feature points is placed in the center plane of the flow channel observation area, with the plane of the calibration plate perpendicular to the optical axis of the schlieren spectrometer. Static images of the calibration plate are acquired using the high-speed imaging component 120. The pixel coordinates of feature points in the calibration plate images are extracted using a corner detection algorithm. And associate with the corresponding known physical coordinates ,in For feature point indexes.

[0036] Solve the homography matrix: Based on the pinhole camera imaging model, the mapping relationship from the physical plane to the image plane is established through the homography matrix. Description. For feature point indexes Corresponding feature point pairs and It satisfies the following mapping equation: ; in, It is a non-zero proportionality coefficient. , , , , , , , , , is a homography matrix elements, The horizontal pixel coordinates of the image. The vertical pixel coordinates of the image. For flow direction coordinates, The coordinates are normal coordinates. Expand the mapping equation and eliminate the non-zero scaling factors. This yields the linear constraint equations for each feature point pair: ; An overdetermined linear system of equations is constructed using multiple feature point pairs, and the homography matrix is ​​solved by singular value decomposition. Optimal solution. Typically, constraints. Or constrained homography matrix The norm is set to 1 to eliminate scaling uncertainty.

[0037] Mapping sensor array position: The homography matrix is ​​obtained by solving the problem. The physical coordinates of each pressure sensor in the pressure sensor array assembly 210 are... Substitute into the mapping equation to calculate the corresponding pixel x-coordinate of each pressure sensor in the image coordinate system. The calculated set of mapped coordinates will be obtained. It is stored in coordinate mapping module 310 and used as a spatial reference for subsequent field constraint construction.

[0038] The specific mathematical operations of corner detection algorithms and singular value decomposition can be implemented by those skilled in the art using OpenCV library functions or standard linear algebra methods. These are well-known technologies in the fields of image processing and numerical computation, and will not be elaborated here.

[0039] For the synchronous acquisition of multimodal data, the specific implementation process includes the following: Constructing a hardware synchronization triggering chain: The data processing subsystem 300 is equipped with a synchronous trigger signal generator, which is connected via shielded coaxial cables to the external trigger input port of the high-speed imaging component 120 and the digital trigger port of the data acquisition component 220. The synchronous trigger signal generator is configured to output a transistor-to-transistor logic level standard trigger pulse sequence. The rising edge of the trigger pulse sequence is set as the global time zero point and the starting reference for subsequent acquisition cycles. The high-speed imaging component 120 is configured in external edge trigger mode, meaning that the high-speed imaging component 120 only initiates exposure when the rising edge of the trigger pulse sequence is detected. The data acquisition component 220 is configured in hardware gating mode, using the trigger pulse sequence to initiate the analog-to-digital conversion process.

[0040] Configure sampling frequency and exposure time window: Set the high-speed imaging component to 120 imaging frame rate The single-frame exposure time is The data acquisition component 220 is set to sample the pressure of the pressure sensor array component 210 at a frequency of [value missing]. To capture the details of high-frequency shock wave oscillations, the pressure sampling frequency was configured. Greater than the imaging frame rate And pressure sampling frequency For imaging frame rate Integer multiples.

[0041] Establish frame-level index mapping relationships: Image frame indexing is established based on a shared clock reference. Index of pressure data points A linear mapping model between them. For image frame indexes. Corresponding Schlieren Image The data processing subsystem 300 calculates the schlieren image according to the following formula. The starting index of the corresponding pressure data segment in the digital pressure timing signal With Termination Index : ; ; in, As the starting index, The image frame index (value is a positive integer). For pressure sampling frequency, For imaging frame rate, This indicates the rounding operation. To terminate the index, This indicates the floor function. This refers to the exposure time for a single frame.

[0042] Extract and align time series data segments: The data processing subsystem 300 calculates the starting index. With Termination Index Extract the corresponding pressure timing data segment from the digital pressure timing signal buffered by the data acquisition component 220. The data processing subsystem 300 will extract pressure time-series data segments. With image frame index Corresponding Schlieren Image Binding, building containment Multimodal data pairs. Multimodal data pairs serve as input units for field constraint construction and operator modulation in subsequent steps, ensuring the physical consistency of visual and aerodynamic information across time scales.

