Method for measuring geometric dimension information of port of thrust vectoring nozzle

Abstract

The invention provides a method for measuring geometric dimension information of a port of a thrust vectoring nozzle, and relates to the technical field of aerospace, and the method comprises the steps: constructing a measurement test bed of a miniature thrust vectoring nozzle model, carrying out the multi-camera synchronous shooting of the miniature thrust vectoring nozzle model moving in the measurement test bed, and then constructing a three-dimensional dense point cloud, according to the method, the three-dimensional dense point cloud is cut and filtered, the point cloud belonging to the end face of the nozzle is identified through the Point Transformer V2 model, and the point cloud of the end face of the nozzle is fitted, so that the point cloud of the end face of the nozzle is identified through the Point Transformer V2 model, compared with an existing visual detection method, the method is more accurate, in the point cloud fitting process, AprilTag coding mark points are introduced, and the point cloud fitting accuracy is improved. Therefore, the problem that the camera needs to be calibrated again after the position of the camera deviates under the complex working condition is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of aerospace technology, in particular to a method for measuring the geometric size information of a vector nozzle port. BACKGROUND

[0002] A vector nozzle is an advanced nozzle structure that can adjust the direction of thrust, widely used in modern fighter jets, missiles and spacecraft. The accurate measurement of the port geometry of the vector nozzle is of great significance to ensure engine performance and improve the maneuverability and stealth performance of the aircraft. Small deviations in the throat diameter of the nozzle will directly affect the combustion chamber pressure and exhaust velocity, while the shape and deflection angle of the nozzle expansion section determine the control accuracy and efficiency of the thrust vector. Traditional measurement methods (such as coordinate measuring machines, calipers, etc.) are not only inefficient and difficult to meet the needs of mass production, but also belong to contact measurement, which can easily cause damage or scratches to the nozzle surface, especially for high-temperature alloys or composite material-made nozzles, which may be irreversible.

[0003] The visual measurement method uses optical sensors, cameras and image processing technology to obtain three-dimensional information of the nozzle surface through a non-contact method. Its advantages include: non-contact to avoid surface damage; modern vision systems can achieve micron-level measurement; high efficiency, suitable for online detection and dynamic measurement; can cope with complex shapes and high-temperature environments. These characteristics make the visual measurement method particularly suitable for vector nozzles, which are complex in structure and harsh in working environment.

[0004] As patent CN117030274A, using binocular stereo vision principle, through high-speed camera to capture the image of glass microbead reflective target on the nozzle surface, and then reconstruct its three-dimensional coordinates, finally calculate the displacement and deformation of the nozzle. But this method obviously cannot be applied to the task of accurately measuring the nozzle area and position at the same time.

[0005] A laser tracker is a high-precision three-dimensional measurement device commonly used for geometric dimension detection of large components. In vector nozzle measurement, it is mainly used for high-precision reconstruction of nozzle profiles under static conditions, especially for quality control in laboratories and manufacturing sites. It obtains the three-dimensional coordinates of the target in space by emitting a laser beam and tracking the position of the reflecting target (usually a spherical reflector). For vector nozzles, the nozzle is usually fixed on a leveling platform, and the target is moved along the inner surface of the nozzle manually or automatically to obtain a large amount of point cloud data, and then the throat diameter, expansion segment shape, and other parameters of the nozzle are fitted. The main advantage of the laser tracker is its extremely high measurement precision (up to millimeter level) and its suitability for large components. However, its limitations are also very obvious: it is only suitable for static measurement and cannot be used for dynamic measurement when the engine is working, making it difficult to capture the true deformation under working conditions; the operation is cumbersome, the measurement process is time-consuming and labor-intensive, and requires professional operators; the environmental requirements are high, sensitive to environmental vibration and temperature changes, and not suitable for on-site online detection.

[0006] High-precision recognition of the front end surface of the nozzle is a key technology that directly determines the measurement accuracy. The goal of this technology is to accurately extract the geometric profile and feature surface of the target from industrial images / point clouds that may contain noise, uneven lighting, or high structural complexity, providing a reliable data foundation for subsequent size measurement.

