Fatigue band recognition and segmentation method of metal material fracture based on deep learning

Through a deep learning-based method, combined with improved RRT algorithm and U-net network, a mathematical model of fracture fatigue bands of metal materials was established, which solved the problems of low recognition efficiency and insufficient accuracy in the prior art, and achieved complexity High-precision fatigue band segmentation in fracture images.

CN113935989BActive Publication Date: 2024-11-05SHENYANG JIANZHU UNIVERSITY
View PDF 0 Cites -1 Cited by

Patent Information

Application Number
CN202111392841.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-23
Publication Date
2024-11-05
Estimated Expiration
2041-11-23

Smart Images

  • Figure CN113935989B_ABST
    Figure CN113935989B_ABST
Patent Text Reader

Abstract

The present invention provides a method for identifying and segmenting fatigue bands on metal material fractures based on deep learning, and relates to the technical field of metal fractures. The method first establishes a mathematical model of fatigue bands on fractures; then simulates the extension trajectory of fatigue bands, and obtains a texture curve image of fatigue bands in combination with the mathematical model of fatigue bands on fractures; collects fracture images, marks the bands of fracture images, obtains labeled data samples, and then establishes a fatigue band sample data set in combination with fatigue band curve images; pre-processes the fatigue band sample data set images; builds a fatigue band identification and segmentation model based on a neural network, and uses the labeled fatigue band sample data for model training; finally, inputs the fatigue band image to be identified and segmented into the trained fatigue band identification model, and performs band identification and segmentation on the fatigue band image on fractures. The method can segment fatigue band feature areas from complex fatigue fractures with a high accuracy rate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of metal fracture technology, and in particular to a method for identifying and segmenting fatigue bands on metal material fractures based on deep learning. Background Art

[0002] In engineering organizations and mechanical equipment, components that have been in service for a long time under alternating stress are prone to fatigue fracture, causing serious economic losses. Therefore, people are increasing their research efforts on fatigue fracture. The study of metal fracture diagrams is an important step in analyzing fatigue fracture, and measuring the spacing of fatigue bands to quantitatively analyze fatigue life and fatigue stress is an important indicator and link in analyzing metal fracture diagrams. However, the complexity of the alternating stresses in the fracture process of components makes the actual component fractures mostly present a variety of mixed forms, which makes the fatigue fracture images collected by the electron microscope present complex multi-feature morphology, which is difficult to observe and analyze. Therefore, it is very meaningful to accurately identify and separate the strips in the complex morphology.

[0003] Early identification methods mainly relied on the experience of researchers in materials and other fields, and used manual visual identification. Although this method uses expert experience with a high accuracy rate, it has a low recognition efficiency, takes a long time, and wastes manpower. This method is basically no longer used. With the continuous development of computer technology, how to use computer image processing and pattern recognition technology to perform fracture morphology analysis, recognition and segmentation has become a hot topic in the field of fracture research.

[0004] In the field of fracture research, the gray level co-occurrence matrix is ​​recognized as a classic and effective method for fatigue fracture morphology recognition and classification in the early days. The gray level co-occurrence matrix (GLCM) is a classic second-order texture statistical method. It not only reflects the distribution characteristics of each gray level, but also reflects the position distribution characteristics between pixels with the same or close gray levels. When using GLCM for texture analysis, the characteristics of texture recognition use a set of statistics calculated by GLCM. These statistics describe the statistical characteristics of image texture from different aspects, but there is a certain correlation between them, which makes the information expression redundant and repetitive, greatly increasing the amount of calculation. The correlation coefficient matrix is ​​obtained for these texture feature parameters, and some feature parameters with large correlation coefficients are removed to obtain relatively independent statistics. This is more accurate for fracture image segmentation to a certain extent, and greatly reduces the amount of calculation and improves the efficiency of image processing.

[0005] However, when this method is used in the specific field of texture segmentation, its effect is difficult to meet the requirements. Due to the complexity of the fracture process in the segmentation of the strips, the actual fracture surface often manifests itself as complex multiple features. A single feature is difficult to accurately segment the fatigue strips. The low accuracy makes it difficult to carry out the next step of fracture morphology analysis.

