In-situ fatigue microcrack automatic identification and measurement method based on multi-task deep learning framework

Abstract

According to the in-situ fatigue microcrack automatic identification and measurement method based on the multi-task deep learning framework, an in-situ microcrack image is obtained through an in-situ fatigue crack propagation test, a data enhancement strategy is provided, and the problems that the number of data sets is limited and the diversity is insufficient are solved. Secondly, a multi-task sharing mechanism and an attention mechanism are utilized, and key features highly related to the microcrack length and scale information are automatically extracted from fatigue test data; then, a multi-task deep learning model is constructed and trained to realize unified processing of micro-crack detection, scale length information identification and scale segmentation tasks, and the effectiveness of the frame is not limited to the magnification times of the in-situ image; after model training is completed, the length of the microcrack is accurately measured and a crack propagation rate curve is drawn by automatically combining a prediction result of the model, so that automatic detection and accurate measurement of the fatigue crack propagation process are ensured. Experimental results show that the method shows excellent precision and stability in in-situ fatigue microcrack detection and measurement, and the crack recognition efficiency in an in-situ fatigue test is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of fatigue experiment crack detection, and mainly relates to an in-situ fatigue microcrack automatic recognition and measurement method based on a multi-task deep learning framework. Background Art

[0002] Studying the evolution behavior and laws of material fatigue microcracks is very important for both the fatigue performance evaluation and anti-fatigue design of materials. Automatically extracting microcracks from microscopic images is an important link in studying the evolution law of fatigue microcracks. Due to the great diversity of the microstructure of materials, the work of segmenting and measuring microcracks from many microstructures is very complex and requires very advanced recognition and segmentation methods to complete.

[0003] In-situ fatigue test is an important means to study the evolution of fatigue microcracks, which can collect microscopic damage evolution images in real time during the fatigue test. In addition, since many small key stressed components such as screws and fasteners are difficult to be processed into macroscopic standard specimens for crack propagation, their crack propagation mechanisms and behaviors need to be studied by processing them into in-situ fatigue small specimens. In recent years, many scholars have carried out research on fatigue microcracks using in-situ fatigue tests. However, during the in-situ fatigue test process, in order to capture the evolution behavior of fatigue microcracks, researchers need to manually calibrate the crack length for a long time. This process is not only cumbersome and time-consuming, but also prone to introducing human errors, directly affecting the accuracy and consistency of the data.

[0004] In recent years, with the rapid development of artificial intelligence technology, deep learning methods have been widely used in fields such as natural language processing and computer vision. Due to its strong feature learning ability and generalization ability, deep learning methods have also begun to gradually extend and develop to other industrial fields. In the field of fatigue fracture, it is widely used for the automatic recognition of cracks. Through the combination of artificial intelligence and the discipline field of fatigue strength, these works have effectively promoted the development of new directions such as intelligent recognition and intelligent evaluation of fatigue cracks. However, most of these studies are aimed at macroscopic crack recognition, and there are deficiencies in the automatic recognition and measurement of microcracks in in-situ fatigue tests. During the fatigue test process, in order to better capture the growing microcracks, it is usually necessary to adjust the magnification of the microscope. In order to achieve the automatic measurement of microcracks, it is necessary to simultaneously identify and segment the crack pixel length information, scale length text information, and scale pixel length information in the microscopic image. Existing technologies usually adopt single-task deep learning models to solve crack detection and segmentation problems. However, these methods often require multiple independent models to process different tasks. This single-task processing method not only increases the complexity of model training and maintenance, but also due to the lack of information sharing between tasks, it leads to obvious deficiencies in recognition efficiency and accuracy, especially in applications that require simultaneous crack detection, scale pixel segmentation, and text recognition.

[0005] In general, although deep learning technology has shown new potential in the automatic identification and measurement of microcracks in in-situ fatigue tests, how to effectively combine multi-task recognition and segmentation, improve the model's ability to share information between tasks, and improve the generalization and accuracy of the model under different experimental conditions are still key issues that need to be urgently addressed in this field. Summary of the invention

[0006] The purpose of the present invention is to solve the problems of low efficiency, insufficient accuracy and low degree of automation in microcrack identification and measurement in existing in-situ fatigue tests. A method for automatic microcrack identification and measurement based on multi-task deep learning is proposed, which can realize intelligent identification and measurement of fatigue microcracks under electron microscope images of different magnifications.

[0007] The technical solution of the present invention is to provide an in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework, which is characterized by:

[0008] Step 1: Data acquisition: The in-situ fatigue test system is used to obtain micro-crack extension images at different cycles. In the data acquisition process, in-situ images at different magnifications are collected to ensure that the detailed information of the cracks is captured, providing a data basis for subsequent data processing and model training;

[0009] Step 2: Data enhancement processing: Use data enhancement technology to generate diverse data samples to enhance the generalization and robustness of the model. Enhance data diversity through operations such as rotation, flipping, and scaling. Divide these enhanced image data into training sets and test sets for subsequent model training and evaluation;

[0010] Step 3: Multi-task deep learning model construction: By constructing a multi-task deep learning framework MT-CrackNet, automatic crack detection, automatic identification of ruler length information, and ruler pixel segmentation tasks are performed simultaneously;

[0011] Step 4: Model training and optimization: Use the enhanced dataset to train the proposed multi-task deep learning framework, optimize the framework parameters, and adjust and optimize the model;

[0012] Step 5: Automatic identification and measurement of microcrack extension: Through the trained MT-CrackNet model, microcracks are automatically identified and segmented, and crack length and scale information are accurately measured. Using the identified crack length data, the aN curve of fatigue crack extension rate is automatically drawn to achieve intelligent identification and measurement of crack extension during fatigue testing.

[0013] Further, in step 1, when collecting data for the in-situ fatigue crack growth test, experimental specimens of the same size and material are selected to conduct the in-situ fatigue crack growth experiment, and an effective crack data set is collected; the data set is labeled. The labeling of microcracks and scale length values uses a rectangular box form, and the data storage format is a txt file in YOLO format. The pixel-level segmentation task of the scale uses the png format of a black-and-white image.