[0007] Currently, a machine vision detection method based on multi-feature fusion and Canny algorithm is provided, which uses a high-precision industrial camera to capture part images and reduces reflections and shadows through a uniform illumination optical system. Then, the image is grayed, denoised by Gaussian filtering, and other preprocessing operations are performed to improve the image quality. The Sobel operator and other operators are used to calculate the gradient intensity and direction of each pixel point in the image, retain the pixel points with the maximum gradient amplitude in the local, thin the edges, and set high and low thresholds. Strong gradient pixel points are directly determined as edges, and weak gradient pixel points need to be connected with strong edge pixels to be retained. The findContours function in OpenCV is used to extract closed contours from the binary edge image, and further calculate the geometric features of the contours (such as area, perimeter, minimum bounding rectangle, etc.), which are used for subsequent screening and measurement. However, the double-threshold mechanism may cause weak edges to be lost or edges to be broken. The lag threshold processing may fail for complex nozzle edge connectivity judgment. The measurement error of key dimensions of the nozzle (such as throat diameter, roundness) may reach several pixels, which cannot meet the micron-level or higher precision measurement requirements in the aerospace field.

[0008] Currently, there are also boundary recognition methods based on point cloud geometric features and deep learning, including normal discontinuity detection and curvature analysis based on geometric features. Such methods are heavily dependent on the selection of threshold values and are sensitive to point cloud density and noise. Deep learning methods use neural network architectures such as PointNet++ and DGCNN that can directly process unordered point clouds. These networks can learn global and local features of point clouds and directly output the probability or label of each point being a boundary point, achieving end-to-end boundary recognition. Deep learning methods can learn more complex and robust feature representations and are more adaptable to noise and changes in point cloud density, but require a large amount of labeled data for training and have relatively poor model interpretability. SUMMARY

[0009] In view of the deficiencies of the prior art, the purpose of the present application is to provide a vector nozzle port geometric size information measurement method, comprising: constructing a measurement test bench for a micro vector nozzle model; synchronously photographing the moving micro vector nozzle model in the measurement test bench with multiple cameras to obtain multi-angle two-dimensional images; based on the multi-angle two-dimensional images, constructing a three-dimensional dense point cloud of the micro vector nozzle model; cropping and filtering the three-dimensional dense point cloud to obtain a key dense point cloud; inputting the key dense point cloud into a category probability prediction model to obtain the confidence of each point cloud, wherein the confidence of the point cloud represents the probability that the point cloud belongs to the nozzle end face; screening the key dense point cloud according to the confidence of the point cloud to obtain the point cloud of the nozzle end face of the micro vector nozzle model; fitting and measuring the point cloud of the nozzle end face and the point cloud of the reference plate to obtain the nozzle diameter, the vector deflection angle, and the vector azimuth angle of the micro vector nozzle model.

[0010] Optionally, the measurement test bench for the micro vector nozzle model comprises a support platform frame, a reference plate, a nozzle base, a nozzle middle section, a nozzle front section, a converging sheet, a first nozzle angle adjusting device, a second nozzle angle adjusting device, and a nozzle opening adjusting device. The reference plate is fixed on a support platform frame, the support platform frame can be automatically leveled by leveling ground corners, so that the reference plate is perpendicular to the ground; one end of the nozzle base is vertically fixed on the reference plate; the first nozzle angle adjusting device is connected by a screw between the other end of the nozzle base and one end of the nozzle middle section, the first nozzle angle adjusting device is used to rotate the nozzle middle section, the second nozzle angle adjusting device is connected by a screw between the other end of the nozzle middle section and one end of the nozzle front section, the second nozzle angle adjusting device is used to rotate the nozzle front section, one end of the nozzle opening adjusting device is connected by a fixed hinge on the other end of the nozzle front section, the other end of the nozzle opening adjusting device is hinged on one end of the converging sheet, the nozzle opening adjusting device is used to control the opening and closing size of the nozzle of the micro-scale vector nozzle model, and the other end of the converging sheet is the nozzle end face.

[0011] Optionally, based on the multi-angle two-dimensional image, a three-dimensional dense point cloud of the micro-scale vector nozzle model is constructed, including: The multi-angle two-dimensional image is processed by using a motion recovery structure method to obtain camera motion parameters and a three-dimensional sparse point cloud, and the three-dimensional sparse point cloud and the camera motion parameters are processed by using a passive multi-view stereo vision method to obtain a three-dimensional dense point cloud.

