Fatigue crack growth identification method for aero-engine components based on machine vision

Abstract

The present invention discloses a method for identifying fatigue crack growth in aircraft engine components based on machine vision, comprising the following steps: S1, acquiring a sequence of image blocks; S2, inputting each image block into a pseudo-stress estimation network to generate an image block pseudo-stress map; S3, inputting the image block and the image block pseudo-stress map into an image feature decoupling encoding module to output a fused feature vector; S4, inputting the fused feature vector into an improved Fox algorithm to generate a crack path guidance map and a direction vector field; S5, inputting the image block pseudo-stress map, the crack path guidance map, the direction vector field, and the fused feature vector into a crack mask generation module to output an image block crack mask; S6, splicing the image block crack masks according to the positions of the image blocks in the surface image to generate a complete crack structure map; S7, outputting crack growth results based on the complete crack structure map. The present invention integrates machine vision with an evolutionary path optimization method to achieve accurate identification of aircraft engine crack growth.

Description

Technical Field

[0001] The present invention relates to the technical field of aircraft engine structural health monitoring, and in particular to a method for identifying fatigue crack propagation of aircraft engine parts based on machine vision. Background Art

[0002] As the power source of aircraft, the structural integrity of aircraft engines is directly related to flight safety and system stability. Under extreme operating conditions such as high temperature, high pressure, and alternating loads, key aircraft engine components (such as turbine blades, compressor discs, and casings) are highly susceptible to microcracks. These cracks gradually propagate under long-term fatigue, potentially leading to structural failure. Therefore, identifying and tracking fatigue crack growth in aircraft engine components is a core technical issue in aircraft maintenance, life prediction, and safety assessment.

[0003] Currently, fatigue crack detection primarily involves physical testing methods such as ultrasonic testing, magnetic particle testing, penetrant testing, X-ray testing, and laser confocal microscopy. However, these traditional nondestructive testing methods often rely on specialized equipment, require a demanding operating environment, and have long testing cycles, making them difficult to rapidly detect large, complex curved surfaces and continuously track crack propagation paths. Furthermore, these methods focus more on determining the presence of cracks rather than modeling and identifying their morphological evolution.

[0004] With the advancement of image processing and machine vision technologies, an increasing number of research efforts are employing computer vision methods for crack identification. Common approaches include image segmentation, crack region detection, and classification based on convolutional neural networks. These methods typically take the entire image as input, employing architectures such as U-Net, ResNet, or Mask R-CNN to extract features and output a crack mask. However, in complex industrial scenarios, these methods still suffer from issues such as unstable accuracy and the tendency to miss detections when dealing with small cracks, background interference, non-uniform lighting, or multi-scale morphological variations.

[0005] Some existing studies have attempted to improve crack detection precision by using image block segmentation and to enhance discrimination capabilities by leveraging physical guidance mechanisms such as pseudo-stress maps. However, current pseudo-stress modeling is often based on shallow CNN regression heatmaps, lacking the support of stress physical meaning. Feature extraction from image blocks also fails to effectively distinguish between image structural features and mechanically sensitive features. Furthermore, existing crack path modeling methods, mostly based on geometric line segment fitting or probabilistic map backtracking, are unable to form a structurally complete crack path topology, let alone extract physically interpretable path direction fields and branch connectivity relationships.

[0006] When generating crack masks, existing techniques often rely on direct mask regression from a single channel or feature map. These techniques lack path-based constraints and pseudo-crack suppression mechanisms, making it easy to misidentify scratches and dust as cracks. Furthermore, when stitching image blocks together, they lack fusion strategies for overlapping areas, leading to problems such as crack breakpoints and morphological distortion, which in turn affects the accuracy of subsequent topological reconstruction.

[0007] Therefore, how to provide a method for identifying fatigue crack growth in aircraft engine components based on machine vision is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] One purpose of the present invention is to propose a method for identifying fatigue crack propagation in aircraft engine components based on machine vision. The present invention integrates machine vision technology with an improved Fox algorithm. Through steps such as image block pseudo-stress modeling, image-stress feature fusion, crack path evolution identification, and topological structure extraction, the method achieves high-precision identification and structured expression of fatigue crack propagation in aircraft engine components. The method has the advantages of strong physical consistency, high identification accuracy, good path continuity, and strong crack morphology interpretability, and is suitable for crack monitoring and life assessment tasks in complex engineering environments.

