Intelligent detection method and system for equipment fatigue cracks based on deep learning

Abstract

This invention provides a deep learning-based intelligent method and system for detecting fatigue cracks in equipment. First, a set of images of the equipment surface is acquired. A first crack detection heatmap is generated using a pre-trained deep learning model. Next, a physical crack propagation constraint model is constructed based on the equipment's design material parameters and real-time operating conditions. Then, based on the first crack detection heatmap and the constructed physical crack propagation constraint model, iterative optimization of the crack region is performed to obtain a second crack detection heatmap. The second crack detection heatmap is then analyzed to determine the crack state parameters on the equipment surface, such as the crack region boundary coordinates, crack propagation direction vector, and crack depth gradient value. Finally, an intelligent fatigue crack detection report is generated, including a schematic diagram of the crack spatial distribution and a crack propagation risk level indicator. This comprehensive approach, combining image data and physical constraints, improves the accuracy and reliability of crack detection.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, and more specifically, to a method and system for intelligent detection of fatigue cracks in equipment based on deep learning. Background Technology

[0002] In the field of equipment manufacturing and maintenance, fatigue crack detection is a crucial step in ensuring the safe and reliable operation of equipment. Traditional methods for fatigue crack detection mainly rely on manual visual inspection and non-destructive testing techniques (such as magnetic particle testing and ultrasonic testing). Manual visual inspection is greatly affected by factors such as the experience and eyesight of the inspectors and the inspection environment, making it prone to missed or false detections, and it is difficult to effectively detect cracks in some hidden areas. While non-destructive testing techniques improve the accuracy of detection to some extent, they usually require specialized equipment and technicians, the inspection process is relatively cumbersome, the inspection efficiency is low, and it is difficult to guarantee the comprehensiveness and accuracy of inspections on complex equipment surfaces.

[0003] With the development of deep learning technology, some deep learning-based image recognition methods have begun to be applied in the field of crack detection. However, most existing deep learning-based crack detection methods rely solely on the image data itself for crack identification, without fully considering the physical properties of equipment materials and the impact of actual operating conditions on crack propagation. This can lead to detection results that do not accurately reflect the actual situation, making it difficult to predict crack propagation trends and risk levels, and thus failing to meet the actual needs of safe equipment operation and maintenance. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a deep learning-based intelligent detection method for equipment fatigue cracks, the method comprising:

[0005] A set of equipment surface detection images is acquired, and a pre-trained deep learning model is called to predict crack regions in the set of equipment surface detection images to generate a first crack detection heat map.

[0006] A crack physical propagation constraint model is constructed based on the material parameters of the equipment design and the real-time operating condition parameters. The crack physical propagation constraint model includes material fatigue limit parameters, stress distribution coefficients, and propagation resistance thresholds.

[0007] Based on the first crack detection heatmap and the crack physical propagation constraint model, iterative optimization processing of the crack region is performed to obtain the second crack detection heatmap.

[0008] The crack state parameters on the equipment surface are determined by analyzing the second crack detection thermogram. The crack state parameters on the equipment surface include the boundary coordinates of the crack region, the crack propagation direction vector, and the crack depth gradient value.

[0009] Based on the surface crack state parameters of the equipment, an intelligent fatigue crack detection report is generated. The intelligent fatigue crack detection report includes a spatial distribution diagram of cracks and a crack propagation risk level indicator.

[0010] In another aspect, embodiments of the present invention also provide a deep learning-based intelligent detection system for equipment fatigue cracks, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention acquires a set of equipment surface inspection images and uses a pre-trained deep learning model to generate a first crack detection heatmap, achieving preliminary and rapid identification of crack areas on the equipment surface. The constructed crack physical propagation constraint model integrates equipment design material parameters and real-time operating condition parameters, accurately reflecting the propagation law and physical characteristics of cracks under actual operating conditions. Based on the first crack detection heatmap and the crack physical propagation constraint model, iterative optimization processing of the crack area is performed to obtain a second crack detection heatmap, effectively improving the accuracy and reliability of crack detection and reducing misjudgments caused by relying solely on image data. The equipment surface crack state parameters determined by analyzing the second crack detection heatmap cover key information such as crack area boundary coordinates, crack propagation direction vector, and crack depth gradient value. The final generated intelligent equipment fatigue crack detection report includes a crack spatial distribution diagram and crack propagation risk level identification, which can intuitively and accurately present the crack situation and potential risks on the equipment surface, improving the intelligence level and detection quality of equipment fatigue crack detection. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the intelligent equipment fatigue crack detection method based on deep learning provided in the embodiments of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of the intelligent equipment fatigue crack detection system based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a deep learning-based intelligent detection method for equipment fatigue cracks, provided in one embodiment of the present invention. The following is a detailed description of this deep learning-based intelligent detection method for equipment fatigue cracks.

[0015] Step S110: Obtain a set of equipment surface detection images, call a pre-trained deep learning model to predict crack regions in the set of equipment surface detection images, and generate a first crack detection heat map.

[0016] In this embodiment, bridge steel structures, a typical piece of equipment in the field of mechanical and civil engineering, are used as the application scenario. Bridge steel structures are exposed to the natural environment for extended periods and are subjected to repeated loads from vehicles, making them prone to fatigue cracks; therefore, their inspection is crucial. When acquiring a set of surface inspection images of the equipment, high-definition industrial cameras or drone-mounted camera equipment can be used to photograph key parts of the bridge steel structure, such as steel beam connection nodes and webs. During the imaging process, it is essential to ensure the accuracy and completeness of the images, avoiding information loss due to improper shooting angles or insufficient lighting. For example, for steel beam connection nodes, images should be taken from multiple directions to ensure that all potential cracks on the node surface are fully captured.

[0017] After acquiring the set of equipment surface detection images, a pre-trained deep learning model needs to be used for processing. This deep learning model can be based on a convolutional neural network, such as the improved U-Net model. During the pre-training process, a large dataset of bridge steel structure crack images was used. This dataset contains crack images of different types, sizes, and locations, all of which have been annotated by professionals.

[0018] Step S111: Adjust the grayscale difference between the potential crack area and the background area in the image unit of the equipment surface detection image set by the local region grayscale stretching algorithm, and retain the edge detail information in the image unit to obtain the enhanced equipment surface image set.

[0019] In this embodiment, the grayscale difference between the potential crack area and the background area in the image unit of the bridge steel structure surface may be small, which is not conducive to subsequent feature extraction and detection. The role of the local region grayscale stretching algorithm is to enhance the above difference. Specifically, each image in the equipment surface detection image set is first divided into several non-overlapping image units. The size of each image unit can be set according to the actual situation, such as 256×256 pixels. For each image unit, the minimum and maximum grayscale values ​​are calculated, and then the grayscale value range is stretched to a preset range, such as 0-255, through linear transformation. This makes the grayscale value difference between the potential crack area and the background area more obvious. At the same time, since it is a local region processing, it can better preserve the edge detail information in the image unit and avoid the loss of details in some areas that may be caused by overall stretching. After this processing, an enhanced equipment surface image set is obtained, in which the images display the crack area more accurately.

[0020] Step S112: Perform multi-resolution feature extraction processing on the enhanced equipment surface image set to obtain a multi-resolution crack feature set.

[0021] In this embodiment, multi-resolution feature extraction is used to obtain crack features at different scales. For each image in the enhanced equipment surface image set, convolution operations are performed using convolution kernels of different sizes. For example, 3×3, 5×5, and 7×7 convolution kernels are used to convolve the images, with each kernel corresponding to a different resolution for feature extraction. During the convolution operation, a sliding window is moved across the image, and the sum of the products of the pixel values ​​within each window and the convolution kernel is calculated to obtain the corresponding feature map. Different sized convolution kernels can capture crack features of different sizes; a 3×3 convolution kernel is suitable for capturing smaller crack details, while a 7×7 convolution kernel is suitable for capturing larger crack outlines. The feature maps obtained from these different convolution kernels are combined to obtain a multi-resolution crack feature set, which contains crack feature information ranging from fine to coarse.

[0022] Step S113: Perform cross-layer feature fusion processing on the multi-resolution crack feature set to generate a fused crack feature map.

[0023] In this embodiment, cross-layer feature fusion effectively combines features from different levels in a multi-resolution crack feature set. The feature maps in the multi-resolution crack feature set come from different convolutional layers. Lower-level feature maps contain more detailed information, such as crack edges and textures, but less semantic information; higher-level feature maps contain more semantic information, such as the overall shape of the crack, but less detailed information. During cross-layer feature fusion, the higher-level feature maps are first upsampled to match their size with the lower-level feature maps. Upsampling can be done using bilinear interpolation, calculating a weighted average of surrounding pixels to obtain new pixel values, thereby expanding the feature map size. Then, the upsampled higher-level feature map is concatenated with the corresponding lower-level feature map. During concatenation, the two feature maps are combined along the channel dimension. For example, if the lower-level feature map has 64 channels and the upsampled higher-level feature map also has 64 channels, the concatenated feature map has 128 channels. Through this method, the fused feature map contains both rich detailed information and sufficient semantic information, generating a fused crack feature map.

[0024] Step S114: Call the deep learning model to calculate the crack region probability of the fused crack feature map, generate the crack existence probability of each pixel, and obtain the initial crack probability map.

