A rotor tip vortex core image recognition method using a U-net neural network

Abstract

The application discloses a kind of to utilize U-net neural network and carry out rotor blade tip vortex vortex core image identification method, belong to vortex core image processing technical field, including the following steps: S1: constructs training U-net neural network model;S2: U-net neural network model predicts blade tip vortex position;S3: obtain vortex core position and area;S4: vortex core position ellipse curve fitting;S5: equal division curve elliptic arc segment and take index;S6: calculate vortex core position and elliptic arc segment distance and belong to nearest elliptic arc segment;S7: the number of vortex core is about arc segment index curve;S8: curve smoothing processing;S9: find peak position;S10: peak position index mapping vortex cluster label, calculate peak index corresponding peak arc segment position;S11: calculate vortex core position and curve peak arc segment distance and belong to nearest peak arc segment, and each vortex is marked with belonging vortex cluster label;The application can accurately identify blade tip vortex, simultaneously without treating PIV image, simple and intuitive operation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of vortex core image processing, and particularly relates to a method for identifying a rotor blade tip vortex core image by using a U-net neural network. BACKGROUND

[0002] The rotor blade tip vortex refers to a spiral trace rolled up behind the blade tip in the process of rotating the rotor of a helicopter, and is one of main flow field structures of a rotor wake, and has an important influence on the flight performance and acoustic performance of the helicopter.

[0003] At present, a method for obtaining the position of the rotor blade tip vortex to study the flight performance and acoustic performance of the helicopter is generally to perform flow visualization experiments and PIV experiments, wherein the flow visualization experiments can intuitively display the flow results of the rotor flow field, but a large amount of tracer smoke causes computer vision to be difficult to accurately identify the position of the rotor blade tip vortex in the photographed picture, and the PIV experiment needs to obtain flow field data by performing processing operations including noise filtering, and then obtain the position of the blade tip vortex by Q criterion and position correction, which is complex and not intuitive.

[0004] In view of this, a method for identifying a rotor blade tip vortex core image by using a U-net neural network is designed to solve the above problems. SUMMARY

[0005] To solve the problems in the background art, the application provides a method for identifying a rotor blade tip vortex core image by using a U-net neural network, which has the characteristics of being able to accurately identify the blade tip vortex and being simple and intuitive without needing to process the PIV image.

[0006] To achieve the above object, the application provides the following technical scheme: a method for identifying a rotor blade tip vortex core image by using a U-net neural network, comprising the following steps:

[0007] S1: constructing a U-net neural network model for training;

[0008] S2: predicting all PIV original images contained in a certain working condition based on the U-net neural network model for training, and outputting a group of output images in png format containing only 0 and 1 elements, wherein 0 represents that the pixel belongs to the image background, and 1 represents that the pixel belongs to the blade tip vortex;

[0009] S3: obtaining the position and area of the vortex core corresponding to each 1-pixel block in the group of output images;

[0010] S4: fitting an elliptic curve to each vortex core position in the group of output images;

[0011] S5: Take the elliptical arc segment between x = 0 and x = 1.1h, divide the elliptical arc segment into 1000 segments from top to bottom according to the arc length, and take the index 1-1000 in sequence, where h represents the rotor height;

[0012] S6: Calculate the distance between each vortex kernel position and the elliptical arc segment in the output image set, and assign the vortex kernel position to the elliptical arc segment that is closest to it;

[0013] S7: Count the number of vortex cores paired for each arc segment and obtain a curve of the number of vortex cores with respect to the arc segment index;

[0014] S8: The curve is smoothed using the moving average method, with a smoothing window size of 30.

[0015] S9: Find the peak position of the smoothed curve using a peak-finding function, with a height threshold of 1.7, a peak distance threshold of 40, and a peak significance threshold of 0.5.

[0016] S10: Map the N peak position indices to vortex cluster labels of [1,2,...,N], and calculate the peak arc position corresponding to the peak index, where the index is 30 index positions before and after a certain peak index;

[0017] S11: Calculate the distance between each vortex core position and the peak arc segment of the curve in the output image of this group, assign the vortex core position to the nearest peak arc segment, and label each vortex with its respective vortex cluster label.

[0018] Furthermore, the specific steps of step S1 include:

[0019] S101: Randomly select 500 images from the original PIV image dataset, and annotate the images to form PNG format annotated images. Use the 500 original PIV images and the corresponding annotated images as the training dataset.

[0020] S102: Adjust the image dimensions in the training dataset to 2304 pixels in length and 1536 pixels in width using nearest neighbor interpolation.