[0043] For the transistor-to-transistor logic level standard and the specific electrical connection method of the hardware trigger circuit, those skilled in the art can refer to the relevant industrial bus standards and instrument manuals for implementation. These are well-known technologies in the field of electronic measurement and will not be elaborated here.

[0044] The specific implementation process for YOLOv8-based region of interest extraction includes the following: Image data preprocessing and tensor transformation: The region generation module 320 receives the time transmitted by the high-speed imaging component 120. Schlieren image frame The region generation module 320 pairs of schlieren image frames. Perform size normalization operation to frame the schlieren image. The image is scaled to the standard resolution required for the YOLOv8 network model input layer. The region generation module 320 normalizes the pixel values ​​of the scaled image, mapping the grayscale values ​​to the [0, 1] interval. If the YOLOv8 network model input layer requires three-channel data, the region generation module 320 copies and expands the single-channel grayscale data along the channel dimension to construct four-dimensional tensor data that meets the input requirements of the YOLOv8 network model.

[0045] Network inference and target selection: The region generation module 320 inputs four-dimensional tensor data into a YOLOv8 neural network model pre-trained based on a shock string sample set labeled with real bounding boxes. Through convolutional neural network feature extraction and multi-scale prediction, the YOLOv8 neural network model outputs a tensor containing multiple candidate bounding box predictions. Each candidate bounding box includes center coordinates, width and height parameters, target confidence, and classification probability. The region generation module 320 performs a non-maximum suppression algorithm, eliminating redundant overlapping boxes based on preset confidence and intersection-over-union (IoU) thresholds, and selecting the shock string structure bounding box with the highest confidence. .

[0046] Boundary coordinate analysis: Region generation module 320 analyzes the shock train structure bounding box Geometric parameters define the boundary box of the shock train structure. Convert to image coordinate system Lower extremum coordinate form. Define the boundary box of the shock train structure. The minimum coordinate in the horizontal direction of the image is The maximum coordinate is The minimum coordinate in the vertical direction of the image is The maximum coordinate is .

[0047] Constructing a binarized region of interest mask: Based on the extreme value coordinates obtained through analysis, the region generation module 320 generates a frame that matches the original schlieren image. Size-consistent binarized region of interest mask matrix The mask matrix generation logic follows the following piecewise function definition: ; in, To binarize the region of interest mask matrix in pixel coordinates Value at this location, The horizontal pixel coordinates of the image. The vertical pixel coordinates of the image. The minimum horizontal coordinates of the bounding box. The maximum horizontal coordinate of the bounding box. The minimum vertical coordinates of the bounding box. This represents the maximum vertical coordinate of the bounding box.

[0048] Applying mask-constrained computational domain: The region generation module 320 binarizes the region of interest mask matrix. The output is sent to the coupling solution module 350. The coupling solution module 350 is configured to binarize the region of interest mask matrix. Element-wise multiplication is performed with the grayscale matrix of the entire image, and the pixel weights outside the boundary box of the shock train structure are reset to zero, while only the core region of the shock train is retained to participate in the subsequent sub-pixel localization calculation.

[0049] For the specific network layer structure of the YOLOv8 neural network and the specific mathematical implementation of the nonmaximum suppression algorithm, those skilled in the art can refer to relevant open-source computer vision frameworks and literature, which are well-known technologies in the field of deep learning, and will not be elaborated here.

[0050] The specific implementation process for deconstructing the spatiotemporal characteristics of the pressure field includes the following: Calculate the time-series mean pressure distribution: The field constraint module 330 receives data from the data processing subsystem 300, which intercepts the current time. Schlieren image frame Corresponding pressure time series data segment Pressure time series data segment It is Matrix, where For the number of pressure sensors, This is the sampling window length. The field constraint module 330 applies pressure time series data segments. The arithmetic mean of each row is calculated to obtain the current time. Average pressure vector distributed along the flow direction ,in Indicates the first The average pressure value of each pressure sensor within the exposure time window.

