An intelligent method for determining the degree of damage and replacement requirements of aircraft brake discs

Abstract

The present invention relates to the technical field of mechanical equipment damage detection, and discloses an intelligent determination method for the degree of damage and replacement requirements of aircraft brake disc pairs. First, the aircraft brake disc moving plate image data is collected and preprocessed, and then the rust, fall-off, disc, and steel plate areas therein are annotated. Based on the annotated data set, the image segmentation network model is trained to obtain the model weights for detecting the degree of damage, and the predicted result image containing the wear information of the brake disc surface is output; the number of pixels contained in each category in the output image is counted to obtain the proportion of the rust and fall-off areas; the actual maximum radial length of the fall-off area of ​​the brake disc pair is calculated, and finally the proportion of the rust and fall-off areas and the maximum radial length of the fall-off area are comprehensively considered to determine whether the brake disc pair should be eliminated. The present invention realizes automatic, intelligent, efficient, and accurate evaluation of the degree of damage of the brake disc pair, avoiding problems such as inconsistent evaluation standards and manual evaluation errors caused by human factors.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical equipment damage detection, and in particular to an intelligent determination method for the damage degree and replacement requirement of an aircraft brake disc. Background Art

[0002] The brake system is the main device in the aircraft braking process, and its working principle is basically the same as that of automobile brakes. However, due to the huge braking energy of civil aircraft, a multi-disc brake system is mainly used, which mainly includes a piston housing assembly, a brake disc pair and a torque cylinder assembly; the brake assembly presses the alternating movable and static plates by hydraulic or electric means, thereby generating friction to slow down and brake the aircraft. The function of the brake disc pair is to convert the huge kinetic energy of the aircraft into heat energy and dissipate it in time. Usually, carbon / carbon composite materials with lighter weight, better performance, high temperature resistance and long life are selected for manufacturing. Of course, as a loss device, the brake disc pair will inevitably be damaged by wear and corrosion during use. Once the degree of damage exceeds the specified standard, it will directly affect the safety of aircraft operation. Therefore, after a certain number of takeoffs and landings, the brake disc pair needs to be replaced in time.

[0003] Due to the limitations of testing equipment and technology, maintenance engineers still use manual visual methods to detect cracks, wear and rust on the brake disc pair, and judge whether the brake disc pair meets the standard for replacement based on experience. This method has the following problems:

[0004] 1) Low detection efficiency. In order to obtain more accurate judgment results, maintenance engineers need to repeatedly check or even measure the damage degree of the brake disc pair, and then give the results based on relevant judgment standards, which leads to high detection intensity, long process and low efficiency.

[0005] 2) The test results are highly subjective. This method relies too much on subjective experience and judgment, which is affected by factors such as knowledge level, professional skills, and work experience. The same degree of damage to the brake disc pair may produce completely different results after being judged by different engineers.

[0006] 3) It is difficult to unify the inspection standards. If the inspection standards implemented by the maintenance engineers are too strict, the brake discs that do not meet the elimination standards will be frequently replaced, resulting in unnecessary waste of resources and economic losses. If the inspection standards implemented by the maintenance engineers are too loose, the brake discs that should be eliminated cannot be scrapped in time, and the operational safety of the aircraft cannot be guaranteed.

[0007] In short, the existing manual detection methods for brake disc damage, which are inefficient, highly subjective, and difficult to unify, can no longer meet the needs of the rapid development of the aviation industry. There is an urgent need to research more convenient, faster, more accurate, and intelligent detection methods.

[0008] In recent years, the increasingly mature computer vision technology can effectively solve the problem of intelligent identification of the damage degree of brake disc assemblies. By accurately identifying the degree of damage, it can determine whether they need to be eliminated. On the one hand, brake disc assemblies with excessive damage can be eliminated in time to ensure operational safety; on the other hand, brake disc assemblies that do not meet the standards can continue to be used while ensuring safety margins, saving operating costs.

[0009] At present, the research on the damage degree of aircraft landing gear brake discs is not in-depth enough. It can only rely on manual visual methods to identify the damage area and radial width and estimate the damage index to determine whether timely replacement is needed. This detection method is time-consuming and labor-intensive, has low detection efficiency, strong subjectivity, and low detection accuracy. On the one hand, if the brake discs that should be replaced are not identified in time, it will affect operational safety. On the other hand, if the brake discs that should not be eliminated are frequently replaced, it will cause a waste of resources and increase operating costs. Summary of the invention

[0010] In view of the above problems, the purpose of the present invention is to provide an intelligent determination method for the damage degree and replacement requirement of aircraft brake discs, so as to achieve automatic, intelligent, efficient and accurate evaluation of the damage degree of brake discs, and avoid problems such as inconsistent evaluation standards and manual evaluation errors caused by human factors. The technical solution is as follows:

[0011] An intelligent method for determining the degree of damage and replacement requirements of aircraft brake discs comprises the following steps:

[0012] Step 1: Collect image data of the aircraft brake disc auxiliary plate and pre-process the collected image data;

[0013] Step 2: Classify the damage of the aircraft brake disc auxiliary plate, mark the rust, detachment, disc, and steel plate areas, form a brake disc auxiliary damage label file, and obtain the aircraft brake disc auxiliary plate surface damage dataset;

