Real-time monitoring method for coaxial rotor spacing in complex environments based on MB network

Abstract

This application relates to a method for real-time monitoring of coaxial rotor blade spacing in complex environments based on Mobile-Blade networks. It includes the following steps: installing an industrial camera and performing machine vision calibration; acquiring blade tip images; constructing a Mobile-Blade network model and training and deploying the model; locating the blade tip reference point and calculating the blade spacing. This invention effectively solves the problems of significant influence from blade aerodynamic characteristics, low accuracy of blade tip marker point detection, and slow detection speed in existing methods. It enables real-time, high-precision monitoring of coaxial rotor blade spacing in complex environments, providing data support for the design and improvement of new coaxial helicopters.

Description

Technical Field

[0001] This application relates to the field of coaxial rotor blade spacing measurement technology, specifically to a method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks. Background Technology

[0002] Coaxial rotor helicopters, due to their compact structure and excellent performance, are the main configuration for next-generation military helicopters. However, the rotor hub drag accounts for more than 50% of the total drag of a coaxial rotor helicopter, and reducing the distance between the upper and lower rotors is an effective way to reduce drag. However, at high speeds, the elastic deformation of the rotors intensifies, and an excessively small distance can easily lead to the risk of rotor tip collision.

[0003] The core problem of real-time monitoring of coaxial rotor blade spacing is: (1) When the upper and lower blades of a coaxial rotor helicopter rotate at high speeds and the monitoring equipment needs to minimize the impact on the aerodynamic characteristics of the blades, traditional contact measurement methods are difficult to implement; (2) Non-contact measurement methods based on traditional image processing methods cannot adapt to all-weather outdoor environments (i.e., morning, noon, evening, and spring, summer, autumn, and winter) and complex lighting (i.e., front lighting, back lighting, and side lighting) during outdoor experiments, and are difficult to meet the real-time monitoring requirements in terms of adaptability, detection accuracy, and detection speed. At present, measurement methods combining general detection networks such as DETR, MobileNet series, and YOLO series are gradually emerging. However, the anchor frame mechanism, multi-scale feature fusion strategy, and receptive field design of such general detection networks are mainly aimed at conventional scale targets. The detection accuracy and time for specific shapes of blade tip markers still cannot meet the actual real-time monitoring requirements of blade spacing. Summary of the Invention

[0004] The purpose of this invention is to provide a method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks. This method can effectively solve the problems of large influence of blade aerodynamic characteristics, low detection accuracy of blade tip markers, and slow detection speed in existing methods. It can achieve real-time high-precision monitoring of coaxial rotor blade spacing in complex environments, providing data support for the design and improvement of new coaxial helicopters.

[0005] The technical solution adopted in this invention is: a method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks, comprising the following steps:

[0006] S1: Install an industrial camera at the same horizontal plane at the center of the coaxial rotor tip intersection and perform machine vision calibration;

[0007] S2: Acquire image of the propeller tip;

[0008] S3: Construct a Mobile-Blade network model, and train and deploy the model; the Mobile-Blade network model includes a backbone network, a neck network, and a head network; the backbone network includes a lightweight bottleneck module for extracting features from the input image to obtain intermediate features. and intermediate features The lightweight bottleneck module includes a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer; the neck network includes a lightweight feature fusion module for processing intermediate features. and intermediate features Perform feature fusion to obtain fused features and fusion features The lightweight feature fusion module includes a 1×1 convolutional layer, an upsampling layer, and a downsampling layer; the head network includes two lightweight detection head modules for processing the fused features. and fusion features Feature decoupling is performed to obtain the confidence, category and bounding box of the paddle tip region. The lightweight detection head module includes a depth-separable convolutional layer, a 1×1 convolutional layer and a two-dimensional convolutional layer.

[0009] S4: Locate the propeller tip reference point and calculate the propeller pitch.

[0010] Furthermore, the industrial camera is installed at the same horizontal plane as the intersection of the optical axis and the coaxial rotor tip, ensuring that its imaging center and the intersection of the coaxial rotor tip motion trajectory are at the same horizontal height.