[0009] A method for identifying fatigue crack growth in aircraft engine components based on machine vision according to an embodiment of the present invention includes the following steps:

[0010] S1. Standardizing a surface image of an aircraft engine component and dividing the surface image into a plurality of fixed-size image blocks with overlapping areas using a sliding window method to obtain an image block sequence;

[0011] S2. constructing a pseudo stress estimation network, and inputting each image block in the image block sequence into the pseudo stress estimation network to generate a corresponding pseudo stress map of the image block;

[0012] S3, inputting the image block and the corresponding image block pseudo stress map into an image feature decoupling coding module, wherein the image feature decoupling coding module includes an image feature extraction branch, a stress feature extraction branch, and a fusion branch, and outputs a fusion feature vector;

[0013] S4. Inputting the fused feature vector into an improved Fox algorithm, wherein the improved Fox algorithm selects an optimal crack path individual from a plurality of candidate crack path individuals based on an evolutionary strategy optimization method, and generates a crack path guidance graph and a direction vector field;

[0014] S5. Inputting the image block pseudo stress map, crack path guidance map, direction vector field, and fused feature vector into a crack mask generation module, wherein the crack mask generation module includes an edge detection submodule, a crack path direction submodule, and a pseudo crack suppression submodule, and outputs an image block crack mask;

[0015] S6. splicing the crack masks of the image blocks according to the spatial positions of the image blocks in the surface image, and fusing the overlapping areas using a confidence-weighted method to generate a complete crack structure map;

[0016] S7. Extracting the crack structure topology relationship based on the complete crack structure diagram and outputting the crack propagation result.

[0017] Optionally, the sliding window method sets the window step size to be smaller than the window size, so that overlapping areas exist between adjacent image blocks, and the overlapping areas are used to enhance feature continuity between image blocks.

[0018] Optionally, the S2 specifically includes:

[0019] S21, constructing a pseudo stress estimation network, wherein the pseudo stress estimation network includes a local feature differential encoding module, a stress concentration area enhancement module, and a simulated stress field decoding module;

[0020] S22, inputting the image block into a local feature differential encoding module, calculating the grayscale gradient difference between each pixel in the image block and its neighboring pixels, and forming a differential feature map reflecting local structural changes;

[0021] S23, the stress concentration area enhancement module designs a stress concentration filter kernel based on the prior knowledge of stress concentration that the stress field gradient is the largest at the crack tip and decreases toward the surrounding area. The stress concentration filter kernel is a two-dimensional Gaussian dynamic convolution kernel with the largest center weight and gradually decreases outward. The module performs nonlinear enhancement on the differential feature map to generate a stress enhancement feature map.

[0022] S24, inputting the stress enhancement feature map into the simulated stress field decoding module, and using the Fourier spectrum domain inverse transform method to obtain an initial simulated stress field distribution map with the same size as the image block;

[0023] S25, performing pixel-by-pixel amplitude normalization processing on the initial simulated stress field distribution map to generate an image block pseudo stress map with a value range between 0 and 1;

[0024] S26: Obtain an image block pseudo stress map corresponding to each image block in the image block sequence.

[0025] Optionally, the S3 specifically includes:

[0026] S31, inputting the image block into an image feature extraction branch, and using a direction-sensitive convolution operator to extract edge and texture features of the crack in the image block, wherein the direction-sensitive convolution operator includes multiple oriented gradient convolution kernels arranged along 0°, 45°, 90°, and 135° directions, and is used to enhance the response of the slender structure in the image block in different directions, thereby generating an image saliency feature map;

[0027] S32. Inputting the pseudo stress map of the image block into a stress feature extraction branch, and using a multi-scale dilated convolution operator to extract the morphology and range features of the stress concentration area. The multi-scale dilated convolution operator includes parallel convolution paths with different dilation rates. The dilation rate refers to the number of pixel intervals between consecutive sampling points in the convolution kernel, and is used to expand the receptive field without changing the convolution kernel size, and output a stress significance feature map.

