Helicopter blade vibration amount monitoring method based on deep learning dichotomy

Abstract

The invention relates to a helicopter blade vibration monitoring method based on a deep learning dichotomy. The helicopter blade vibration monitoring method comprises the following steps: establishing a mapping relation from a two-dimensional pixel coordinate to a three-dimensional world coordinate; collecting a real-time blade vibration image; constructing an improved YOLOv8n network, and using a deep learning dichotomy to complete the circle detection of the blade mark points, specifically, using the improved YOLOv8n network to position the blade mark points and using an elliptical segmentation network to segment the blade mark points, and obtaining the circle center pixel coordinates of the blade mark points; collecting the initial position of the paddle mark point; and resolving the blade vibration amount in real time. According to the method, the problems of poor adaptability, low positioning precision and insufficient real-time processing performance of a traditional visual monitoring method in a complex environment can be solved, high-robustness real-time monitoring of millimeter-level vibration quantity is realized, and a reliable technical means is provided for life cycle fatigue evaluation of the helicopter blade.

Description

Technical Field

[0001] The present application relates to the technical field of computer vision real-time monitoring, and specifically to a method for monitoring helicopter blade vibration based on deep learning dichotomy. Background Art

[0002] As a core power component, the dynamic characteristics of helicopter blades directly affect flight safety. Traditional blade vibration monitoring methods are mainly divided into contact and non-contact methods: contact measurement usually uses strain gauges or accelerometers directly fixed on the blade surface. Although local strain information can be obtained, there are problems such as additional mass changing the structural dynamic characteristics and complex wiring affecting the operation of rotating parts; in non-contact measurement, optical measurement based on machine vision has gradually become a research hotspot due to its advantages such as full-domain measurement and no additional mass.

[0003] Among the existing visual measurement methods, the commonly used one is the marker tracking technology based on traditional image processing. This method arranges high-contrast circular markers on the surface of the blade, uses Canny edge detection combined with the least squares ellipse fitting algorithm to extract the coordinates of the center of the circle, and then calculates the three-dimensional vibration through coordinate mapping. However, this method has significant defects: 1. It relies on ideal lighting conditions. Under complex ambient light or shadow interference, edge detection is easily affected by noise, resulting in increased ellipse fitting errors; 2. When the markers are blurred or partially blocked due to high-speed rotation, it is difficult for traditional algorithms to accurately segment the marker area, and the accuracy of center positioning drops sharply, with the error often exceeding 2 pixels; 3. The processing flow lacks adaptive capabilities, and the threshold parameters need to be adjusted repeatedly by humans. The single-frame processing time is >30ms, which cannot meet the real-time requirements when the vibration frequency exceeds 5Hz.

[0004] Although some studies in recent years have attempted to introduce general detection networks such as Faster R-CNN network and YOLOv5 network into vibration monitoring, there are still the following shortcomings: 1. The network structure is redundant. The detection head is designed for multi-scale targets, while the blade marker point size is single, resulting in a waste of computing resources; 2. The traditional convolution module has a large number of parameters, making it difficult to achieve real-time inference in embedded devices; 3. The end-to-end method of directly regressing the center coordinates lacks sub-pixel positioning capabilities, and the measurement error is generally higher than 1.5mm, which cannot meet the needs of high-precision fatigue testing. Summary of the invention

[0005] The purpose of the present invention is to provide a helicopter blade vibration monitoring method based on deep learning dichotomy, which can solve the problems of poor adaptability of traditional visual monitoring methods in complex environments, low positioning accuracy and insufficient real-time processing performance, and realize highly robust real-time monitoring of millimeter-level vibration, providing a reliable technical means for fatigue assessment of helicopter blades throughout their life cycle.

[0006] The technical solution adopted by this application is: a helicopter blade vibration monitoring method based on deep learning dichotomy, comprising the following steps: Step S1: Establishing a mapping relationship from two-dimensional pixel coordinates to three-dimensional world coordinates; Step S2: collecting real-time blade vibration images; Step S3: construct an improved YOLOv8n network, use lightweight convolution modules to replace traditional convolution modules in all Bottleneck modules in the YOLOv8n network, and delete the large target detection head of the YOLOv8n network; use deep learning dichotomy to complete the circular detection of the blade markers, specifically including using the improved YOLOv8n network to locate the blade markers and using the ellipse segmentation network to segment the blade markers, and obtain the pixel coordinates of the center of the blade markers; Step S4: collecting the initial position of the blade marking point; Step S5: Calculate the blade vibration in real time.