[0014] Further, in step 2, the data augmentation process includes the following steps:

[0015] Step 2.1: Augmentation of the data set for the detection task of microcrack and scale length information; First, the original image is scaled to a unified pixel size, and a specific area in the image is cropped to cover the area where the scale length information is located, simulating the random changes that numbers and letters may appear under different experimental conditions; Second, the scaled image is randomly cropped into areas of a specific pixel size to increase the diversity of the positions and sizes of the key information in the image; Then, new number and letter labels are generated in the randomly cropped image, and the symbol "μm" is labeled as a separate detection target; Finally, the generated label information is saved to construct a detection data set containing numbers, letters, and cracks.

[0016] Step 2.2: Augmentation of the data set for the pixel-level segmentation task of the scale; For the pixel-level segmentation task of the scale, a similar data augmentation strategy is adopted. First, the original image is scaled to a unified pixel size; Then, randomly crop areas of a specific pixel size and randomly generate scale areas of different lengths on the cropped areas; Finally, generate the labels corresponding to the scale areas to complete the construction of the pixel segmentation data set.

[0017] Further, in step 3, the construction of the multi-task deep learning framework MT-CrackNet includes the following steps:

[0018] Step 3.1: Construct the MT-CrackNet neural network model; The MT-CrackNet neural network model mainly consists of four modules: a backbone network module, a neck module, a detection module, and a segmentation module; The MT-CrackNet neural network model takes the in-situ fatigue crack growth data of the specimen, that is, the in-situ fatigue microcrack image, as input, and the output is the position and true length of the fatigue microcracks in the image.

[0019] Step 3.2: Establishment of the evaluation metrics for the MT-CrackNet neural network model; For the object detection tasks of microcracks and scale length information, precision, recall, and mean average precision are used as evaluation metrics. In the segmentation task, IoU is used as the evaluation metric. Precision and recall can be expressed as:

[0020] P = TP / (TP + FP) (8)

[0021] R = TP / (TP + FN) (9) Where: P represents precision, R represents recall; TP represents true positive, which refers to the class samples correctly identified as crack and scale length information by the model; FP represents false positive, which refers to the class samples mislabeled as crack and scale length information; FN represents false negative, which refers to the class samples of microcracks and scale length information missed by the model. The average precision is the average precision when the overlap rate between the detection box and the true label is above 50%, and it is used to evaluate the robustness of the model when dealing with different complexity and unbalanced data sets.

[0022] In the segmentation task, IoU is used as an evaluation metric, mainly to measure the overlap degree between the model detection result and the actual label. It can be expressed as:

[0023]

[0024] Where: U is the scale segmentation result of the model, is the true label. The larger the IoU value, the higher the overlap degree between the scale region detection result and the true label, and the smaller the detection error.

[0025] Step 3.3: Establishment of the loss function of the MT-CrackNet neural network model; The loss function of the MT-CrackNet model consists of three parts, the microcrack detection loss L det_Crack , the scale length information detection loss L det_Num , and the scale pixel length segmentation loss L seg .

[0026] In the model detection task, the loss function is composed of the sum of the microcrack detection loss and the scale length information detection loss. The specific loss functions mainly include the object confidence loss, the class confidence loss, and the coordinate regression loss. L det_Crack and L det_Num can be expressed as:

[0027] L det_Crack = L obj_Crack + L class_Crack + L box_Crack (11)

[0028] L det_Num = L obj_Num + L class_Num + L box_Num (12)

[0029] Where: L obj_Crack and L obj_Num represent the object confidence loss, L class_Crack and Lclass_Num represents the class confidence loss, L box_Crack and L box_Num represents the coordinate regression loss. The object confidence loss and the class loss use binary cross-entropy loss; while the localization loss uses CIoU loss.

[0030] In the segmentation task of the model, the loss function mainly consists of two parts: cross-entropy loss and Dice loss. The segmentation loss L seg can be expressed as:

[0031] L seg = L ce + L dice (13)

[0032] where: L ce represents the cross-entropy loss, and L dice represents the Dice loss.

[0033] The MT-CrackNet neural network model is trained by minimizing the loss function and updating the parameters by the stochastic gradient descent method.

[0034] Furthermore, the backbone network module is mainly composed of a series of convolutional layers, ELAN layers, and MP layers, which are used to extract features from the input micro-crack images. Among them, a series of convolutional layers are used to extract the features of micro-cracks, scale length information, and scale region from the input pictures. The ELAN layer structure optimizes the processing of long-range dependencies through multiple convolutions and attention mechanisms, significantly reducing the computational complexity and improving the computational efficiency and the ability to extract features of micro-cracks, scale length information, and scale regions. The MP layer in the backbone network module effectively reduces the spatial dimension of the feature maps of micro-cracks, scale length information, and scale regions to enhance the model's ability to capture target features.

[0035] The neck module is mainly composed of a spatial pyramid pooling layer of convolutional sparse coding, a convolutional block attention layer, and multiple basic convolutional layers and upsampling, which are mainly used to further extract the feature information of the image and fuse the features of micro-cracks, scale length information, and scale regions generated by the backbone network. The spatial pyramid pooling layer of convolutional sparse coding combines the spatial pyramid pooling technology and the cross-stage partial connection technology to achieve the efficient extraction of multi-scale features, so as to improve the accuracy and efficiency of micro-crack, scale length information, and scale region detection and segmentation. The attention layer enhances the feature expression ability through the attention mechanisms in the channel and spatial dimensions, improving the adaptive adjustment effect of the feature maps. The feature maps of micro-cracks, scale length information, and scale regions output by the neck module are then used in the detection and segmentation parts of the model.

[0036] The detection module adopts a multi-scale detection scheme. The micro-crack and scale length information features processed by the neck module will be input into decoupled heads with three different resolutions, and each decoupled head consists of two basic convolutional layers. In multi-scale detection, prior anchor points with three different aspect ratios are assigned to the grid points of each feature map. The detection head detects the target position offset and size scaling based on these anchor points, and outputs a tensor containing class detection and bounding boxes to optimize the model performance.

[0037] The segmentation module restores the high-dimensional feature map of the scale region output by the neck module to the original image size through upsampling and convolution operations, and gradually generates the segmentation result of the scale pixel values. The segmentation module splices the feature map of a specific layer in the backbone network module with the corresponding feature map in the segmentation head, integrating low-level spatial details and high-level semantic information to improve the accuracy and detail retention ability of segmentation. By combining feature information of different depths, the proposed model can perform image segmentation more effectively in complex scenarios, providing more detailed and accurate segmentation results.