[0011] Furthermore, the specific method for machine vision calibration is as follows:

[0012] A circular calibration template was placed on the vertical plane where the tips of the coaxial rotor blades intersected. Target parameters were set, and k images of the circular calibration template were acquired using an industrial camera. The EDCircle algorithm was used to process each circular calibration template image to obtain the center coordinates and radii of all circular markers. The centers of the circular markers at the four corners of each circular calibration template image were connected to form a quadrilateral. The quadrilateral was then divided into two triangles using the line connecting the centers of the upper right and lower left corner markers as its diagonal. The pixel equivalent result of the machine vision calibration was obtained by solving the calibration using a machine vision calibration formula. The specific expression of the machine vision calibration formula is as follows:

[0013] ;

[0014] in, , and Let represent the lengths of the three sides of the i-th triangle in the j-th calibration template image. ; This represents the semi-perimeter of the i-th triangle in the j-th calibration template image; Let represent the area of ​​the i-th triangle in the j-th calibration template image; This indicates the center distance of the calibration template. This indicates the number of rows in the template. Indicates the number of columns in the calibration template; This indicates the total number of calibration template images collected; Let j represent the pixel equivalent of the j-th calibration template image, where j = 1, 2, ..., k; This represents the pixel equivalent result of machine vision calibration.

[0015] Furthermore, the specific process for acquiring propeller tip images is as follows: a reflective marker is attached to the propeller tip and illuminated with a strobe light, and then an industrial camera is used to acquire the propeller tip image.

[0016] Furthermore, the training and deployment process of the Mobile-Blade network model is as follows: paddle tip images are collected in a real environment and a virtual environment of the software; the collected paddle tip images are labeled using a labeling tool to obtain corresponding label files; the paddle tip images and corresponding label files are combined to form a paddle tip image dataset; the Mobile-Blade network model is trained using the paddle tip image dataset to obtain the optimal model weight file; the optimal model weight file is converted into a .engine format sequence file and the sequence file is deployed through the TensorRT platform.

[0017] Furthermore, the specific steps for locating the propeller tip reference point are as follows:

[0018] S401: The propeller tip image acquired in step S2 is detected using the Mobile-Blade network model to obtain the information of all propeller tip detection boxes in the image. The information of each propeller tip detection box includes the X-axis and Y-axis coordinates of the upper left corner of the detection box, the width W and the height H of the detection box; and it is determined whether the number of propeller tip detection boxes is equal to the number of detection boxes. ,in, This indicates the number of propeller tips that need to be detected correctly. If the number of blade tip detection frames is not equal to Then return to step S2 and re-acquire the propeller tip image. If the number of propeller tip detection boxes is equal to... Then proceed to the next step;

[0019] S402: Keeps the center position of each propeller tip detection frame unchanged, while increasing the area. ,in The grayscale values ​​outside the propeller tip detection box region are set to 0 to obtain an extended mask image. The propeller tip detection box region in the extended mask image is then processed using the Otsu thresholding algorithm to obtain a thresholded segmentation image. A morphological closing operation is performed on the thresholded segmentation image to obtain a morphological closing image. All propeller tip contours are extracted from the morphological closing image and sorted by area from largest to smallest. The n propeller tip contours with the largest areas are output. ;

[0020] S403: Use the minimum and maximum Y-axis tip profiles as the upper and lower tip profiles, respectively; connect the starting and ending points of the upper and lower tips, and set the aerodynamic center at three-quarters of the distance from the ending point to the connecting line; perform error correction on the starting and ending points of the upper and lower tips relative to the aerodynamic center to obtain the tip reference point. Paddle tip reference point Paddle tip reference point Paddle tip reference point Paddle tip reference point and paddle tip reference point .