[0051] Calculate the pressure gradient along the friction line and locate the point of sudden rise: Field constraint module 330 for mean pressure vector Calculate the first-order central difference to obtain the pressure gradient vector along the flow direction. Pressure gradient vector The Middle element The calculation formula follows the following segmented definition: ; in, For the first Physical coordinates of a pressure sensor For the first Average pressure value of each pressure sensor This represents the total number of pressure sensors. The field constraint module 330 searches for the pressure gradient vector. The maximum value is used to identify the sensor index corresponding to the maximum pressure gradient as the center index of the pressure surge. .

[0052] Retrieve the image coordinates corresponding to the center of pressure: The field constraint module 330 calls the coordinate mapping module 310 to store the mapped coordinate set. According to the index of the center of pressure surge Retrieve the x-coordinate of the corresponding image pixels Image pixel x-coordinate This represents the approximate position of the leading edge of the shock train in the image coordinate system from the perspective of aerodynamic measurement.

[0053] Construct a one-dimensional Gaussian space weight function: To integrate aerodynamic prior information to suppress visual noise, the field constraint module 330 uses the image pixel horizontal coordinate... Centered on the horizontal axis of the image, construct a structure defined on the horizontal axis. Weighting function in the upper-dimensional Gaussian space One-dimensional Gaussian space weighting function Used to characterize the spatial probability distribution of the shock wave leading edge: ; in, The horizontal pixel coordinates of the image. The standard deviation represents the spatial weights. The field constraint module 330 calculates the average pixel spacing between adjacent pressure sensors on the image based on the mapped coordinate set. And set spatial weight standard deviation ,in This is an empirical coefficient.

[0054] Generate the spatial probability weight matrix: Field constraint module 330 will use the one-dimensional Gaussian space weight function Copy and expand along the vertical direction of the image to generate a schlieren image frame. Size-consistent spatial probability weight matrix Spatial probability weight matrix In any pixel The value depends only on the horizontal pixel coordinates of the image. ,Right now Spatial probability weight matrix The image is then transmitted to the coupling solution module 350 for weighted modulation of the image gradient features.

[0055] For operator adaptive modulation, the following specific implementation process is included in the frequency domain feature extraction stage of the pressure signal: Local time series data extraction: Operator modulation module 340 outputs pressure surge center index based on field constraint module 330 From the pressure time series data segment Extract the first Row data. The extracted data is defined as a local pressure fluctuation sequence. ,in For discrete-time indexing, , This is the sampling window length. Local pressure fluctuation sequence. It reflects the change in pressure values ​​in the core region of the shock wave front within the current exposure time window.

[0056] DC removal and windowing preprocessing: To eliminate the influence of the DC component and reduce spectral leakage, the operator modulation module 340 first calculates the local pressure fluctuation sequence. Arithmetic mean Then, a mean-reduction operation is performed to obtain a zero-mean sequence. Subsequently, the operator modulation module 340 utilizes the Hanning window function. The zero-mean sequence is weighted to obtain the preprocessed sequence. The preprocessing calculation formula is as follows: ; in, For preprocessed sequences, This is a local pressure fluctuation sequence. This is the arithmetic mean of the local pressure fluctuation sequence. For the Hanning window function, This is the length of the sampling window.

[0057] Fast Fourier Transform: Operator modulation module 340 pairs preprocessed sequences Perform a Fast Fourier Transform to convert the time-domain signal into a complex spectrum in the frequency domain. The transformation formula is as follows: ; in, For the first Complex spectral coefficients at each frequency point , The imaginary unit is used. The operator modulation module 340 calculates the one-sided power spectral density based on the complex spectral coefficients. The frequency axis resolution , This is the pressure sampling frequency.