[0025] In this embodiment, the invoked deep learning model processes the received fused crack feature map through a series of neural network layers. First, the fused crack feature map enters multiple convolutional layers to further extract more abstract features. Each convolutional layer consists of multiple convolutional kernels and activation functions. The convolutional kernels continue to extract features from the feature map, and activation functions such as ReLU are used to introduce non-linearity and enhance the model's expressive power. After processing by the convolutional layers, the feature map enters a fully connected layer. The fully connected layer connects all pixel values ​​of the feature map and uses matrix operations to obtain the probability that each pixel belongs to a crack region. For example, for each pixel in the fused crack feature map, the model outputs a value between 0 and 1, representing the probability that the pixel has a crack. Combining the probability values ​​of all pixels yields an initial crack probability map, where the value at each position reflects the likelihood of a crack existing at that position.

[0026] Step S115: Spatial smoothing is performed on the initial crack probability map. The mean filtering algorithm is used to eliminate isolated noise points in the probability map and retain the probability distribution characteristics of the continuous crack region to generate a first crack detection heat map. Each pixel value of the first crack detection heat map represents the probability that a crack exists at the corresponding pixel location.

[0027] In this embodiment, the initial crack probability map may contain some isolated high-probability points. These points are not actual crack areas but rather noise points. The mean filtering algorithm can effectively eliminate these noise points. The window size for mean filtering can be selected according to the actual situation, such as a 3×3 window. During processing, the window slides across the initial crack probability map. For each window, the average probability of all pixels within the window is calculated, and this average value replaces the probability value of the pixel at the center of the window. In this way, isolated noise points, due to the lower probability values ​​of their surrounding pixels, will have their probability values ​​lowered after mean filtering. Meanwhile, continuous crack areas, due to the higher probability values ​​of their surrounding pixels, will still maintain a high probability value after filtering, thus preserving the probability distribution characteristics of continuous crack areas. After spatial smoothing, a first crack detection heatmap is generated, which can more accurately reflect the distribution of crack areas.

[0028] Step S120: Construct a crack physical propagation constraint model based on the equipment design material parameters and real-time operating condition parameters. The crack physical propagation constraint model includes material fatigue limit parameters, stress distribution coefficients, and propagation resistance thresholds.

[0029] In this embodiment, for the bridge steel structure equipment, the design material parameters can be obtained from its design manual, including the type of material, such as Q345 steel, as well as its elastic modulus, Poisson's ratio, fracture toughness value, and fatigue strength coefficient. Real-time operating parameters are collected through a sensor network installed on the bridge. For example, the operating load value can be obtained from the vehicle load on the bridge using a weighing sensor. Operating speed is not a suitable parameter for bridges; it can be replaced by parameters such as vehicle traffic frequency and ambient temperature. Constructing a crack physical propagation constraint model requires comprehensively considering the influence of these parameters on crack propagation. The material fatigue limit parameter is the maximum cyclic stress amplitude that the material can withstand; exceeding this value makes crack propagation easier. The stress distribution coefficient reflects the degree of stress concentration in different areas of the equipment surface. The propagation resistance threshold represents the material resistance that crack propagation needs to overcome.

[0030] Step S121: Extract the equipment design material parameters from the equipment design manual, determine the standard value range of each equipment design material parameter through the material mechanical property test report, and obtain the set of basic material parameters. The equipment design material parameters include the material elastic modulus, Poisson's ratio, fracture toughness value and material fatigue strength coefficient.

[0031] In this embodiment, parameters such as the elastic modulus, Poisson's ratio, fracture toughness, and fatigue strength coefficient of the material are extracted from the design manual of bridge steel structures. Then, the mechanical property test report of the material is consulted to determine the standard value range of each parameter. For example, the standard value range of the elastic modulus of Q345 steel is within a certain interval, and the standard value range of Poisson's ratio is within another interval, etc. These parameters and their standard value ranges are compiled to form a set of basic material parameters.

[0032] Step S122: Collect real-time operating condition parameters of the equipment through the equipment sensor network. The real-time operating condition parameters include the operating load value and operating speed of the equipment.

[0033] In this embodiment, the sensor network for the bridge steel structure includes sensors installed at various key locations on the bridge. Operating load values ​​are collected via weighing sensors. When a vehicle crosses the bridge, the weighing sensors convert the vehicle's load into an electrical signal, which is then processed to obtain the specific operating load value. For operating speed, since the bridge itself does not involve rotating parts, parameters such as vehicle speed and number of vehicles passing through can be collected instead. These parameters also reflect the bridge's real-time operating conditions. The data collected by the sensors is transmitted wirelessly to a data processing center for further processing and analysis. When collecting the above data, if potentially private data such as vehicle information is involved, data anonymization techniques are used, such as blurring vehicle license plate information, retaining only non-privacy data related to the operating conditions, such as vehicle load and speed, to protect privacy.

[0034] Step S123: Based on the ambient temperature in the real-time operating condition parameters, query the material performance temperature influence curve, adjust the material elastic modulus and fracture toughness values ​​in the material basic parameter set, and obtain the material correction parameter set. The values ​​of the material correction parameter set decrease linearly with the increase of ambient temperature.

[0035] In this embodiment, the ambient temperature in the real-time operating condition parameters is collected by a temperature sensor. The material property temperature effect curve is obtained experimentally, reflecting the relationship between the material's elastic modulus and fracture toughness values ​​and ambient temperature. For example, for Q345 steel, its elastic modulus and fracture toughness values ​​decrease linearly with increasing ambient temperature. Based on the collected ambient temperature, the corresponding correction values ​​for the elastic modulus and fracture toughness values ​​are found on the material property temperature effect curve, and then these two parameters in the material's basic parameter set are adjusted. The adjusted parameters form a material correction parameter set, making the parameters more consistent with the actual influence of environmental conditions on material properties.

[0036] Step S124: Based on the set of material correction parameters and the real-time operating condition parameters, the stress field distribution on the surface of the equipment under the current load and rotation speed is simulated using the finite element analysis method. The stress concentration factor of the potential crack region is extracted, and the ratio of the stress concentration factor to the material fatigue strength coefficient is used as the stress distribution coefficient.

[0037] In this embodiment, parameters such as the elastic modulus and Poisson's ratio from the material correction parameter set, as well as the operating load value from the real-time operating condition parameters, are input into the finite element analysis software. In the finite element analysis software, a finite element model of the bridge steel structure is constructed, mesh elements are generated, and boundary conditions, such as fixed constraints and operating loads, are applied. Then, a stress field distribution simulation is performed to obtain the stress field distribution on the equipment surface under the current operating conditions. The stress concentration factor for potential crack regions is extracted from the stress field distribution results. Potential crack regions are typically areas with high stress or abrupt changes in structural geometry. Finally, the ratio of the stress concentration factor to the material fatigue strength coefficient is calculated to obtain the stress distribution coefficient, which reflects the likelihood of crack propagation in that region.

[0038] Step S1241: Based on the geometric dimensions and material parameters in the equipment design drawings, divide the equipment surface into quadrilateral mesh elements to obtain the finite element mesh model of the equipment surface. The mesh element size is refined in the crack potential region and sparse in the non-crack potential region.

[0039] In this embodiment, the precise geometric dimensions of the bridge steel structure, such as the length, width, and thickness of the steel beams, as well as material parameters such as the elastic modulus and Poisson's ratio, are obtained based on the design drawings. A geometric model of the equipment surface is then constructed using finite element analysis software according to these dimensions. The geometric model is then meshed using quadrilateral mesh elements. For potential crack areas, such as the connection nodes of the steel beams, where stress distribution is complex and cracks are easily generated, the mesh element size is reduced to a smaller size (i.e., densification) to improve the accuracy of stress calculation. In non-potential crack areas, such as the straight sections of the steel beams, the mesh element size can be larger to reduce computational load. After meshing, the finite element mesh model of the equipment surface is obtained.

[0040] Step S1242: Use the operating load value and operating speed in the real-time operating condition parameters as the boundary conditions for finite element analysis, apply the corresponding uniformly distributed load and centrifugal force in the finite element mesh model, and set the equipment fixed constraint boundary.

[0041] In this embodiment, for the bridge steel structure, the operating load value in the real-time operating condition parameters is applied as a uniformly distributed load to the bridge deck portion of the finite element mesh model to simulate the load effect of vehicles on the bridge. Since the bridge does not have rotating parts, the centrifugal force corresponding to the operating speed is not applicable here; instead, a corresponding dynamic load can be applied based on parameters such as vehicle traffic frequency. Simultaneously, fixed constraint boundaries are set for the equipment; for example, the connection between the bridge piers and the foundation is set as a fixed constraint to limit its displacement in all directions. By applying these boundary conditions, the finite element analysis more closely reflects the actual stress conditions of the bridge.

[0042] Step S1243: Call the finite element solver to calculate the stress field of the finite element mesh model, solve the linear elasticity equilibrium equation, obtain the stress components of each mesh element on the equipment surface, and generate a stress field distribution cloud map of the equipment surface. The stress components include normal stress and shear stress.

[0043] In this embodiment, the finite element solver solves the linear elasticity equilibrium equation based on the input finite element mesh model, material parameters, and boundary conditions. This linear elasticity equilibrium equation describes the relationship between internal stress and external forces in an object under static equilibrium. During the solution process, the normal stress and shear stress of each mesh element are calculated using methods such as matrix operations. Normal stress is the stress perpendicular to the cross-section, and shear stress is the stress parallel to the cross-section. The magnitude of these stress components is represented by different colors on the geometric model of the equipment surface, generating a stress field distribution cloud map. The darker the color, the greater the stress. This stress field distribution cloud map allows for a visual observation of the stress distribution on the equipment surface.