[0014] Step 3: Based on the labeled aircraft brake disc auxiliary plate surface damage dataset, the image segmentation network model is trained through multiple iterations using the transfer training plus frozen training method to obtain the model weights for detecting the surface damage degree of the aircraft brake disc auxiliary plate;

[0015] Step 4: Input the aircraft brake disc auxiliary plate image data into the constructed image segmentation network model, use the trained model weights for forward reasoning, and output a prediction result image with the same size as the input image and containing the brake disc auxiliary plate surface wear information;

[0016] Step 5: Count the number of pixels in each category in the output image to obtain the proportion of the rusted and peeled areas on the secondary surface of the brake disc;

[0017] Step 6: Further process the output image to calculate the pixel width of the annular portion of the brake disc auxiliary plate and the maximum radial pixel length of the detached area, so as to obtain the actual maximum radial length of the detached area of ​​the brake disc auxiliary plate;

[0018] Step 7: Determine whether to eliminate the brake disc pair based on the proportion of the rusted and peeled areas on the brake disc pair's surface and the maximum radial length of the peeled area.

[0019] Furthermore, the step 2 also includes data enhancement of the aircraft brake disc auxiliary plate surface damage dataset, specifically:

[0020] Perform mosaic data enhancement and mixed data enhancement on the labeled image data;

[0021] The mosaic data enhancement is as follows: 4 pictures are randomly read from the aircraft brake disc auxiliary plate surface damage data set each time, and the 4 pictures are flipped, scaled, and color gamut changed respectively. After the operation is completed, the 4 original pictures are arranged in the form of the first picture at the upper left, the second picture at the lower left, the third picture at the lower right, and the fourth picture at the upper right, and then the four pictures are spliced ​​into a new picture, and the content of the annotation box needs to be retained on the new picture;

[0022] The mixed data enhancement is to mix images between different classes to expand the training data set. The formula is as follows:

[0023] (1)

[0024] (2)

[0025] (3)

[0026] in, , is the mixed image and corresponding label, Beta is the Beta distribution, α and β are all constants; is the random sample number calculated from the Beta distribution; ( , )、( , ) are two randomly selected samples and corresponding labels from the same batch.

[0027] Furthermore, the step 3 specifically includes the following steps:

[0028] Step 3.1: Construct an image segmentation network model, including an encoder for the downsampling path and a decoder for the upsampling path; the encoder performs multiple downsampling of the image through convolutional layers and pooling layers, gradually reducing the size and number of channels of the image, extracting features to obtain abstract semantic features; the decoder gradually restores the resolution of the image through upsampling operations, while introducing feature information from the encoder; a direct connection is established between each layer of the encoder and the corresponding layer of the decoder to transmit detailed spatial information; in the last layer of the network, a 1×1 convolution is used to fuse feature maps of different resolutions to produce the final segmentation result and achieve pixel-level classification;

[0029] Step 3.2: Based on the pre-trained weights of the VOC dataset, further freeze the image segmentation network model, freeze some common pre-trained weights, use the backbone network, and put more resources on the network parameter calculation in the later part of the training. When the loss value exceeds the set number of times and does not decrease, the frozen training is automatically ended.

[0030] Step 3.3: After the freeze training is completed, start the unfreeze training and train all the layers of the image segmentation network model together. When the loss value exceeds the set number of times and does not decrease, the training is terminated to obtain the final image segmentation network model weights used to detect the surface damage of the brake disc auxiliary rotor.

[0031] Furthermore, the specific process of step 4 is as follows:

[0032] Step 4.1: When the input image size is 512*512*3, the specific execution method is as follows:

[0033] 1) conv1: perform two [3,3] 64-channel convolutions to obtain a preliminary effective feature layer of [512,512,64], and then perform 2*2 maximum pooling to obtain a feature layer of [256,256,64];

[0034] 2) conv2: perform two [3,3] 128-channel convolutions to obtain a preliminary effective feature layer of [256,256,128], and then perform 2*2 maximum pooling to obtain a feature layer of [128,128,128];

[0035] 3) conv3: perform three [3,3] 256-channel convolutions to obtain a preliminary effective feature layer of [128,128,256], and then perform 2*2 maximum pooling to obtain a feature layer of [64,64,256];

[0036] 4) conv4: perform three [3,3] 512-channel convolutions to obtain a preliminary effective feature layer of [64,64,512], and then perform 2*2 maximum pooling to obtain a feature layer of [32,32,512];

[0037] 5) conv5: perform three [3,3] 512-channel convolutions to obtain a preliminary effective feature layer of [32,32,512];

[0038] Step 4.2: For the five preliminary effective feature layers obtained in step 4.1, use the enhanced feature extraction network to perform feature fusion. The feature fusion method is to upsample and stack the feature layers. When upsampling, directly perform two times upsampling and then perform feature fusion. The final feature layer has the same height and width as the input image.

[0039] Step 4.3: Use a 1×1 convolution kernel to adjust the channels of the features obtained in step 4.2, adjust the number of channels of the final feature layer to the number of categories, and output the predicted result image.