[0007] Furthermore, the mapping relationship from the two-dimensional pixel coordinates to the three-dimensional world coordinates in step S1 is: ; in,( u , v ) is the two-dimensional pixel coordinate of the marking point on the blade, z c is the distance from the marking point to the origin along the optical axis in the camera coordinate system, K is the camera’s intrinsic parameter matrix, R is the rotation matrix from the camera's 2D pixel coordinate system to the vibration plane's 3D world coordinate system, t is the translation matrix from the camera's 2D pixel coordinate system to the vibration plane's 3D world coordinate system, ( x w , y w , z w ) is the three-dimensional world coordinate of the marking point in the three-dimensional world coordinate system of the vibration plane.

[0008] Furthermore, the lightweight convolution module includes a channel splitting module, a two-dimensional convolution module and a channel splicing module; the input feature map is first split into feature maps according to the proportion by the channel splitting module. Figure 1 and Features Figure 2 ,feature Figure 1 After feature extraction by two-dimensional convolution module and feature Figure 2 Channel concatenation is performed to obtain the output feature map.

[0009] Furthermore, the specific steps of step S3 are: Step S301: using the LabelImg annotation tool to annotate the marking points on the real-time blade vibration image collected in step S2, obtaining a corresponding label file, and combining the blade vibration image and the corresponding label file into a marking point detection data set; Step S302: training the improved YOLOv8n marker detection network, inputting the marker detection data set into the trained improved YOLOv8n marker detection network for recognition, and outputting the position information of the markers; Step S303: cropping a blade marker point image from the blade vibration image, and performing binary segmentation on the cropped marker point image to obtain a marker point binary image of the same size as the cropped blade marker point image, and forming a marker point segmentation data set with the blade marker point image and the corresponding marker point binary image; Step S304: constructing an ellipse segmentation network, the ellipse segmentation network includes three parts: an encoding network D1, a decoding network D2 and an output network D3; the encoding network D1 includes an input layer and an encoding layer, the encoding layer includes three double-layer convolutional layers and three pooling layers that are alternately arranged; the decoding network D2 includes three upsampling layers and three double-layer convolutional layers that are alternately arranged, and each upsampling layer is jump-connected to the double-layer convolutional layer in the encoding network D1; the output network D3 includes a point-by-point convolutional layer; wherein the double-layer convolutional layer includes an input layer, two convolutional layers that are alternately arranged, two batch normalization layers, two activation function layers and an output layer; Step S305: training the ellipse segmentation network, using binary cross entropy loss as a loss function to calculate the segmentation error of the marker points during the training process, back-propagating the error, updating the network parameters, achieving network convergence, and continuously training the network until the error between the predicted marker point center coordinates and the true center coordinates is within 0.5 pixels, and based on the calculated segmentation error, saving a set of model parameters with the minimum loss value or the minimum average center error during the training process; Step S306: Load the ellipse segmentation network model and set the network's hyperparameters, and crop all circular marker point images from the blade vibration image according to the position information of the marker points obtained in step S302, and scale them to the same size as the cropped blade marker point images; then input them into the ellipse segmentation network for processing, the ellipse segmentation network outputs a binary image of each marker point and performs ellipse fitting on the binary image to obtain the center coordinates of all marker points; according to the scaling relationship when the marker points are cropped, the center coordinates are transformed into the original image to obtain the two-dimensional pixel coordinates of the center of each circular marker point in the original image.

[0010] Furthermore, the calculation formula of the loss function is as follows: ; in, YP Represents the predicted value output by the ellipse segmentation network; YT Represents the true value corresponding to the input image; Loss ( YT,YP ) represents the loss function, that is, the segmentation error between the network prediction value and the true value, yp ij Represents the predicted image i OK j Pixel values ​​on columns; yt ij Represents the real image i OK j The pixel value on the column.

[0011] Furthermore, the specific steps of step S4 are: collecting a blade image in a static state as an initial state, using a circle center positioning method based on deep learning dichotomy to obtain the center two-dimensional pixel coordinates of all marking points in the initial state, and according to the mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates, converting each blade marking point from the center two-dimensional pixel coordinates to the three-dimensional world coordinates as ( XS k , YS k , 0) and save, where, k =1, 2, …, n , n Indicates the number of marking points on the blade. XS k Indicates k The initial three-dimensional horizontal coordinates of the marker points, YS k Indicates k The initial three-dimensional vertical coordinates of the marker points.