[0044] Step S1244: Locate the potential crack region in the stress field distribution cloud map, and extract the stress concentration factor of each grid cell in the potential crack region. The potential crack region is the region where the curvature change of the equipment surface is greater than a set change threshold or where stress is concentrated. The stress concentration factor is the ratio of the maximum stress value of the grid cell to the nominal stress value.

[0045] In this embodiment, potential crack regions are located by analyzing the curvature changes and stress magnitudes of the equipment surface in the stress field distribution cloud map. When the curvature change of the equipment surface exceeds a set threshold, stress concentration is likely to occur in this region, which is considered a potential crack region. Simultaneously, areas with darker colors in the stress field distribution cloud map, indicating higher stress, are also potential crack regions. For each located potential crack region, the maximum stress value and nominal stress value of each grid cell are extracted. The nominal stress value is the average stress calculated based on a simple stress model. The stress concentration factor is the ratio of the maximum stress value to the nominal stress value of a grid cell. The larger this ratio, the higher the stress concentration in the region where the grid cell is located, and the more likely cracks are to occur.

[0046] Step S1245: Read the material fatigue strength coefficient from the set of material correction parameters, calculate the stress distribution coefficient of each potential crack region, and generate a set of stress distribution coefficients containing the stress distribution coefficients of all potential crack regions. The stress distribution coefficient is the ratio of the stress concentration factor of the potential crack region to the material fatigue strength coefficient. A stress distribution coefficient greater than 1 indicates that there is a risk of crack propagation in the potential crack region.

[0047] In this embodiment, the material fatigue strength coefficient is obtained from the material correction parameter set. For each potential crack region, its stress concentration factor is divided by the material fatigue strength coefficient to obtain the stress distribution coefficient of that region. The stress distribution coefficients of all potential crack regions are collected to form a set of stress distribution coefficients. When the stress distribution coefficient is greater than 1, it indicates that the stress concentration in that region exceeds the material fatigue strength coefficient, and there is a risk of crack propagation; when the stress distribution coefficient is less than or equal to 1, the risk of crack propagation in that region is relatively low.

[0048] Step S125: Based on the fracture toughness value in the material correction parameter set and the vibration acceleration in the real-time operating condition parameters, the fracture toughness value is converted into the critical stress intensity factor required for crack propagation. Combined with the promoting effect of vibration acceleration on crack propagation, the crack propagation resistance threshold is determined. The crack propagation resistance threshold decreases linearly with the increase of vibration acceleration.

[0049] In this embodiment, the fracture toughness value in the material correction parameter set is an indicator of the material's ability to resist crack propagation. According to fracture mechanics theory, the fracture toughness value is converted into a critical stress intensity factor required for crack propagation, which represents the intensity of the stress field at the crack tip. Vibration acceleration in real-time operating parameters can be collected by an acceleration sensor, such as the vibration acceleration of a bridge caused by vehicle movement. Vibration acceleration promotes crack propagation; therefore, its influence needs to be considered when determining the crack propagation resistance threshold. Specifically, as vibration acceleration increases, the crack propagation resistance threshold decreases linearly; that is, the greater the vibration, the easier it is for the crack to overcome material resistance and propagate. By combining the critical stress intensity factor obtained from the fracture toughness value conversion and the influence of vibration acceleration, the final crack propagation resistance threshold is determined.

[0050] Step S126: Input the material fatigue limit parameter, stress distribution coefficient and propagation resistance threshold into the constraint model construction module to establish a quantitative relationship between the three and generate a crack physical propagation constraint model containing parameter interaction terms.

[0051] In this embodiment, the constraint model construction module is a software module used to construct a mathematical model. After the material fatigue limit parameter, stress distribution coefficient, and crack propagation resistance threshold are input into this module, it establishes a quantitative relationship between the three through statistical analysis, regression analysis, and other methods. For example, there may be parameter interaction terms such as the product of the material fatigue limit parameter and the stress distribution coefficient, and the ratio of the stress distribution coefficient to the crack propagation resistance threshold. These parameter interaction terms can more accurately reflect the combined influence of the three parameters on crack propagation. By establishing mathematical expressions that include these parameter interaction terms, a crack physical propagation constraint model is generated. This crack physical propagation constraint model can be used to predict crack propagation under different parameter combinations.

[0052] Step S130: Based on the first crack detection heat map and the crack physical propagation constraint model, perform iterative optimization processing of the crack region to obtain the second crack detection heat map.

[0053] In this embodiment, the first crack detection heatmap provides an image-based probability distribution of crack regions, while the crack physical propagation constraint model provides physical constraints on crack propagation. The purpose of iterative optimization is to combine the information from both to more accurately locate and correct the crack regions. By continuously iteratively adjusting the probability distribution of crack regions, the final second crack detection heatmap is made to conform to both image features and physical laws.

[0054] Step S131: Initialize the iterative optimization parameters, which include the maximum number of iterations, the heat map probability update step size, and the constraint model parameter adjustment coefficient. The maximum number of iterations is determined based on the resolution of the equipment surface image; the higher the resolution, the more iterations are performed. The initial value of the heat map probability update step size is positively correlated with the crack physical propagation constraint model's propagation resistance threshold.

[0055] In this embodiment, the setting of the maximum number of iterations for the surface image of the bridge steel structure needs to consider the image resolution. If the image resolution is high and contains more pixels, more iterations are needed to stabilize the probability distribution of the crack region. For example, for an image with a resolution of 2048×2048 pixels, the maximum number of iterations can be set to a larger value; for images with lower resolution, the maximum number of iterations can be reduced accordingly. The probability update step size of the heatmap determines the magnitude of probability value adjustment in each iteration. Its initial value is positively correlated with the propagation resistance threshold, that is, the larger the propagation resistance threshold, the larger the initial update step size. This is because when the material's resistance to crack propagation is large, a larger adjustment magnitude is needed to respond to the physical constraints. The constraint model parameter adjustment coefficient is used to adjust the parameters of the crack physical propagation constraint model during the iteration process. Its value ranges from 0 to 1, and the specific value can be set according to the actual situation to ensure that the adjustment of model parameters is stable and effective.

[0056] Step S132: Input the first crack detection heatmap into the iterative optimization module, extract regions with probability values ​​higher than the initial probability threshold as candidate crack regions, record the boundary pixel coordinates and probability value distribution of the candidate crack regions, and obtain the initial candidate crack region set.

[0057] In this embodiment, after receiving the first crack detection heatmap, the iterative optimization module first sets an initial probability threshold, which can be determined based on experience or experimental data. Then, it traverses each pixel in the heatmap, identifying connected regions containing pixels with probability values ​​higher than the threshold as candidate crack regions. For each candidate crack region, its boundary pixel coordinates are extracted using an edge detection algorithm; these boundary pixel coordinates are represented by the row and column positions of the pixels. Simultaneously, the probability value distribution of all pixels within this region is recorded, forming a probability matrix. The boundary pixel coordinates and probability value distributions of all candidate crack regions are then combined to obtain the initial set of candidate crack regions.

[0058] Step S133: Based on the crack physical propagation constraint model, perform physical feasibility verification on the initial candidate crack region set, calculate the theoretical propagation path of the candidate crack region under the current stress distribution coefficient, compare the theoretical propagation path with the actual distribution of the candidate crack region, and calculate the path deviation value. The path deviation value is the average Euclidean distance between the theoretical propagation path and the actual distribution boundary.

[0059] In this embodiment, for each candidate crack region in the initial set of candidate crack regions, the theoretical propagation path of the region under the current working condition is calculated based on the stress distribution coefficient and material fatigue limit parameters in the crack physical propagation constraint model, combined with the crack propagation theory in fracture mechanics. The calculation of the theoretical propagation path takes into account the direction and magnitude of stress, and typically propagates along the direction of maximum stress. Then, the calculated theoretical propagation path is compared with the actual distribution boundary of the candidate crack region. Specifically, in the image coordinate system, points on the theoretical propagation path and points on the actual distribution boundary are taken respectively, and the Euclidean distance between each pair of corresponding points is calculated. The average of these distances is then calculated to obtain the path deviation value. The smaller the path deviation value, the more closely the actual distribution of the candidate crack region matches the theoretical propagation path, and the greater the probability that the candidate crack region is a real crack region.

[0060] Step S134: Update the probability value of the first crack detection heatmap according to the path deviation value. For candidate crack regions with path deviation values ​​less than the deviation threshold, increase the probability value of their internal pixels. For candidate crack regions with path deviation values ​​not less than the deviation threshold, decrease the probability value of their internal pixels. Generate an updated crack heatmap.

[0061] In this embodiment, a deviation threshold is set, which can be determined according to the detection accuracy requirements. For candidate crack regions with path deviation values ​​less than the deviation threshold, it indicates that they conform to the theoretical propagation law, so the probability values ​​of all pixels within this region are increased. The magnitude of the increase is determined by the heatmap probability update step size, for example, increasing the probability value of each pixel by the update step size and the product of (deviation threshold - path deviation value) / deviation threshold. Thus, the smaller the path deviation value, the greater the probability increase. For candidate crack regions with path deviation values ​​not less than the deviation threshold, they do not conform to the theoretical propagation law and may be noise or false detection regions, so the probability values ​​of pixels within this region are reduced. The magnitude of the reduction is also determined by the update step size, for example, decreasing the probability value of each pixel by the update step size and the product of (path deviation value - deviation threshold) / path deviation value. After the above update processing, an updated crack heatmap is generated, and the probability distribution of the updated heatmap is more reasonable.