[0021] S103: Construct the U-net neural network model, including convolution, downsampling, and upsampling. The input image size for convolution is set to 2304 pixels in length and 1536 pixels in width. Convolution, downsampling, and upsampling are all performed twice consecutively. The convolution kernel size is 3×3, with 1 pixel of mirrored padding outside the four edges of the image. The convolution stride is 1. Batch normalization is performed after convolution. Simultaneously, 30% of neurons are randomly deactivated, and the Leakey ReLU function is used for non-linear activation. During each downsampling process... The size is halved, the kernel size is 3×3, the padding is 1 pixel mirrored outside the four edges of the image, the stride is 2, the feature map channels are reduced by half during each upsampling process, the kernel size is 1×1, no padding, the stride is 1, the feature map size is doubled by nearest neighbor bilinear interpolation, after each upsampling, it is concatenated with the downsampled feature map of the same layer, after upsampling, the convolution adjusts the predicted image channels to two channels, the kernel size is 3×3, the padding is 1 pixel mirrored outside the four edges of the image, the stride is 1;

[0022] S104: A U-net neural network model built based on images with a length of 2304 pixels and a width of 1536 pixels in the training dataset. After predicting the predicted value during the forward propagation process, the loss function value between the output image and the labeled image is calculated through the cross-entropy loss function. During the back propagation process, the partial derivative of the loss function value with respect to each parameter is calculated layer by layer from the output through the chain rule. The parameters are optimized by the Adam optimizer based on the partial derivatives.

[0023] S105: Repeat step S104 until the U-net neural network model reaches its optimal state after 120 rounds of optimization. Save the generated weight file. The U-net neural network model training is now complete.

[0024] Furthermore, the specific steps of step S2 include:

[0025] S201: Read each image from the original PIV image dataset and resize the images to 2304 pixels long and 1536 pixels wide using nearest neighbor interpolation.

[0026] S202: The U-net neural network model predicts images in the original PIV image dataset based on the weight file obtained from training, and obtains dual-channel predicted images.

[0027] S203: Compare the probability values ​​between the two-channel predicted images, and take the index corresponding to the larger predicted value as the pixel value of the output image. At this time, the output image containing only 0 and 1 elements is obtained, where 0 represents that the pixel belongs to the image background and 1 represents that the pixel belongs to the paddle tip vortex. Save the output image as a PNG image.

[0028] Furthermore, in step S9, during the process of finding the peak position of the smoothed curve, if the product of the equivalent width of a peak and the peak height is greater than 1.7 times the average value of all peaks, it is considered that a tip vortex combination has occurred at the peak position. If a tip vortex combination phenomenon occurs at a peak position, two new peaks are defined.

[0029] Compared with the prior art, the beneficial effects of the present invention are:

[0030] This invention first identifies the location of tip vortices based on a trained U-net neural network model, and then classifies the tip vortex locations based on line-following clustering. Compared with existing technologies, it can accurately identify tip vortices without processing PIV images, and the operation is simple and intuitive. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the U-net neural network model of the present invention;

[0032] Figure 2 This is a schematic diagram of the method of the present invention;

[0033] Figure 3 This is a flowchart of the method of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] This invention provides the following technical solution: a method for rotor tip vortex core image recognition using a U-net neural network, comprising the following steps:

[0036] S1: Construct and train the U-net neural network model;

[0037] S2: Based on the U-net neural network model trained by the construction, predict all the original PIV images contained in a certain working condition, and output a set of png format output images containing only 0 and 1 elements, where 0 represents that the pixel belongs to the image background and 1 represents that the pixel belongs to the tip vortex.

[0038] S3: Obtain the centroid position and area of ​​each 1-pixel block in the output image of this group, which are the vortex core position and area;

[0039] S4: Fit an elliptic curve to each vortex core position of the output images in this group;

[0040] S5: Take the elliptical arc segment between x = 0 and x = 1.1h, divide the elliptical arc segment into 1000 segments from top to bottom according to the arc length, and take the index 1-1000 in sequence, where h represents the rotor height;

[0041] S6: Calculate the distance between each vortex kernel position and the elliptical arc segment in the output image set, and assign the vortex kernel position to the elliptical arc segment that is closest to it;

[0042] S7: Count the number of vortex cores paired for each arc segment and obtain a curve of the number of vortex cores with respect to the arc segment index;

[0043] S8: The curve is smoothed using the moving average method, with a smoothing window size of 30.

[0044] S9: Find the peak position of the smoothed curve using a peak-finding function, with a height threshold of 1.7, a peak distance threshold of 40, and a peak significance threshold of 0.5.