[0006] The recognition and segmentation of fracture images based on wavelet packet transform is also a typical method. Wavelet packet transform is an extension and improvement of wavelet transform. The multi-resolution theory of wavelet analysis further divides the frequency into binary subspaces, further improving the resolution of the frequency domain time period. The essence of wavelet packet transform is to approximate the original signal with a linear combination of orthogonal wavelet packet basis functions. It is a more sophisticated signal analysis method that can more freely select the corresponding frequency band according to the characteristics of the analyzed signal, thereby matching the signal spectrum, improving the time domain resolution, and better extracting the corresponding features.

[0007] However, as a time-frequency analysis tool, wavelet packet transform can process the medium frequency and medium-high frequency segments in the image texture details. However, the details of the fracture image also include very rich strip texture information. It is difficult to express the characteristics of the details with a single time-frequency analysis, so the extracted single texture feature is difficult to accurately segment the fatigue strips from the complex texture morphology.

[0008] With the continuous development of science and technology, traditional machine learning and feature recognition technologies are gradually being replaced by deep learning. First of all, image recognition and segmentation based on deep learning are significantly better than traditional image processing methods in processing complex and redundant samples. Since the deep neural network itself can also be regarded as a "high-dimensional feature processor", when faced with complex multi-feature fracture images, the use of deep neural networks can very efficiently identify and segment the features of the fracture with high accuracy. Moreover, the deep learning method also has the advantages of strong learning ability, wide coverage, strong adaptability, good portability, etc., which are very suitable for complex feature fracture images.

[0009] At present, image recognition methods based on deep learning have made outstanding contributions in the field of face recognition. When segmenting highly complex images, it is of utmost importance to consider how to improve the segmentation and recognition accuracy, how to improve the efficiency of segmentation, and how to improve the simplicity of the method. Therefore, it is urgent to develop a new accurate and efficient fatigue band recognition and segmentation method to meet the requirements of researchers in the field of materials and effectively improve the recognition accuracy of fracture images for bands. Summary of the invention

[0010] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and provide a method for identifying and segmenting fatigue bands in metal material fractures based on deep learning, which can accurately identify and segment fatigue bands in fracture images.

[0011] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for identifying and segmenting fatigue bands on metal material fractures based on deep learning, comprising the following steps:

[0012] Step 1: According to the fracture fatigue mechanism of metal materials, a mathematical model of fracture fatigue band is established;

[0013] (1) According to the fatigue band morphology mechanism, determine the relationship between fatigue stress amplitude and fatigue band spacing;

[0014] When the number of stress cycles is set to N and the fatigue crack extension length is set to a, the fatigue crack extension μ under each load cycle is expressed as follows:

[0015] μ=△S=da / dN

[0016] Among them, △S is the fatigue strip interval;

[0017] According to Paris's formula: The relationship between fatigue stress amplitude and fatigue band spacing is obtained as shown in the following formula:

[0018]

[0019] Among them, C and m are metal material parameters, Y is the shape factor, and Δσ is the fatigue stress amplitude;

[0020] (2) Obtain the arc direction of the fatigue strip by using the square difference accumulation method;

[0021] According to the minimum or maximum value of the image grayscale difference in each direction of the fatigue stripe, the arc direction of the fatigue stripe is obtained;

[0022] The grayscale differences of the fatigue strip along the four directions of 0°, 45°, 90°, and 135° are set to be d 0° ,d 45° ,d 90° ,d 135° ,in:

[0023] d 0° (x,y)=[I(x-1,y)-I(x+1,y)] 2

[0024] d 45° (x,y)=[I(x-1,y+1)-I(x+1,y-1)] 2 ×0.5

[0025] d 90° (x,y)=[I(x,y-1)-I(x,y+1)] 2

[0026] d 45° (x,y)=[I(x-1,y-1)-I(x+1,y+1)] 2 ×0.5

[0027] Where I(x,y) is the image grayscale of point (x,y) on the fatigue strip diagram;

[0028] Then, according to the image grayscale differences of the fatigue strip along the four directions of 0°, 45°, 90°, and 135°, the mean square difference of the image grayscale of the fatigue strip along these four directions is calculated, and the mean square difference of the image grayscale in these four directions is the arc direction of the fatigue strip;

[0029] Step 2: Use the improved RRT algorithm to simulate the extension trajectory of the fatigue band, and combine the mathematical model of the fracture fatigue band to obtain the texture curve image of the fatigue band;