[0012] Optionally, the three-dimensional dense point cloud is cropped and filtered to obtain a key dense point cloud, including: The three-dimensional dense point cloud is cropped based on point cloud density to obtain a cropped dense point cloud, the cropped dense point cloud is voxel down-sampled to obtain a down-sampled dense point cloud, and the down-sampled dense point cloud is filtered to obtain the key dense point cloud.

[0013] Optionally, the category probability prediction model is obtained by training a PointTransformer V2 model based on training samples and a loss function, the training samples include input samples and output samples, the input samples are point clouds, and the output samples are probabilities that the point clouds belong to the nozzle end face; The loss function is represented as: ; Wherein, is a loss function value, is the probability that the point cloud belongs to the nozzle end face, α is a category balance factor, and γ is a focusing factor.

[0014] Optionally, the key dense point cloud is screened according to the confidence of the point cloud to obtain the point cloud of the nozzle end face of the micro-scale vector nozzle model, including: All point clouds are sorted from high to low according to the confidence to obtain a confidence list, and the point clouds with a confidence greater than or equal to a preset threshold are obtained according to the preset threshold to form an initial Endface point cloud. fitting the initial Endface point cloud into a spatial circle, measuring the diameter of the spatial circle to obtain the initial nozzle diameter of the scaled vector nozzle; determining the number of point clouds corresponding to the initial nozzle diameter based on a functional relationship between the number of point clouds and the nozzle diameter, multiplying the number of point clouds by 1.25 to obtain the final number of point clouds N; obtaining the first N point clouds in the confidence list as the point clouds of the nozzle end face.

[0015] Optionally, the point clouds of the nozzle end face and the point clouds of the reference plate are fitted and measured to obtain the nozzle diameter, the vector deflection angle and the vector azimuth angle of the scaled vector nozzle model, comprising: fitting the point clouds of the reference plate by using the RANSAC method; fitting the point clouds of the nozzle end face to obtain a fitting plane, projecting the point clouds of the nozzle end face onto the fitting plane, fitting a circle on the projected point clouds on the fitting plane, measuring the diameter of the fitted circle to obtain the nozzle diameter in the point clouds, multiplying the nozzle diameter in the point clouds by a size factor to obtain the nozzle diameter of the scaled vector nozzle model, and measuring the vector deflection angle and the vector azimuth angle based on the normal of the reference plate and an arbitrary direction of the edge of the reference plate as a standard.

[0016] Optionally, the size factor is obtained by the following method: An encoding marker point is arranged at each of the four corners of the reference plate of the scaled vector nozzle model, the distance between any two encoding marker points is obtained as a real distance, and the distance between the corresponding two encoding marker points in the three-dimensional dense point cloud is obtained as a point cloud distance, and the ratio of the real distance to the point cloud distance is the scale factor.

[0017] The beneficial effects produced by the above technical solutions are: The application constructs a measurement test bench for the scaled vector nozzle model, and synchronously photographs the moving scaled vector nozzle model in the measurement test bench by multiple cameras, thereby constructing a three-dimensional dense point cloud, cropping and filtering the three-dimensional dense point cloud, identifying the point clouds belonging to the nozzle end face by the Point Transformer V2 model, and fitting the point clouds of the nozzle end face. Thus, the application realizes the identification of the point clouds of the nozzle end face by the Point Transformer V2 model, which is more accurate than the existing visual detection method. In the point cloud fitting process, the scale factor is obtained by introducing the AprilTag encoding marker point, thereby avoiding the problem of re-calibration after the camera position is offset under complex working conditions. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is a schematic diagram of the nozzle diameter Φ, the vector deflection angle α and the vector azimuth angle β in the embodiment of the application. Figure 2 A flow chart of a vectoring nozzle port geometry information measurement method in an embodiment of the present application; Figure 3 A structural schematic diagram of a measurement test bench of a micro vectoring nozzle model in an embodiment of the present application, in which, 1 is a support platform frame, 2 is a reference plate, 3 is a nozzle base, 4 is a nozzle middle section, 5 is a nozzle front section, 6 is a convergent piece, 7 is a first nozzle angle adjusting device, 8 is a second nozzle angle adjusting device, and 9 is a nozzle opening adjusting device; Figure 4 A functional relationship diagram of the number of GT endface points and nozzle diameter Φ in an embodiment of the present application; Figure 5 A schematic diagram of a mark coding point in an embodiment of the present application. DETAILED DESCRIPTION