[0028] S33, the fusion branch introduces a cross-domain attention mechanism, and generates a pixel-level cross-attention weight matrix by calculating the similarity between the image saliency feature map and the stress saliency feature map in the spatial dimension;

[0029] S34, based on the pixel-level cross attention weight matrix, performing pixel-by-pixel weighted fusion on the image saliency feature map and the stress saliency feature map to obtain a fused feature map;

[0030] S35. Perform a flattening operation on the fused feature map, wherein the flattening operation includes expanding the three-dimensional tensor structure with height, width and number of channels in the fused feature map into a one-dimensional vector in sequence according to channel priority, forming a fused feature vector and outputting it.

[0031] Optionally, the S4 specifically includes:

[0032] S41, through the improved fox algorithm, the fused feature vector is 2D projected and coded to construct the initial population set of candidate crack paths , each candidate crack path individual Represented as a sequence of continuous pixel coordinates ,in Indicates the The coordinates of pixels are used to describe the crack propagation path within the image block;

[0033] S42. Define a fitness function for each candidate crack path individual:

[0034] ;

[0035] in, Represents the candidate crack path individuals The fitness value of 、 and represents the weight coefficient, In the pseudo stress map of the image block corresponding to the fused feature vector, the pixel The stress intensity value at the coordinate, Represents the candidate crack path individuals The path length, Represents the candidate crack path individuals The set of angles between the paths, represents the variance of the path direction change;

[0036] S43, arrange the fitness values ​​of all candidate crack path individuals in descending order, and select the first The candidate crack path individuals with fitness values ​​constitute the elite path individual set ;

[0037] S44, execute the elite path individual Perform crossover operations on the parent path individuals and generate new path individuals by alternately extracting path segments:

[0038] ;

[0039] in, represents the new path individual, is the parent path individual, express From the first point to the The coordinate sequence of the points, express No. Point to The coordinate sequence of the points, Indicates path segment concatenation, Indicates the total number of waypoints;

[0040] S45, adding Gaussian noise perturbation to each coordinate point in the new path individual to achieve path variation;

[0041] S46, repeating steps S42 to S45 until the number of evolutionary rounds reaches a preset maximum value, and finally selecting the candidate crack path individual with the largest fitness as the optimal crack path individual;

[0042] S47, drawing path pixels on a mask of the same size as the original image block according to the continuous pixel coordinate sequence of the optimal crack path individual, to construct a crack path guidance map;

[0043] S48. Calculate the direction vectors of adjacent coordinate points of the optimal crack path individual, normalize them into direction unit vectors, map each direction unit vector to a corresponding pixel position, and construct a direction vector field.

[0044] Optionally, the S5 specifically includes:

[0045] S51, the image block pseudo stress map is input to the edge detection submodule, using the Canny edge detection algorithm to extract the edge of the image block pseudo stress map to obtain an edge response map;

[0046] S52: Input the crack path guidance map and the direction vector field into the crack path direction submodule, define the effective range of the edge response map using the crack path guidance map, perform a direction consistency check on each pixel within the effective range using the direction vector field, retain only edge pixels that are consistent with the path direction, and output a path direction enhancement map;

[0047] S53: Input the fused feature vector to a pseudo-crack suppression submodule, calculate the Euclidean distance between the feature value of each pixel in the fused feature vector and the edge pixel value in the corresponding path direction enhancement image, use the Euclidean distance as a judgment basis, filter out pixel areas whose Euclidean distance is greater than a set distance threshold, and output a pseudo-crack suppression image;

[0048] S54, fusing the edge response map, the path direction enhancement map, and the pseudo crack suppression map into a preliminary crack mask map by using a pixel-by-pixel weighted summation method;

[0049] S55 , performing binarization processing on the preliminary crack mask image, eliminating small holes and noise areas, and outputting an image block crack mask.

[0050] Optionally, the S6 specifically includes:

[0051] S61, re-mapping the crack mask of each image block to the corresponding surface image position according to the initial division coordinates of the image block in the sliding window method, and aligning and splicing the image blocks according to the original coordinates in a blank matrix of the original image size;

[0052] S62: For the overlapping areas between the image blocks, multiple pixel values ​​from the crack masks of different image blocks at the same pixel position are fused, and a confidence-weighted average method is used to perform fusion calculation to obtain the final mask value of each overlapping pixel point:

[0053] ;

[0054] in, Represents pixel points The final mask value of represents the total number of image blocks, Indicates the Image blocks at position The crack mask value on Indicates the Image blocks at position Confidence weights for the crack mask values ​​on :

[0055] ;

[0056] in, represents the exponential function, Represents pixel points To The Euclidean distance between the centers of the image patches, represents the Gaussian distance attenuation coefficient;

[0057] S63. The final mask value of each pixel after all fusion is combined into a complete crack structure map.