[0012] Furthermore, the specific steps of step S5 are: using an industrial camera to take a blade vibration image in real time, using a circle center positioning method based on deep learning dichotomy to obtain the two-dimensional pixel coordinates of the center of all marked points in the vibration image, and according to the mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates, converting all the two-dimensional pixel coordinates of the center of the circle into three-dimensional world coordinates ( XE k , YE k , 0), where k =1, 2, …, n , k Indicates the sequence number of the marking point. n Indicates the number of marking points on the blade. XE k Indicates k The three-dimensional horizontal coordinate of a marked point at a certain moment in the vibration process, YE kIndicates k The three-dimensional vertical coordinates of a marking point at a certain moment in the vibration process; the blade vibration amount at each marking point position is calculated =| YE k - YS k |, complete the calculation of the vibration amount at each position of the blade.

[0013] The beneficial effects of the present application are as follows: the present invention adopts a circle center positioning method based on deep learning dichotomy, which can locate the center of the marked point in a complex lighting environment with a positioning accuracy of 99%; the present invention adopts monocular vision technology, and realizes contactless measurement by acquiring the mapping relationship between the two-dimensional pixel coordinates on the camera image and the three-dimensional world coordinates on the blade vibration plane, while controlling the measurement error within 1mm; when the blade vibration frequency is about 6Hz, real-time monitoring of the blade vibration amount can be achieved; the present invention combines target detection and target segmentation technology based on deep learning with computer vision technology, and has the advantages of non-contact, strong adaptability and high precision in the case of real-time monitoring of vibration amount, and can be used for monitoring the blade vibration amount in fatigue tests, which has important practical value for the design and development of helicopter blades. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0015] Figure 1 is a flow chart of a method according to an embodiment of the present invention; Figure 2 Schematic diagram of the structure of a real-time acquisition device for blade vibration images in an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of a lightweight convolution module in an embodiment of the present invention; Figure 4 is a structural diagram of an ellipse segmentation network in an embodiment of the present invention; Figure 5 is a schematic diagram of blade vibration in an embodiment of the present invention; Figure 6 is a marker point segmentation result diagram of an embodiment of the present invention, wherein (a), (c) and (e) are marker images under different background illumination environments, and (b), (d) and (f) are segmentation result diagrams under different background illumination environments; Figure 7 Schematic diagram of the structure of the vibration detection and verification device in an embodiment of the present invention.

[0016] Explanation of reference numerals: 101 - processing computer, 102 - industrial camera, 103 - blade, 104 - blade marking point. DETAILED DESCRIPTION

[0017] In order to more clearly understand the above-mentioned purpose, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention can also be implemented in other ways different from those described herein, and therefore, the present invention is not limited to the limitations of the specific embodiments disclosed below.