[0021] Furthermore, the formula for calculating the propeller pitch is:

[0022] ;

[0023] in, This represents the measured blade tip distance. Pixel equivalent results calibrated for machine vision Indicates the propeller tip reference point Paddle tip reference point or propeller tip reference point The Y-axis coordinate of any reference point in the middle, Indicates the propeller tip reference point Paddle tip reference point or propeller tip reference point The Y-axis coordinate of any reference point in the equation.

[0024] The beneficial effects of this invention are as follows: The real-time blade pitch monitoring system built by this invention can realize non-contact measurement of coaxial rotor blade pitch; through the Mobile-Blade network model, it can achieve rapid blade tip positioning in complex outdoor environments with an accuracy of 99%; based on this, image processing operations are performed to accurately locate the reference point coordinates of the blade tips of the upper and lower blades, and combined with the pixel equivalent results calibrated by machine vision, the actual blade pitch is calculated with a measurement accuracy of less than 1mm; the average detection time per image on the TensorRT platform is 3.2ms, and the maximum frame rate is 300FPS. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of hardware device connections according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of machine vision calibration according to an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of the paddle tip image dataset according to an embodiment of the present invention; wherein, (a) is a paddle tip image collected under direct sunlight at noon on a sunny day, (b) is a paddle tip image collected under direct sunlight in the afternoon on a sunny day, and (c) is a paddle tip image collected under direct sunlight at noon on a cloudy day.

[0030] Figure 5 This is a schematic diagram of the Mobile-Blade network model in an embodiment of the present invention;

[0031] Figure 6 This is a schematic diagram of the lightweight bottleneck module of the Mobile-Blade network model in an embodiment of the present invention;

[0032] Figure 7 This is a schematic diagram of the lightweight feature fusion module of the Mobile-Blade network model in an embodiment of the present invention;

[0033] Figure 8 This is a schematic diagram of the lightweight detection head module of the Mobile-Blade network model in an embodiment of the present invention;

[0034] Figure 9 This is a detection effect diagram of the propeller tip reference point in an embodiment of the present invention;

[0035] Figure 10 This is a schematic diagram of the results of a generalization experiment under a simulated environment in an embodiment of the present invention;

[0036] Figure 11 This is a schematic diagram showing the average error results of each group under the simulated environment in an embodiment of the present invention. Detailed Implementation

[0037] To better understand the above-described objects, features, and advantages of the present invention, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the invention; however, the invention may be practiced in other ways different from those described herein, and therefore, the invention is not limited to the specific embodiments disclosed below.

[0038] like Figure 1 As shown, this embodiment of the invention provides a method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks. The hardware device connection and layout settings are as follows. Figure 2 As shown, the hardware device includes a data processing unit 101, a signal triggering device 102, a switch 103, an industrial camera 104, a strobe light 105, a photoelectric sensor 106, and a coaxial rotor 107. The data processing unit 101, signal triggering device 102, switch 103, industrial camera 104, strobe light 105, and photoelectric sensor 106 are electrically connected, enabling the transmission of the rotor tip image data of the coaxial rotor 107 acquired by the industrial camera 104 to the data processing unit 101. The real-time monitoring method for coaxial rotor blade spacing in complex environments based on an MB network includes the following steps:

[0039] S1: Install the industrial camera 104 and perform machine vision calibration. The industrial camera 104 is installed on the same horizontal plane as the intersection center of the optical axis and the tip of the coaxial rotor 107, ensuring that its imaging center and the intersection center point of the propeller tip motion trajectory of the coaxial rotor 107 are at the same horizontal height. In this embodiment of the invention, the industrial camera 104 is model UI-5040CP-M-GLRev.2.2, with a camera lens focal length of 25mm, and is equipped with a data processing unit 101, a signal triggering device 102, a switch 103, a strobe light 105, and a circular calibration template, constituting a machine vision calibration device.

[0040] The specific method for machine vision calibration is as follows:

[0041] After the machine vision calibration device is installed, adjust the focal length and aperture of the industrial camera 104 lens to produce a clear image of the paddle tip.