[0038] Furthermore, in step 4, the model training and optimization include the following steps:

[0039] Step 4.1: Training and optimization of the object detection task for micro-crack and scale length information; First, train the feature extraction module, neck module, and object detection module of the model to optimize the model's feature extraction ability and the performance of the object detection task, so as to capture and recognize the deep information in the image and improve the accuracy of micro-crack and digital letter detection. After the training of the object detection part is completed and stable detection results are obtained, freeze the parameters of the model's feature extraction module and object detection module to retain the trained feature representation. Ensure that these modules will not be disturbed in the subsequent segmentation task training and maintain their superior performance in the object detection task.

[0040] Step 4.2: Training and optimization of the scale segmentation task; After the model training and optimization in step 4.1, freeze the parameters of the model's feature extraction module, neck module, and object detection module, and conduct independent training of the model segmentation module. The training of the segmentation task focuses on the segmentation optimization of the scale pixel length, ensuring that the model can accurately segment the scale region in the image, reducing the conflict between different tasks, and improving the overall performance of the model.

[0041] Furthermore, in step 5, the data is recognized and measured by the MT-CrackNet neural network model trained in step 4, and a variety of deep learning-based crack target detection models and scale segmentation models are introduced for comparative evaluation. The comparison models and the MT-CrackNet neural network model are both trained and parameter-updated with the same dataset, that is, the dataset obtained after data augmentation based on the method in step 2. After the model training is completed, from an original SEM image, the model can obtain the position and pixel length of the microcracks, the actual length and pixel length of the scale. Finally, the actual length of the microcracks can be calculated by formula (14):

[0042]

[0043] Where: S Crack represents the actual length of the fatigue microcrack, S CrackPixel represents the pixel length of the microcrack, S Scale represents the actual length of the scale, S ScalePixel represents the pixel length of the scale.

[0044] After detecting the actual length of the microcracks, by using the cycle number corresponding to the in-situ microcrack image, the a-N curve of the fatigue crack growth rate is automatically drawn to realize the intelligent recognition and measurement of crack growth during the fatigue test.

[0045] The beneficial effects achieved by the present invention are as follows:

[0046] (1) The present invention proposes a multi-task deep learning framework for the automatic recognition and measurement of in-situ fatigue microcracks, which can intelligently realize the intelligent recognition and measurement of microcracks during the in-situ fatigue test. This method does not require manual annotation, greatly reducing the experimental operation time, lowering the operation complexity, ensuring the efficient measurement of the fatigue crack growth rate, and is particularly suitable for the fatigue crack growth test in the SEM in-situ fatigue test system.

[0047] (2) The multi-task deep learning framework proposed by the present invention, through task sharing mechanism, attention mechanism and multi-scale strategy, etc., enhances the model's ability to handle long-distance dependence relationships and the ability to retain detailed information, and demonstrates an efficient detection ability for microcracks under complex backgrounds in the crack recognition task.

[0048] (3) The present invention can simultaneously complete the tasks of microcrack detection, scale information recognition and scale length segmentation, greatly improving the experimental efficiency and the model's multi-task processing ability.

[0049] (4) In the fatigue crack growth experiment, the R between the microcrack length detected by the model of the present invention and the actual annotated length 2The value reaches 0.99, verifying the efficiency and reliability of the model in the automated testing of fatigue crack propagation. This method can achieve the automated measurement of the fatigue crack propagation process and provide accurate crack length measurement data. Description of the Drawings

[0050] Figure 1 It is a schematic diagram of the framework of a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0051] Figure 2 It is a schematic diagram of the in-situ fatigue expansion test piece and test system in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0052] Figure 3 It is a schematic diagram of an example of the test data set in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0053] Figure 4 It is a schematic diagram of the multi-task model framework in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0054] Figure 5 It is a schematic diagram of the model training verification loss in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0055] Figure 6 It is a schematic diagram of the object detection accuracy-recall curve in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0056] Figure 7 It is a schematic diagram of the object detection visualization in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0057] Figure 8 It is a schematic diagram of the visualization of different model segmentation results in the test area of the object detection model in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0058] Figure 9 It is a schematic diagram of the post-processing of the model segmentation results in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0059] Figure 10 It is a schematic diagram of the comparison of the IoU values of different segmentation models in a multi-task deep learning method for in-situ fatigue microcrack automatic identification and measurement;

[0060] Figure 11Schematic diagram of visualizing the segmentation results of different models in the segmentation model test area of a multi-task deep learning method for in-situ fatigue microcrack automatic recognition and measurement;

[0061] Figure 12 Schematic diagram of comparing the model detection results with the ground truth in a multi-task deep learning method for in-situ fatigue microcrack automatic recognition and measurement; Detailed implementation manners

[0062] The technical solutions of the present invention will be described in more detail below with reference to the accompanying drawings. The present invention includes but is not limited to the following embodiments.

[0063] Figure 1 As shown, this embodiment provides a multi-task deep learning framework for in-situ fatigue microcrack automatic recognition and measurement. The detection framework includes the following steps:

[0064] Step 1, data acquisition: Obtain microcrack propagation images at different cycle numbers through an in-situ fatigue test system. During the data acquisition process, in-situ images at different magnifications are collected to ensure capturing the detailed information of the cracks, providing a data basis for subsequent data processing and model training.

[0065] In this embodiment, the specimen size in step 1 is as shown in Figure 3 (a). The thickness is 1 mm, and the material used is 316L steel. The specific parameters of the material tensile test are shown in Table 1. Before the test, the surface of the specimen is polished successively with silicon carbide sandpapers with particle sizes of 500, 800, 1000, 2000, and 3000, and then finely polished with 3μm, 0.5μm diamond suspensions and 0.02μm silica suspensions. The prefabricated notch length is 0.4 mm. Before the in-situ fatigue crack propagation test, a crack is prefabricated by applying a force equal to 100% of the material yield strength, and then the force is adjusted to 80% of the yield strength for crack propagation operation.

[0066] The test uses an in-situ fatigue test system SEM-SERVO PULSER, as shown in Figure 3 (b). This system integrates a servo-hydraulic fatigue test system and a scanning electron microscope (SEM), enabling high-resolution SEM observations of the specimen during uniaxial or cyclic loading. Through three experiments, a total of 341 effective test images are collected, which record the process of fatigue microcracks initiating from the prefabricated notch of the specimen and gradually expanding until the specimen reaches the fatigue life limit. To verify the generalization performance of the subsequent model, 305 images from the first and second test images are used to train the proposed MT-CrackNet model, and 36 images from the third test images are used to evaluate the performance of the model on unseen data.