[0018] like Figure 1 As shown, the embodiment of the present invention proposes a helicopter blade vibration monitoring method based on deep learning dichotomy, comprising the following steps: Step S1: Establishing a mapping relationship between two-dimensional pixel coordinates and three-dimensional world coordinates. Figure 2 The blade vibration image real-time acquisition device shown in the figure has a CPU model of Intel(R) Core (TM) i7-8750H, a GPU model of GTX 1070, a video memory of 8G, a memory of 24G, an operating system of WIN10, and a disk capacity of a solid state drive of 256GB + a mechanical hard disk of 1T. The industrial camera 102 is a German IDS digital industrial camera, and the lens is a Japanese RICOH lens with a focal length of 12mm. The circular marking points in the calibration plate are 8 rows and 7 columns, and the center distance is 35mm. The size of the collected image is fixed to 1280×1024. The calibration plate is kept on the blade vibration plane and within the camera field of view. The mapping relationship from the two-dimensional pixel coordinates of the imaging plane to the three-dimensional world coordinates of the vibration plane is calculated, that is, the rotation matrix and the translation matrix. The specific method is: Place the calibration plate on the vibration plane of the blade 103, and keep the calibration plate within the shooting field of view of the industrial camera 102. Input the calibration parameters such as the center distance, number of circle rows, number of circle columns, etc. into the processing computer 101, and use the camera to collect 10 calibration plate images. The 10 calibration plate images obtained by shooting are calibrated by Zhang's calibration method to obtain the corresponding reprojection error, rotation matrix and translation matrix. Select the image with the smallest calibration reprojection error, and take its corresponding rotation matrix and translation matrix as the rotation matrix and translation matrix of the camera's two-dimensional pixel coordinate system to the vibration plane's three-dimensional world coordinate system. The mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates is: ; in,( u , v) is the two-dimensional pixel coordinate of the blade marking point 104, z c is the distance from the marking point to the origin along the optical axis in the camera coordinate system, K is the camera’s intrinsic parameter matrix, R is the rotation matrix from the camera's 2D pixel coordinate system to the vibration plane's 3D world coordinate system, t is the translation matrix from the camera's 2D pixel coordinate system to the vibration plane's 3D world coordinate system, ( x w , y w , z w ) is the three-dimensional world coordinate of the marking point in the three-dimensional world coordinate system of the vibration plane. Zhang’s calibration method stipulates that the calibration plate plane z w = 0, z w = 0 into the formula and eliminate z c , solve to get the three-dimensional world coordinates of the marker point ( x w , y w , 0), completing the mapping from two-dimensional pixel coordinates to three-dimensional world coordinates.

[0019] Step S2: Collect real-time blade vibration images. In the present example, a plastic sheet is cut into the shape of a helicopter blade as a simulated blade, and 5 circular marking points are pasted on the simulated blade, i.e., the number of marking points is n =5, the diameter of the marking point is 3.5 cm, and the color is silver-white. The trigger mode of the industrial camera 102 is set to internal trigger, the camera acquisition frame rate is set to 50 frames per second, and the camera imaging plane is at a 45-degree angle to the blade vibration plane. During the blade vibration process, the industrial camera 102 continuously captures the vibration image.

[0020] Step S3: construct an improved YOLOv8n network, use lightweight convolution modules to replace the traditional convolution modules in all Bottleneck modules in the YOLOv8n network, and delete the large target detection head of the YOLOv8n network; use deep learning dichotomy to complete the circular detection of the blade marker points, specifically including using the improved YOLOv8n network to locate the blade marker points and using the ellipse segmentation network to segment the blade marker points, and obtain the pixel coordinates of the center of the blade marker points.

[0021] The specific steps of step S3 are: Step S301: Use the LabelImg annotation tool to annotate the marking points on the real-time blade vibration image collected in step S2 to obtain the corresponding label file, and form a marking point detection data set with the blade vibration image and the corresponding label file. Figure 2 The blade vibration image real-time acquisition device shown is placed in different lighting environments to construct a blade vibration test simulation environment, collects blade vibration images under different lighting environments and obtains corresponding label files, and puts the blade vibration images and the corresponding label files into different folders under the same directory to form a marker point detection data set, which is divided into a training set and a test set in a 4:1 ratio.

[0022] Step S302: Train the improved YOLOv8n marker detection network, input the marker detection data set into the trained improved YOLOv8n marker detection network for recognition, and output the location information of the markers. Figure 3 As shown in the figure, b in the figure represents the number of images input to the network at one time, h represents the image height, w represents the image width, and c represents the number of image channels. The lightweight convolution module includes a channel splitting module, a two-dimensional convolution module, and a channel splicing module; the input feature map is first split into feature maps according to the proportion by the channel splitting module. Figure 1 and Features Figure 2 ,feature Figure 1 After feature extraction by two-dimensional convolution module and feature Figure 2 Channel concatenation is performed to obtain an output feature map. In the embodiment of the present invention, the input feature map is split into feature maps in a ratio of 1:3. Figure 1 and Features Figure 2 .

[0023] Step S303: crop a blade marker point image from the blade vibration image, adjust the image size to 120×120 pixels, and perform binary segmentation on the cropped blade marker point image to obtain a marker point binary image of the same size as the cropped blade marker point image, and put the blade marker point image and the corresponding marker point binary image into different folders in the same directory to form a marker point segmentation data set, and randomly divide it into a training set and a test set in a ratio of 4:1.