[0062] Step S135: Adjust the parameters of the crack physical propagation constraint model based on the updated crack thermogram, extract the crack depth gradient value of the probability region in the updated crack thermogram that the probability value is greater than the set probability value, and adjust the stress distribution coefficient and propagation resistance threshold in the crack physical propagation constraint model according to the matching degree between the crack depth gradient value and the material fatigue limit parameter, and generate the adjusted constraint model.

[0063] In this embodiment, a predetermined probability value is set, and regions with probability values ​​greater than this predetermined probability value are extracted from the updated crack thermogram as probability regions. For these probability regions, the crack depth gradient values ​​obtained in previous steps are used to analyze their depth changes. Then, the matching degree between the crack depth gradient values ​​and the material fatigue limit parameters is calculated. The higher the matching degree, the better the current material parameters reflect the crack propagation in that region. Based on the matching degree, the stress distribution coefficient and propagation resistance threshold in the crack physical propagation constraint model are adjusted. For example, if the matching degree is low, the stress distribution coefficient is appropriately increased or the propagation resistance threshold is decreased to make the model more consistent with the actual crack depth changes. After adjustment, an adjusted constraint model is generated, which can better adapt to the crack situation reflected by the updated crack thermogram.

[0064] Step S136: Determine whether the current iteration count has reached the maximum iteration count or whether the path deviation value is less than the minimum deviation threshold. If not, return to the step of inputting the updated crack heat map into the iterative optimization module to continue the iteration. If the threshold is reached, stop the iteration and determine the current updated crack heat map as the second crack detection heat map.

[0065] In this embodiment, after each iteration, the current iteration count and path deviation value are recorded. The current iteration count is compared with the maximum iteration count, and the path deviation value is compared with the minimum deviation threshold. If the current iteration count has not yet reached the maximum iteration count, and the path deviation value is not less than the minimum deviation threshold, then the process returns to step S132, and the updated crack heatmap is input into the iterative optimization module again for the next iteration. If the current iteration count has reached the maximum iteration count, or the path deviation value is less than the minimum deviation threshold, then the iteration process stops. At this point, the updated crack heatmap can better balance image features and physical constraints, and it is determined as the second crack detection heatmap.

[0066] Step S140: Analyze the second crack detection heat map to determine the crack state parameters on the equipment surface. The crack state parameters on the equipment surface include the boundary coordinates of the crack region, the crack propagation direction vector, and the crack depth gradient value.

[0067] In this embodiment, the second crack detection heatmap accurately reflects the distribution of the crack region. Analyzing this heatmap involves extracting specific parameters that describe the crack state. The crack region boundary coordinates determine the location and shape of the crack; the crack propagation direction vector indicates the direction in which the crack may continue to propagate; and the crack depth gradient value reflects the change of the crack in the depth direction. These parameters together constitute the crack state parameters of the equipment surface.

[0068] Step S141: Binarize the second crack detection heatmap, use an adaptive threshold segmentation algorithm to determine the segmentation threshold between the crack region and the background region, mark the region with a probability value higher than the segmentation threshold as the crack region, and obtain a binary image of the crack on the equipment surface.

[0069] In this embodiment, the adaptive threshold segmentation algorithm can automatically determine the segmentation threshold based on the grayscale features of local image regions, making it suitable for images with uneven grayscale distribution in the second crack detection heatmap. Through this algorithm, regions in the heatmap with probability values ​​higher than the segmentation threshold are marked as crack regions and assigned a value of 1; regions with probability values ​​lower than or equal to the segmentation threshold are marked as background regions and assigned a value of 0, thus obtaining a binary image of the equipment surface crack. Obtaining the binary image of the equipment surface crack makes the distinction between crack regions and background regions clearer, facilitating subsequent feature extraction.

[0070] Step S1411: Divide the second crack detection heatmap into multiple non-overlapping image sub-blocks.

[0071] In this embodiment, the second crack detection heatmap is divided into multiple non-overlapping image sub-blocks according to a set size, for example, each sub-block is 64×64 pixels in size. The division begins from the top left corner of the image and proceeds row by row and column by column, ensuring that each sub-block is complete and non-overlapping. Dividing the image into sub-blocks is to calculate the segmentation threshold within local regions, improving the threshold's adaptability.

[0072] Step S1412: Calculate the mean gray value and standard deviation of gray value for each image sub-block. Use the product of the mean gray value and the standard deviation of gray value as the local threshold of the sub-block. The local threshold of the sub-block increases with the increase of the mean gray value of the sub-block and increases with the increase of the standard deviation of gray value.

[0073] In this embodiment, for each image sub-block, the gray-level mean and gray-level standard deviation of the probability values ​​of all pixels within it are calculated. The gray-level mean reflects the overall brightness level of the image sub-block, while the gray-level standard deviation reflects the dispersion of pixel values ​​within the sub-block. Multiplying the gray-level mean and gray-level standard deviation yields the local threshold for that sub-block. Since a larger gray-level mean and gray-level standard deviation indicate a greater likelihood that the image sub-block may contain cracked areas, the local threshold for the sub-block increases with both, thereby better adapting to the characteristics of different sub-blocks.

[0074] Step S1413: Smooth the local thresholds of all image sub-blocks to obtain a smoothed local threshold matrix. Compare the probability value of each pixel in the second crack detection heatmap with the local threshold at the corresponding position in the smoothed local threshold matrix. If the probability value is higher than the local threshold, it is marked as a crack region pixel and assigned a value of 1; otherwise, it is marked as a background region pixel and assigned a value of 0, thus obtaining a binary image.

[0075] In this embodiment, since the local thresholds of adjacent image sub-blocks may differ significantly, directly using them would result in block artifacts in the binary image. Therefore, it is necessary to smooth the local thresholds of all sub-blocks, for example, by using Gaussian filtering to make the threshold transition between adjacent sub-blocks smoother, resulting in a smoothed local threshold matrix. Then, the probability value of each pixel in the second crack detection heatmap is compared with the local threshold at the corresponding position in the smoothed local threshold matrix. Based on the comparison result, it is marked as a crack region or a background region, thus obtaining a binary image.

[0076] Step S1414: Perform morphological processing on the binary image to obtain a binary image of surface cracks in the equipment. The size of the structural element in the morphological processing is determined based on the average width of the crack region.

[0077] In this embodiment, the binary image may contain small noise points and pores within the crack area, which need to be eliminated through morphological processing. Morphological processing includes erosion and dilation operations. Erosion can eliminate small noise points, while dilation can fill pores within the crack area. The size of the structuring element is determined based on the average width of the crack area. For example, if the average width of the crack area is 5 pixels, the size of the structuring element can be selected as 3×3 pixels to ensure that the crack area is not excessively eroded or excessively dilated during processing. After morphological processing, a more accurate and complete binary image of the equipment surface crack is obtained.

[0078] Step S142: Extract the boundary pixels of the crack region in the binary image of the crack on the surface of the equipment, connect the edge pixels of the crack region through the contour tracking algorithm to generate a closed crack region boundary contour, and arrange the coordinates of each pixel on the crack region boundary contour in a clockwise direction to obtain the crack region boundary coordinate sequence.

[0079] In this embodiment, in the binary image of the surface crack, the pixel value of the crack area is 1, and the background area is 0. A contour tracking algorithm is used to track adjacent edge pixels sequentially in a predetermined direction (e.g., clockwise) starting from a certain edge pixel in the crack area, until the starting point is reached, forming a closed crack area boundary contour. During the tracking process, the coordinates (row and column numbers) of each boundary pixel are recorded. These coordinates are then arranged clockwise to form a crack area boundary coordinate sequence. This sequence accurately describes the shape and location of the crack area.

[0080] Step S143: Perform curve fitting processing on the boundary coordinate sequence of the crack region. Use a polynomial fitting algorithm to fit the boundary coordinate sequence, calculate the first derivative of the fitted curve to determine the tangent direction of the curve, and take the tangent direction vector at both ends of the crack region as the crack propagation direction vector. The crack propagation direction vector includes the direction angle and the direction cosine value.

[0081] In this embodiment, a polynomial fitting algorithm is used to fit the coordinate sequence of the crack region boundary. For example, a cubic polynomial is selected to fit the x and y coordinates in the coordinate sequence, respectively, to obtain the polynomials of x and y with respect to parameter t, where t is the sequence index. After fitting, the first derivative of the fitted curve is calculated to obtain the slope of the tangent at each point, thereby determining the tangent direction. At the two endpoints of the crack region, the direction angle and direction cosine value are calculated based on the tangent direction. The direction angle is the angle between the tangent and the positive x-axis, and the direction cosine value includes the cosine components of the tangent in the x and y directions. These parameters together constitute the crack propagation direction vector, indicating the possible propagation direction of the crack.

[0082] Step S144: Extract the probability value distribution inside the crack region from the second crack detection heat map, map the probability value distribution to the corresponding crack depth value, and generate a crack region depth distribution map. The crack depth value is positively correlated with the probability value.

[0083] In this embodiment, after determining the extent of the crack region in the second crack detection heatmap, the probability values ​​of all pixels within that region are extracted to form a probability value distribution. Since a higher probability value indicates a greater likelihood of a crack existing at that location, and typically also implies a deeper crack, the probability value distribution is mapped to crack depth values. The mapping relationship can be determined using experimental data; for example, probability value 0 can be mapped to depth 0, probability value 1 to the maximum possible depth, and intermediate probability values ​​can be mapped to their corresponding depth values ​​according to a linear relationship. Through this mapping, a crack region depth distribution map is generated, which visually reflects the distribution of cracks in the depth direction.