[0067] Step 2. Data augmentation: Use data augmentation techniques to generate diverse data samples, enhancing the generalization ability and robustness of the model. Enhance data diversity through operations such as rotation, flipping, and scaling. Divide these augmented image data into a training set and a test set for subsequent model training and evaluation.

[0068] Step 2.1: During the acquisition of SEM images, due to the requirements of high-resolution imaging and the complexity of operations, large-scale acquisition of image resources is both time-consuming and costly. Therefore, the number of original test images is usually limited.

[0069] To effectively overcome this limitation and improve the generalization ability of the model in multiple tasks such as the detection of microcracks and scale length information, this example designs different data augmentation strategies for the detection tasks of microcracks and scale length information and the scale segmentation task respectively.

[0070] For the detection tasks of microcracks and scale length information, in the data augmentation process, this example randomly adjusts the positions of these numbers and letters to simulate diverse scenarios that may occur under different experimental conditions, enabling the model to identify these key information from multiple angles and different scenarios without relying on the fixed positions of the numbers and letters. The specific dataset augmentation strategy is as follows: First, uniformly scale the original images to 2560×1920 pixels, and crop specific regions of the images to cover the scale length information regions of the original images. In SEM images, "scale length information" usually refers to the information text indicating the scale length dimension, such as "100μm", and in this example, this type of information is simply referred to as "numbers and letters". Second, randomly crop the scaled images to 500×500 pixels; Third, randomly generate numbers and letters on the randomly cropped images and the covered original images to enhance the diversity of the target positions and sizes and improve the generalization ability of the model; Finally, generate the labels for the numbers, letters, and cracks. Among them, the "μm" symbol is used as a separate type of detection target for training to better complete the recognition of scale length information.

[0071] During the data augmentation process, the crack regions and scale information are scaled proportionally to ensure the consistency of crack length measurement between the augmented images and the original images. Since image scaling only involves resolution adjustment and does not affect the physical size of the cracks, the relative size of the cracks does not change, thus avoiding length errors introduced by scaling. Through data augmentation operations, the object detection dataset is expanded from 305 images to 1460 images, including 305 images of 2560×1920 pixels and 1119 images of 500×500 pixels. The new dataset is defined as "Dataset A".

[0072] Step 2.2: For the task of pixel-level segmentation of the scale, in this example, data augmentation operations are performed on the original test images, and the specific augmentation strategy is as follows. First, the original images are uniformly scaled to 2560×1920 pixels; second, the images are randomly cropped to 1280×960 pixels; third, multiple scale regions with random lengths are generated at random positions in each cropped image; finally, the labels of the scale regions are generated. After data augmentation, the pixel segmentation dataset is expanded from 305 images to 978 images, and the new dataset is defined as "Dataset B".

[0073] An example of the dataset after data augmentation is shown in the appendix Figure 3 as follows. After the above data augmentation and annotation, the types and quantities of different datasets are shown in Table 2, and the training set, validation set, and test set are divided according to 8:1:1.

[0074] Table 1 Types and Quantities of Test Datasets

[0075]

[0076] Step 3: Construction of a multi-task deep learning model: By constructing a multi-task deep learning framework MT-CrackNet, the tasks of automatic crack detection, automatic recognition of scale length information, and scale pixel segmentation are performed simultaneously;

[0077] Step 3.1: As shown in the appendix Figure 4 MT-CrackNet mainly consists of four modules: a backbone network module, a neck module, a detection module, and a segmentation module. Different from the traditional single-task processing framework, the core of the MT-CrackNet model designed in this example is that the model can effectively process the micro-crack detection task, scale length information recognition task, and scale segmentation task with a unified architecture through the mutual coordination and integration of each module in the network, thereby effectively improving the efficiency and accuracy of data processing in in-situ fatigue tests.

[0078] The backbone network module mainly consists of a series of convolutional layers, ELAN layers, and MP layers, which are used to extract features from the input micro-crack images. Among them, a series of convolutional layers are used to extract the feature information of micro-cracks, scale length information, and scale regions from the input pictures. Its basic convolutional module integrates batch normalization, SiLU activation function, and three types of convolutional layers, namely 1×1 convolution, 3×1 convolution, and 3×2 convolution. The ELAN layer optimizes the processing of long-range dependencies through multiple convolutions and attention mechanisms, significantly reducing the computational complexity and improving the computational efficiency and the ability to extract features of micro-cracks, scale length information, and scale regions. In addition, the MP layer in the backbone network module effectively reduces the spatial dimension of the feature maps of micro-cracks, scale length information, and scale regions, enhancing the model's ability to capture target features.

[0079] The neck module is used to further extract the feature information of the image and fuse the features of micro-cracks, scale length information, and scale area generated by the backbone network. The neck module of this example mainly consists of a spatial pyramid pooling layer of convolutional sparse coding, a convolutional block attention module, and multiple basic convolutional layers and upsampling. The spatial pyramid pooling layer of convolutional sparse coding realizes the efficient extraction of multi-scale features by combining spatial pyramid pooling technology and cross-stage partial connection technology, improving the accuracy and efficiency of micro-crack, scale length information, and scale area detection and segmentation. The convolutional block attention module enhances the feature expression ability through the attention mechanism in the channel and spatial dimensions, improving the adaptive adjustment effect of the feature map. The feature maps of micro-cracks, scale length information, and scale area output by the neck module are then used for the detection and segmentation parts of the model.

[0080] To more accurately identify fatigue micro-cracks of different lengths, the detection module of the MT-CrackNet model adopts a multi-scale detection scheme. The micro-crack and scale length information features processed by the neck module are input into decoupled heads with three different resolutions, and each decoupled head consists of two basic convolutional layers. This structure improves the model's processing ability for targets of different scales. In multi-scale detection, three prior anchor points with different aspect ratios are assigned to each grid point of the feature map. The sizes of the prior anchor points in the MT-CrackNet model are 40×40×69, 80×80×69, and 160×160×69 to more accurately detect small cracks. The detection head detects the target position offset and size scaling based on these anchor points and outputs a tensor containing class detection and bounding boxes to optimize the model performance.