[0030] Step 2.1: Define the starting point, end point, and number of sampling points of the fatigue strip arc, and set the step length between each sampling point to t; define the starting point and end point of the fatigue strip arc as the root node and target node of the random tree, and define the obstacle position coordinates;

[0031] Step 2.2: The spacing and curvature of the fatigue strip obtained by the fatigue strip mathematical model are used as constraints and set as obstacles for the RRT algorithm to adjust the extension direction of the next random point of the RRT algorithm;

[0032] Step 2.3: Use the generated new random point as the parent point, input the constraint parameters, and generate the downward child tree branch;

[0033] Step 2.4: Add all new random points to the random tree set, and connect all sub-tree branches into a trajectory line to obtain the final fatigue strip texture curve;

[0034] Step 3: Use a scanning electron microscope to collect fracture images, use a labeling tool to mark the strips of the fracture images, obtain labeled data samples, and then combine the fatigue strip texture curve image obtained by the improved RRT algorithm to establish a fatigue strip sample data set and perform data set enhancement;

[0035] Step 4: Preprocess the fatigue stripe sample data set images by denoising and geometric transformation;

[0036] Step 4.1: Use Gaussian filtering to perform image denoising to remove random noise from the image in the dataset and reduce noise interference in the image;

[0037] Step 4.2: Use the nearest neighbor interpolation method to perform geometric transformation of the image to correct the instrument's own error and the random error of the instrument's position during fracture image acquisition;

[0038] Step 5: Build a fatigue band recognition and segmentation model based on a neural network, and use the labeled fatigue band sample data to train the model, adjust parameters, and optimize;

[0039] The fatigue band recognition model adopts the U-net network structure. The overall structure of the U-net network adopts an encoding-decoding structure, including a downsampling module, an upsampling module, and a jump link, which can realize end-to-end image segmentation; the downsampling module includes four layers of double convolution sub-blocks, and the incoming data is downsampled and features are extracted after passing through each double convolution sub-block. The purpose of the downsampling module is to extract feature maps by continuously compressing the input image; the upsampling module is symmetrical with the downsampling module in structure, and also includes four layers of double convolution sub-blocks, and a deconvolution is added after each layer of double convolution sub-block to restore the size of the feature map; the jump link part is to fuse the features obtained after convolution with the features obtained by deconvolution in the upsampling module each time downsampling is prepared, and the features are spliced ​​together in the channel dimension;

[0040] The U-net network structure uses a multi-layer double convolution structure and multiple downsampling to form a contraction path, continuously downsamples to extract features, expands feature channels, and obtains feature maps; after obtaining the feature map, the multi-layer double convolution structure is used again with multiple upsampling to restore the size of the feature map;

[0041] The double convolution structure is composed of two convolution kernels and two linear rectification functions. A maximum pooling layer is added after the double convolution structure for downsampling to ensure the accuracy of feature extraction. The deconvolution method is used to upsample the feature map to expand the image information.

[0042] The U-net network structure adds a skip connection method to the first three downsampling layers and upsampling layers to connect the feature maps, completes the combination of downsampling depth information and upsampling shallow layer information, completes the compensation of image missing information, and restores image pixel information;

[0043] U-net network structure At the end of the feature recognition process, a softmax function is added for feature classification;

[0044] The fatigue band recognition model uses a cross entropy function as a loss function;

[0045] Step 6: Input the fatigue stripe image to be identified and segmented into the trained fatigue stripe recognition model, perform stripe recognition and segmentation on the fracture fatigue stripe image, and finally obtain an end-to-end semantic segmentation result.

[0046] The beneficial effects of the above technical solution are: the metal material fracture fatigue band recognition and segmentation method based on deep learning provided by the present invention performs data set enhancement through the improved RRT algorithm, effectively solving the problem of unsatisfactory learning effect caused by too little network learnable data in the segmentation task, and effectively expanding the number of samples. Through the network model for metal fracture fatigue band feature segmentation, the fatigue band feature area can be segmented from a complex fatigue fracture with a high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flowchart of a method for identifying and segmenting fatigue bands on metal material fractures based on deep learning provided in an embodiment of the present invention;

[0048] Figure 2 A schematic diagram of a strip trajectory generated by an improved RRT algorithm provided in an embodiment of the present invention;

[0049] Figure 3 A real fracture image collected by a scanning electron microscope and a camera provided in an embodiment of the present invention;