[0042] A circular calibration template is placed on the vertical plane where the tips of the coaxial rotor blades meet. The center distance of the circular markers on the circular calibration template is... 80.00mm, number of rows Number of columns: 6 The target parameters are set to 5, and k circular calibration template images are acquired using an industrial camera. The EDCircle algorithm is used to process each circular calibration template image to obtain the center coordinates and radii of all circular markers. All circular markers are sorted from top to bottom and left to right, and labeled from 1 to 30. In this embodiment, k is 3. For the j-th circular calibration template image, the centers of the circular markers at the four corners are connected to form a quadrilateral. The quadrilateral is then divided into two triangles by using the line connecting the centers of the upper right and lower left circular markers as a diagonal. The upper triangle is denoted as S. 1j The triangle located below is denoted as S. 2j The first triangle S 1j The side lengths are a 1j b 1j and c 1j The second triangle S 2j The side lengths are a 2j b 2j and c 2j Then, the area of ​​the quadrilateral in the circular calibration template image is calculated; finally, the machine vision calibration formula is used to solve for the pixel equivalent result of the machine vision calibration. ,like Figure 3 As shown, the specific expression of the machine vision calibration formula is:

[0043] ;

[0044] in, , and Let represent the lengths of the three sides of the i-th triangle in the j-th calibration template image. ; This represents the semi-perimeter of the i-th triangle in the j-th calibration template image; Let represent the area of ​​the i-th triangle in the j-th calibration template image; This indicates the center distance of the calibration template. This indicates the number of rows in the template. Indicates the number of columns in the calibration template; This indicates the total number of calibration template images collected; Let j represent the pixel equivalent of the j-th calibration template image, where j = 1, 2, ..., k; This represents the pixel equivalent result of machine vision calibration.

[0045] S2: Acquiring propeller tip images. The specific process is as follows: A reflective marker is attached to the propeller tip, and the marker is illuminated by a strobe light 105. Then, an industrial camera 104 acquires the propeller tip image. In this embodiment, a photoelectric sensor 106 (model SM312LVMHSQD) generates a trigger signal. The photoelectric sensor 106 is fixed below the rotor, and a reflective sticker is attached to the tip of the coaxial rotor 107 to ensure that the industrial camera 104 can capture the propeller tip image effectively during rotor-to-rotor interaction. When the coaxial rotor 107 rotates to the front of the industrial camera 104, the photoelectric sensor 106 located below the propeller blades synchronously triggers the industrial camera 104 and the strobe light 105 to acquire the propeller tip image in real time. The acquired propeller tip image is transmitted to the data processing unit 101 via a switch 103, where the trigger signals are sequentially numbered, with the number modulo the number of propeller blades. In this embodiment of the invention, the number of blades is 4. After taking the remainder of the number, the number with a remainder of 0 corresponds to the image of blade A0, the number with a remainder of 1 corresponds to the image of blade A1, the number with a remainder of 2 corresponds to the image of blade A2, and the number with a remainder of 3 corresponds to the image of blade A3. This achieves the recognition of blade images.