[0081] The segmentation module restores the high-dimensional feature map of the scale area output by the neck module to the original image size through upsampling and convolutional operations and gradually generates the segmentation result of the scale pixel values. To utilize the feature information of different depths, the segmentation module adopts a feature skip design, splicing the feature maps of specific layers in the backbone network module with the corresponding feature maps in the segmentation head, fusing the low-level spatial details and high-level semantic information, and improving the accuracy and detail retention ability of segmentation. This design can effectively improve the segmentation performance of the MT-CrackNet model, enhance the ability to capture scale pixel value information, and ensure the accuracy and integrity of the segmentation result. By combining the feature information of different depths, the model proposed in this example can perform image segmentation more effectively in complex scenarios and provide more detailed and accurate segmentation results.

[0082] Step 3.2: To more effectively evaluate the performance of the proposed model, this example uses multiple performance metrics. For the micro-crack and scale length information object detection tasks, accuracy, recall, and average precision are used as evaluation metrics. In the segmentation task, IoU is used as the evaluation metric. Accuracy and recall can be expressed as:

[0083] P = TP / (TP + FP) (15)

[0084] R = TP / (TP + FN) (16) Where: P represents precision, R represents recall; TP represents true positive, which refers to the category samples correctly identified as crack and scale length information by the model; FP represents false positive, which refers to the category samples mislabeled as crack and scale length information; FN represents false negative, which refers to the category samples of microcracks and scale length information missed by the model.

[0085] The average precision is the average precision when the overlap rate between the detection box and the true label is above 50%, and it is used to evaluate the robustness of the model when dealing with different complexity and unbalanced data sets. In the segmentation task, IoU is used as an evaluation metric, mainly to measure the overlap degree between the model detection result and the actual label. It can be expressed as:

[0086]

[0087] Where: U is the scale segmentation result of the model, is the true label. The larger the IoU value, the higher the overlap degree between the scale region detection result and the true label, and the smaller the detection error.

[0088] Step 3.3: To obtain a suitable model, in this example, a loss function is constructed to measure the difference between the model detection result and the true label. The loss function of the MT-CrackNet model consists of three parts, L det_Crack represents the microcrack detection loss, L det_Num represents the scale length information detection loss, L seg represents the scale pixel length segmentation loss.

[0089] In the detection task of this example, the loss function is composed of the sum of the microcrack detection loss and the scale length information detection loss. The specific loss function mainly includes the object confidence loss, the class confidence loss, and the coordinate regression loss. Therefore, L det_Crack and L det_Num can be expressed as:

[0090] L det_Crack = L obj_Crack + L class_Crack + L box_Crack (18)

[0091] L det_Num = L obj_Num + L class_Num + L box_Num (19)

[0092] In the formula, Lobj_Crack and L obj_Num represents the object confidence loss, L class_Crack and L class_Num represents the class confidence loss, L box_Crack and L box_Num represents the coordinate regression loss. Among them, the object confidence loss and the class loss use binary cross-entropy loss, which helps the model better distinguish crack, scale length information and background during the training process; while the localization loss uses CIoU loss, aiming to optimize the position, size, overlap and aspect ratio of the detection box simultaneously.

[0093] For the segmentation task of this instance, in order to balance the detection accuracy of the scale area and the quality of the segmentation result, the loss function mainly consists of two parts: cross-entropy loss and Dice loss. Therefore, the segmentation loss L seg can be expressed as:

[0094] L seg = L ce + L dice (20)

[0095] In the formula, L ce represents the cross-entropy loss, L dice represents the Dice loss. In the image segmentation task, the cross-entropy loss is widely used to measure the difference between the model detection and the true label, and the performance of the model is evaluated by calculating the difference between the detection probability distribution and the actual distribution of each pixel. The Dice loss pays more attention to the overlap of the target area, so it performs better in dealing with segmentation tasks with a large difference in the number of foreground and background pixels. The smaller its value, the higher the overlap degree between the detected scale area segmentation result and the actual segmentation result.

[0096] Step 4, Model training and optimization: Use the enhanced dataset to train the proposed multi-task deep learning framework, optimize the framework parameters, and adjust and optimize the model;

[0097] Step 4.1: The MT-CrackNet model proposed in this instance includes the object detection tasks of micro-cracks and scale length information, as well as the segmentation task of the scale pixel length. The model is constructed in a way of sharing weight parameters. Specifically, the training process of the model adopts a staged training method, and the training strategy is as follows: First, train the feature extraction part and the object detection part of the model. By optimizing the feature extraction ability of the model, the model can better capture and identify the deep information in the image, and improve the accuracy of micro-crack and digital letter detection.

[0098] Step 4.2: After the training of the target detection part is completed, freeze the parameters of the feature extraction module (Backbone), the neck module, and the target detection module (Detection Head) of the MT-CrackNet model, and train the segmentation task part of the model separately, so as to avoid conflicts between different tasks during training and improve the overall performance of the model.

[0099] During the model training process, the batch-size is set to 16. The model parameters are updated using the stochastic gradient descent method. The initial learning rate for training is set to 0.01, the weight decay is 0.0005, and the momentum is 0.937. The learning rate is adjusted using the cosine annealing strategy. The detailed parameters of each module of the model are shown in the table. The number of training epochs in the target detection training stage of the model is set to 200, and the number of training epochs in the segmentation training stage is set to 20.

[0100] The loss change during the training of the model detection part is as shown in the appendix Figure 5 (a). It can be seen that the loss of the model gradually decreases as the number of training epochs increases, and the average precision gradually increases, indicating that the model gradually learns and improves its target detection performance for microcracks and scale length information during training. Subsequently, in the later stage of training (about 40 training epochs), the loss value tends to level off and finally stabilizes near 0.001; the average precision gradually increases and finally stabilizes at about 98.9%, indicating that the model has reached a certain stability. The loss of the model segmentation module training is as shown in the appendix Figure 5 (b). The loss rapidly drops from 1.1 to about 0.2 within 5 training epochs, and then gradually decreases and finally stabilizes near 0.1. This indicates that the model gradually learns effective features during training and continuously optimizes its segmentation ability, resulting in a continuous reduction in loss. In the training stage of the model segmentation task, since the model has learned the deep feature information in the image through the target detection task, the segmentation task can achieve significant results with only a small amount of training (at the 3rd training epoch, the IoU rapidly rises from 0.1 to 0.99).