[0050] Figure 4 A schematic diagram of the U-net network structure provided by an embodiment of the present invention;

[0051] Figure 5 A schematic diagram of a dual convolution structure provided by an embodiment of the present invention;

[0052] Figure 6 A schematic diagram of a deconvolution method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0054] In this embodiment, a method for identifying and segmenting fatigue bands on metal material fracture surfaces based on deep learning is provided. Figure 1 As shown, the following steps are included:

[0055] Step 1: According to the fracture fatigue mechanism of metal materials, a mathematical model of fracture fatigue band is established;

[0056] (1) According to the fatigue band morphology mechanism, determine the relationship between fatigue stress amplitude and fatigue band spacing;

[0057] When the number of stress cycles is set to N and the fatigue crack extension length is set to a, the fatigue crack extension μ under each load cycle is expressed as follows:

[0058]

[0059] Among them, △S is the fatigue strip interval;

[0060] According to Paris's formula: The relationship between fatigue stress amplitude and fatigue band spacing is obtained as shown in the following formula:

[0061]

[0062] Among them, C and m are metal material parameters, Y is the shape factor, and Δσ is the fatigue stress amplitude;

[0063] From this formula, it can be seen that when different stress amplitudes are set, the corresponding fatigue strip interval can be obtained. As the stress amplitude setting changes, the corresponding strip interval can be used in the next step of trajectory generation.

[0064] (2) Obtain the arc direction of the fatigue stripe (that is, the approximate direction of the fatigue stripe) by using the square difference accumulation method;

[0065] The basic idea is: for fatigue stripe images, the image data difference along the stripe tangent direction is the smallest, while the image data difference along the stripe normal direction is the largest. Therefore, the arc direction of the fatigue stripe is obtained according to the minimum or maximum value of the image grayscale difference in each direction of the fatigue stripe;

[0066] The image grayscale differences of the fatigue strip along the four directions of 0°, 45°, 90°, and 135° are set to be d 0° ,d 45° ,d 90° ,d 135° ,in:

[0067] d 0° (x,y)=[I(x-1,y)-I(x+1,y)] 2

[0068] d 45° (x,y)=[I(x-1,y+1)-I(x+1,y-1)] 2 ×0.5

[0069] d 90° (x,y)=[I(x,y-1)-I(x,y+1)] 2

[0070] d 45° (x,y)=[I(x-1,y-1)-I(x+1,y+1)] 2 ×0.5

[0071] Where I(x,y) is the image grayscale of point (x,y) on the fatigue strip diagram;

[0072] Then, according to the image grayscale differences of the fatigue strip along the four directions of 0°, 45°, 90°, and 135°, the mean square difference of the image grayscale of the fatigue strip along these four directions is calculated, and the mean square difference of the image grayscale in these four directions is the arc direction of the fatigue strip;

[0073] Step 2: Use the improved RRT algorithm to simulate the extension trajectory of the fatigue band, and combine the mathematical model of the fracture fatigue band to obtain the texture curve image of the fatigue band;

[0074] As a relatively traditional path planning method, RRT has a simple principle and can be used regardless of the number of degrees of freedom and the complexity of the constraints. This is one of the main reasons why it is popular in the field of robotics. The basic steps of RRT are:

[0075] 1. The starting point is a seed, from which branches begin to grow;

[0076] 2. Generate a random point in space;

[0077] 3. Find the closest point on the tree and record it as new;

[0078] 4. Grow in the new direction. If there are no obstacles, add the grown branches and endpoints to the tree.

[0079] Random points are generally evenly distributed, so when there are no obstacles, the branches grow approximately evenly in all directions, which allows for rapid exploration of space. If you have information about the area where the path is most likely to be found, you can focus on exploring that area, and at this time, uniform distribution is no longer sufficient.