[0019] The specific embodiments of the present application will be further described in detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0020] In view of the problems of low efficiency, large error, and dynamic parameter distortion in the operation of the existing vectoring nozzle attitude parameter measurement technology, the present application provides a vectoring nozzle port geometry information measurement method, which can complete the vectoring nozzle port geometry information measurement tasks such as nozzle diameter Φ, vector deflection angle α (the "amplitude" of the nozzle end face normal relative to the original position deflection), and vector azimuth angle β (the "direction" of the nozzle end face normal relative to the original position deflection) with high precision through a passive vision method, like Figure 1 A schematic diagram of nozzle diameter Φ, vector deflection angle α, and vector azimuth angle β.

[0021] Based on this, the present application provides a vectoring nozzle port geometry information measurement method, which combines Figure 2 may include the following contents: constructing a measurement test bench of a micro vectoring nozzle model; combined with Figure 3 The measurement test bench of the micro vectoring nozzle model includes a support platform frame 1, a reference plate 2, a nozzle base 3, a nozzle middle section 4, a nozzle front section 5, a convergent piece 6, a first nozzle angle adjusting device 7, a second nozzle angle adjusting device 8, and a nozzle opening adjusting device 9; The reference plate 2 is fixed on the support platform frame 1, the support platform frame 1 can be automatically leveled by leveling the corner, so that the reference plate 2 is perpendicular to the ground level; one end of the nozzle base 3 is vertically fixed on the reference plate 2; the first nozzle angle adjusting device 7 is connected by screw between the other end of the nozzle base 3 and one end of the nozzle middle section 4, the first nozzle angle adjusting device 7 is used to rotate the nozzle middle section 4, the second nozzle angle adjusting device 8 is connected by screw between the other end of the nozzle middle section 4 and one end of the nozzle front section 5, the second nozzle angle adjusting device 8 is used to rotate the nozzle front section 5, one end of the nozzle opening adjusting device 9 is connected by fixed hinge on the other end of the nozzle front section 5, the other end of the nozzle opening adjusting device 9 is hinged on one end of the convergent piece 6, the nozzle opening adjusting device 9 is used to control the opening and closing size of the nozzle of the vector nozzle, the other end of the convergent piece 6 is the nozzle end face.

[0022] Wherein the first nozzle angle adjusting device 7 and the second nozzle angle adjusting device 8 are composed of gear structure, motor and photoelectric limit stop, the nozzle opening adjusting device 9 is composed of cylindrical recessed groove slider, actuator ring and connecting rod, the nozzle opening and closing blade is connected by hinge on the nozzle front end, the laser ranging sensor is arranged on the nozzle opening and closing blade, the attitude solution sensor is arranged at the inside center point of the nozzle front section.

[0023] The moving micro vector nozzle model in the measurement test bed is synchronously photographed by multiple cameras, and multiple angle two-dimensional images are obtained; Based on the multiple angle two-dimensional images, the three-dimensional dense point cloud of the micro vector nozzle model is constructed; Specifically, the motion recovery structure method SFM is used to process the multiple angle two-dimensional images, and the camera motion parameters and the three-dimensional sparse point cloud are obtained, and the passive multi-view stereo vision method MVS is used to process the three-dimensional sparse point cloud and the camera motion parameters, and the three-dimensional dense point cloud is obtained.

[0024] SFM aims to recover camera motion parameters (including intrinsic and extrinsic parameters) from a set of uncalibrated two-dimensional images and generate a three-dimensional sparse point cloud of the scene. The basic principle is to extract image feature points (such as SIFT, ORB, etc.) and match them, estimate the relative pose between cameras using matched points, calculate the three-dimensional space position of the feature points through triangulation principle, and finally use bundle adjustment to jointly optimize camera parameters and three-dimensional point coordinates to minimize the re-projection error. MVS is based on the known camera parameters and sparse point cloud provided by SFM, and aims to generate a dense three-dimensional point cloud of the scene. Its core principle is to estimate the depth value of a large number of pixels or even each pixel in the image through multi-view photo-consistency comparison using epipolar geometry constraints, and finally obtain a dense scene geometry through depth map fusion.