[0084] Step S145: Perform gradient calculation processing on the depth distribution map of the crack region. Use the Sobel operator to calculate the gradient values ​​of the depth distribution map in the horizontal and vertical directions. Combine the gradient values ​​into a crack depth gradient vector. Extract the magnitude and direction of the crack depth gradient vector to obtain the crack depth gradient value. The magnitude of the crack depth gradient value represents the rate of change of depth, and the direction represents the direction of increasing depth.

[0085] In this embodiment, the Sobel operator includes horizontal and vertical convolution kernels, used to calculate the gradients of the depth distribution map in the horizontal and vertical directions, respectively. The horizontal convolution kernel is convolved with the depth distribution map to obtain the horizontal gradient value; the vertical convolution kernel is convolved with the depth distribution map to obtain the vertical gradient value. For each pixel, the horizontal and vertical gradient values ​​are combined to form a crack depth gradient vector. Then, the magnitude of this crack depth gradient vector is calculated, where the magnitude represents the rate of change of depth; the direction of the crack depth gradient vector is also calculated, where the direction represents the direction of increasing depth. These parameters constitute the crack depth gradient value, reflecting the spatial variation of the crack depth.

[0086] Step S146: Combine the crack region boundary coordinate sequence, crack propagation direction vector, and crack depth gradient value into equipment surface crack state parameters. The crack region boundary coordinate sequence is represented by pixel coordinates, the crack propagation direction vector is represented by a unit vector, and the crack depth gradient value is represented by gradient magnitude and gradient direction angle.

[0087] In this embodiment, the previously obtained crack region boundary coordinate sequence, crack propagation direction vector, and crack depth gradient value are integrated to form the equipment surface crack state parameters. The crack region boundary coordinate sequence is represented by the row and column numbers of pixels; the crack propagation direction vector is normalized to a unit vector with a magnitude of 1 for easy subsequent calculation and analysis; the crack depth gradient value is explicitly represented as the gradient magnitude and gradient direction angle, reflecting the depth change rate and depth increase direction, respectively. The equipment surface crack state parameters comprehensively describe the characteristics of the crack.

[0088] Step S150: Generate an intelligent fatigue crack detection report for the equipment based on the surface crack state parameters. The intelligent fatigue crack detection report includes a schematic diagram of crack spatial distribution and a crack propagation risk level indicator.

[0089] In this embodiment, a detailed intelligent detection report of equipment fatigue cracks is generated using various information from the surface crack state parameters. The crack spatial distribution diagram visually displays the location and shape of the cracks on the equipment surface; the crack propagation risk level indicator assesses the degree of risk of crack propagation based on the characteristics of the cracks and related parameters.

[0090] Step S151: Analyze the crack region boundary coordinate sequence in the surface crack state parameters of the equipment, and convert the crack region boundary coordinate sequence from the image coordinate system to the equipment physical coordinate system to obtain the crack region boundary coordinates in the physical coordinate system. The equipment physical coordinate system takes the lower left corner of the bottom surface of the equipment as the origin, the X-axis is along the length direction of the equipment, and the Y-axis is along the width direction of the equipment.

[0091] In this embodiment, the image coordinate system has its origin at the top left corner of the image, with the row number as the y-axis and the column number as the x-axis, while the equipment physical coordinate system has its origin at the bottom left corner of the bridge steel structure, with the X-axis along the bridge length and the Y-axis along the bridge width. A transformation relationship needs to be established between the two coordinate systems. This transformation relationship can be obtained through camera calibration, including scaling factors and coordinate offsets. Based on the transformation relationship, the coordinates of each pixel in the crack region boundary coordinate sequence are converted to coordinates in the equipment physical coordinate system, resulting in the crack region boundary coordinates in the physical coordinate system. The converted coordinates are expressed in actual length units (such as millimeters), which is more consistent with practical engineering applications.

[0092] Step S152: Draw a schematic diagram of the spatial distribution of cracks based on the boundary coordinates of the crack region in the physical coordinate system. Use a polygon filling algorithm to mark the position and shape of the crack region in the simplified two-dimensional structure of the equipment. Mark the geometric center coordinates and maximum size parameter of the crack region. The maximum size parameter is the maximum distance between two points in the boundary coordinate sequence of the crack region.

[0093] In this embodiment, a simplified two-dimensional structural diagram of the bridge steel structure is first obtained, reflecting the overall structure and key component locations of the bridge. Then, based on the boundary coordinates of the crack region in the physical coordinate system, the location of the crack region is determined within the simplified two-dimensional structural diagram. A polygon filling algorithm is used to fill the polygonal region formed by connecting the boundary coordinates with a specific color to distinguish it from other regions, thus marking the shape of the crack region. The geometric center coordinates of the crack region are calculated, i.e., the average X-axis and Y-axis coordinates of all boundary coordinates, and marked at the corresponding positions in the simplified two-dimensional structural diagram. Simultaneously, the distance between any two points in the boundary coordinate sequence is calculated, and the maximum distance is identified as the maximum dimension parameter, which is then marked on the simplified two-dimensional structural diagram along with the corresponding two-point positions. Through these annotations, a spatial distribution diagram of the cracks is generated, allowing relevant personnel to intuitively understand the specific distribution of cracks on the bridge steel structure.

[0094] Step S1521: Extract the X-axis and Y-axis coordinate values ​​of each coordinate point in the crack region boundary coordinate sequence under the physical coordinate system, and perform coordinate calibration with the coordinate system of the equipment two-dimensional structural diagram to determine the corresponding position of each boundary coordinate point in the equipment two-dimensional structural diagram, thereby obtaining the crack region boundary coordinate sequence in the equipment two-dimensional structural diagram.

[0095] In this embodiment, the X-axis and Y-axis coordinates of each point are extracted one by one from the crack region boundary coordinate sequence in the physical coordinate system. Since the simplified two-dimensional structural diagram of the equipment has its own coordinate system, coordinate calibration is required to ensure accurate positioning of the crack region boundary coordinates within the simplified diagram. During coordinate calibration, corresponding reference points in both coordinate systems are found, such as the coordinates of a fixed bolt position in a bridge steel structure in both the physical and simplified diagram coordinate systems. The scaling ratio and offset are determined by calculating the transformation relationship between the reference points. Then, based on this transformation relationship, each boundary coordinate point in the physical coordinate system is transformed to the coordinate system of the simplified two-dimensional structural diagram of the equipment, resulting in the crack region boundary coordinate sequence in the simplified diagram, ensuring that each coordinate point accurately corresponds to its corresponding position in the simplified diagram.

[0096] Step S1522: Based on the coordinate sequence of the crack region boundary in the simplified two-dimensional structure diagram of the equipment, connect the coordinate points in sequence to form a closed polygonal outline of the crack region, record the vertex connection order of the polygonal outline, and generate the polygonal outline data of the crack region.

[0097] In this embodiment, based on the coordinate sequence of the crack region boundary in the simplified two-dimensional structural diagram of the equipment, adjacent coordinate points are connected sequentially with straight lines according to their order in the sequence. Starting from the first coordinate point, the connection is made to the second coordinate point, then to the third coordinate point, and so on, until the last coordinate point is connected to the first coordinate point, forming a closed polygonal outline. This polygonal outline represents the shape of the crack region in the simplified two-dimensional structural diagram of the equipment. During the connection process, the connection order of each vertex is recorded in detail, including the coordinates of each vertex and its connection relationship with the preceding and following vertices. This information is then integrated to generate the polygonal outline data of the crack region.

[0098] Step S1523: Call the polygon filling algorithm to perform internal filling processing on the polygon outline data of the crack area, select a fill color that has a grayscale difference from the background of the equipment two-dimensional structural diagram, and generate the filled crack area graphic data.

[0099] In this embodiment, a polygon filling algorithm is invoked. This algorithm fills the area within the polygonal contour data of the crack region, starting from the inside of the contour and proceeding row by row or column by column. The fill color must have a significant grayscale difference from the background color of the simplified 2D structural diagram of the equipment. For example, if the background of the diagram is light-colored, a dark fill color can be selected to ensure that the filled crack region can be accurately displayed for observation and identification. During the filling process, the algorithm determines whether each pixel is located inside the polygonal contour. For pixels located inside, the selected fill color is assigned, ultimately generating the filled crack region graphic data. This crack region graphic data completely presents the shape and location of the crack region.

[0100] Step S1524: Perform region analysis on the filled crack region graphic data, calculate the average X-axis coordinate and average Y-axis coordinate of all coordinate points, and combine them to form the geometric center coordinates of the crack region.

[0101] In this embodiment, a region analysis is performed on the filled crack region graphic data, traversing all coordinate points contained in the graphic data. For each coordinate point, its X-axis and Y-axis coordinate values ​​are extracted, and then the average of all X-axis coordinate values ​​and the average of all Y-axis coordinate values ​​are calculated. These two averages are combined to obtain a new coordinate point, which is the geometric center coordinate of the crack region. The geometric center coordinate reflects the approximate location center of the crack region. In subsequent annotations, this coordinate point is clearly marked on the simplified two-dimensional structure diagram of the equipment, which helps to quickly locate the center position of the crack region.

[0102] Step S1525: Traverse all coordinate points in the boundary coordinate sequence of the crack region under the physical coordinate system, calculate the straight-line distance between every two coordinate points, select the maximum straight-line distance as the maximum size parameter of the crack region, and record the corresponding positions of the two coordinate points corresponding to the maximum straight-line distance in the simplified two-dimensional structure diagram of the equipment.