[0080] According to this series of excellent characteristics of the RRT algorithm, we can improve it on its basis to enhance the corresponding data set. Although the RRT algorithm planning is very efficient, it is too random, so we need to give it corresponding direction guidance during the tree extension process so that it can quickly and accurately complete the process from the starting point to the end point, so that the algorithm can converge accurately. The improvement of RRT should be comprehensively considered in terms of random sampling method, nearest point definition method, and tree expansion form. Here we propose an improved RRT algorithm. The above problems can be perfectly solved by establishing a fatigue strip mathematical model. After obtaining the various parameters in the fatigue strip mathematical model, the corresponding parameters can be set as the constraints and obstacles in the fatigue strip extension process, that is, the RRT algorithm process, and the direction of path extension is obtained by assuming the position of the constraints. Then, according to the step size set by the algorithm, the corresponding random points can be determined, and then new branches can be generated. After several steps of constraints, a series of random points that conform to the fatigue strip mechanism law can be obtained, and finally the entire path of the fatigue strip will be obtained, that is, the texture curve image of the fatigue strip. The specific implementation process is:

[0081] Step 2.1: Initialize the RRT algorithm working environment, define the starting point, end point, and number of sampling points of the fatigue strip arc, and set the step length between each sampling point to t; define the starting point and end point of the fatigue strip arc as the root node and target node of the random tree, and define the obstacle position coordinates;

[0082] Step 2.2: The spacing and curvature of the fatigue strip obtained by the fatigue strip mathematical model are used as constraints and set as obstacles for the RRT algorithm to adjust the extension direction of the next random point of the RRT algorithm;

[0083] Step 2.3: Use the generated new random point as the parent point, input the constraint parameters, and generate the downward child tree branch;

[0084] Step 2.4: Add all new random points to the random tree set, and connect all sub-tree branches into a trajectory line to obtain the final fatigue strip texture curve; the fatigue strip texture curve obtained by the improved RRT algorithm in this embodiment is as follows Figure 2 shown.

[0085] Step 3: Use a scanning electron microscope to collect fracture images, use a labeling tool to mark the strips of the fracture images, obtain labeled data samples, and then combine the fatigue strip curve image obtained by the improved RRT algorithm to establish a fatigue strip sample data set and enhance the data set;

[0086] In this embodiment, the specific method of collecting the fracture image using the scanning electron microscope is as follows: prepare a suitable fatigue tensile specimen, use a fatigue testing machine to perform a tensile test on the fatigue tensile specimen to obtain a fracture specimen, and then use a scanning electron microscope to collect the fracture image;

[0087] Data set enhancement also requires the support of fracture images. Often, a data set formed by combining real samples with artificial data samples often has better results when training neural networks, so it is necessary to collect fracture images. The scanning electron microscope used in the fracture image collection of this embodiment is very suitable for fracture research because of its high resolution, large depth of field and ability to directly observe samples. Another important feature of the scanning electron microscope is that the magnification range is relatively wide, and fracture images of different magnifications can be collected. The fatigue fracture image obtained in this embodiment is as follows. Figure 3 shown.

[0088] Regarding the preparation of fatigue fracture specimens, the internationally accepted three-point bending specimen preparation method is generally used. When the fracture is formed after the specimen is broken, if various lubricating oils, oxide layers and corrosion products remain on the fracture, they may interfere with the effect of the picture, so it is necessary to deal with them. In order to reduce their influence, ultrasonic cleaning can be performed in an organic solvent first, and then the replica can be repeatedly torn off. If there is a fracture oxide layer that is difficult to remove, a chemical formula can be used for rust removal. After the fracture specimen is prepared, it is scanned and imaged by a scanning electron microscope to obtain a fracture image.

[0089] Since the training of neural networks requires prominent features to achieve better training results, the fracture images need to be pre-processed by labeling. The strip morphology in the fracture is divided out by labeling. This embodiment uses via for image annotation. First, the fracture image that needs to be labeled is imported. The labeling operation of the strip area is performed based on the understanding of the mechanism of fatigue stripes. A series of polylines are used to distinguish the strip features from other features in the edge area of ​​the strip, and mark them. Obtain a precisely labeled fracture image. Then import the labeled fracture image into the fatigue strip curve graph obtained by the improved RRT method, and merge the two to obtain an enhanced data set.

[0090] Step 4: Preprocess the fatigue stripe sample data set images by denoising and geometric transformation;

[0091] Step 4.1: Use Gaussian filtering to remove random noise from the image in the data set, reduce noise interference in the image, and obtain sample images with high recognition. In the process of collecting fatigue fracture images, due to force majeure, the image will inevitably be interfered by noise. First, the fracture image should be preprocessed, mainly using Gaussian filtering to eliminate random noise in the image, so that the image is more recognizable.