[0101] This embodiment effectively improves the quality of model feature sharing and training stability by adopting a phased training strategy, reduces the competition and mutual interference between tasks, and accelerates the convergence speed and detection performance of the model.

[0102] Step 5: Automatic identification and measurement of microcrack propagation: Automatically identify and segment microcracks through the trained MT-CrackNet model, and accurately measure the crack length and scale information. Using the identified crack length data, automatically draw the a-N curve of the fatigue crack propagation rate to achieve intelligent identification and measurement of crack propagation during the fatigue test.

[0103] The performance evaluation of the MT-CrackNet model's detection part under different categories, including accuracy, recall, and average precision, is shown in Table 3.

[0104] Among them, in terms of digital detection, the model accuracy reaches 0.993, the recall rate is 0.991, and the average precision value is 0.992, indicating that there are almost no false positives and false negatives when the model identifies numbers. In terms of letter detection, the model accuracy reaches 0.997, the recall rate is 0.992, and the average precision value reaches 0.994, indicating the high accuracy and stability of the model in letter recognition. In terms of microcrack detection, the model detection accuracy is 0.985, and the recall rate is slightly reduced to 0.842. However, the average precision value of microcrack detection is still relatively high, at 0.959, showing the model's efficient recognition ability for microcracks in complex backgrounds. Generally speaking, the model demonstrates good generalization performance in various detection tasks.

[0105] Table 2 Target Detection Evaluation Metrics

[0106]

[0107] In actual detection, it is desired that the model can accurately detect fatigue microcracks and at the same time capture all fatigue microcracks as much as possible. As a comprehensive evaluation tool, the precision-recall curve can consider both precision and recall, providing a more comprehensive perspective on performance evaluation. The precision-recall curve of the detection model is as shown in the appendix Figure 6 As shown, the abscissa of this curve is the recall rate, and the ordinate is the precision. The closer the precision-recall curve is to the upper right corner, the more it means that the model can maintain high precision while achieving a high recall rate. In this example, the detection curves of numbers and letters are close to the upper right corner, indicating that the MT-CrackNet model has extremely accurate recognition ability for these two types of targets. Due to the influence of the scale and morphology diversity of fatigue microcracks, its detection curve is slightly lower than the other two types. Generally speaking, all categories can achieve a high recall rate of over 0.9 while having a high precision rate (above 0.9), showing the excellent performance of the model in overall comprehensive detection performance.

[0108] The visualization results of the MT-CrackNet model's target detection results are as shown in the appendix Figure 7As shown in the figure. The process of obtaining the actual length of the scale is as follows: First, identify and locate the "μm" symbol; then, identify the numbers in the area in front of it; finally, output the reading of the actual length of the scale according to the coordinate positions of the numbers. Therefore, whether the "μm" category and its position can be correctly identified is crucial for the subsequent identification of the actual length of the scale. It can be seen from the visualization results that the MT-CrackNet model proposed in the present invention can accurately identify each category of numbers and letters, thereby ensuring the accurate identification of the actual length of the scale. In addition, since the detection part of the model adopts a multi-scale decoupled head design, the ability to identify microcracks of different scales is enhanced. This design enables the model to accurately identify and locate microcracks of various sizes at different test stages and accurately obtain the numerical value of the pixel length of the microcracks. It can be seen from the figure that whether it is small cracks or long cracks, the detection results of the MT-CrackNet model have a small difference from the true values, and the average relative error is within 5%, showing good generalization performance.

[0109] The present invention introduces several classical deep learning-based object detection models for comparison, including the Faster RCNN model and the YOLOV7 model. These models and the MT-CrackNet model proposed in this example will be trained and tested on the same dataset A, and accuracy, recall, and mean average precision are used as the evaluation metrics for the models.

[0110] Table 3 shows the comparison of the results of the object detection models. It can be seen from the comparison that the detection part of the MT-CrackNet model proposed in this example has obvious advantages in terms of accuracy, recall, or mean average precision. It shows that the proposed MT-CrackNet model can not only accurately identify and measure the pixel length of microcracks, but also accurately distinguish various numbers and letters, so as to obtain the actual length of the scale.

[0111]

[0112]

[0113] Table 3 Comparison of the Results of Object Detection Models

[0114]

[0115] The detection results and visualization of the object detection part of the model in the test area are shown in the appendix Figure 8 ​As shown in the figure. Since the detection task in this example mainly focuses on the accurate identification of fatigue microcracks and the actual length of the scale, the result visualization only shows the detection results of these two parts. It can be seen from the visualization results that due to the use of decoupled heads at three different scales in the detection part, the object detection part of the MT-CrackNet model in this example shows better recognition results and the best stability compared with other models in the test area results. Even in complex backgrounds, it can still accurately identify crack and scale length information and has good generalization performance. In contrast, the generalization ability of the remaining models is relatively low, and they cannot accurately identify microcracks and numbers well.

[0116] The output result of the segmentation part of the MT-CrackNet model is as shown in the appendix Figure 9 (b). It can be observed that there are some false positive regions in the segmentation result, that is, the background information is wrongly segmented into the scale region. This is because the shape and color of the numbers and letters in the original microscopic image Figure 9 (a) and the complex background information of the image will affect the model recognition result. Therefore, in this example, morphological erosion and dilation operations are introduced for post-processing the results to eliminate the false positive regions in the image and improve the segmentation effect. Morphological erosion is a technique that erodes the edges of an image, removing small-size noise by eliminating small pixels at the edges, thereby clearing false positive segmentation regions. Correspondingly, morphological dilation fills in breaks and gaps in the segmentation result by expanding the edges of the object, promoting the closure of the region.

[0117] The image after post-processing operation is as shown in the appendix Figure 9 (c). The erosion operation successfully removes most of the small false positive regions caused by model misjudgment, and the dilation process effectively connects those segmentation lines that should be continuous but are broken due to model limitations, thus ensuring the high accuracy and reliability of the segmentation result. As shown in the appendix Figure 9 (d) is the detection visualization diagram of the MT-CrackNet model in the scale pixel segmentation task. It can be seen from the figure that the model can accurately segment scale regions of different scales, thereby obtaining the scale pixel length. And the average IoU of the detection part also reaches 98.6%, indicating the effectiveness of the model segmentation part.