[0103] In this embodiment, each coordinate point in the crack region boundary coordinate sequence under the physical coordinate system is traversed, and each coordinate point is paired with all other coordinate points. For each pair of coordinate points, the distance between the two points is calculated using the straight-line distance calculation formula based on their X-axis and Y-axis coordinate values. After completing the distance calculation for all pairs, the largest straight-line distance is selected from these distance values ​​and determined as the maximum size parameter of the crack region. At the same time, the two coordinate points corresponding to this maximum straight-line distance are found, and their corresponding positions in the equipment's two-dimensional structural diagram are determined according to the previously determined coordinate transformation relationship and recorded, in preparation for subsequently marking the maximum size parameter in the equipment's two-dimensional structural diagram.

[0104] Step S1526: Integrate the filled crack area graphic data, geometric center coordinates, and coordinate point positions corresponding to the maximum size parameter into the equipment two-dimensional structural diagram, mark the coordinate values ​​at the geometric center coordinates, draw a straight line between the two coordinate points corresponding to the maximum size parameter and mark the maximum size parameter, and generate a crack spatial distribution diagram containing complete annotation information.

[0105] In this embodiment, the filled crack region graphic data is superimposed onto a simplified two-dimensional structural diagram of the equipment, so that the position and shape of the crack region correspond to the bridge structure in the diagram. The specific values ​​of the geometric center coordinates are labeled with text, such as (X1, Y1), to accurately determine the center position of the crack region. A straight line is drawn between the two coordinate points corresponding to the maximum size parameter, visually showing the maximum size range of the crack region. The specific value of the maximum size parameter, such as L1, is labeled next to the line. Through the above integration and labeling operations, a crack spatial distribution diagram containing complete labeling information is generated. This crack spatial distribution diagram accurately and precisely presents the spatial distribution of cracks on the bridge steel structure.

[0106] Step S153: Extract the crack propagation direction vector and crack depth gradient value from the crack state parameters on the equipment surface. Combine the stress distribution coefficient in the crack physical propagation constraint model to calculate the crack propagation risk index. The crack propagation risk index is the product of the modulus of the crack depth gradient value, the stress distribution coefficient, and the cosine of the angle between the propagation direction vector and the stress direction.

[0107] In this embodiment, the crack propagation direction vector and crack depth gradient value are extracted from the crack state parameters on the equipment surface, and the stress distribution coefficient is obtained from the crack physical propagation constraint model. The magnitude of the crack depth gradient value reflects the rate of change of crack depth; the larger the magnitude, the more drastic the change in crack depth. The stress distribution coefficient reflects the degree of stress concentration in the crack area; the larger the coefficient, the more obvious the stress concentration. The cosine of the angle between the propagation direction vector and the stress direction indicates the consistency between the crack propagation direction and the stress direction; the closer the cosine value is to 1, the more consistent the two directions are, and the easier it is for the crack to propagate under stress. The crack propagation risk index is obtained by comprehensively weighting these three parameters. This crack propagation risk index comprehensively reflects the potential risk of crack propagation.

[0108] Step S154: Determine the crack propagation risk level identifier based on the crack propagation risk index, divide the risk index into multiple continuous intervals, each interval corresponds to a risk level, and each risk level corresponds to a preset risk description text.

[0109] In this embodiment, the crack propagation risk index is divided into multiple continuous intervals based on its possible value range. For example, it can be divided into low-risk, medium-risk, and high-risk intervals. Each interval corresponds to a specific risk level, such as low-risk, medium-risk, or high-risk. Simultaneously, a corresponding risk description text is preset for each risk level. The description text for a low-risk level might be "The crack propagation speed is slow, and the impact on equipment safety is relatively small in the short term"; the description text for a medium-risk level might be "The crack has a certain propagation trend, and monitoring needs to be strengthened"; and the description text for a high-risk level might be "The crack propagation risk is high, which may endanger equipment safety, and immediate measures need to be taken." Through this method, the abstract risk index is transformed into intuitive risk level identifiers and specific description texts, enabling relevant personnel to understand and take appropriate countermeasures.

[0110] Step S155: Integrate the crack spatial distribution diagram, crack propagation risk level identifier, and risk description text into the inspection report template. The inspection report template includes a basic equipment information column, a crack distribution column, a risk assessment column, and a treatment suggestion column. The basic equipment information column is filled with the equipment model and inspection time. The treatment suggestion column matches the preset maintenance treatment suggestions according to the risk level identifier to generate an intelligent inspection report of equipment fatigue cracks.

[0111] In this embodiment, the inspection report template is a pre-designed document format containing multiple sections, including basic equipment information, crack distribution, risk assessment, and treatment recommendations. The previously generated crack spatial distribution diagram is placed in the crack distribution section to accurately show the location and shape of the cracks. The crack propagation risk level identifier and corresponding risk description text are placed in the risk assessment section to explain the risk of crack propagation. In the basic equipment information section, information such as the bridge steel structure model and the specific inspection time is entered. The treatment recommendations section matches preset maintenance and treatment suggestions based on the crack propagation risk level identifier. For example, a low-risk level corresponds to "regular inspection, no special treatment required"; a medium-risk level corresponds to "increase inspection frequency and closely monitor crack changes"; and a high-risk level corresponds to "immediately arrange repair or replacement of components." All this information is integrated into the inspection report template to ultimately generate an intelligent inspection report on equipment fatigue cracks.

[0112] The method further includes a model training step for training the deep learning model, specifically including the following steps:

[0113] Step S210: Collect surface crack image samples of the equipment. The surface crack image samples of the equipment include crack images of different types, sizes and locations. At the same time, collect the corresponding crack-free image samples as negative samples.

[0114] In this embodiment, crack image samples of the bridge steel structure surface are collected through various channels. These samples should cover different types of cracks, such as transverse cracks, longitudinal cracks, and diagonal cracks; cracks of different sizes, from micro-cracks to significant large cracks; and cracks in different locations, such as at steel beam connection nodes, webs, and flanges. Simultaneously, a sufficient number of crack-free image samples are collected as negative samples to ensure the balance of the training samples. During the image sample collection process, if privacy data such as personnel information in the bridge's surrounding environment is involved, the images are processed, such as using image cropping techniques to retain only the relevant areas of the bridge steel structure surface and remove parts containing privacy information to protect privacy.

[0115] Step S211: Preprocess the collected equipment surface crack image samples to obtain a preprocessed image sample set. The preprocessing includes image denoising, size normalization and data augmentation.

[0116] In this embodiment, image denoising employs methods such as Gaussian filtering and median filtering to remove noise interference from the image, making crack features more accurate. Size standardization adjusts all image samples to a uniform size, such as 512×512 pixels, for model input and processing. Data augmentation increases sample diversity and expands the number of training samples by performing operations such as rotation, flipping, scaling, and brightness adjustment on the images, thus avoiding model overfitting. For example, rotating the image clockwise by a set angle, flipping it horizontally or vertically, scaling it by different proportions, and increasing or decreasing brightness, etc., yields a preprocessed image sample set.

[0117] Step S212: Label each image sample in the preprocessed image sample set, marking the location and boundary of the crack area, and generating corresponding label data. The label data adopts the same size as the image sample, wherein the pixel value of the crack area is 1 and the pixel value of the background area is 0.

[0118] In this embodiment, professionals use image annotation tools to annotate the preprocessed image samples. During the annotation process, the outline of the crack is precisely delineated along the boundary of the crack region to determine its location and extent. The generated label data has the same size as the corresponding image sample and is in binary form, i.e., the pixel value of the crack region is marked as 1, and the pixel value of the background region is marked as 0. The above label data is used to calculate the loss function of the model during training, guiding the model to learn how to accurately identify crack regions.

[0119] Step S213: Divide the preprocessed image sample set and corresponding label data into a training set, a validation set and a test set. The training set is used for learning model parameters, the validation set is used for parameter adjustment during model training, and the test set is used to evaluate the final performance of the model.

[0120] In this embodiment, the image sample set and label data are divided according to a set ratio, for example, into a training set, a validation set, and a test set in a 7:2:1 ratio. The training set contains most of the sample data and is used by the model to learn the characteristics and patterns of cracks during training, continuously adjusting the model's parameters. The validation set is used to evaluate the model's performance after each epoch of training, and the model's hyperparameters, such as the learning rate and the number of iterations, are adjusted based on the evaluation results to improve the model's generalization ability. The test set is used to comprehensively evaluate the model's final performance after training, including metrics such as detection accuracy and recall, to verify the model's effectiveness.

[0121] Step S214: Construct a deep learning model. The deep learning model adopts an improved U-Net network structure, including an encoder part and a decoder part. The encoder part is used to extract features from the image and contains multiple convolutional layers and pooling layers. The decoder part is used to upsample and fuse the features and output a crack probability map with the same size as the input image.

[0122] In this embodiment, the improved U-Net network structure is optimized based on the traditional U-Net. The encoder part consists of multiple convolutional blocks, each containing two convolutional layers and one pooling layer. The convolutional layers use 3×3 convolutional kernels to extract local features of the image, and each convolutional layer is followed by a ReLU activation function to increase the model's non-linear expressive power. The pooling layer uses 2×2 max pooling to reduce the size of the feature map, reduce computation, and retain important features. The decoder part also consists of multiple convolutional blocks, each containing an upsampling layer and a convolutional layer. The upsampling layer uses transposed convolution to restore the feature map size to the same size as the feature map of the corresponding stage of the encoder, and then concatenates and fuses it with the feature map of the corresponding stage of the encoder to make full use of low-level and high-level features. Finally, a 1×1 convolutional layer outputs a crack probability map with the same size as the input image, where each pixel value in the probability map represents the probability of a crack existing at that location.