[0040] Furthermore, the labeling categories corresponding to the rust, fall-off, disc, and steel sheet areas in step 2 are: xiushi, tuoluo, diepian, and gangpian respectively;

[0041] The step 5 is specifically as follows: traverse the regions marked with diepian, gangpian, tuoluo, and xiushi respectively, count the total number of pixels in each part, and calculate the ratio of all the pixels in the marked tuoluo and xiushi parts to the sum of all the pixels in the brake disc auxiliary moving plate. ratio ; The formula is as follows:

[0042] (4)

[0043] in, P m For the dropped pixels, P x is the rusted pixel, T It is the total number of pixels of the brake disc.

[0044] Furthermore, the step 6 specifically includes:

[0045] Step 6.1: Use the openCV computer vision processing library to calculate the maximum radial pixel length of the peeling area; first, perform color filtering on the image output by the image segmentation network model to separate the red part representing the peeling area, the green part representing the normal part, and the blue part representing the rusted area; use the findContours function to find the inner diameter contour of the brake disc auxiliary rotor according to the rules, and use the approxPolyDP (fitting polygon) method to approximate the contour curve with line segments, so that the contour curve becomes line segments;

[0046] Step 6.2: Traverse the image segmentation network model to process all the red points on the brake disc auxiliary plate, and calculate the minimum distance direction from each red point to the corresponding line segment of the contour curve; then use the Bresenham straight line algorithm to find the red pixel length of all red points in the minimum distance direction, and then get the maximum value of the pixel length, which is the maximum pixel length of the detached area;

[0047] Step 6.3: Use the openCV computer vision processing library to calculate the width of the brake disc auxiliary plate ring; first perform color filtering to separate the red part representing the detached area, the green part representing the normal part, and the blue part representing the rusted area; then use the dilate and erode functions to dilate and erode to remove the five small circles to find the outer and inner diameter contours of the brake disc auxiliary plate ring; then use the least squares method and Gaussian elimination method to fit concentric ellipses; finally, calculate the major and minor diameters of the two ellipses according to the ellipse formula to obtain the ellipse arc width;

[0048] Step 6.4: According to the obtained maximum radial pixel length of the falling-off area and the width of the elliptical arc, the ratio of the maximum radial pixel length of the falling-off area to the width of the elliptical arc is obtained, and the actual maximum radial length of the falling-off area is further obtained according to the actual width of the elliptical arc.

[0049] Furthermore, the specific calculation process of calculating the minimum distance direction from the red point to the corresponding line segment of the contour curve in step 6.2 is as follows:

[0050] Assume that the coordinates of the red point are P(x0, y0), and the two end points of the line segment are P1(x1, y1) and P2(x2, y2).

[0051] Then the parametric coordinate equation of the line segment P1P2 is:

[0052] (5)

[0053] In the formula, t is the parameter of the line, which is a variable. Since this is a line segment, it limits t is 0 to 1; ( x ,y ) is the foot of the perpendicular from the red point P to the line segment P1P2, which is the dependent variable;

[0054] The square of the distance from the red point P to the point on the line segment P1P2 is expressed as:

[0055] (6)

[0056] make = 0 Solve to get the variable t Specific values:

[0057] (7)

[0058] So we get the direction of minimum distance:

[0059] (8)

[0060] In the formula, It is the direction of the minimum distance from the red point P to the corresponding line segment of the contour curve; norm is a regularization function, the goal is to change the length of the vector to 1; ( , ) is the specific coordinate value of the foot of the perpendicular from the red point P to the line segment P1P2.

[0061] Furthermore, the specific process of fitting concentric ellipses by the least square method and Gaussian elimination method in step 6.3 is as follows:

[0062] The outer diameter and the inner diameter are two concentric circles. Due to the shooting problem, they become two concentric ellipses with the same angle and curvature after affine transformation. Therefore, these two ellipses are fitted according to the points on the outer diameter and inner diameter contours.

[0063] Assume that the equations of the two ellipses are:

[0064] (9)

[0065] In the formula, A, B, C, D, E1 and E2 are all constants, and A 2 -4B>0;

[0066] Then the objective function is:

[0067] (10)

[0068] In the formula, N 1 and N 2 is the total number of points on the inner diameter and outer diameter respectively; and are the horizontal and vertical coordinates of the points on the inner diameter, and are the horizontal and vertical coordinates of the points on the outer diameter;

[0069] Taking partial derivatives of A, B, C, D, E1, E2 we get:

[0070] (11)

[0071] (12)

[0072] (13)

[0073] (14)

[0074] Use Gaussian elimination method to solve A, B, C, D, E1, E2 to get the equations of the two ellipses; according to the ellipse formula, get the major diameter a and short path b :

[0075] (15)

[0076] Where, E is a constant;

[0077] Finally, we get the width of the ellipse arc:

[0078] (16)

[0079] In the formula, and are the major and minor diameters of the outer ellipse, and are the major and minor diameters of the inner ellipse respectively.

[0080] The beneficial effects of the present invention are as follows: the present invention takes the landing gear brake disc pair as the research object, analyzes the wear characteristics of the landing gear disc brake system, establishes a semantic segmentation model of the landing gear brake disc image, forms an intelligent determination method for the damage degree and replacement demand of the aircraft brake disc pair by calculating the rust rate, shedding rate and maximum radial width index of the shedding area of ​​the brake disc pair surface, and designs a computer program to realize automatic, rapid and accurate evaluation and identification of the damage degree of the brake disc pair, thereby avoiding the problems of inconsistent evaluation standards and manual evaluation errors caused by human factors. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 The present invention is a flow chart of an intelligent method for determining the damage degree and replacement requirements of aircraft brake discs.