[0045] S10: Map the N peak position indices to vortex cluster labels of [1,2,...,N], and calculate the peak arc position corresponding to the peak index, where the index is 30 index positions before and after a certain peak index;

[0046] S11: Calculate the distance between each vortex core position and the peak arc segment of the curve in the output image of this group, assign the vortex core position to the nearest peak arc segment, and label each vortex with its respective vortex cluster label.

[0047] Specifically, step S1 includes the following steps:

[0048] S101: Randomly select 500 images from the original PIV image dataset, and annotate the images to form PNG format annotated images. Use the 500 original PIV images and the corresponding annotated images as the training dataset.

[0049] PNG format refers to a bitmap image format that uses a lossless compression algorithm. The name is an abbreviation for "Portable Network Graphics". It has the characteristic that no image data is lost during compression and decompression. PNG format images can occupy less storage space than uncompressed image formats while maintaining high quality.

[0050] S102: Adjust the image dimensions in the training dataset to 2304 pixels in length and 1536 pixels in width using nearest neighbor interpolation.

[0051] Nearest neighbor interpolation is a basic and simple image scaling algorithm, also known as nearest neighbor interpolation. The principle is to find the source image pixel coordinates that correspond to the target image pixel coordinates, and then use rounding or discarding decimal places to convert this floating-point coordinates into integer coordinates. Finally, the source image pixel value corresponding to the integer coordinates is taken as the value of the target image pixel.

[0052] S103: Construct the U-net neural network model, including convolution, downsampling, and upsampling. The input image size for convolution is set to 2304 pixels in length and 1536 pixels in width. Convolution, downsampling, and upsampling are all performed twice consecutively. The convolution kernel size is 3×3, with 1 pixel of mirrored padding outside the four edges of the image. The convolution stride is 1. Batch normalization is performed after convolution. Simultaneously, 30% of neurons are randomly deactivated, and the Leakey ReLU function is used for non-linear activation. During each downsampling process... The size is halved, the kernel size is 3×3, the padding is 1 pixel mirrored outside the four edges of the image, the stride is 2, the feature map channels are reduced by half during each upsampling process, the kernel size is 1×1, no padding, the stride is 1, the feature map size is doubled by nearest neighbor bilinear interpolation, after each upsampling, it is concatenated with the downsampled feature map of the same layer, after upsampling, the convolution adjusts the predicted image channels to two channels, the kernel size is 3×3, the padding is 1 pixel mirrored outside the four edges of the image, the stride is 1;

[0053] S104: A U-net neural network model built based on images with a length of 2304 pixels and a width of 1536 pixels in the training dataset. After predicting the predicted value during the forward propagation process, the loss function value between the output image and the labeled image is calculated through the cross-entropy loss function. During the back propagation process, the partial derivative of the loss function value with respect to each parameter is calculated layer by layer from the output through the chain rule. The parameters are optimized by the Adam optimizer based on the partial derivatives.

[0054] The cross-entropy loss function measures the difference between two probability distributions and is often used to evaluate the difference between the model's predicted distribution and the true distribution.

[0055] The chain rule is a fundamental theorem in calculus that allows the calculation of the derivative of a composite function. In model training, it can effectively calculate the gradient of each parameter with respect to the loss function. The gradient is necessary for optimizing the model and updating parameters.

[0056] S105: Repeat step S104 until the U-net neural network model reaches its optimal state after 120 rounds of optimization. Save the generated weight file. The U-net neural network model training is now complete.

[0057] Specifically, step S2 includes the following steps:

[0058] S201: Read each image from the original PIV image dataset and resize the images to 2304 pixels long and 1536 pixels wide using nearest neighbor interpolation.

[0059] S202: The U-net neural network model predicts images in the original PIV image dataset based on the weight file obtained from training, and obtains dual-channel predicted images.

[0060] S203: Compare the probability values ​​between the two-channel predicted images, and take the index corresponding to the larger predicted value as the pixel value of the output image. At this time, the output image containing only 0 and 1 elements is obtained, where 0 represents that the pixel belongs to the image background and 1 represents that the pixel belongs to the paddle tip vortex. Save the output image as a PNG image.

[0061] Specifically, in step S9, during the process of finding the peak position of the smoothed curve, if the product of the equivalent width of a peak and the peak height is greater than 1.7 times the average value of all peaks, it is considered that a tip vortex combination has occurred at the peak position. If a tip vortex combination phenomenon occurs at a peak position, two new peaks are defined.