[0046] S3: Construct the Mobile-Blade network model, also known as the MB network model, and train and deploy the model. For example... Figure 5 As shown, the Mobile-Blade network model includes a backbone network, a neck network, and a head network. The backbone network includes a lightweight bottleneck module used to extract features from the input image to obtain intermediate features. and intermediate features .like Figure 6 As shown, the lightweight bottleneck module includes a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer; the lightweight bottleneck module includes a left main path and a right branch; the left main path contains two blocks, each block first performs a convolution operation on the input image with a stride of 1 and a kernel of 3×3, then passes it through a batch normalization layer, and finally processes it through a ReLU activation layer; the right branch processes the input image through a batch normalization layer and a ReLU activation layer, and then performs a matrix addition operation with the output of the left main path to obtain the output of the lightweight bottleneck module. The neck network includes a lightweight feature fusion module, used for intermediate features... and intermediate features Perform feature fusion to obtain fused features and fusion features .like Figure 7 As shown, the lightweight feature fusion module includes a 1×1 convolutional layer, an upsampling layer, and a downsampling layer; the lightweight feature fusion module includes an upper branch and a lower branch; the upper branch converts the intermediate features output by the backbone network... After passing through a 1×1 convolutional layer, the first and second branches are formed; the lower branches take the intermediate features output by the backbone network. After passing through a 1×1 convolutional layer, the third and fourth branches are formed; the second branch passes through an upsampling layer and a 1×1 convolutional layer, and is then added to the third branch to output the fused features. The fourth branch, after passing through a downsampling layer and a 1×1 convolutional layer, is added to the first branch to output the fused features. The head network includes two lightweight detection head modules for processing fused features. and fusion features Feature decoupling is performed to obtain the confidence score, category, and bounding box of the paddle tip region. For example... Figure 8 As shown, the lightweight detection head module includes a depthwise separable convolutional layer, a 1×1 convolutional layer, and a 2D convolutional layer; both lightweight detection head modules consist of three branches, each branch including a depthwise separable convolutional layer, a 1×1 convolutional layer, and a 2D convolutional layer; each branch first integrates the fused features output by the neck network. or fusion features The system passes through a depthwise separable convolutional layer, two consecutive 1×1 convolutional layers, and then a 2D convolutional layer. The three branches output confidence, class, and bounding box, respectively.

[0047] The training and deployment process of the Mobile-Blade network model is as follows: Paddle tip images are acquired in both real and virtual environments. The acquired paddle tip images are labeled using the LabelImg annotation tool to obtain corresponding label files. The labeled information files are then converted into text files. The images and text files are stored in two separate files in the same directory. The text file contains the number of all paddle tips, their coordinate positions, and their length and width information. The paddle tip images and their corresponding label files are combined to form a paddle tip image dataset. This dataset is then randomly divided into training, testing, and validation sets in an 8:1:1 ratio. During real-world experiments, if any paddle tip images are found to have incorrect positioning, they are added to the paddle tip image dataset for data augmentation. The paddle tip image dataset in this embodiment is as follows: Figure 4As shown, the Mobile-Blade network model is trained using the paddle tip image dataset to obtain the optimal model weight file. A .yaml file is created to record the attributes of the paddle tip image dataset, including the paths to the training, testing, and validation sets, as well as the labeled object category information. Model training hyperparameters are configured, including dataset attributes, model batch size, learning rate, and pre-trained model. The Mobile-Blade network is trained using the configured file and parameters. A regression loss function is used to calculate the error between the predicted and ground truth boxes, and the error is backpropagated to update the network weights, achieving network convergence and obtaining the optimal detection model. The optimal detection model is exported as a .onnx file, and the .onnx file is reconstructed into a .engine inference file format using the TensorRT inference engine. The .engine file is loaded into the inference engine, controlling the graphics card to perform high-speed inference to detect paddle tips in the images and obtain the location information of the paddle tip detection boxes in the images.

[0048] S4: Locate the propeller tip reference point and calculate the propeller pitch. The specific steps are as follows:

[0049] S401: The propeller tip image acquired in step S2 is detected using the Mobile-Blade network model to obtain the information of all propeller tip detection boxes in the image. The information of each propeller tip detection box includes the X-axis and Y-axis coordinates of the upper left corner of the detection box, the width W and the height H of the detection box; and it is determined whether the number of propeller tip detection boxes is equal to the number of detection boxes. ,in, This indicates the number of propeller tips that need to be detected correctly. If the number of blade tip detection frames is not equal to Then return to step S2 and re-acquire the propeller tip image. If the number of propeller tip detection boxes is equal to... Then proceed to the next step.

[0050] S402: Keeps the center position of each propeller tip detection frame unchanged, while increasing the area. ,in The grayscale values ​​outside the propeller tip detection box region are set to 0 to obtain an extended mask image. The propeller tip detection box region in the extended mask image is then processed using the Otsu thresholding algorithm to obtain a thresholded segmentation image. A morphological closing operation is performed on the thresholded segmentation image to obtain a morphological closing image. All propeller tip contours are extracted from the morphological closing image and sorted by area from largest to smallest, with the largest area being output. The outline of the paddle tip, among which In this embodiment of the invention, The value of is 15, and the value of n is 2.