[0058] Calculate the oscillation index: Operator modulation module 340 based on single-sided power spectral density Calculate the oscillation index characterizing the dynamic properties of shock trains Oscillator Defined as the ratio of energy within a characteristic frequency band to the total energy across the entire frequency band, it is used to quantify the degree of unsteadiness in the flow field. The calculation formula is as follows: ; in, It is an oscillation index. The lower limit frequency of the characteristic frequency band. The upper limit frequency of the characteristic frequency band, For pressure sampling frequency, To prevent the denominator from being a zero, extremely small positive number. Characteristic frequency band. The frequency range of the shock wave train in the wind tunnel is preset according to its inherent oscillation frequency.

[0059] Oscillation state determination and operator output: The operator modulation module 340 will calculate the oscillation index. Compared with the preset stability threshold Compare them.

[0060] If the oscillation index Less than or equal to the stability threshold The current flow field is determined to be in a quasi-steady state, and the operator modulation module 340 outputs the steady-state differential operator identifier.

[0061] If the oscillation index Greater than the stability threshold The current flow field is determined to be in an unsteady oscillation state, and the operator modulation module 340 outputs the dynamic integral fitting operator identifier.

[0062] The specific butterfly operation algorithm for Fast Fourier Transform can be implemented by those skilled in the art using digital signal processing chips or standard mathematical libraries. It is a well-known technology in the field of signal processing and will not be elaborated here.

[0063] For operator adaptive modulation, the following specific implementation process is included in the dynamic decision-making stage of image solution operators: Operator branch decision: The coupled solution module 350 receives the operator identifier output by the operator modulation module 340. If the operator identifier is a steady-state differential operator, the coupled solution module 350 executes the steady-state differential operator sub-pixel positioning logic. If the operator identifier is a dynamic integral fitting operator, the coupled solution module 350 executes the dynamic integral fitting operator positioning logic.

[0064] Subpixel localization using steady-state differential operators: The coupled solution module 350 uses the Sobel edge detection operator to process schlieren image frames. Convolution operations are performed to extract the horizontal grayscale gradient map. The coupled solution module 350 utilizes the binarized region of interest mask matrix. With spatial probability weight matrix The weighted gradient energy distribution curve is obtained by performing a weighted calculation on the horizontal grayscale gradient image. Weighted gradient energy distribution curve The calculation formula is as follows: ; in, Image horizontal pixel coordinates Accumulated gradient energy, The vertical pixel coordinates of the image. Image height, For schlieren image frames, This represents the convolution operator. The horizontal Sobel convolution kernel is used. This indicates the absolute value operation. This is a binary mask matrix for the region of interest. This is the spatial probability weight matrix.

[0065] Coupled solution module 350 searches for weighted gradient energy distribution curve Integer coordinates of peak position Integer coordinates of peak position This serves as the initial search benchmark for subsequent multi-field coupled sub-pixel accurate solutions.

[0066] Dynamic integral fitting operator localization: To address the motion blur phenomenon caused by high-speed shock wave oscillations, the coupled solution module 350 employs a Gaussian function to calculate the weighted gradient energy distribution curve. Nonlinear least squares fitting is performed. Coupled solution module 350 constructs the objective function. The position parameters of the shock train leading edge are solved through iterative optimization. Objective function The definition is as follows: ; in, The set of parameters to be optimized To fit the peak amplitude, To fit the position of the peak center, To fit the standard deviation of the peak width, Integer coordinates of peak position The local fitting neighborhood is centered. The coupled solution module 350 will output the integer coordinates of the peak position. Set as the center position of the fitted peak Initial values, weighted gradient energy distribution curve Integer coordinates at the peak position The value is set to the fitted peak amplitude. Initial values. Coupled solution module 350 minimizes the objective function using the Levenberg-Marquardt algorithm. The optimal parameters are obtained by solving. This refers to the sub-pixel position of the leading edge of the shock train at the current moment. .

[0067] The specific mathematical iteration process of the Sobel operator convolution operation and the Levenberg-Marquardt optimization algorithm can be implemented by those skilled in the art using the OpenCV computer vision library or the SciPy scientific computing library. These are well-known techniques in the field of numerical computing and will not be elaborated here.