[0123] Step S215: Set the model training parameters, including learning rate, batch size, number of iterations, etc., use the cross-entropy loss function to calculate the loss between the model prediction and the label data, and use the Adam optimizer to optimize the model parameters.

[0124] In this embodiment, the initial learning rate is set to a suitable value, such as 0.001, and can be dynamically adjusted during training based on changes in the loss on the validation set. The batch size is set according to the performance of the hardware device, such as 16 or 32, to improve training efficiency. The number of iterations is set to a sufficiently large value, such as 100 epochs, to ensure that the model can fully learn the sample features. The cross-entropy loss function can effectively measure the difference between the crack probability map predicted by the model and the label data; the smaller the loss value, the more accurate the model prediction. The Adam optimizer combines the advantages of momentum gradient descent and adaptive learning rate, enabling it to quickly and stably optimize model parameters, continuously reducing the loss function value.

[0125] Step S216: Input the training set into the constructed deep learning model for training. After each iteration cycle, use the validation set to evaluate the model and adjust the model parameters according to the evaluation results. When the loss of the validation set no longer decreases or reaches the preset number of iterations, stop training and save the trained model parameters.

[0126] In this embodiment, image samples and corresponding label data from the training set are input into the deep learning model. The model performs forward computation based on the input image samples to obtain the predicted crack probability map. Then, the loss between the predicted value and the label data is calculated using the cross-entropy loss function. Based on the loss value, the Adam optimizer is used for backpropagation to update the model parameters, continuously reducing the loss. After each iteration cycle, the validation set is input into the model, and the loss and evaluation metrics, such as accuracy and IoU, are calculated. If the loss of the validation set does not decrease for several cycles, it indicates that the model may be overfitting, and training is stopped at this point; alternatively, training is stopped when a preset number of iterations is reached. Finally, the trained model parameters are saved to obtain the pre-trained deep learning model, which is used for subsequent equipment surface crack detection.

[0127] Step S217: Use the test set to evaluate the performance of the trained model, calculate the model's accuracy, recall, F1 score and other evaluation metrics. If the evaluation metrics meet the preset standards, the model training is complete; if not, adjust the model structure or training parameters and retrain.

[0128] In this embodiment, image samples from the test set are input into the trained model to obtain the model's prediction results. The prediction results are compared with the labeled data of the test set to calculate the model's accuracy (the proportion of correctly predicted crack pixels and background pixels to the total number of pixels), recall (the proportion of correctly predicted crack pixels to the actual crack pixels), and F1 score (the harmonic mean of accuracy and recall), which comprehensively reflects the model's performance. If these evaluation metrics meet the preset standards, such as an accuracy of 0.9 or higher, a recall of 0.85 or higher, and an F1 score of 0.88 or higher, the model performance is considered good, and training is complete. If the preset standards are not met, the reasons need to be analyzed. This could be due to an unreasonable model structure or improper training parameter settings. In this case, the model structure should be adjusted, such as increasing the number of convolutional layers or changing the size of the convolutional kernels, or adjusting training parameters such as the learning rate and batch size. Then, training and evaluation should be repeated until the model performance reaches the preset standards.

[0129] Figure 2 The illustration shows exemplary hardware and software components of a deep learning-based intelligent equipment fatigue crack detection system 100, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the deep learning-based intelligent equipment fatigue crack detection system 100 and to perform the functions in this application.

[0130] For example, the deep learning-based intelligent equipment fatigue crack detection system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the deep learning-based intelligent equipment fatigue crack detection system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The deep learning-based intelligent equipment fatigue crack detection system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0131] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned intelligent detection method for equipment fatigue cracks based on deep learning is implemented.

[0132] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A deep learning-based intelligent detection method for fatigue cracks in equipment, characterized in that, The method includes: A set of equipment surface detection images is acquired, and a pre-trained deep learning model is called to predict crack regions in the set of equipment surface detection images to generate a first crack detection heat map. A crack physical propagation constraint model is constructed based on the material parameters of the equipment design and the real-time operating condition parameters. The crack physical propagation constraint model includes material fatigue limit parameters, stress distribution coefficients, and propagation resistance thresholds. Based on the first crack detection heatmap and the crack physical propagation constraint model, iterative optimization processing of the crack region is performed to obtain the second crack detection heatmap. The crack state parameters on the equipment surface are determined by analyzing the second crack detection thermogram. The crack state parameters on the equipment surface include the boundary coordinates of the crack region, the crack propagation direction vector, and the crack depth gradient value. Based on the surface crack state parameters of the equipment, an intelligent fatigue crack detection report is generated. The intelligent fatigue crack detection report includes a spatial distribution diagram of cracks and a crack propagation risk level indicator. The step of performing iterative optimization processing of the crack region based on the first crack detection heatmap and the crack physical propagation constraint model to obtain a second crack detection heatmap includes: Initialize the iterative optimization parameters, which include the maximum number of iterations, the heat map probability update step size, and the constraint model parameter adjustment coefficient. The maximum number of iterations is determined based on the resolution of the equipment surface image; the higher the resolution, the more iterations are performed. The initial value of the heat map probability update step size is positively correlated with the crack physical propagation constraint model's propagation resistance threshold. The first crack detection heatmap is input into the iterative optimization module, and regions with probability values ​​higher than the initial probability threshold are extracted as candidate crack regions. The boundary pixel coordinates and probability value distribution of the candidate crack regions are recorded to obtain the initial set of candidate crack regions. Based on the crack physical propagation constraint model, the physical feasibility of the initial candidate crack region set is verified. The theoretical propagation path of the candidate crack region under the current stress distribution coefficient is calculated. The theoretical propagation path is compared with the actual distribution of the candidate crack region, and the path deviation value is calculated. The path deviation value is the average Euclidean distance between the theoretical propagation path and the actual distribution boundary. The probability value of the first crack detection heatmap is updated based on the path deviation value. For candidate crack regions with path deviation values ​​less than the deviation threshold, the probability value of their internal pixels is increased. For candidate crack regions with path deviation values ​​not less than the deviation threshold, the probability value of their internal pixels is decreased, and an updated crack heatmap is generated. Based on the updated crack thermogram, the parameters of the crack physical propagation constraint model are adjusted. The crack depth gradient value of the probability region in the updated crack thermogram with a probability value greater than a set probability value is extracted. According to the matching degree between the crack depth gradient value and the material fatigue limit parameter, the stress distribution coefficient and propagation resistance threshold in the crack physical propagation constraint model are adjusted to generate the adjusted constraint model. Determine whether the current iteration count has reached the maximum iteration count or whether the path deviation value is less than the minimum deviation threshold. If not, return to the step of inputting the updated crack heat map into the iterative optimization module to continue the iteration. If the maximum iteration count has been reached, stop the iteration and determine the current updated crack heat map as the second crack detection heat map.

2. The intelligent detection method for equipment fatigue cracks based on deep learning according to claim 1, characterized in that, The process of acquiring a set of equipment surface detection images, calling a pre-trained deep learning model to predict crack regions in the set of equipment surface detection images, and generating a first crack detection heatmap includes: By adjusting the grayscale difference between the potential crack area and the background area in the image unit of the equipment surface detection image set through the local area grayscale stretching algorithm, the edge detail information in the image unit is preserved, and an enhanced equipment surface image set is obtained. Multi-resolution feature extraction processing is performed on the set of surface images of the enhanced equipment to obtain a multi-resolution crack feature set; The multi-resolution crack feature set is subjected to cross-layer feature fusion processing to generate a fused crack feature map. The deep learning model is invoked to calculate the probability of crack regions in the fused crack feature map, generating the probability of crack existence for each pixel, and obtaining an initial crack probability map. The initial crack probability map is spatially smoothed, and a mean filtering algorithm is used to eliminate isolated noise points in the probability map while retaining the probability distribution characteristics of continuous crack regions, thereby generating a first crack detection heatmap. Each pixel value in the first crack detection heatmap represents the probability that a crack exists at the corresponding pixel location.

3. The intelligent detection method for equipment fatigue cracks based on deep learning according to claim 1, characterized in that, The construction of the crack physical propagation constraint model based on equipment design material parameters and real-time operating condition parameters includes: Extract the equipment design material parameters from the equipment design manual, determine the standard value range of each equipment design material parameter through the material mechanical property test report, and obtain the set of basic material parameters. The equipment design material parameters include the material elastic modulus, Poisson's ratio, fracture toughness value and material fatigue strength coefficient. The equipment's real-time operating condition parameters are collected through the equipment's sensor network, including the equipment's operating load value and operating speed. Based on the ambient temperature in the real-time operating condition parameters, the material performance temperature influence curve is queried, and the material elastic modulus and fracture toughness values ​​in the material basic parameter set are adjusted to obtain the material correction parameter set. The values ​​of the material correction parameter set decrease linearly with the increase of ambient temperature. Based on the set of material correction parameters and the real-time operating condition parameters, the stress field distribution on the surface of the equipment under the current load and rotation speed is simulated by the finite element analysis method. The stress concentration factor of the potential crack region is extracted, and the ratio of the stress concentration factor to the material fatigue strength coefficient is used as the stress distribution coefficient. Based on the fracture toughness value in the material correction parameter set and the vibration acceleration in the real-time operating condition parameters, the fracture toughness value is converted into the critical stress intensity factor required for crack propagation. Combined with the promoting effect of vibration acceleration on crack propagation, the crack propagation resistance threshold is determined. The crack propagation resistance threshold decreases linearly with the increase of vibration acceleration. The material fatigue limit parameter, stress distribution coefficient, and propagation resistance threshold are input into the constraint model construction module to establish a quantitative relationship among the three and generate a crack physical propagation constraint model that includes parameter interaction terms.