[0025] Crop and filter the three-dimensional dense point cloud to obtain a key dense point cloud; Specifically, the three-dimensional dense point cloud is cropped based on point cloud density to obtain a cropped dense point cloud. Specifically, the region where the point cloud density is greater than or equal to a preset point cloud density and the neighborhood of the region are retained, for example, the radius of the region is 1, and the radius of the region and its neighborhood can be 1.5. The cropped dense point cloud is voxel down-sampled to reduce the point cloud density, and a down-sampled dense point cloud is obtained. The down-sampled dense point cloud is filtered, specifically, statistical analysis-based filtering, spatial density-based filtering, spatial down-sampling-based filtering, and smoothing processing-based filtering can be used to remove outliers and noise points, and a key dense point cloud is obtained.

[0026] Since the number of endface class points (i.e. points on the nozzle end face) is much less than the number of other class points (i.e. other points), direct point cloud segmentation will result in serious false detection and missed detection, which will seriously affect the final measurement result. The Focal Loss module is introduced into the PointTransformer V2 model to improve the weight of the points on the end face in the loss calculation part. When the endface class is misclassified as the other class, the model loss will be significantly improved, which highly avoids the occurrence of the model missing detection phenomenon.

[0027] The key dense point cloud is input into a category probability prediction model to obtain the confidence of each point cloud, wherein the confidence of the point cloud represents the probability that the point cloud belongs to the nozzle end face. The category probability prediction model is obtained by training the Point Transformer V2 model based on training samples and a loss function. The training samples include input samples and output samples. The input samples are point clouds, and the output samples are the probabilities that the point clouds belong to the nozzle end face. wherein the loss function is represented as: ; wherein, is a loss function value, is a probability of the point cloud belonging to the nozzle end face, and a = [0.999, 0.001] is a class balance factor and g = 2 is a focus factor.

[0028] Considering that the “correct” points are closer to the nozzle end face and the probability of belonging to the endface class is greater than that of the “false detection” points, the points in the point cloud are sorted according to the confidence (i.e., the class probability p t of the endface class) in this strategy, and the closer a point is to GTendface, the higher its prediction score and the higher its ranking. A confidence “threshold” is set for the point cloud of each angle, and the points with a confidence higher than the threshold are classified as the endface class. The specific process is as follows: According to the confidence of the point cloud, the key dense point cloud is screened to obtain the point cloud of the nozzle end face of the scaled vector nozzle model; Specifically, all the point clouds are sorted from high to low according to the confidence to obtain a confidence list, and the point clouds with a confidence greater than or equal to a preset threshold are obtained according to the preset threshold to form an initial Endface point cloud. The initial Endface point cloud is fitted into a spatial circle, which can be realized by an iterative optimization method based on a mathematical model, a method based on robust statistics and sampling, a method based on surface reconstruction and interpolation, and a method based on deep learning. The diameter of the spatial circle is measured to obtain the initial nozzle diameter of the scaled vector nozzle. Since the number of GTendface points has a certain functional relationship with the nozzle diameter F, as shown below, Figure 4 Since the point cloud density set by the reconstructed point cloud is certain, during the opening and closing of the nozzle, the nozzle diameter is smallest to the nozzle opening to all the blades to form a complete circle and realize non-interference contact. The number of points on the reconstructed end face depends on the nozzle circumference at this time, i.e., it is linearly related to the nozzle diameter, and the number of point clouds also increases linearly with the increase of the diameter. After that, the nozzle continues to open, and since the blade ports are separated from each other, the point cloud does not appear in the gap, and the number of points on the end face does not increase, and the number of point clouds remains constant.

[0029] Based on the functional relationship between the number of point clouds and the nozzle diameter, the number of point clouds of the initial nozzle diameter is determined, and the number of point clouds is multiplied by 1.25 to obtain the final number of point clouds N. The first N point clouds in the confidence list are obtained and used as the point cloud of the nozzle end face.