[0082] Figure 2 Schematic diagram of brake disc secondary surface damage classification.

[0083] Figure 3 (a) is the original image of the aircraft brake disc auxiliary plate.

[0084] Figure 3(b) shows the image after annotation.

[0085] Figure 4 This is the structure diagram of the image segmentation network model.

[0086] Figure 5 It is the backbone feature extraction network.

[0087] Figure 6 The main feature extraction process.

[0088] Figure 7 To enhance the feature extraction process.

[0089] Figure 8 For the prediction results.

[0090] Figure 9 (a) shows Example 1 of edge contour processing of the detached area.

[0091] Figure 9 (b) shows Example 2 of edge contour processing of the detached area.

[0092] Fig.10 Schematic diagram of the distance from the red point to the inner diameter contour.

[0093] FIG11( a ) is a schematic diagram of a binary image for finding the inner and outer diameter contours of the brake disc auxiliary rotor.

[0094] FIG11( b ) is a diagram showing the markings for finding the inner and outer diameter contours of the brake disc auxiliary rotor. DETAILED DESCRIPTION

[0095] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0096] An intelligent method for determining the damage degree and replacement requirements of aircraft brake discs, such as Figure 1 As shown, it includes six steps: landing gear brake disc image data collection and classification, data set preparation and data enhancement, model training, landing gear brake disc damage degree detection and identification, detection result evaluation, and brake disc detachment part maximum radial width detection, among which:

[0097] Step S1: The maintenance engineer uses a camera or other image acquisition device to collect photos of the aircraft brake disc auxiliary plate, and then cleans and organizes the collected image data.

[0098] The specific process is:

[0099] Step S1.1: Collect pictures of the surface of the aircraft brake disc auxiliary rotor at different angles, different distances, and different brightness, and extract appropriate samples frame by frame from the collected video data.

[0100] Step S1.2: During the image collection process of the brake disc auxiliary rotor surface, the number of samples of different damage categories should be roughly equal to avoid the model being biased towards the category with a larger number. At the same time, for each category, various samples should be included as much as possible to avoid only samples of a specific type.

[0101] Step S2: sorting the collected aircraft landing gear brake disc auxiliary image data, classifying the damage of the landing gear brake disc auxiliary rotor, marking the rusted, detached, disc, and steel plate areas, such as Figure 2 As shown; and annotate as required to form a brake disc secondary damage label file to obtain a data set; after the annotation is completed, the target data set is enhanced to increase the diversity of the training data, using Mosaic data enhancement and Mix Up data enhancement methods.

[0102] The specific process is:

[0103] Step S2.1: Classify the damage of the landing gear brake disc auxiliary rotor. The surface damage of the brake disc auxiliary rotor should be divided into two categories: rust and shedding. Rust is caused by oxidation reaction on the surface of the brake disc when it is in contact with water or in a humid environment. Its shape and distribution are highly random. Shedding of the rotor surface is caused by long-term wear and tear and overheating of the brake disc. The shedding area is usually distributed on the inner and outer diameters of the rotor and the thermal expansion groove.

[0104] Step S2.2: After sorting and organizing the sample data, various types of data are annotated to form the data set required for training. The annotated categories are: xiushi, tuoluo, gangpian, and diepian. Figure 3 (a) shows the original image collected, and Figure 3 (b) shows the annotated image. The blue area is the rusted area, annotated as "xiushi", the red area is the detached area, annotated as "tuoluo", the green area is the normal part, annotated as "diepian", and the yellow area is the steel sheet, annotated as "gangpian". The annotated image data is divided into training set, validation set, and test set in a ratio of 7:2:1.

[0105] Step S2.3: Mosaic data enhancement and Mix Up data enhancement are performed on the image data annotated in step S2.1. Specifically, mosaic data enhancement is to randomly read 4 pictures from the data set prepared in step S2.1 each time, and then flip (flip the original picture left and right), scale (scale the original picture), and change the color gamut (change the brightness, saturation, and hue of the original picture) and other operations on the 4 pictures. After the operation is completed, the original pictures are placed in four directions with the first picture on the upper left, the second picture on the lower left, the third picture on the lower right, and the fourth picture on the upper right. Then, the fixed areas of the four pictures are cut out using a matrix method and spliced ​​into a new picture, which contains a series of contents such as a labeling box. This operation enriches the background information of the image, and stitching the four images together also increases the number of samples in the data set in disguise. When batch normalization is performed, it is equivalent to calculating the four images at the same time, which greatly improves the training efficiency.

[0106] Hybrid data enhancement is an algorithm used in computer vision to enhance images by mixing different classes. It can mix images of different classes to expand the training data set. The formula is as follows:

[0107] (1)

[0108] (2)

[0109] (3)

[0110] in , is the mixed image and corresponding label, Beta is the Beta distribution, α , β are all constants, is the number of random samples calculated from a Beta distribution, ( , )、( , ) are two randomly selected samples and corresponding labels in the same batch.

[0111] Step S3: Using the data set obtained in step S2, the image segmentation model is trained through multiple iterations to obtain the model weights for detecting the surface damage of the brake disc auxiliary rotor.