[0068] For accurate sub-pixel calculation of multi-field coupling, when determining that the flow field is in a steady-state mode, the following specific implementation process is included: Constructing the centroid computational neighborhood: The coupled solution module 350 calculates the weighted gradient energy distribution curve based on the obtained curve. and its peak position integer coordinates Define the local integral neighborhood for subpixel centroid solving. Local integral neighborhood Integer coordinates of peak position The radius is calculated using the preset centroid as the center. It is half-width, and the coverage area is Centroid radius calculation The value is usually set to 3 to 5 pixels. This range can cover the small jitter range of the shock leading edge in steady state, while excluding background noise interference in non-shock areas.

[0069] Perform weighted gradient centroid calculation: The coupled solution module 350 will calculate the weighted gradient energy distribution curve. In the local integral neighborhood The values ​​within the range are used as weights to perform a weighted average calculation on the pixel coordinates, resulting in sub-pixel-level image coordinates of the shock wave leading edge. By introducing gradient energy based on multi-field coupling as weights, this calculation process is mathematically equivalent to solving for the physical centroid of a one-dimensional grayscale gradient field, effectively smoothing discretization errors at single pixels. The centroid calculation formula is as follows: ; in, The subpixel-level image coordinates of the shock wave leading edge These are integer coordinates of the peak position. This represents the local offset relative to the peak value. Calculate the radius for the centroid. This represents the weighted gradient energy value at the corresponding coordinates.

[0070] Steady-state subpixel coordinate cache: The coupled solution module 350 calculates the sub-pixel image coordinates of the shock wave leading edge. This is marked as the steady-state solution result. Because the coordinates at this point... Still within the image pixel space containing optical distortion, the coupling solution module 350 will... The data is then transmitted to the subsequent physical coordinate inversion module for distortion correction and spatial mapping, instead of directly outputting the physical position.

[0071] Data smoothing and output: To further suppress high-frequency electronic noise, the coupling solution module 350 is equipped with a moving average filter. The coupling solution module 350 calculates the physical position of the shock wave leading edge at multiple consecutive sampling times. The process involves performing a time-domain moving average to output the final steady-state shock wave position measurement. This measurement is transmitted via a data interface to a host computer display terminal as feedback for flow field state monitoring.

[0072] For accurate sub-pixel calculation of multi-field coupling, when the flow field is determined to be in dynamic mode, the following specific implementation process is included: Constructing an error function density field reconstruction model: The coupling solution module 350 establishes the physical relationship between the light intensity distribution and the density gradient. The shock wave leading edge shifts within an extremely short exposure time, causing the density field on the imaging sensor to exhibit a grayscale distribution in the form of an error function. This, in turn, makes the grayscale gradient field mathematically represent a Gaussian distribution. The coupling solution module 350 calls the dynamic integral fitting operator localization logic to calculate the optimal parameter set. Among them, the optimal parameters The corresponding inflection point of the error function distribution, i.e., the instantaneous geometric center of the shock wave leading edge; optimal parameters The blur width is caused by the combined effect of shock wave velocity and exposure time.

[0073] Calculate the goodness of fit and parameter confidence: To eliminate erroneous fitting results caused by extreme flow field turbulence or optical obstruction, the coupled solution module 350 calculates the ratio of the sum of squares of the fitted residuals to the sum of squares of the total deviations, and constructs the fitting coefficients of determination. Fitting coefficients of determination The calculation formula is as follows: ; in, To determine the coefficients for fitting, For local fitting neighborhood Inner image horizontal pixel coordinates, To accumulate gradient energy for actual observation, To substitute into the optimal parameter set Theoretical value of post-Gaussian model For local fitting neighborhood The arithmetic mean of the internal cumulative gradient energy.

[0074] Coupled solution module 350 determines the coefficient of determination of the fit. Does it exceed the preset confidence threshold? If the fitted coefficient of determination... If the result is greater than the confidence threshold, the coupled solution module 350 considers the solution valid; otherwise, the coupled solution module 350 discards the solution result of the current frame and marks it as invalid data.