[0118] This invention introduces several deep learning-based segmentation models for comparison, including the classic U-Net model and U2-Net model, the PSPnet model, TwinLiteNet model, and Fast SCNN model for lane line segmentation. These models and the model proposed in this example will be trained and tested on the same dataset B, and IoU is used as the evaluation metric for the models. Figure 10The IoU comparison of the pixel segmentation model results is shown. Through the comparison, it can be directly seen that the MT-CrackNet model has a relatively obvious advantage in the pixel segmentation results. The IoU value can reach 98.6%, and it can accurately segment the length of the pixel area to obtain the pixel length value of the scale.

[0119] The detection results and visualization of the model segmentation part in the test area are as attached Figure 11 shown. From the visualization results, it can be seen that due to the effective feature extraction of the input image and the adoption of the skip design idea, the MT-CrackNet model shows better generalization ability in the results of the test area compared with other models. It can accurately identify the scale area and correctly distinguish it from the digital and letter areas and interference information. In contrast, the generalization ability of the remaining models is relatively low, and they cannot accurately segment the scale area.

[0120] These data prove that the MT-CrackNet model framework proposed in the present invention enhances the model's ability to process long-distance dependencies and retain detailed information through task sharing, attention mechanism, and multi-scale strategy, etc. It can accurately identify the microcrack length in the In-situ fatigue microcrack image and perform well in generalization performance. This result not only verifies the innovation of this embodiment in technology but also demonstrates its significant advantages in practical applications.

[0121] The fatigue crack growth rate da / dN is the change rate of the crack length a with respect to the number of cycles N under the action of fatigue loads, which can reflect the speed of crack growth. When conducting the fatigue microcrack growth experiment, the MT-CrackNet model proposed in the present invention is used to detect and measure the microcrack images at different numbers of cycles, and the actual length of the microcrack, that is, the a value, is calculated according to formula (14). As attached Figure 12 shown, which shows the comparison results between the microcrack length detected by the model and the actual microcrack length, where Figure 12 (a) shows the data results of the first two in-situ fatigue microcrack growth experiments, Figure 12 (c) shows the data results of the third experiment, which are data not seen by the model, so as to evaluate the generalization performance of the model. From Figure 12 (a) and (c), it can be directly seen that whether on the known dataset or the unknown dataset, the difference between the microcrack length detected by the model and the actually labeled microcrack length is relatively small, thus verifying the good generalization performance of the model framework proposed in the present invention.

[0122] And Figure 12 (b) and Figure 12(d) shows the regression curve analysis between the actual length and the model-detected length of microcracks in three tests. This figure shows a high consistency between the microcrack lengths detected by the model and the actual microcrack lengths, where the R2 value reaches 0.99, indicating that the model has extremely high detection accuracy and reflects the strong robustness of the model in the task of fatigue microcrack length detection.

[0123] Through the method proposed above, the in-situ automatic recognition and measurement of fatigue microcracks can be achieved. This method extracts and optimizes features through task sharing, attention mechanism and multi-scale strategy, and can capture the global dependence and key detail information of microcracks under electron microscope images with different magnifications, ensuring the accuracy and stability of microcrack recognition. At the same time, this method does not require manual annotation after scanning the pictures, greatly saving the test operation time and reducing the operation difficulty. Therefore, when conducting in-situ fatigue crack propagation tests on the in-situ fatigue test system, the fatigue microcrack propagation rate can be measured intelligently. Compared with traditional crack recognition techniques, the present invention has significantly improved in terms of crack recognition accuracy, efficiency and model generalization ability, and is particularly suitable for fatigue crack propagation tests under complex backgrounds.

[0124] The present invention is not limited to the above specific embodiments. Those of ordinary skill in the art can implement the present invention in many other specific embodiments according to the embodiments and the disclosed content of the drawings. Therefore, any design that adopts the design structure and idea of the present invention and makes some simple transformations or changes falls within the protection scope of the present invention.

Claims

1. An in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework, characterized by: Step 1: Data acquisition: The in-situ fatigue test system is used to obtain micro-crack extension images at different cycles. In the data acquisition process, in-situ images at different magnifications are collected to ensure that the detailed information of the cracks is captured, providing a data basis for subsequent data processing and model training; Step 2: Data enhancement processing: Use data enhancement technology to generate diverse data samples to enhance the generalization and robustness of the model. Divide these enhanced image data into training sets and test sets for subsequent model training and evaluation; Step 3: Multi-task deep learning model construction: By constructing a multi-task deep learning framework MT-CrackNet, the automatic crack detection, automatic identification of ruler length information, and ruler pixel segmentation tasks are completed simultaneously; Step 4: Model training and optimization: Use the enhanced dataset to train the proposed multi-task deep learning framework, optimize the framework parameters, and adjust and optimize the model; Step 5: Automatic identification and measurement of microcrack extension: Through the trained MT-CrackNet model, microcracks are automatically identified and segmented, and crack length and scale information are accurately measured. Using the identified crack length data, the aN curve of fatigue crack extension rate is automatically drawn to achieve intelligent identification and measurement of crack extension during fatigue testing.

2. The in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework according to claim 1 is characterized in that: In step 1, when collecting data for in-situ fatigue crack growth experiments, experimental specimens of uniform size and material are selected to conduct in-situ fatigue crack growth experiments and collect valid microcrack data sets; and the data sets are labeled.

3. The in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework according to claim 1 is characterized in that: In step 2, the data enhancement process includes the following steps: Step 2.1: Enhance the dataset of the detection task of microcracks and scale length information. The original images are uniformly scaled and cropped, and new numeric and letter labels are generated in the cropped images. Step 2.2: Enhance the dataset for the ruler pixel-level segmentation task. The original image is uniformly scaled and cropped, and ruler regions of different lengths are randomly generated on the cropped image.