[0024] Step S304: construct an ellipse segmentation network, the structure of which is as follows: Figure 4As shown, it includes three parts: an encoding network D1, a decoding network D2 and an output network D3; the encoding network D1 includes an input layer and an encoding layer, and the encoding layer includes three double-layer convolutional layers and three pooling layers arranged alternately; the decoding network D2 includes three upsampling layers and three double-layer convolutional layers arranged alternately, and each upsampling layer is jump-connected with the double-layer convolutional layer in the encoding network D1; the output network D3 includes a point-by-point convolutional layer; wherein the double-layer convolutional layer includes an input layer, two convolutional layers arranged alternately, two batch normalization layers, two activation function layers and an output layer, in the figure indicates the number of input tensor channels, mid indicates the number of channels in the tensor processing process, and out indicates the number of output tensor channels. In the embodiment of the present invention, the scaling factor of the upsampling layer is set to 2, and the upsampling algorithm selects bilinear interpolation.

[0025] Step S305: The ellipse segmentation network is trained. During the training process, the number of training rounds is set to 150, the learning rate is 1E-4, and the training batch is set to 20. The binary cross entropy loss is used as the loss function to calculate the segmentation error of the marker point, and the error is back-propagated to update the network parameters to achieve network convergence. The network is continuously trained until the error between the predicted marker point center coordinates and the true center coordinates is within 0.5 pixels, and according to the calculated segmentation error, a set of model parameters with the minimum loss value or the minimum average center error during the training process is saved.

[0026] The calculation formula of the loss function is as follows: ; in, YP Represents the predicted value output by the ellipse segmentation network; YT Represents the true value corresponding to the input image; Loss ( YT,YP ) represents the loss function, that is, the segmentation error between the network prediction value and the true value, yp ij Represents the predicted image i OK j Pixel values ​​on columns; yt ij Represents the real image i OK j The pixel value on the column.

[0027] Step S306: Load the ellipse segmentation network model and set the network's hyperparameters, and crop all circular marker point images from the blade vibration image according to the position information of the marker points obtained in step S302, and scale them to the same size as the cropped blade marker point image, i.e., 120×120 pixels; then input them into the ellipse segmentation network for processing, the ellipse segmentation network outputs a binary image of each marker point and performs ellipse fitting on the binary image to obtain the center coordinates of all marker points; according to the scaling relationship when the marker points are cropped, the center coordinates are transformed into the original image to obtain the two-dimensional pixel coordinates of the center of each circular marker point in the original image.

[0028] Step S4: Collect the initial position of the blade marking point. The specific steps are: collect the blade image in a static state as the initial state, use the circle center positioning method based on deep learning dichotomy to obtain the two-dimensional pixel coordinates of the center of all marking points in the initial state, and convert all the two-dimensional pixel coordinates of the center of the circle into three-dimensional world coordinates according to the mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates ( XS k , YS k , 0) and save, where, k =1, 2, …, n , n Indicates the number of marking points on the blade. XS k Indicates k The initial three-dimensional horizontal coordinates of the marker points, YS k Indicates k The initial 3D vertical coordinates of the marker points.

[0029] Step S5: Real-time calculation of blade vibration, the specific steps are as follows: Figure 5 As shown, an industrial camera 102 is used to capture blade vibration images in real time, and a circle center positioning method based on deep learning dichotomy is used to obtain the two-dimensional pixel coordinates of the centers of all marked points in the vibration image. According to the mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates, all the two-dimensional pixel coordinates of the centers of the circles are converted into three-dimensional world coordinates ( XE k , YE k , 0), where k =1, 2, …, n , k Indicates the sequence number of the marking point. n Indicates the number of marking points on the blade. XE k Indicates k The three-dimensional horizontal coordinate of a marked point at a certain moment in the vibration process, YE kIndicates k The three-dimensional ordinate of a marking point at a certain moment in the vibration process; the hollow circle in the figure represents the initial position of the blade marking point, and the solid circle represents the position of the blade marking point at a certain moment. The blade vibration amount at each marking point position is obtained by calculation =| YE k - YS k |, complete the calculation of the vibration amount at each position of the blade.

[0030] The performance indicators of the monitoring method described in the embodiment of the present invention are explained below through experiments.

[0031] 1. Robustness test Place the simulated paddle in Figure 6 The experiment was conducted under the test images with different brightness and complex backgrounds shown in the figure. (a), (c) and (e) in the figure are the marker point images with different brightness, and (b), (d) and (f) are the segmentation result images of the corresponding marker points. Through the experiment, it can be seen that under different brightness and complex background conditions, the embodiment of the present invention can well segment the marker points and fit the center of the marker points, and has strong robustness.