[0051] S403: Use the minimum and maximum Y-axis tip profiles as the upper and lower tip profiles, respectively; connect the starting and ending points of the upper and lower tips, and set the aerodynamic center at three-quarters of the distance from the ending point to the connecting line; perform error correction on the starting and ending points of the upper and lower tips relative to the aerodynamic center to obtain the tip reference point. Paddle tip reference point Paddle tip reference point Paddle tip reference point Paddle tip reference point and paddle tip reference point The positioning result of the propeller tip reference point in this embodiment of the invention is as follows: Figure 9 As shown.

[0052] The formula for calculating propeller pitch is:

[0053] ;

[0054] in, This represents the measured blade tip distance. Pixel equivalent results calibrated for machine vision Indicates the propeller tip reference point Paddle tip reference point or propeller tip reference point The Y-axis coordinate of any reference point in the middle, Indicates the propeller tip reference point Paddle tip reference point or propeller tip reference point The Y-axis coordinate of any reference point in the equation.

[0055] The performance indicators of the embodiments of the present invention were tested in the hardware environment shown in Table 1 below.

[0056] Table 1 Experimental Hardware Platform

[0057] equipment Model parameters CPU CPU Intel(R) Core(TM) i7-12700 GPU NVIDIA GeForce GTX 3060 Memory 64GB disk Samsung SSD 990 PRO 2TB

[0058] 1. Model Comparison Test

[0059] Model comparison tests were conducted using a self-made paddle tip image dataset. The metrics of average precision, accuracy, recall, number of parameters, and number of floating-point operations were validated. Average precision was defined as mAP@50-95, accuracy as mAP@50, recall R as the proportion of correctly predicted targets to the total number of correct targets in the entire image, number of parameters as Paramas, and number of floating-point operations as GFLOPs. The test results are shown in Table 2. The results show that the present invention achieves the highest average precision and accuracy with a small number of parameters and a small number of floating-point operations.

[0060] Table 2 Model Comparison Results

[0061] network Average accuracy accuracy Recall rate Parameter quantity / M Number of floating-point operations / G MobileNetV3-small 0.4418 0.3562 0.5000 1.68 6.60 RT-DETR-L 0.6700 0.3100 0.4600 9.40 21.50 YOLOV5 0.9999 0.9955 1.0000 47.03 118.70 YOLOv8s 0.9999 0.9962 1.0000 11.20 28.40 Embodiments of the present invention 1.0000 0.9970 0.9999 2.50 8.40

[0062] 2. Adaptability Test

[0063] Reflective markers were affixed to cardboard to simulate the position and shape of the real propeller tip in an outdoor simulated environment. Images of the propeller tip were collected at locations similar to the actual experimental site, under different lighting conditions at different times of day (morning, noon, and evening). A certain number of images were selected for testing. The generalization experiment results under the simulated environment are as follows: Figure 10 As shown, where, Figure 10 (a1) in the image represents an image taken in the morning under backlighting conditions. Figure 10 (a2) in the text represents Figure 10 Positioning of the propeller tip region in (a1) Figure 10 (a3) in the text represents Figure 10 The aerodynamic center of the propeller tip reference point is located in (a1); Figure 10 (b1) in the image represents an image taken in the afternoon under backlighting conditions. Figure 10 (b2) in the text represents Figure 10 Positioning of the propeller tip region in (b1) Figure 10 (b3) in the text represents Figure 10 The aerodynamic center of the propeller tip reference point is located in (b1); Figure 10 (c1) in the image represents an image taken under backlighting conditions at midday. Figure 10 (c2) in the text represents Figure 10 Positioning of the propeller tip region in (c1) Figure 10 (c3) in the text represents Figure 10 The aerodynamic center of the propeller tip reference point is located in (c1); Figure 10 (d1) in the image represents an image taken in the morning under direct sunlight. Figure 10 (d2) in the text represents Figure 10 Positioning of the propeller tip region in (d1) Figure 10 (d3) in the text represents Figure 10 The aerodynamic center of the propeller tip reference point is located in (d1); Figure 10 (e1) in the image represents an image taken in the afternoon under direct sunlight. Figure 10 (e2) in the text represents Figure 10 Positioning of the propeller tip region in (e1) Figure 10 (e3) in the text represents Figure 10 The aerodynamic center of the propeller tip reference point is located in (e1); Figure 10 (f1) in the image represents an image taken under conditions of direct sunlight at midday. Figure 10 (f2) in the middle represents Figure 10 Positioning of the propeller tip region in (f1) Figure 10 (f3) in the text represents Figure 10 The aerodynamic center of the propeller tip reference point in (f1) is located; the simulation results show that the recognition rate reaches 100%.