[0092] Step 4.2: Image geometric transformation is also called image space transformation. The collected image is processed by geometric transformations such as translation, transposition, mirroring, rotation, and scaling. The present invention uses the nearest neighbor interpolation method to perform geometric transformation of the image to correct the instrument's own errors during fracture image collection and the random errors of the instrument's position (imaging angle, perspective relationship, and even the lens itself);

[0093] Step 5: Build a fatigue band recognition and segmentation model based on a neural network, and use the labeled fatigue band sample data to train the model, adjust parameters, and optimize;

[0094] The fatigue band recognition model adopts the U-net network structure. The overall structure of the U-net network adopts an encoding-decoding structure, including a downsampling module, an upsampling module, and a jump link, which can realize end-to-end image segmentation; the downsampling module includes four layers of double convolution sub-blocks, and the incoming data is downsampled and extracted features after passing through each double convolution sub-block. The purpose of the downsampling module is to extract feature maps by continuously compressing the input image; the upsampling module is symmetrical with the downsampling module in structure, and also includes four layers of double convolution sub-blocks, and a deconvolution is added after each layer of double convolution sub-block to restore the size of the feature map. The design significance of the entire upsampling module is to restore the feature map to the size of the original image. The jump link part is to fuse the features obtained after convolution with the features obtained by deconvolution in the upsampling module each time downsampling is prepared, and the features are spliced ​​together in the channel dimension, so as to better combine global features with local features and improve segmentation accuracy;

[0095] The U-net network is modified and expanded based on the fully convolutional neural network (FCN). Compared with FCN, it improves the conventional expansion network. Its specific network structure is as follows: Figure 4 As shown in the figure, the pooling operation is replaced by the upsampling operation, thereby improving the resolution of the output. In order to better compensate for the missing information of the image, the deep features obtained from the contraction network are combined with the shallow features in the upsampling, so that the continuous convolutional layers can output more accurately. In the U-net network, there are a large number of feature channels, allowing the network to propagate range information to the high-resolution layer.

[0096] The U-net network structure uses a multi-layer double convolution structure and multiple downsampling to form a contraction path, continuously downsamples to extract features, expands feature channels, and obtains feature maps; after obtaining the feature map, the multi-layer double convolution structure is used again with multiple upsampling to restore the feature map size, and the final output image has the same size as the original input image; the double convolution structure is as follows Figure 5 As shown in the figure, the structure is composed of two convolution kernels plus two linear rectification functions. A maximum pooling layer is added after the double convolution structure for downsampling to ensure the accuracy of feature extraction. The deconvolution method is used to upsample the feature map to expand the image information.

[0097] In this embodiment, the contraction path includes two 3×3 convolutions, each of which is followed by a ReLU and a maximum pooling layer (with a step size of 2) for downsampling. Each downsampling reduces the number of feature channels by half. The expansion path includes several upsampling (2×2 deconvolution). The schematic diagram of the deconvolution is shown in Figure 6As shown in Figure 1, its main function is to halve the number of feature channels while restoring the image size. The output image is passed through a 1×1 convolution to map the 64-element feature vector to a class label.

[0098] The U-net network structure adds a skip connection method to the first three downsampling layers and upsampling layers to connect the feature maps, completes the combination of downsampling depth information and upsampling shallow layer information, completes the compensation of image missing information, and restores image pixel information;

[0099] U-net network structure At the end of the feature recognition process, a softmax function is added for feature classification;

[0100] The softmax function formula is as follows:

[0101]

[0102] Among them, a k (x) represents the score of each pixel (x) corresponding to the feature channel (k), K is the number of classes, p k (x) is the classification result of pixel x for class k.

[0103] The fatigue band recognition model uses the cross entropy function as the loss function, which is defined as follows:

[0104] E=∑ x∈Ω ω(x)log(P l(x) (x)

[0105] Where E represents the loss function, ω(x) represents the weight mapping (weight score introduced based on experience), l(x) represents the true label of pixel x on the image, Ω is the set of pixel positions of the entire image, and P l (x) is the classification result of pixel x. The ultimate goal is to maximize the segmentation accuracy while controlling the probability of other category features, that is, to minimize the loss;

[0106] In this embodiment, the weight of each real sample is pre-calculated to compensate for the different frequencies of pixels of a certain class in the training dataset and guide the network to learn the separation boundary we introduced in the strip area.