[0075] Extracting sub-pixel centers and motion blur: After confirming the solution is valid, the coupled solution module 350 will assign the optimal parameters. (i.e., the fitted peak center position) is extracted as sub-pixel-level image coordinates of the shock wave leading edge. The coupled solution module 350 is based on optimal parameters. (i.e., the standard deviation of the fitted peak width) Calculate the shock wave motion ambiguity Shock wave motion ambiguity Defined as the full width at half maximum (FWHM) of a Gaussian distribution, it is used to quantify the intensity of shock wave oscillations within the current exposure window: ; in, Shock wave motion blur (unit: pixels). To fit the standard deviation of the peak width, It is the natural logarithm constant.

[0076] Dynamic spatiotemporal coordinate caching: The coupling solution module 350 extracts the timestamp of the current schlieren image frame. The coupled solution module 350 will use timestamps. subpixel-level image coordinates of the shock front Correlate and generate spatiotemporal coordinate pairs to be corrected. The coupled solution module 350 stores the coordinate pair in a cache queue, awaiting subsequent steps for geometric distortion correction and physical coordinate inversion to ensure the spatial accuracy of the dynamic trajectory reconstruction.

[0077] For accurate sub-pixel solving of multi-field coupled systems, the following specific implementation process is included in the physical coordinate inversion and output stages: Geometric distortion nonlinear correction: To eliminate the impact of radial distortion introduced by the imaging lens system in the edge region on sub-pixel positioning accuracy, the coupled calculation module 350 calculates the sub-pixel level image coordinates of the output shock wave leading edge. Perform nonlinear correction. The coupled solution module 350 calls the camera intrinsic parameter data stored in the calibration database, including the optical center coordinates. and radial distortion coefficient .

[0078] Considering the position of the shock wave leading edge Using one-dimensional coordinates, the coupled solution module 350 converts the image's vertical center coordinates... Using the vertical coordinate reference of this point, calculate the normalized radial distance of this point relative to the optical center. The coordinates of the corrected image are calculated based on the polynomial distortion model. .

[0079] Normalized radial distance The correction calculation formula is as follows: ; ; in: The horizontal coordinates of the output subpixel-level image of the shock front; The horizontal pixel coordinates of the camera's optical center; The vertical pixel coordinates of the camera's optical center are used here as the reference height for shock wave calculation; The denominator is the normalized radial distance relative to the optical center. This serves as a reference value for the half-diagonal length of the image used for normalization. The first-order radial distortion coefficient; It is the second-order radial distortion coefficient; These are the horizontal pixel coordinates of the image after distortion correction.

[0080] The aforementioned nonlinear correction logic can effectively compensate for optical aberrations introduced by wide-angle lenses or window glass, ensuring the geometric fidelity of physical inversion.

[0081] Solution of kinematic parameters in all states: The coupled solution module 350 calculates the physical location of the shock train based on the inverse transformation of the homography matrix and constructs a complete physical state vector. The coupled solution module 350 calls the obtained homography matrix H and calculates its inverse matrix Hinverse. -1 .

[0082] Coupled solution module 350 constructs the corrected image coordinate vector Using the inverse matrix H -1Mapping image coordinate vectors back to homogeneous coordinates in physical space : ; in, For the corrected horizontal coordinates, This represents the vertical coordinates of the image center. Here, we choose... This is based on the assumption that the shock train structure is representative of the flow channel center plane, and the two-dimensional image coordinates ( , The coordinates are mapped to the central axis of the physical space flow channel.

[0083] Subsequently, the time was calculated using homogeneous coordinate normalization. Physical location of the leading edge of the shock train : ; in, The actual position value along the flow direction in physical space (unit: millimeter or meter).

[0084] Subsequently, the coupling solution module 350 uses a five-point center difference algorithm to determine the physical location of the shock train leading edge. The sequence is numerically differentiated to calculate the instantaneous velocity of the shock train. : ; in, For a moment Instantaneous velocity of the shock train (unit: meters per second). This corresponds to the physical position of the leading edge of the shock train at that moment. This represents the sampling interval of the high-speed camera. Compared to the two-point difference algorithm, the five-point center difference algorithm can more effectively suppress the amplification effect of position measurement noise on velocity calculation.