[0058] Optionally, the S7 specifically includes:

[0059] S71, performing a binarization operation on the complete crack structure image, marking pixels whose final mask values ​​are greater than a preset mask threshold as crack pixels, and marking the rest as background pixels, to form a binary crack mask image;

[0060] S72, extracting all connected regions in the binary crack mask image, and defining each connected region as a crack substructure;

[0061] S73, performing a skeleton extraction operation on each crack substructure, and extracting a crack path centerline with a single pixel width using a distance transform;

[0062] S74, based on the center line of the crack path, identifying the coordinates of the starting point and the ending point in the crack path, and calculating the direction vector of the crack path, defining the direction vector as a normalized vector from the starting point to the ending point of the path;

[0063] S75. If a single crack path has multiple branches, record the pixel positions of all branch points and represent the connection relationship between each branch and the main path in the form of a graph structure;

[0064] S76. Output the crack propagation result, where the crack propagation result includes the main crack path, branch structure, direction vector, and coordinates of the starting point and the ending point.

[0065] The beneficial effects of the present invention are:

[0066] First, the present invention introduces a stress perception mechanism through a pseudo-stress estimation network. It does not rely on finite element simulation input, but instead uses local differential coding within image blocks and prior knowledge of stress concentration areas to construct a two-dimensional Gaussian filter kernel, thereby improving the physical consistency and mechanical interpretability of the pseudo-stress map, making the model sensitive to the stress response differences between real cracks and surface artifacts, and effectively enhancing the distinguishability of cracks.

[0067] Secondly, in the feature encoding stage, the image feature extraction branch and the stress feature extraction branch are decoupled and modeled, and the spatial consistency fusion of multi-source features is achieved by introducing a cross-domain attention mechanism. This overcomes the problem that image texture and mechanical distribution cannot be effectively fused in traditional methods, and improves the ability of the fused features to express the true morphology of the crack.

[0068] In addition, during the crack path modeling process, the improved Fox algorithm designed in the present invention defines a fitness function that jointly optimizes the fusion stress intensity, path length and direction consistency, and combines the path individual crossover and mutation strategy to realize the structured evolutionary modeling of the crack path. It can not only output the spatial coordinate sequence of the main crack path, but also simultaneously generate the direction vector field, providing physical guidance for subsequent mask generation and topological analysis.

[0069] Furthermore, in the mask generation stage, by introducing the crack path guidance map and direction vector field, the candidate area is limited based on edge detection, and the fusion features and Euclidean distance threshold are combined to suppress pseudo-crack interference, realizing the multi-source information constraint generation of the crack contour, effectively improving the detection accuracy and anti-interference ability.

[0070] Finally, during the block-level crack mask fusion stage, the present invention proposes a confidence-weighted fusion strategy based on the Gaussian distance decay function. This effectively addresses the issues of information redundancy and boundary fractures in overlapping image blocks, resulting in a complete crack structure map with enhanced coherence and structural fidelity. Furthermore, skeleton extraction and topological analysis of the complete crack map identify the main crack path, branching structure, and direction vector, outputting crack propagation results with structural semantics, supporting higher-level structural integrity assessment and crack growth prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0072] Figure 1 This is the overall flow chart of the machine vision-based fatigue crack growth identification method for aircraft engine parts proposed by the present invention;

[0073] Figure 2 This is a schematic diagram of the structure of the pseudo-stress estimation network of the machine vision-based fatigue crack growth identification method for aircraft engine parts proposed in the present invention;

[0074] Figure 3 This is a flowchart of the individual evolution of crack paths using the improved Fox algorithm for the machine vision-based fatigue crack propagation identification method for aircraft engine components proposed in the present invention. DETAILED DESCRIPTION

[0075] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0076] refer to Figure 1-Figure 3 , a method for identifying fatigue crack growth of aircraft engine parts based on machine vision, including the following steps:

[0077] S1. Standardizing a surface image of an aircraft engine component and dividing the surface image into a plurality of fixed-size image blocks with overlapping areas using a sliding window method to obtain an image block sequence;

[0078] S2. constructing a pseudo stress estimation network, and inputting each image block in the image block sequence into the pseudo stress estimation network to generate a corresponding pseudo stress map of the image block;