[0107]

[0108] Among them, ω(x) represents the weight mapping mentioned above, ω c (x) is the weight for balancing the category ratio, ω0 represents the initial empirical weight, d1 is the distance from a pixel to the nearest strip, d2 is the distance from a pixel to the second nearest strip, σ 2 It represents the variance of two variables d1 and d2.

[0109] Weight initialization is very important. Here we use Gaussian distribution to initialize weights with a standard deviation of Where N represents the number of input nodes of a neuron. For example, for a 3×3 convolution, the previous layer has 64 feature channels, then N=3×3×64=576.

[0110] U-net is a relatively simple network among many networks. It is easy to apply and improve in actual use. It has a lot of room for improvement and can be transformed in the architecture according to different needs. Therefore, it is very practical in image segmentation.

[0111] Step 6: Input the fatigue stripe image to be identified and segmented into the trained fatigue stripe recognition model, perform stripe recognition and segmentation on the fracture fatigue stripe image, and finally obtain the end-to-end semantic segmentation result. The end-to-end semantic segmentation result means that the network input is a complete image, and the output is also a complete visualization image instead of a feature map, that is, the input and output are consistent.

[0112] After the fatigue band recognition model is trained, the U-net network can recognize the bands based on the learned features. More specifically, some fracture images are selected as test sets to input into the neural network for image segmentation. After the computer completes the function convergence process, the U-net network can accurately identify the band features and accurately segment the band area.

[0113] U-net is suitable for completing semantic segmentation tasks. Standard semantic segmentation is also called full-pixel semantic segmentation, which is the process of classifying each pixel as belonging to an object class. In the fracture image, according to the pixel-level classification, the strip features can be distinguished from holes, dimples, tear edges and other morphologies, and the corresponding division of different types of morphologies can be obtained.

[0114] Semantic segmentation is more suitable for extracting target features from complex shapes. More specifically, semantic segmentation is similar to processing classified label data, using pixel-level onehot encoding for predicting classified targets, that is, creating an output channel for each classification category.

[0115] In summary, the task of semantic segmentation is to process the input image through a deep learning algorithm to obtain an output image of the same size with semantic labels.

[0116] Assume that the input image size is W×W, the convolution kernel size is F×F, the stride is S, and P layers are added to the outermost of the convolution layer. The output image size after the convolution layer is N×N, and the formula is as follows:

[0117]

[0118] For the pooling operation, let the input image size be W×H, the convolution kernel size be F×F, S be the step size, and the output image size after pooling be:

[0119]

[0120] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present invention.