[0030] The point cloud of the nozzle end face and the point cloud of the reference plate are fitted and measured to obtain the nozzle diameter, the vector deflection angle, and the vector azimuth angle of the scaled vector nozzle model.

[0031] Considering that RANSAC can tolerate more outliers than least squares, the point cloud of the reference plate is fitted by using the RANSAC method; the point cloud of the nozzle end face is fitted to obtain a fitting plane, the point cloud of the nozzle end face is projected onto the fitting plane, the point cloud projected on the fitting plane is fitted with a circle, the diameter of the fitted circle is measured to obtain the nozzle diameter in the point cloud, the nozzle diameter in the point cloud is multiplied by the size factor to obtain the nozzle diameter of the scaled vector nozzle model, and the normal of the reference plate and the edge of the reference plate in any direction are taken as the standard to measure the vector deflection angle and the vector azimuth angle.

[0032] The size factor is obtained by the following method: An encoded marker point is arranged at each of the four corners of the reference plate of the scaled vector nozzle model, wherein the marker encoded point is as shown in Figure 5 The distance between any two encoded marker points is taken as the real distance, and the distance between the corresponding two encoded marker points in the three-dimensional dense point cloud is taken as the point cloud distance, and the ratio of the real distance to the point cloud distance is the scale factor.

[0033] The present application can accurately measure the angle change and nozzle opening size of the vector nozzle in a laboratory environment, the real scale information of the point cloud is obtained by introducing AprilTag encoded marker points, thereby avoiding the problem of re-calibration after the camera position is offset under complex working conditions, the Point Transformer V2 model is improved to complete the accurate cutting of the vector nozzle end face, the cutting accuracy is as shown in Table 1, and the final visual measurement accuracy reaches an error ≤±1.5% (diameter).

[0034] Table 1 Experimental accuracy IoU Acc Training set 0.5194 0.9972 Test set 0.4501 0.9282 The above description is only the preferred embodiments of the present disclosure and the explanation of the applied technical principles. Those skilled in the art should understand that the scope of the application involved in the embodiments of the present disclosure is not limited to the technical solutions formed by the specific combinations of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or equivalent features without departing from the above inventive concept. For example, the above features are replaced with the technical features disclosed in the embodiments of the present disclosure (but not limited to) having similar functions to form technical solutions.

Claims

1. A method for measuring the geometric dimensions of a vector nozzle port, characterized in that, include: A measurement test bench for constructing a miniature vector nozzle model; Multi-camera synchronous imaging was performed on a miniature vector nozzle model moving in the measurement test rig to obtain multi-angle two-dimensional images; Based on multi-angle two-dimensional images, a three-dimensional dense point cloud of a miniature vector nozzle model is constructed. The key dense point cloud is obtained by cropping and filtering the 3D dense point cloud. The key dense point cloud is input into the category probability prediction model to obtain the confidence level of each point cloud. The confidence level of the point cloud represents the probability that the point cloud belongs to the nozzle end face. Based on the confidence level of the point cloud, the key dense point cloud is filtered to obtain the point cloud of the nozzle end face of the miniature vector nozzle model. By fitting and measuring the point cloud of the nozzle end face and the point cloud of the reference plate in the measurement test rig, the nozzle diameter, vector deflection angle and vector azimuth angle of the miniaturized vector nozzle model are obtained.

2. The method for measuring the geometric dimensions of a vector nozzle port according to claim 1, characterized in that, The measurement test bench for the miniature vector nozzle model includes a support platform frame, a reference plate, a nozzle base, a nozzle middle section, a nozzle front section, a convergent plate, a first nozzle angle adjustment device, a second nozzle angle adjustment device, and a nozzle opening adjustment device. The reference plate is fixed on the support platform frame, which can be automatically leveled by adjusting the ground angle to make the reference plate perpendicular to the ground. One end of the nozzle base is vertically fixed on the reference plate. The first nozzle angle adjustment device is connected to the other end of the nozzle base and one end of the nozzle middle section by screws. The first nozzle angle adjustment device is used to rotate the nozzle middle section. The second nozzle angle adjustment device is connected to the other end of the nozzle middle section and one end of the nozzle front section by screws. The second nozzle angle adjustment device is used to rotate the nozzle front section. One end of the nozzle opening adjustment device is connected to the other end of the nozzle front section by a fixed hinge. The other end of the nozzle opening adjustment device is hinged to one end of the convergent plate. The nozzle opening adjustment device is used to control the opening and closing size of the nozzle of the miniature vector nozzle model. The other end of the convergent plate is the nozzle end face.