[0112] The specific steps include:

[0113] Step S3.1: Construct an image segmentation network model, which consists of an encoder (downsampling path) and a decoder (upsampling path), with a "U"-shaped structure connecting the two in the middle. This structure allows the network to learn local features and global features at the same time, which is conducive to improving segmentation accuracy. The downsampling (encoder) part uses convolutional layers and pooling layers to perform multiple downsampling of the image, gradually reducing the size and number of channels of the image, and extracting features. The upsampling (decoder) part gradually restores the resolution of the image through upsampling operations (usually using transposed convolution or upsampling plus convolution), while introducing feature information from the encoder. A direct connection is established between each layer of the encoder and the corresponding layer of the decoder. These connections convey detailed spatial information, which helps to improve segmentation accuracy, especially the accuracy of target edge segmentation. In the last layer of the network, 1×1 convolution is used to fuse feature maps of different resolutions to produce the final segmentation result and achieve pixel-level classification.

[0114] like Figure 4 As shown, the structure of the image segmentation network model includes a backbone feature extraction network, an enhanced feature extraction network and a prediction network.

[0115] like Figure 5 As shown in the figure, the backbone feature extraction part of the image segmentation network model consists of convolutional layers and maximum pooling layers, and the overall structure is similar to VGG (Visual Graphics Generator).

[0116] Step S3.2: Pre-train based on the VOC (Visual Object Classes) dataset to obtain pre-training weights, with an initial learning rate of 0.001, using the adam optimizer and setting the momentum to 0.9, the weight decay to 0, the learning rate reduction method used to be "cos", and dual-threaded data reading to speed up data reading. When the loss value does not decrease for multiple times, the training is automatically terminated, and the unfreeze training of step S3.3 is performed.

[0117] The VOC dataset used in this embodiment is a widely used computer vision dataset, mainly used for tasks such as target detection, image segmentation, and image classification; it contains labeled images of various categories, each of which has an associated bounding box and an object category label. The VoC dataset covers releases from multiple years, and each year's dataset contains a training set, a validation set, and a test set. The training set is used for model training and parameter optimization, the validation set is used for model parameter adjustment and performance evaluation, and the test set is used for performance evaluation and comparison of the final model.

[0118] Step S3.3: While performing step S3.1, freeze training is used to freeze some common pre-trained weights, such as the backbone network, and put more resources into the network parameter calculation in the later part of the training. This can effectively reduce the training time and greatly improve resource utilization. When the loss value does not decrease for many times, the frozen training is automatically ended, and then the unfrozen training is started, and all layers of the model are trained together, and finally the model weights for detecting the surface damage of the brake disc auxiliary rotor are obtained. The frozen training parameters are score threshold of 0.5, batch_size of 8, and initial learning rate of 0.0001.

[0119] Step S3.4: Evaluate the model performance using the test dataset.

[0120] 1) Pixel Accuracy: used to evaluate the classification accuracy of the model for each pixel.

[0121] (4)

[0122] 2) Precision: It is used to measure the proportion of positive examples correctly predicted by the model relative to the total number of positive examples predicted by the model.

[0123] (5)

[0124] 3) Recall: It is used to measure the proportion of positive examples correctly predicted by the model relative to the actual number of positive examples.

[0125] (6)

[0126] 4) Mean Intersection over Union (mIoU): used to measure the segmentation performance of the model for each category.

[0127] (7)

[0128] 5) Mean Pixel Accuracy: used to measure the classification accuracy at the pixel level.

[0129] (8)

[0130] 6) Average Precision (mPrecision): It is used to measure the average accuracy of the model, that is, the proportion of positive examples correctly predicted by the model.

[0131] (9)

[0132] 7) Mean recall (mRecall): It is used to measure the average recall of the model, that is, the proportion of positive examples correctly predicted by the model relative to the proportion of actual positive examples.

[0133] (10)

[0134] Among them, TN is true negative, indicating that the model correctly classifies negative samples as negative; TP is true positive, indicating that the model correctly classifies positive samples as positive; FP is false positive, indicating that the model incorrectly classifies negative samples as positive; FN is false negative, indicating that the model incorrectly classifies positive samples as negative; C represents the total number of categories. is the intersection-and-union ratio, ; is the pixel accuracy, is the precision value, is the recall rate.

[0135] Step S4: input the image data into the constructed model, perform forward reasoning using the trained model weights, and output a prediction result image that is the same size as the input image and contains the surface wear information of the brake disc auxiliary pad.

[0136] The specific steps include:

[0137] Step S4.1: When the input image size is 512*512*3, the specific execution method is as follows Figure 6 The main feature extraction process is shown in the figure, as follows:

[0138] 1) conv1: Perform two [3,3] 64-channel convolutions to obtain a preliminary effective feature layer of [512,512,64], and then perform 2*2 maximum pooling to obtain a feature layer of [256,256,64].

[0139] 2) conv2: Perform two [3,3] 128-channel convolutions to obtain a preliminary effective feature layer of [256,256,128], and then perform 2*2 maximum pooling to obtain a feature layer of [128,128,128].

[0140] 3) conv3: Perform three [3,3] 256-channel convolutions to obtain a preliminary effective feature layer of [128,128,256], and then perform 2*2 maximum pooling to obtain a feature layer of [64,64,256].