Claims

1. A method for identifying and segmenting fatigue bands on metal material fracture surfaces based on deep learning, characterized in that: The following steps are involved: Step 1: According to the fracture fatigue mechanism of metal materials, a mathematical model of fracture fatigue band is established; Step 2: simulate the extension trajectory of the fatigue band and obtain the texture curve image of the fatigue band by combining the mathematical model of the fracture fatigue band; Step 3: Use a scanning electron microscope to collect fracture images, use a labeling tool to mark the strips of the fracture images, obtain labeled data samples, and then combine the fatigue stripe texture curve image obtained in step 2 to establish a fatigue stripe sample data set and perform data set enhancement; Step 4: Preprocess the fatigue stripe sample data set images by denoising and geometric transformation; Step 5: Build a fatigue band recognition and segmentation model based on a neural network, and use the labeled fatigue band sample data to train the model, adjust parameters, and optimize; Step 6: Input the fatigue stripe image to be identified and segmented into the trained fatigue stripe recognition model, perform stripe recognition and segmentation on the fracture fatigue stripe image, and finally obtain an end-to-end semantic segmentation result; The specific method of step 1 is: (1) According to the fatigue band morphology mechanism, determine the relationship between fatigue stress amplitude and fatigue band spacing; When the number of stress cycles is set to N and the fatigue crack extension length is set to a, the fatigue crack extension μ under each load cycle is expressed as follows: μ=△S=da / dN Among them, △S is the fatigue strip interval; According to Paris's formula: The relationship between fatigue stress amplitude and fatigue band spacing is obtained as shown in the following formula: Among them, C and m are metal material parameters, Y is the shape factor, and Δσ is the fatigue stress amplitude; (2) Obtain the arc direction of the fatigue strip by using the square difference accumulation method; According to the minimum or maximum value of the image grayscale difference in each direction of the fatigue stripe, the arc direction of the fatigue stripe is obtained; The grayscale differences of the fatigue strip along the four directions of 0°, 45°, 90°, and 135° are set to be d 0° ,d 45° ,d 90° ,d 135° ,in: d 0° (x,y)=[I(x-1,y)-I(x+1,y)] 2 d 45° (x,y)=[I(x-1,y+1)-I(x+1,y-1)] 2 ×0.5 d 90° (x,y)=[I(x,y-1)-I(x,y+1)] 2 d 135° (x,y)=[I(x-1,y-1)-I(x+1,y+1)] 2 ×0.5 Where I(x,y) is the image grayscale of point (x,y) on the fatigue strip diagram; Then, according to the image grayscale differences of the fatigue strip along the four directions of 0°, 45°, 90°, and 135°, the mean square difference of the image grayscale of the fatigue strip along these four directions is calculated, and the mean square difference of the image grayscale in these four directions is the arc direction of the fatigue strip; The step 2 uses the improved RRT algorithm to simulate the extension trajectory of the fatigue strip, and combines the mathematical model of the fracture fatigue strip to obtain the texture curve image of the fatigue strip. The specific method is: Step 2.1: Define the starting point, end point, and number of sampling points of the fatigue strip arc, and set the step length between each sampling point to t; define the starting point and end point of the fatigue strip arc as the root node and target node of the random tree, and define the obstacle position coordinates; Step 2.2: The spacing and curvature of the fatigue strip obtained by the fatigue strip mathematical model are used as constraints and set as obstacles for the RRT algorithm to adjust the extension direction of the next random point of the RRT algorithm; Step 2.3: Use the generated new random point as the parent point, input the constraint parameters, and generate the downward child tree branch; Step 2.4: Add all new random points to the random tree set, and connect all sub-tree branches into a trajectory line to obtain the final fatigue strip texture curve; The fatigue band recognition model in step 5 uses a U-net network structure. The overall structure of the U-net network adopts an encoding-decoding structure, including a downsampling module, an upsampling module, and a jump link, which can realize end-to-end image segmentation; the downsampling module includes four layers of double convolution sub-blocks, and the incoming data is downsampled and extracted features after passing through each double convolution sub-block. The purpose of the downsampling module is to extract feature maps by continuously compressing the input image; the upsampling module is symmetrical in structure to the downsampling module, and also includes four layers of double convolution sub-blocks, and a deconvolution is added after each layer of double convolution sub-block to restore the size of the feature map; the jump link part is to fuse the features obtained after convolution with the features obtained by deconvolution in the upsampling module each time downsampling is prepared, and the features are spliced ​​together in the channel dimension; The U-net network structure uses a multi-layer double convolution structure and multiple downsampling to form a contraction path, continuously downsamples to extract features, expands feature channels, and obtains feature maps; after obtaining the feature map, the multi-layer double convolution structure is used again with multiple upsampling to restore the size of the feature map; The U-net network structure adds a skip link method to connect the feature maps in the first three downsampling layers and upsampling layers, completes the combination of downsampling depth information and upsampling shallow layer information, completes the compensation of image missing information, and restores image pixel information; The U-net network structure adds a softmax function for feature classification after obtaining the feature map and restoring the feature map size.

2. The method for identifying and segmenting fatigue bands on metal material fracture surfaces based on deep learning according to claim 1 is characterized in that: The specific method of step 4 is: Step 4.1: Use Gaussian filtering to perform image denoising to remove random noise from the image in the dataset and reduce noise interference in the image; Step 4.2: Use the nearest neighbor interpolation method to perform geometric transformation of the image to correct the instrument's own error and random error of the instrument's position during fracture image acquisition.

3. The method for identifying and segmenting fatigue bands on metal material fracture surfaces based on deep learning according to claim 2 is characterized in that: The double convolution structure is composed of two convolution kernels and two linear rectification functions. A maximum pooling layer is added after the double convolution structure for downsampling to ensure the accuracy of feature extraction. The deconvolution method is used to upsample the feature map to expand the image information.

4. The method for identifying and segmenting fatigue bands on metal material fracture surfaces based on deep learning according to claim 3 is characterized in that: The fatigue band recognition model uses a cross entropy function as a loss function.