[0064] 3. Detection accuracy test

[0065] Five groups of 100 images each were randomly selected from the simulated propeller tip images captured in the experiment. The calibration result was 0.7646 mm / pixel. The average error of the propeller spacing in each group of images was calculated, and the results are as follows: Figure 11 As shown in the figure, the maximum average error is 0.61 mm and the minimum is 0.09 mm, and the overall curve shows little fluctuation, indicating that the embodiment of the present invention is stable and highly accurate.

[0066] 4. Detection time test

[0067] 200 propeller pitch measurement experiments were conducted in a simulated environment. The average time for each step, including propeller tip recognition, propeller pitch calculation, and image storage, was statistically analyzed and calculated. The results are shown in Table 3. The table shows that the average total time for propeller pitch measurement per image in the simulated environment was 3.2 ms. Different models were deployed on this hardware platform, and the test results are shown in Table 4. The results indicate that the embodiment of this invention has the shortest average processing time for propeller tip recognition per image, requiring only 1.2 ms.

[0068] Table 3. Average time consumption of each step in the simulation experiment algorithm.

[0069] Processing steps Paddle tip recognition Pitch calculation Image storage Total time Average time / ms 1.2 1.0 1.0 3.2

[0070] Table 4. Test results of average processing time for different models

[0071] Model Average processing time / ms RT-DETR-L 5.2 YOLOV5 8.6 YOLOv8s 3.3 Embodiments of the present invention 1.2

[0072] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks, characterized in that, Includes the following steps: S1: Install an industrial camera at the same horizontal plane as the center of convergence of the coaxial rotor tips and perform machine vision calibration. The specific method for machine vision calibration is as follows: A circular calibration template was placed on the vertical plane where the tips of the coaxial rotor blades intersected. Target parameters were set, and k images of the circular calibration template were acquired using an industrial camera. The EDCircle algorithm was used to process each circular calibration template image to obtain the center coordinates and radii of all circular markers. The centers of the circular markers at the four corners of each circular calibration template image were connected to form a quadrilateral. The quadrilateral was then divided into two triangles using the line connecting the centers of the upper right and lower left corner markers as its diagonal. The pixel equivalent result of the machine vision calibration was obtained by solving the calibration using a machine vision calibration formula. The specific expression of the machine vision calibration formula is as follows: ； in, , and Let represent the lengths of the three sides of the i-th triangle in the j-th calibration template image. ; This represents the semi-perimeter of the i-th triangle in the j-th calibration template image; Let represent the area of ​​the i-th triangle in the j-th calibration template image; This indicates the center distance of the calibration template. This indicates the number of rows in the template. Indicates the number of columns in the calibration template; This indicates the total number of calibration template images collected; Let j represent the pixel equivalent of the j-th calibration template image, where j = 1, 2, ..., k; This represents the pixel equivalent result of machine vision calibration; S2: Acquire image of the propeller tip; S3: Construct a Mobile-Blade network model, and train and deploy the model; the Mobile-Blade network model includes a backbone network, a neck network, and a head network; the backbone network includes a lightweight bottleneck module for extracting features from the input image to obtain intermediate features. and intermediate features The lightweight bottleneck module includes a 3×3 convolutional layer, a batch normalization layer, and a ReLU activation layer; the neck network includes a lightweight feature fusion module for processing intermediate features. and intermediate features Perform feature fusion to obtain fused features and fusion features The lightweight feature fusion module includes a 1×1 convolutional layer, an upsampling layer, and a downsampling layer; the head network includes two lightweight detection head modules for processing the fused features. and fusion features Feature decoupling is performed to obtain the confidence, category and bounding box of the paddle tip region. The lightweight detection head module includes a depth-separable convolutional layer, a 1×1 convolutional layer and a two-dimensional convolutional layer. S4: Locate the propeller tip reference point and calculate the propeller pitch.