[0141] 4) conv4: Perform three [3,3] 512-channel convolutions to obtain a preliminary effective feature layer of [64,64,512], and then perform 2*2 maximum pooling to obtain a feature layer of [32,32,512].

[0142] 5) conv5: Perform three [3,3] 512-channel convolutions to obtain a preliminary effective feature layer of [32,32,512].

[0143] Step S4.2: For the five preliminary effective feature layers obtained in step S4.1, use the enhanced feature extraction network to perform feature fusion. The feature fusion method is to upsample and stack the feature layers. When upsampling, directly perform two times upsampling and then perform feature fusion. The final feature layer has the same height and width as the input image. Figure 7 shown.

[0144] Step S4.3: Use a 1×1 convolution kernel to adjust the channels of the features obtained in step S4.2, adjust the number of channels of the final feature layer to the number of categories (num_classes), and output the predicted result image.

[0145] The prediction results obtained based on the input image are as follows Figure 8 shown.

[0146] Step S5: Count the number of pixels in each category in the output image to obtain the proportion of the rusted and peeled areas on the surface of the brake disc auxiliary rotor.

[0147] Traverse all categories including "diepian", "gangpian", "tuoluo", and "xiushi" areas, count the total number of pixels in each part, and calculate the ratio of all pixels in the "tuoluo" and "xiushi" parts to the sum of all pixels in the brake disc. ratio The formula is as follows:

[0148] (11)

[0149] in, P m For the dropped pixels, P x is the rusted pixel, T It is the total number of pixels of the brake disc rotor.

[0150] Step S6: further process the output image, calculate the pixel width of the annular part of the brake disc auxiliary plate and the maximum radial pixel length of the detached area, so as to obtain the actual maximum radial length of the detached area of ​​the brake disc auxiliary plate. Combined with the proportion of the rust and detached area on the surface of the brake disc auxiliary plate and the maximum radial length of the detached area, it is finally determined whether the brake disc pair should be eliminated.

[0151] The step S6 should also include:

[0152] Step S6.1: Use the openCV computer vision processing library to calculate the maximum radial pixel length of the detached area. First, perform color filtering on the image output by the image segmentation network model to separate the red, green, and blue parts. Use the findContours function to find the inner diameter contour of the brake disc auxiliary rotor according to the rules, and use the approxPolyDP method to approximate the contour with line segments. After the contour curve is approxified, the contour becomes a line segment.

[0153] Examples of edge contour processing of the fall-off area are shown in Figure 9 (a) and Figure 9 (b)

[0154] Step S6.2: Traverse all the red points on the brake disc auxiliary plate after processing the image segmentation network model, and calculate the direction of their minimum distance to the obtained curve. Then use the Bresenham straight line algorithm to find the red pixel length of all red points in the direction of the minimum distance, and finally find the maximum value of these pixel lengths, which is the maximum pixel length of the detached area. Among them, the specific calculation process for calculating the minimum distance direction from the red point to the curve is as follows: Assume that the coordinates of the red point are P(x0, y0), and the two end points of the line segment are P1(x1, y1) and P2(x2, y2). The distance from the red point to the contour is as follows: Fig.10 shown.

[0155] Then the parametric coordinate equation of the line segment P1P2 is:

[0156] (12)

[0157] The square of the distance from P to the point on the line segment P1P2 is expressed as:

[0158] (13)

[0159] make = 0 Solving for:

[0160] (14)

[0161] Therefore, the minimum distance direction can be obtained:

[0162] (15)

[0163] Step S6.3: Use the openCV computer vision processing library to calculate the width of the ring. First, perform color filtering to separate the red, green, and blue parts. Then use the dilate and erode functions to dilate and erode to remove the five small circles to find the outer and inner diameter contours. Next, use the least squares method and Gaussian elimination method to fit concentric ellipses. Finally, calculate the major and minor diameters of the two ellipses according to the ellipse formula to obtain the ellipse arc width. The specific process of fitting concentric ellipses by the least squares method and Gaussian elimination method is as follows: The outer diameter and inner diameter are two concentric circles. Due to the shooting problem, they become two concentric, same-angle, and same-curvature ellipses after affine transformation. Therefore, it is necessary to fit these two ellipses according to the points on the outer and inner diameter contours, as shown in Figures 11 (a) and (b).

[0164] Assume that the equations of the two ellipses are:

[0165] (16)

[0166] Then the objective function is:

[0167] (17)

[0168] In the formula, N 1 and N 2 is the total number of points on the inner diameter and outer diameter respectively; and are the horizontal and vertical coordinates of the points on the inner diameter, and are the horizontal and vertical coordinates of the point on the outer diameter.

[0169] Among them, (X1, Y1) is the set of points on the outer diameter, and (X2, Y2) is the set of points on the inner diameter; taking partial derivatives of A, B, C, D, E1, E2, we get:

[0170] (18)

[0171] (19)

[0172] (20)

[0173] (twenty one)

[0174] Use Gaussian elimination method to solve A, B, C, D, E1, E2 to get the equations of the two ellipses. According to the ellipse formula, the major diameter is obtained. a and short path b :

[0175] (twenty two)

[0176] Finally, we get the width of the ellipse arc:

[0177] (twenty three)

[0178] Step S6.4: Based on the obtained maximum pixel length of the dropout area and the elliptical arc width, the ratio of the maximum dropout area length to the elliptical arc width can be obtained, and the actual maximum radial length of the dropout area can be further calculated based on the actual width of the elliptical arc.