[0032] 2. Measurement accuracy test The training effect of the ellipse segmentation network in the example of the present invention is shown in Table 1.

[0033] Table 1 Ellipse segmentation network training index data

[0034] During the network training process, 5 sets of data were taken on average from 0 to 150 training times. The higher the pixel accuracy and the smaller the pixel error, the more accurate the marker point segmentation network is for the marker point and the more accurate the circle center detection is. From the data in Table 1, it can be seen that on the validation set, the circle center pixel error is maintained within 0.5 pixels, indicating that the marker point segmentation model has a high accuracy rate.

[0035] Use Figure 7 The device with known circle center distance marking points attached thereto is used to verify the measurement accuracy of the embodiment of the present invention, and the test results shown in Table 2 can be obtained.

[0036] Table 2 Verification of measurement accuracy of vibration monitoring method

[0037] It can be seen from the data in Table 2 that the embodiment of the present invention can achieve high-precision detection of the center of the circular marking point, and the average error of the vibration amount can be controlled within 1 mm, which can fully meet the monitoring requirements of the blade vibration amount.

[0038] 3. Center positioning time test.

[0039] Considering that the vibration frequency of the helicopter blades during fatigue testing is about 6 Hz, in order to achieve real-time vibration monitoring, the acquisition, processing, and result visualization of each vibration image and other operation times need to be controlled within 20 ms, and the processing frame rate is controlled above 50 frames. In the embodiment of the present invention, the number of blade marking points is 5, and the experimental center positioning time data is shown in Table 3.

[0040] Table 3 Center positioning time

[0041] Among them, the GTX 1070 GPU calls the improved YOLOv8n marker detection network and ellipse segmentation network, which takes 15ms from input image to output of 5 circle center coordinates, while the RTX 4060Ti GPU takes 6ms in total, which can fully meet the real-time monitoring requirements of blade vibration.

[0042] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A helicopter blade vibration monitoring method based on deep learning dichotomy, characterized in that: The steps include: Step S1: Establishing a mapping relationship from two-dimensional pixel coordinates to three-dimensional world coordinates; Step S2: collecting real-time blade vibration images; Step S3: construct an improved YOLOv8n network, use lightweight convolution modules to replace traditional convolution modules in all Bottleneck modules in the YOLOv8n network, and delete the large target detection head of the YOLOv8n network; use deep learning dichotomy to complete the circular detection of the blade markers, specifically including using the improved YOLOv8n network to locate the blade markers and using the ellipse segmentation network to segment the blade markers, and obtain the pixel coordinates of the center of the blade markers; Step S4: collecting the initial position of the blade marking point; Step S5: Calculate the blade vibration in real time.

2. A helicopter blade vibration monitoring method based on deep learning dichotomy according to claim 1, characterized in that: The mapping relationship from the two-dimensional pixel coordinates to the three-dimensional world coordinates in step S1 is: ； Where (u, v) is the 2D pixel coordinate of the marking point on the blade, z c is the distance from the marking point to the origin along the optical axis in the camera coordinate system, K is the camera's intrinsic parameter matrix, R is the rotation matrix from the camera's 2D pixel coordinate system to the vibration plane's 3D world coordinate system, t is the translation matrix from the camera's 2D pixel coordinate system to the vibration plane's 3D world coordinate system, (x w ,y w , z w ) is the three-dimensional world coordinate of the marking point in the three-dimensional world coordinate system of the vibration plane.

3. A helicopter blade vibration monitoring method based on deep learning dichotomy according to claim 2, characterized in that: The lightweight convolution module includes a channel splitting module, a two-dimensional convolution module and a channel splicing module; the input feature map is first split into feature map 1 and feature map 2 according to the proportion by the channel splitting module, and the feature map 1 is subjected to feature extraction by the two-dimensional convolution module and then channel spliced ​​with the feature map 2 to obtain the output feature map.