4. The intelligent equipment fatigue crack detection method based on deep learning according to claim 3, characterized in that, Based on the set of material correction parameters and the real-time operating condition parameters, the finite element analysis method is used to simulate the stress field distribution on the equipment surface under the current load and rotational speed, extract the stress concentration factor of the potential crack region, and use the ratio of the stress concentration factor to the material fatigue strength coefficient as the stress distribution coefficient, including: Based on the geometric dimensions and material parameters in the equipment design drawings, the equipment surface is divided into quadrilateral mesh elements to obtain the finite element mesh model of the equipment surface. The mesh element size is refined in the crack potential region and sparse in the non-crack potential region. The operating load value and operating speed in the real-time operating condition parameters are used as the boundary conditions for finite element analysis. The corresponding uniformly distributed load and centrifugal force are applied to the finite element mesh model, and the equipment fixed constraint boundary is set. The finite element solver is invoked to calculate the stress field of the finite element mesh model, solve the linear elasticity equilibrium equation, obtain the stress components of each mesh element on the equipment surface, and generate a stress field distribution cloud map of the equipment surface. The stress components include normal stress and shear stress. Locate the potential crack region in the stress field distribution cloud map, and extract the stress concentration factor of each grid cell in the potential crack region. The potential crack region is the region where the curvature change of the equipment surface is greater than a set change threshold or where stress is concentrated. The stress concentration factor is the ratio of the maximum stress value of the grid cell to the nominal stress value. The material fatigue strength coefficient is read from the set of material correction parameters, the stress distribution coefficient of each potential crack region is calculated, and a set of stress distribution coefficients containing the stress distribution coefficients of all potential crack regions is generated. The stress distribution coefficient is the ratio of the stress concentration factor of the potential crack region to the material fatigue strength coefficient. The stress distribution coefficient being greater than 1 indicates that there is a risk of crack propagation in the potential crack region.

5. The intelligent equipment fatigue crack detection method based on deep learning according to claim 1, characterized in that, The step of analyzing the second crack detection thermogram to determine the surface crack state parameters of the equipment includes: The second crack detection heatmap is binarized, and an adaptive threshold segmentation algorithm is used to determine the segmentation threshold between the crack region and the background region. Regions with a probability value higher than the segmentation threshold are marked as crack regions, thus obtaining a binary image of the crack on the equipment surface. Extract the boundary pixels of the crack region in the binary image of the crack on the surface of the equipment, connect the edge pixels of the crack region through the contour tracking algorithm to generate a closed crack region boundary contour, and arrange the coordinates of each pixel on the crack region boundary contour in a clockwise direction to obtain the crack region boundary coordinate sequence. The boundary coordinate sequence of the crack region is subjected to curve fitting. A polynomial fitting algorithm is used to fit the boundary coordinate sequence. The first derivative of the fitted curve is calculated to determine the tangent direction of the curve. The tangent direction vectors at both ends of the crack region are used as the crack propagation direction vectors. The crack propagation direction vectors include the direction angle and the direction cosine value. The probability value distribution inside the crack region is extracted from the second crack detection heat map, and the probability value distribution is mapped to the corresponding crack depth value to generate a crack region depth distribution map. The crack depth value is positively correlated with the probability value. The depth distribution map of the crack region is processed by gradient calculation. The Sobel operator is used to calculate the gradient values ​​of the depth distribution map in the horizontal and vertical directions. The gradient values ​​are combined into a crack depth gradient vector. The magnitude and direction of the crack depth gradient vector are extracted to obtain the crack depth gradient value. The magnitude of the crack depth gradient value represents the rate of change of depth, and the direction represents the direction of increasing depth. The crack region boundary coordinate sequence, crack propagation direction vector, and crack depth gradient value are combined to form the crack state parameters of the equipment surface. The crack region boundary coordinate sequence is represented by pixel coordinates, the crack propagation direction vector is represented by a unit vector, and the crack depth gradient value is represented by gradient magnitude and gradient direction angle.

6. The intelligent equipment fatigue crack detection method based on deep learning according to claim 5, characterized in that, The process of binarizing the second crack detection heatmap, using an adaptive threshold segmentation algorithm to determine the segmentation threshold between the crack region and the background region, and marking regions with probability values ​​higher than the segmentation threshold as crack regions, yields a binary image of the crack on the equipment surface, including: The second crack detection heatmap is divided into multiple non-overlapping image sub-blocks; Calculate the mean gray value and standard deviation of gray value for each image sub-block. Use the product of the mean gray value and the standard deviation of gray value as the local threshold of the sub-block. The local threshold of the sub-block increases as the mean gray value of the sub-block increases and as the standard deviation of gray value increases. The local thresholds of all image sub-blocks are smoothed to obtain a smoothed local threshold matrix. The probability value of each pixel in the second crack detection heatmap is compared with the local threshold at the corresponding position in the smoothed local threshold matrix. If the probability value is higher than the local threshold, it is marked as a crack region pixel and assigned a value of 1; otherwise, it is marked as a background region pixel and assigned a value of 0, thus obtaining a binary image. The binary image is subjected to morphological processing to obtain a binary image of cracks on the equipment surface. The size of the structural element in the morphological processing is determined based on the average width of the crack region.

7. The intelligent equipment fatigue crack detection method based on deep learning according to claim 1, characterized in that, The process of generating an intelligent fatigue crack detection report based on the surface crack state parameters of the equipment includes: The crack region boundary coordinate sequence in the surface crack state parameters of the equipment is analyzed, and the crack region boundary coordinate sequence is converted from the image coordinate system to the equipment physical coordinate system to obtain the crack region boundary coordinates in the physical coordinate system. The equipment physical coordinate system takes the lower left corner of the bottom surface of the equipment as the origin, the X-axis is along the length direction of the equipment, and the Y-axis is along the width direction of the equipment. A schematic diagram of the spatial distribution of cracks is drawn based on the boundary coordinates of the crack region in the physical coordinate system. The position and shape of the crack region are marked in the two-dimensional structural diagram of the equipment using a polygon filling algorithm. The geometric center coordinates and maximum size parameter of the crack region are marked. The maximum size parameter is the maximum distance between two points in the boundary coordinate sequence of the crack region. Extract the crack propagation direction vector and crack depth gradient value from the crack state parameters on the equipment surface, and combine them with the stress distribution coefficient in the crack physical propagation constraint model to calculate the crack propagation risk index. The crack propagation risk index is the product of the modulus of the crack depth gradient value, the stress distribution coefficient, and the cosine of the angle between the propagation direction vector and the stress direction. The crack propagation risk level identifier is determined based on the crack propagation risk index. The risk index is divided into multiple consecutive intervals, each interval corresponds to a risk level, and each risk level corresponds to a preset risk description text. The crack spatial distribution diagram, crack propagation risk level identifier, and risk description text are integrated into the inspection report template. The inspection report template includes a basic equipment information column, a crack distribution column, a risk assessment column, and a treatment suggestion column. The basic equipment information column is filled with the equipment model and inspection time. The treatment suggestion column matches preset maintenance and treatment suggestions according to the risk level identifier to generate an intelligent inspection report of equipment fatigue cracks.

8. The intelligent equipment fatigue crack detection method based on deep learning according to claim 7, characterized in that, The step involves drawing a schematic diagram of the crack spatial distribution based on the crack region boundary coordinates in the physical coordinate system, and using a polygon filling algorithm to mark the location and shape of the crack region in the simplified two-dimensional structural diagram of the equipment, as well as marking the geometric center coordinates and maximum size parameters of the crack region, including: Extract the X-axis and Y-axis coordinate values ​​of each point in the crack region boundary coordinate sequence under the physical coordinate system, and perform coordinate calibration with the coordinate system of the equipment two-dimensional structural diagram to determine the corresponding position of each boundary coordinate point in the equipment two-dimensional structural diagram, thereby obtaining the crack region boundary coordinate sequence in the equipment two-dimensional structural diagram. Based on the coordinate sequence of the crack region boundary in the simplified two-dimensional structure diagram of the equipment, the coordinate points are connected in sequence to form a closed polygonal outline of the crack region. The vertex connection order of the polygonal outline is recorded to generate the polygonal outline data of the crack region. The polygon filling algorithm is called to perform internal filling processing on the polygon outline data of the crack area, and a fill color with a grayscale difference from the background of the equipment two-dimensional structural diagram is selected to generate the filled crack area graphic data. Perform region analysis on the filled crack region graphic data, calculate the average X-axis coordinate and average Y-axis coordinate of all coordinate points, and combine them to form the geometric center coordinates of the crack region; Traverse all coordinate points in the boundary coordinate sequence of the crack region under the physical coordinate system, calculate the straight-line distance between every two coordinate points, select the maximum straight-line distance as the maximum size parameter of the crack region, and record the corresponding positions of the two coordinate points corresponding to the maximum straight-line distance in the simplified two-dimensional structure diagram of the equipment. The filled crack area graphic data, geometric center coordinates, and coordinate point positions corresponding to the maximum size parameter are integrated into the equipment's two-dimensional structural diagram. The coordinate values ​​are marked at the geometric center coordinates, and a straight line is drawn between the two coordinate points corresponding to the maximum size parameter, with the maximum size parameter marked, to generate a crack spatial distribution diagram containing complete annotation information.

9. A deep learning-based intelligent detection system for equipment fatigue cracks, characterized in that, The device includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the deep learning-based intelligent detection method for equipment fatigue cracks as described in any one of claims 1-8.

Images