[0179] If the maximum radial width of the worn and detached connected area is greater than 0.25 inches (6.4mm), or the sum of the detachment and rust ratio is greater than 0.25, then the brake disc assembly is judged to have exceeded the standard of damage and needs to be replaced in time; otherwise, it is within the standard and can continue to be used.

Claims

1. An intelligent method for determining the damage degree and replacement requirements of aircraft brake discs, characterized in that: The following steps are involved: Step 1: Collect image data of the aircraft brake disc auxiliary plate and pre-process the collected image data; Step 2: Classify the damage of the aircraft brake disc auxiliary plate, mark the rust, detachment, disc, and steel plate areas, form a brake disc auxiliary damage label file, and obtain the aircraft brake disc auxiliary plate surface damage dataset; Step 3: Based on the labeled aircraft brake disc auxiliary plate surface damage dataset, the image segmentation network model is trained through multiple iterations using the transfer training plus frozen training method to obtain the model weights for detecting the surface damage degree of the aircraft brake disc auxiliary plate; Step 4: Input the aircraft brake disc auxiliary plate image data into the constructed image segmentation network model, use the trained model weights for forward reasoning, and output a prediction result image with the same size as the input image and containing the brake disc auxiliary plate surface wear information; Step 5: Count the number of pixels in each category in the output prediction result image to obtain the proportion of the rusted and peeled areas on the brake disc secondary surface; Step 6: Further process the output image to calculate the pixel width of the annular portion of the brake disc auxiliary plate and the maximum radial pixel length of the detached area, so as to obtain the actual maximum radial length of the detached area of ​​the brake disc auxiliary plate; Step 7: Comprehensively consider the proportion of the rusted and peeled areas on the brake disc pair moving plate surface and the actual maximum radial length of the peeled area to determine whether to eliminate the brake disc pair; The step 6 specifically includes: Step 6.1: Use the openCV computer vision processing library to calculate the maximum radial pixel length of the peeling area; first, perform color filtering on the image output by the image segmentation network model to separate the red part representing the peeling area, the green part representing the normal part, and the blue part representing the rusted area; Use the findContours function to find the inner diameter contour of the brake disc auxiliary rotor according to the rules, and use the approxPolyDP method to approximate the contour curve with line segments to convert the contour curve into line segments; Step 6.2: Traverse the image segmentation network model to process all the red points on the brake disc auxiliary plate, and calculate the minimum distance direction from each red point to the corresponding line segment of the contour curve; then use the Bresenham straight line algorithm to find the red pixel length of all red points in the minimum distance direction, and then get the maximum value of the pixel length, which is the maximum radial pixel length of the detached area; Step 6.3: Use the openCV computer vision processing library to calculate the width of the brake disc auxiliary plate ring; first perform color filtering to separate the red part representing the detached area, the green part representing the normal part, and the blue part representing the rusted area; then use the dilate and erode functions to dilate and erode to remove the five small circles to find the outer and inner diameter contours of the brake disc auxiliary plate ring; then use the least squares method and Gaussian elimination method to fit concentric ellipses; finally, calculate the major and minor diameters of the two ellipses according to the ellipse formula to obtain the ellipse arc width; Step 6.4: According to the obtained maximum radial pixel length of the falling-off area and the width of the elliptical arc, the ratio of the maximum radial pixel length of the falling-off area to the width of the elliptical arc is obtained, and the actual maximum radial length of the falling-off area is further obtained according to the actual width of the elliptical arc.

2. The intelligent determination method for the damage degree and replacement requirement of aircraft brake disc according to claim 1 is characterized in that: The step 2 also includes data enhancement of the aircraft brake disc auxiliary plate surface damage dataset, specifically: Perform mosaic data enhancement and mixed data enhancement on the labeled image data; The mosaic data enhancement is as follows: 4 pictures are randomly read from the aircraft brake disc auxiliary plate surface damage data set each time, and the 4 pictures are flipped, scaled, and color gamut changed respectively. After the operations are completed, the 4 original pictures are arranged in the form of the first picture at the upper left, the second picture at the lower left, the third picture at the lower right, and the fourth picture at the upper right, and then the four pictures are spliced ​​into a new picture, and the content of the annotation box is retained on the new picture; The mixed data enhancement is to mix images between different classes to expand the training data set. The formula is as follows: (1) (2) (3) in, , is the mixed image and corresponding label, Beta is the Beta distribution, α and β are all constants; is the random sample number calculated from the Beta distribution; ( , )、( , ) are two randomly selected samples and corresponding labels in the same batch.

3. The intelligent determination method for the damage degree and replacement requirement of aircraft brake disc according to claim 1 is characterized in that: The step 3 specifically includes the following steps: Step 3.1: Construct an image segmentation network model, including an encoder for the downsampling path and a decoder for the upsampling path; the encoder performs multiple downsampling of the image through convolutional layers and pooling layers, gradually reducing the size and number of channels of the image, extracting features to obtain abstract semantic features; the decoder gradually restores the resolution of the image through upsampling operations, while introducing feature information from the encoder; a direct connection is established between each layer of the encoder and the corresponding layer of the decoder to transmit detailed spatial information; in the last layer of the network, a 1×1 convolution is used to fuse feature maps of different resolutions to produce the final segmentation result and achieve pixel-level classification; Step 3.2: Based on the pre-trained weights of the VOC dataset, further freeze the image segmentation network model, freeze some common pre-trained weights, use the backbone network, and put more resources on the network parameter calculation in the later part of the training. When the loss value exceeds the set number of times and does not decrease, the frozen training is automatically terminated; Step 3.3: After the freeze training is completed, start the unfreeze training and train all the layers of the image segmentation network model together. When the loss value exceeds the set number of times and does not decrease, the training is terminated to obtain the final image segmentation network model weights used to detect the surface damage of the brake disc auxiliary plate.