2. The method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks according to claim 1, characterized in that, The industrial camera is installed at the same horizontal plane as the intersection of the optical axis and the coaxial rotor tip, ensuring that its imaging center and the intersection of the coaxial rotor tip motion trajectory are at the same horizontal height.

3. The method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks according to claim 1, characterized in that, The specific process for acquiring propeller tip images is as follows: a reflective marker is attached to the propeller tip and illuminated with a strobe light, and then an industrial camera is used to acquire the propeller tip image.

4. The method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks according to claim 1, characterized in that, The training and deployment process of the Mobile-Blade network model is as follows: Paddle tip images are collected in both real and virtual environments using the software; the collected paddle tip images are labeled using an annotation tool to obtain corresponding label files; the paddle tip images and their corresponding label files are combined to form a paddle tip image dataset; the Mobile-Blade network model is trained using the paddle tip image dataset to obtain the optimal model weight file; the optimal model weight file is converted into a .engine format sequence file and deployed using the TensorRT platform.

5. The method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks according to claim 1, characterized in that, The specific steps for positioning the propeller tip reference point are as follows: S401: The propeller tip image acquired in step S2 is detected using the Mobile-Blade network model to obtain the information of all propeller tip detection boxes in the image. The information of each propeller tip detection box includes the X-axis and Y-axis coordinates of the upper left corner of the detection box, the width W and the height H of the detection box; and it is determined whether the number of propeller tip detection boxes is equal to the number of detection boxes. ,in, This indicates the number of propeller tips that need to be detected correctly. If the number of blade tip detection frames is not equal to Then return to step S2 and re-acquire the propeller tip image. If the number of propeller tip detection boxes is equal to... Then proceed to the next step; S402: Keeps the center position of each propeller tip detection frame unchanged, while increasing the area. ,in The grayscale values ​​outside the propeller tip detection box region are set to 0 to obtain an extended mask image. The propeller tip detection box region in the extended mask image is then processed using the Otsu thresholding algorithm to obtain a thresholded segmentation image. A morphological closing operation is performed on the thresholded segmentation image to obtain a morphological closing image. All propeller tip contours are extracted from the morphological closing image and sorted by area from largest to smallest. The n propeller tip contours with the largest areas are output. ; S403: Use the minimum and maximum Y-axis tip profiles as the upper and lower tip profiles, respectively; connect the starting and ending points of the upper and lower tips, and set the aerodynamic center at three-quarters of the distance from the ending point to the connecting line; perform error correction on the starting and ending points of the upper and lower tips relative to the aerodynamic center to obtain the tip reference point. Paddle tip reference point Paddle tip reference point Paddle tip reference point Paddle tip reference point and paddle tip reference point .

6. The method for real-time monitoring of coaxial rotor blade spacing in complex environments based on MB networks according to claim 1, characterized in that, The formula for calculating propeller pitch is: ； in, This represents the measured paddle pitch. Pixel equivalent results calibrated for machine vision Indicates the propeller tip reference point Paddle tip reference point or propeller tip reference point The Y-axis coordinate of any reference point in the middle, Indicates the propeller tip reference point Paddle tip reference point or propeller tip reference point The Y-axis coordinate of any reference point in the equation.

Images