4. A helicopter blade vibration monitoring method based on deep learning dichotomy according to claim 3, characterized in that: The specific steps of step S3 are: Step S301: using the LabelImg annotation tool to annotate the marking points on the real-time blade vibration image collected in step S2, obtaining a corresponding label file, and combining the blade vibration image and the corresponding label file into a marking point detection data set; Step S302: training the improved YOLOv8n marker detection network, inputting the marker detection data set into the trained improved YOLOv8n marker detection network for recognition, and outputting the position information of the markers; Step S303: cropping a blade marker point image from the blade vibration image, and performing binary segmentation on the cropped marker point image to obtain a marker point binary image of the same size as the cropped blade marker point image, and forming a marker point segmentation data set with the blade marker point image and the corresponding marker point binary image; Step S304: constructing an ellipse segmentation network, the ellipse segmentation network includes three parts: an encoding network D1, a decoding network D2 and an output network D3; the encoding network D1 includes an input layer and an encoding layer, the encoding layer includes three double-layer convolutional layers and three pooling layers that are alternately arranged; the decoding network D2 includes three upsampling layers and three double-layer convolutional layers that are alternately arranged, and each upsampling layer is jump-connected to the double-layer convolutional layer in the encoding network D1; the output network D3 includes a point-by-point convolutional layer; wherein the double-layer convolutional layer includes an input layer, two convolutional layers that are alternately arranged, two batch normalization layers, two activation function layers and an output layer; Step S305: training the ellipse segmentation network, using binary cross entropy loss as a loss function to calculate the segmentation error of the marker points during the training process, back-propagating the error, updating the network parameters, achieving network convergence, and continuously training the network until the error between the predicted marker point center coordinates and the true center coordinates is within 0.5 pixels, and based on the calculated segmentation error, saving a set of model parameters with the minimum loss value or the minimum average center error during the training process; Step S306: Load the ellipse segmentation network model and set the network's hyperparameters, and crop all circular marker point images from the blade vibration image according to the position information of the marker points obtained in step S302, and scale them to the same size as the cropped blade marker point images; then input them into the ellipse segmentation network for processing, the ellipse segmentation network outputs a binary image of each marker point and performs ellipse fitting on the binary image to obtain the center coordinates of all marker points; according to the scaling relationship when the marker points are cropped, the center coordinates are transformed into the original image to obtain the two-dimensional pixel coordinates of the center of each circular marker point in the original image.

5. A helicopter blade vibration monitoring method based on deep learning dichotomy according to claim 4, characterized in that: The calculation formula of the loss function is as follows: ； Among them, YP represents the predicted value output by the ellipse segmentation network; YT represents the true value corresponding to the input image; Loss(YT,YP) represents the loss function, that is, the segmentation error between the network prediction value and the true value, yp ij Represents the pixel value on the i-th row and j-th column of the predicted image; yt ij Represents the pixel value on the i-th row and j-th column of the real image.

6. A helicopter blade vibration monitoring method based on deep learning dichotomy according to claim 5, characterized in that: The specific steps of step S4 are: collecting the blade image in a static state as the initial state, using the circle center positioning method based on deep learning dichotomy to obtain the two-dimensional pixel coordinates of the center of all the marking points in the initial state, and according to the mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates, converting each blade marking point from the two-dimensional pixel coordinates of the circle center to the three-dimensional world coordinates (XS k , YS k , 0) and save, where, k =1, 2, …, n , n Indicates the number of marking points on the blade, XS k YS represents the initial three-dimensional horizontal coordinate of the kth marker point. k Represents the initial three-dimensional vertical coordinate of the kth marker point.

7. A helicopter blade vibration monitoring method based on deep learning dichotomy according to claim 6, characterized in that: The specific steps of step S5 are: using an industrial camera to take a real-time image of the blade vibration, using a circle center positioning method based on deep learning dichotomy to obtain the two-dimensional pixel coordinates of the center of all the marked points in the vibration image, and according to the mapping relationship between the two-dimensional pixel coordinates and the three-dimensional world coordinates, converting all the two-dimensional pixel coordinates of the center into three-dimensional world coordinates (XE k , YE k , 0), where k =1, 2, …, n , k Indicates the sequence number of the marking point. n Indicates the number of marking points on the blade, XE k YE represents the three-dimensional horizontal coordinate of the kth mark point at a certain moment in the vibration process. k Indicates the three-dimensional ordinate of the kth mark point at a certain moment in the vibration process; calculates the blade vibration amount at each mark point =| YE k -YS k |, complete the calculation of the vibration amount at each position of the blade.

Images