4. The intelligent determination method for the damage degree and replacement requirement of aircraft brake disc according to claim 1 is characterized in that: The specific process of step 4 is as follows: Step 4.1: When the input image size is 512*512*3, the specific execution method is as follows: 1) conv1: perform two [3,3] 64-channel convolutions to obtain a preliminary effective feature layer of [512,512,64], and then perform 2*2 maximum pooling to obtain a feature layer of [256,256,64]; 2) conv2: perform two [3,3] 128-channel convolutions to obtain a preliminary effective feature layer of [256,256,128], and then perform 2*2 maximum pooling to obtain a feature layer of [128,128,128]; 3) conv3: perform three [3,3] 256-channel convolutions to obtain a preliminary effective feature layer of [128,128,256], and then perform 2*2 maximum pooling to obtain a feature layer of [64,64,256]; 4) conv4: perform three [3,3] 512-channel convolutions to obtain a preliminary effective feature layer of [64,64,512], and then perform 2*2 maximum pooling to obtain a feature layer of [32,32,512]; 5) conv5: perform three [3,3] 512-channel convolutions to obtain a preliminary effective feature layer of [32,32,512]; Step 4.2: For the five preliminary effective feature layers obtained in step 4.1, use the enhanced feature extraction network to perform feature fusion. The feature fusion method is to upsample and stack the feature layers. When upsampling, directly perform two times upsampling and then perform feature fusion. The final feature layer has the same height and width as the input image. Step 4.3: Use a 1×1 convolution kernel to adjust the channels of the features obtained in step 4.2, adjust the number of channels of the final feature layer to the number of categories, and output the predicted result image.

5. The intelligent determination method for the damage degree and replacement requirement of aircraft brake disc according to claim 1 is characterized in that: The marking categories corresponding to the rust, fall-off, disc, and steel sheet areas in step 2 are: xiushi, tuoluo, diepian, and gangpian; The step 5 is specifically as follows: traverse the regions marked with diepian, gangpian, tuoluo, and xiushi respectively, count the total number of pixels in each part, and calculate the ratio of all the pixels in the marked tuoluo and xiushi parts to the sum of all the pixels in the brake disc auxiliary moving plate. ratio ; The formula is as follows: (4) in, P m For the dropped pixels, P x is the rusted pixel, T It is the total number of pixels of the brake disc rotor.

6. The intelligent determination method for the damage degree and replacement requirement of aircraft brake disc according to claim 1 is characterized in that: The specific calculation process of calculating the minimum distance direction from the red point to the corresponding line segment of the contour curve in step 6.2 is as follows: Assume the coordinates of the red point are P(x0, y0), and the two end points of the line segment are P1(x1, y1) and P2 (x2, y2); Then the parametric coordinate equation of line segment P1 P2 is: (5) In the formula, t is the parameter of the line segment, is a variable; ( x , y ) is the foot of the perpendicular from the red point P to the line segment P1 P2, which is the dependent variable; The square of the distance from the red point P to the point on the line segment P1 P2 is expressed as: (6) make = 0 Solve to get the variable t Specific values: (7) So we get the direction of minimum distance: (8) In the formula, It is the direction of the minimum distance from the red point P to the corresponding line segment of the contour curve; norm is a regularization function, the goal is to change the length of the vector to 1; ( , ) is the specific coordinate value of the foot of the perpendicular from the red point P to the line segment P1 P2.

7. The intelligent determination method for the damage degree and replacement requirement of aircraft brake disc according to claim 1 is characterized in that: The specific process of fitting concentric ellipses by the least squares method and Gaussian elimination method in step 6.3 is as follows: The outer diameter and the inner diameter are two concentric circles. Due to the shooting problem, they become two concentric ellipses with the same angle and curvature after affine transformation. Therefore, these two ellipses are fitted according to the points on the outer diameter and inner diameter contours. Assume that the equations of the two ellipses are: (9) In the formula, A, B, C, D, E1 and E2 are all constants, and A 2 -4B>0; Then the objective function is: (10) In the formula, N 1 and N 2 is the total number of points on the inner diameter and outer diameter respectively; and are the horizontal and vertical coordinates of the points on the inner diameter, and are the horizontal and vertical coordinates of the points on the outer diameter; Taking partial derivatives of A, B, C, D, E1, E2 we get: (11) (12) (13) (14) Use Gaussian elimination method to solve A, B, C, D, E1, E2 to get the equations of the two ellipses; according to the ellipse formula, get the major diameter a and short path b : (15) Where, E is a constant; Finally, we get the width of the ellipse arc: (16) In the formula, and are the major and minor diameters of the outer ellipse, and are the major and minor diameters of the inner ellipse respectively.

Images