3. The method for measuring the geometric dimensions of a vector nozzle port according to claim 1, characterized in that, Based on multi-angle two-dimensional images, a three-dimensional dense point cloud of a miniature vector nozzle model is constructed, including: The structure-of-motion method is used to process multi-angle two-dimensional images to obtain camera motion parameters and three-dimensional sparse point clouds. Then, the three-dimensional sparse point clouds and camera motion parameters are processed by a passive multi-view stereo vision method to obtain three-dimensional dense point clouds.

4. The method for measuring the geometric dimensions of a vector nozzle port according to claim 1, characterized in that, The 3D dense point cloud is cropped and filtered to obtain the key dense point cloud, including: The 3D dense point cloud is cropped based on the point cloud density to obtain the cropped dense point cloud. The cropped dense point cloud is then downsampled using voxels to obtain the downsampled dense point cloud. Finally, the downsampled dense point cloud is filtered to obtain the key dense point cloud.

5. The method for measuring the geometric dimensions of a vector nozzle port according to claim 1, characterized in that, The category probability prediction model is obtained by training the Point Transformer V2 model based on training samples and a loss function. The training samples include input samples and output samples. The input samples are point clouds, and the output samples are the probability that the point cloud belongs to the nozzle end face. The loss function is expressed as: ； in, The value of the loss function. Let α be the probability that the point cloud belongs to the nozzle end face, α be the class balance factor, and γ be the focusing factor.

6. The method for measuring the geometric dimensions of a vector nozzle port according to claim 1, characterized in that, Based on the confidence level of the point cloud, key dense point clouds are filtered to obtain the point cloud of the nozzle end face of the miniature vector nozzle model, including: Sort all point clouds from high to low confidence to obtain a confidence list. Based on a preset threshold, obtain point clouds with confidence scores greater than or equal to the preset threshold to form the initial Endface point cloud. The initial Endface point cloud is fitted into a spatial circle, and the diameter of the spatial circle is measured to obtain the initial nozzle diameter of the miniaturized vector nozzle. Based on the functional relationship between the number of point clouds and the nozzle diameter, the number of point clouds with the initial nozzle diameter is determined, and the number of point clouds is multiplied by 1.25 to obtain the final number of point clouds N. Obtain the top N point clouds from the confidence list and use them as the point cloud of the nozzle end face.

7. The method for measuring the geometric dimensions of a vector nozzle port according to claim 1, characterized in that, By fitting and measuring the point cloud of the nozzle end face and the point cloud of the reference plate, the nozzle diameter, vector deflection angle, and vector azimuth angle of the miniaturized vector nozzle model are obtained, including: The RANSAC method is used to fit the point cloud of the reference plate; the point cloud of the nozzle end face is also fitted to obtain a fitting plane. The point cloud of the nozzle end face is projected onto the fitting plane, and the projected point cloud on the fitting plane is fitted with a circle. The diameter of the fitted circle is measured to obtain the nozzle diameter in the point cloud. The nozzle diameter in the point cloud is multiplied by the size factor to obtain the nozzle diameter of the miniaturized vector nozzle model. The normal of the reference plate and any direction of the reference plate edge are used as standards to measure the vector deflection angle and vector azimuth angle.

8. The method for measuring the geometric dimensions of a vector nozzle port according to claim 7, characterized in that, The size factor is obtained in the following way: In the baseline plate of the miniature vector nozzle model, an coded marker point is set at each of the four corners. The distance between any two coded marker points is taken as the true distance, and the distance between the corresponding two coded marker points in the 3D dense point cloud is taken as the point cloud distance. The ratio of the true distance to the point cloud distance is the scale factor.

Images