4. The in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework according to claim 1 is characterized in that: In step 3, the construction of the multi-task deep learning framework MT-CrackNet includes the following steps: Step 3.1: Construct the MT-CrackNet neural network model; the MT-CrackNet neural network model mainly consists of four modules: backbone network module, neck module, detection module and segmentation module; the MT-CrackNet neural network model takes the in-situ fatigue crack growth data of the specimen, that is, the in-situ fatigue microcrack image as input, and outputs the position and true length of the fatigue microcrack in the image; Step 3.2: Establishment of evaluation indicators for the MT-CrackNet neural network model; for the microcrack and scale length information target detection tasks, accuracy, recall, and average precision are used as evaluation indicators. In the segmentation task, IoU is used as the evaluation indicator. Accuracy and recall can be expressed as: P=TP / (TP+FP) (1) R=TP / (TP+FN) (2) Among them: P represents precision, R represents recall; TP represents true positive examples, which refers to the category samples correctly identified by the model as cracks and ruler length information; FP represents false positive examples, which refers to the category samples that are incorrectly labeled as cracks and ruler length information; FN represents false negative examples, which refers to the category samples of microcracks and ruler length information that are missed by the model. The average precision is the average precision when the overlap rate between the detection box and the true label is above 50%. It evaluates the robustness of the model when dealing with different complexities and unbalanced data sets. In the segmentation task, IoU is used as an evaluation indicator to measure the degree of overlap between the model detection results and the actual labels. It can be expressed as: Among them: U is the scale segmentation result of the model, is the true label. The larger the IoU value, the higher the overlap between the detection result of the scale area and the true label, and the smaller the detection error. Step 3.3: Establishment of the loss function of the MT-CrackNet neural network model; The loss function of the MT-CrackNet model consists of three parts: microcrack detection loss L det_Crack , scale length information detection loss L det_Num , scale pixel length segmentation loss L seg . In the model detection task, the loss function is composed of the sum of microcrack detection loss and scale length information detection loss, where the specific loss function mainly includes target confidence loss, category confidence loss and coordinate regression loss. det_Crack and L det_Num It can be expressed as: L det_Crack =L obj_Crack +L class_Crack +L box_Crack (4) L det_Num =L obj_Num +L class_Num +L box_Num (5) Where: L obj_Crack and L obj_Num represents the target confidence loss, L class_Crack and L class_Num represents the category confidence loss, L box_Crack and L box_Num Represents the coordinate regression loss. The target confidence loss and category loss use binary cross entropy loss; while the positioning loss uses CIoU loss. In the segmentation task of the model, the loss function mainly consists of two parts: cross entropy loss and Dice loss. seg It can be expressed as: L seg =L ce +L dice (6) Where: L ce represents the cross entropy loss, L dice Denotes Dice loss.

5. The in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework according to claim 3 is characterized in that: The backbone network module is mainly composed of a series of convolutional layers, ELAN layers and MP layers, which are used to extract features from the input microcrack image. Among them, a series of convolutional layers are used to extract feature information of microcracks, scale length information and scale area from the input image. The ELAN layer structure optimizes the processing of long-distance dependencies through multiple convolutions and attention mechanisms, significantly reducing the computational complexity, and improving the computational efficiency and feature extraction capabilities of microcracks, scale length information and scale area. The MP layer in the backbone network module effectively reduces the spatial dimensions of the feature maps of microcracks, scale length information and scale area to enhance the model's ability to capture target features. The neck module is mainly composed of a spatial pyramid pooling layer of convolutional sparse coding, a convolutional block attention layer, and multiple basic convolutional layers and upsampling. It is mainly used to further extract the feature information of the image and fuse the features of microcracks, scale length information, and scale area generated by the backbone network. The spatial pyramid pooling layer of convolutional sparse coding combines spatial pyramid pooling technology and cross-stage partial connection technology to achieve efficient extraction of multi-scale features to improve the accuracy and efficiency of microcrack, scale length information, and scale area detection and segmentation. The convolutional block attention layer enhances the feature expression ability through the attention mechanism of the channel and spatial dimensions, and improves the adaptive adjustment effect of the feature map. The microcrack, scale length information, and scale area feature map output by the neck module are then used in the detection and segmentation parts of the model. The detection module adopts a multi-scale detection scheme. The microcrack and scale length information features processed by the neck module will be input into three decoupling heads with different resolutions, each of which consists of two basic convolutional layers. In multi-scale detection, three prior anchor points with different aspect ratios are assigned to the grid points of each feature map. The detection head detects the target position offset and size scaling based on these anchor points, and outputs a tensor containing category detection and bounding box to optimize model performance. The segmentation module restores the high-dimensional feature map of the scale area output by the neck module to the original image size through upsampling and convolution operations, and gradually generates the segmentation results of the scale pixel values. The segmentation module splices the feature map of a specific layer in the backbone network module with the corresponding feature map in the segmentation head, integrating low-level spatial details and high-level semantic information to improve the accuracy of segmentation and the ability to retain details. By combining feature information at different depths, the proposed model can perform image segmentation more effectively in complex scenes and provide more detailed and accurate segmentation results.

6. The in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework according to claim 1 is characterized in that: In step 4, model training and optimization includes the following steps: Step 4.1: Optimization of target detection task training for microcracks and scale length information; First, train the model’s feature extraction module (backbone network module), neck module, and target detection module (detection module). After the target detection part is trained and stable detection results are obtained, freeze the parameters of the model’s feature extraction module and target detection module to retain the trained feature representation. Step 4.2: Scale segmentation task training optimization; After the model training optimization in step 4.1, freeze the parameters of the model's feature extraction module (backbone network module), neck module, and target detection module (detection module), and perform independent training of the model segmentation module (segmentation module).

7. The in-situ fatigue microcrack automatic identification and measurement method based on a multi-task deep learning framework according to claim 1 is characterized in that: In step 5, the data is identified and measured based on the MT-CrackNet neural network model trained in step 4, and a variety of crack target detection models based on deep learning and the scale segmentation model are introduced for comparative evaluation. The comparative model and the MT-CrackNet neural network model are trained and parameter updated using the same data set, i.e., the data set obtained after data enhancement based on the method in step 2. After the model training is completed, the model can obtain the location and pixel length of the microcrack, the actual length of the scale and the pixel length from an original SEM image, and finally the actual length of the microcrack can be calculated using formula (7): Where: S Crack Indicates the actual length of fatigue microcrack, S CrackPixel represents the pixel length of microcrack, S Scale Indicates the actual length of the ruler, S ScalePixel Indicates the ruler pixel length. After detecting the actual length of the microcrack, the aN curve of the fatigue crack growth rate is automatically drawn by using the cycle number corresponding to the in-situ microcrack image, thereby realizing intelligent identification and measurement of crack growth during fatigue testing.

Images