[0062] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for rotor tip vortex core image recognition using a U-net neural network, characterized in that, Includes the following steps: S1: Construct and train the U-net neural network model; S2: Based on the U-net neural network model trained by the construction, predict all the original PIV images contained in a certain working condition, and output a set of png format output images containing only 0 and 1 elements, where 0 represents that the element belongs to the image background and 1 represents that the element belongs to the tip vortex. S3: Obtain the centroid position and area of ​​each 1-element block in the output image of this group, which are the vortex core position and area; S4: Fit an elliptic curve to each vortex core position of the output images in this group; S5: Take the elliptical arc segment between x = 0 and x = 1.1h, divide the elliptical arc segment into 1000 segments from top to bottom according to the arc length, and take the index 1-1000 in sequence, where h represents the rotor height; S6: Calculate the distance between each vortex kernel position and the elliptical arc segment in the output image set, and assign the vortex kernel position to the elliptical arc segment that is closest to it; S7: Count the number of vortex cores paired for each arc segment and obtain a curve of the number of vortex cores with respect to the arc segment index; S8: The curve is smoothed using the moving average method, with a smoothing window size of 30. S9: Find the peak position of the smoothed curve using a peak-finding function, with a height threshold of 1.7, a peak distance threshold of 40, and a peak significance threshold of 0.

5. S10: Map the N peak position indices to vortex cluster labels of [1,2,...,N], and calculate the peak arc position corresponding to the peak index, where the index is 30 index positions before and after a certain peak index; S11: Calculate the distance between each vortex core position and the peak arc segment of the curve in the output image of this group, assign the vortex core position to the nearest peak arc segment, and label each vortex with its respective vortex cluster label.

2. The method for rotor tip vortex core image recognition using U-net neural network according to claim 1, characterized in that: The specific steps of step S1 include: S101: Randomly select 500 images from the original PIV image dataset, and annotate the images to form PNG format annotated images. Use the 500 original PIV images and the corresponding annotated images as the training dataset. S102: Adjust the image dimensions in the training dataset to 2304 pixels in length and 1536 pixels in width using nearest neighbor interpolation. S103: Construct the U-net neural network model, including convolution, downsampling, and upsampling. The input image size for convolution is set to 2304 pixels in length and 1536 pixels in width. Convolution, downsampling, and upsampling are all performed twice consecutively. The convolution kernel size is 3×3, with 1 pixel of mirrored padding outside the four edges of the image. The convolution stride is 1. Batch normalization is performed after convolution. Simultaneously, 30% of neurons are randomly deactivated, and the Leakey ReLU function is used for non-linear activation. During each downsampling process... The size is halved, the kernel size is 3×3, the padding is 1 pixel mirrored outside the four edges of the image, the stride is 2, the feature map channels are reduced by half during each upsampling process, the kernel size is 1×1, no padding, the stride is 1, the feature map size is doubled by nearest neighbor bilinear interpolation, after each upsampling, it is concatenated with the downsampled feature map of the same layer, after upsampling, the convolution adjusts the predicted image channels to two channels, the kernel size is 3×3, the padding is 1 pixel mirrored outside the four edges of the image, the stride is 1; S104: A U-net neural network model built based on images with a length of 2304 pixels and a width of 1536 pixels in the training dataset. After predicting the predicted value during the forward propagation process, the loss function value between the output image and the labeled image is calculated through the cross-entropy loss function. During the back propagation process, the partial derivative of the loss function value with respect to each parameter is calculated layer by layer from the output through the chain rule. The parameters are optimized by the Adam optimizer based on the partial derivatives. S105: Repeat step S104 until the U-net neural network model reaches its optimal state after 120 rounds of optimization. Save the generated weight file. The U-net neural network model training is now complete.

3. The method for rotor tip vortex core image recognition using U-net neural network according to claim 1, characterized in that: The specific steps of step S2 include: S201: Read each image from the original PIV image dataset and resize the images to 2304 pixels long and 1536 pixels wide using nearest neighbor interpolation. S202: The U-net neural network model predicts images in the original PIV image dataset based on the weight file obtained from training, and obtains dual-channel predicted images. S203: Compare the probability values ​​between the two-channel predicted images, and take the index corresponding to the larger predicted value as the pixel value of the output image. At this time, the output image containing only 0 and 1 elements is obtained, where 0 represents that the element belongs to the image background and 1 represents that the element belongs to the tip vortex. Save the output image as a PNG image.

4. The method for rotor tip vortex core image recognition using U-net neural network according to claim 1, characterized in that: In step S9, during the process of finding the peak position of the smoothed curve, if the product of the equivalent width and peak height of a certain peak is greater than 1.7 times the average value of the product of the equivalent width and peak height of all peaks, it is considered that a blade tip vortex combination has occurred at the peak position. If a blade tip vortex combination phenomenon occurs at a certain peak position, two new peaks are defined.

Images