[0085] Data encapsulation and real-time transmission: The coupled calculation module 350 encapsulates the calculated physical parameters into data packets conforming to industrial Ethernet protocols (such as UDP or EtherCAT). The data packet structure includes: a frame synchronization header, a 64-bit high-precision timestamp, and the physical location of the leading edge of the shock train. instantaneous velocity of shock train Flow field modal identifiers (steady-state / dynamic), fitting coefficients of determination And checksums. The Coupled Solver Module 350 sends data packets to the aircraft ground test control console via a gigabit Ethernet interface to adjust the total wind tunnel pressure or Mach number in real time, thereby achieving closed-loop active control of the wind tunnel flow field.

[0086] For the truncation error characteristics and numerical stability of the five-point central difference algorithm, those skilled in the art can refer to relevant textbooks on numerical analysis. It is a well-known technique in the field of computational mathematics and will not be elaborated here.

Claims

1. A method for accurate detection of the leading edge position of shock trains based on YOLOv8, characterized in that, Includes the following steps: Establish a spatial mapping relationship between the physical coordinates of the pressure sensor array component (210) and the image pixel coordinates; Simultaneously acquire flow field schlieren images and wall pressure time-series data; The flow field schlieren image is input into the YOLOv8 neural network model to generate a binarized computational mask for the region of interest of the shock train. Based on the path distribution characteristics of the wall pressure time series data, a spatial probability field is constructed to constrain the weights for image feature extraction. The frequency domain characteristics of the wall pressure time series data are analyzed to determine the flow field stability, and the steady-state differential operator or dynamic integral fitting operator is adaptively selected accordingly. Within the region defined by the binarized computation mask, the flow field schlieren image is subjected to multi-field coupling solution by combining the spatial probability field and using the selected steady-state differential operator or the dynamic integral fitting operator to obtain the sub-pixel-level image coordinates of the shock train leading edge. Based on the spatial mapping relationship, the sub-pixel level image coordinates are inverted into physical spatial locations and output.

2. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The specific steps for establishing the spatial mapping relationship between the physical coordinates of the pressure sensor array assembly (210) and the image pixel coordinates include: The homography matrix between the physical space coordinate system and the image pixel coordinate system is calculated using the feature points of the calibration plate; The physical installation position of each sensor in the pressure sensor array assembly (210) is mapped to a pixel position index in the image pixel coordinate system using the homography matrix, and the pixel position index is used as the spatial reference for constructing the spatial probability field.

3. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The steps for constructing a spatial probability field to constrain image feature extraction weights based on the along-path distribution characteristics of the wall pressure time series data specifically include: Calculate the pressure gradient along the wall pressure time series data and locate the sensor node corresponding to the maximum pressure gradient value; A one-dimensional Gaussian space weighting function is constructed using the mapping position of the sensor node in the image pixel coordinate system as the center. The one-dimensional Gaussian spatial weight function is extended into a spatial probability weight matrix with the same size as the flow field schlieren image, which is then used as the spatial probability field.

4. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The steps for analyzing the frequency domain characteristics of the wall pressure time series data to determine the stability of the flow field specifically include: Extract the local pressure time series data of the sensor node corresponding to the maximum pressure gradient, and use the local pressure time series data as the object of frequency domain analysis; The local pressure time series data is subjected to a fast Fourier transform to generate the corresponding power spectral density distribution; The ratio of characteristic band energy to full band energy is calculated based on the power spectral density distribution to obtain the oscillation index characterizing the dynamic stability of the flow field. Operator selection is performed based on the comparison result between the oscillation index and the preset threshold: when the oscillation index is less than or equal to the preset threshold, the flow field is determined to be in steady state and the steady-state differential operator is invoked; When the oscillation index is greater than the preset threshold, the flow field is determined to be dynamic and the dynamic integral fitting operator is invoked.

5. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The steps of performing multi-field coupled solution on the flow field schlieren image using the selected steady-state differential operator specifically include: An edge detection operator is applied to the flow field schlieren image to perform a convolution operation, generating a grayscale gradient map in the horizontal direction; The binarized computation mask and the spatial probability field are used as weight matrices to perform weighted processing on the gray-level gradient map, resulting in a weighted gradient energy distribution. Centroid calculation is performed within the peak neighborhood of the weighted gradient energy distribution, and the calculated weighted centroid is used as the subpixel-level image coordinates of the shock train leading edge.

6. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The step of performing multi-field coupled solution of the flow field schlieren image using the selected dynamic integral fitting operator specifically includes: The flow field schlieren image is multiplied element-wise with the binarized computation mask and the spatial probability field to generate weighted image feature data used as the fitting object. A nonlinear least squares fitting objective function based on an error function distribution model is constructed, and the objective function is used to solve for the distribution parameters of the weighted image feature data; The weighted image feature data is substituted into the nonlinear least squares fitting objective function for iterative optimization to obtain the optimal fitting peak center parameters. The optimally fitted peak center parameters are used as subpixel-level image coordinates of the shock train leading edge.

7. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The specific steps of inputting the flow field schlieren image into the YOLOv8 neural network model include: The flow field schlieren image is preprocessed and its dimensions are transformed to generate tensor data that meets the input requirements of the YOLOv8 neural network model; The tensor data is input into the YOLOv8 neural network model for inference, and the output is a prediction result containing multiple candidate bounding boxes; The non-maximum suppression algorithm is applied to the prediction results to select the optimal bounding box, and the geometric coordinates of the optimal bounding box are analyzed. The binary computation mask is constructed based on the geometric coordinates, retaining only the numerical values ​​of the region inside the optimal bounding box.

8. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, The steps for simultaneously acquiring flow field schlieren images and wall pressure time-series data specifically include: By using a hardware-triggered pulse sequence to simultaneously drive the optical observation subsystem (100) and the aerodynamic measurement subsystem (200), a shared time reference is established between the flow field schlieren image and the wall pressure time series data; Based on the ratio between the sampling frequency of the aerodynamic measurement subsystem (200) and the imaging frame rate of the optical observation subsystem (100), a linear mapping model is constructed that associates the frame index of the flow field schlieren image with the data point index of the wall pressure time series data. Substitute the frame index of the flow field schlieren image into the linear mapping model to calculate the starting pressure data point index and the ending pressure data point index of the exposure window covering the flow field schlieren image. Based on the starting pressure data point index and the ending pressure data point index, corresponding data segments are extracted from the wall pressure time series data, and the data segments are used as pressure data segments that are time-aligned with the flow field schlieren image.

9. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 2, characterized in that, The step of inverting the sub-pixel level image coordinates into physical spatial locations based on the spatial mapping relationship specifically includes: By calling camera intrinsic parameter data, the normalized radial distance of the subpixel-level image coordinates relative to the image optical center is calculated; Substitute the normalized radial distance into the polynomial distortion model to calculate the corrected image coordinates after eliminating optical distortion. The physical spatial location is obtained by performing a projection transformation on the coordinates of the corrected image using the inverse of the homography matrix.

10. The method for accurate detection of the leading edge position of shock trains based on YOLOv8 according to claim 1, characterized in that, After inverting the sub-pixel-level image coordinates into physical spatial locations based on the spatial mapping relationship and outputting them, the method further includes a step of calculating the instantaneous moving velocity of the shock wave train, specifically including: The physical spatial locations output at multiple consecutive sampling times are cached in time sequence to construct a time sequence of physical locations to be calculated; The sampling interval time of the flow field schlieren image is obtained, and the sampling interval time is used as the time step parameter for the differential operation; Substitute the physical location time sequence and the sampling interval into the five-point center difference calculation model for numerical differentiation; The calculation result of the numerical differentiation operation is output as the instantaneous moving speed of the shock train leading edge.

Images