[0079] S3, inputting the image block and the corresponding image block pseudo stress map into an image feature decoupling coding module, wherein the image feature decoupling coding module includes an image feature extraction branch, a stress feature extraction branch, and a fusion branch, and outputs a fusion feature vector;

[0080] S4. Inputting the fused feature vector into an improved Fox algorithm, wherein the improved Fox algorithm selects an optimal crack path individual from a plurality of candidate crack path individuals based on an evolutionary strategy optimization method, and generates a crack path guidance graph and a direction vector field;

[0081] S5. Inputting the image block pseudo stress map, crack path guidance map, direction vector field, and fused feature vector into a crack mask generation module, wherein the crack mask generation module includes an edge detection submodule, a crack path direction submodule, and a pseudo crack suppression submodule, and outputs an image block crack mask;

[0082] S6. splicing the crack masks of the image blocks according to the spatial positions of the image blocks in the surface image, and fusing the overlapping areas using a confidence-weighted method to generate a complete crack structure map;

[0083] S7. Extracting the crack structure topology relationship based on the complete crack structure diagram and outputting the crack propagation result.

[0084] The present invention constructs a bottom-up crack propagation identification mechanism by introducing a multi-stage process including pseudo-stress estimation, image block decoupling encoding, crack path modeling, and topological structure extraction. Unlike traditional black-box detection based on image segmentation, this method embeds mechanics-inspired pseudo-stress modeling and crack growth trend reasoning while retaining image details, so that the recognition results not only have contour morphology, but also have directional consistency and structural logic. The entire process forms a closed-loop recognition channel from image blocks to topological structures, significantly improving the continuity, accuracy, and interpretability of crack propagation modeling, and is particularly suitable for engineering scenarios with extremely high structural safety requirements, such as aircraft engines.

[0085] In this embodiment, the sliding window method sets the window step size to be smaller than the window size, so that overlapping areas exist between adjacent image blocks. The overlapping areas are used to enhance the feature continuity between the image blocks.

[0086] Using a sliding approach with a window step size smaller than the window size effectively introduces overlapping regions between image blocks, allowing crack information from adjacent areas to reappear in multiple image blocks, thereby enhancing the model's ability to learn crack boundary regions. By capturing redundant features and fusing confidence scores within overlapping regions, false detections of fractures and splits can be reduced during crack recognition, improving the spatial coherence and precision of recognition. This strategy achieves local redundancy enhancement while ensuring complete image coverage, making it suitable for high-fidelity reconstruction of high-resolution crack structure maps.

[0087] In this embodiment, S2 specifically includes:

[0088] S21, constructing a pseudo stress estimation network, wherein the pseudo stress estimation network includes a local feature differential encoding module, a stress concentration area enhancement module, and a simulated stress field decoding module;

[0089] S22, inputting the image block into a local feature differential encoding module, calculating the grayscale gradient difference between each pixel in the image block and its neighboring pixels, and forming a differential feature map reflecting local structural changes;

[0090] S23, the stress concentration area enhancement module designs a stress concentration filter kernel based on the prior knowledge of stress concentration that the stress field gradient is the largest at the crack tip and decreases toward the surrounding area. The stress concentration filter kernel is a two-dimensional Gaussian dynamic convolution kernel with the largest center weight and gradually decreases outward. The module performs nonlinear enhancement on the differential feature map to generate a stress enhancement feature map.

[0091] S24, inputting the stress enhancement feature map into the simulated stress field decoding module, and using the Fourier spectrum domain inverse transform method to obtain an initial simulated stress field distribution map with the same size as the image block;

[0092] S25, performing pixel-by-pixel amplitude normalization processing on the initial simulated stress field distribution map to generate an image block pseudo stress map with a value range between 0 and 1;

[0093] S26: Obtain an image block pseudo stress map corresponding to each image block in the image block sequence.

[0094] A pseudo-stress estimation network is constructed within the image block, and a pseudo-stress map is generated by combining differential encoding with stress concentration priors. This eliminates the need for finite element simulation inputs to infer the stress response from the image itself. A Gaussian stress filter kernel significantly enhances the crack tip stress region, improving the model's ability to focus on high-risk areas. The introduction of Fourier frequency-domain inversion methods makes the stress field representation more continuous and physically consistent. The resulting pseudo-stress map for the image block has morphological features that closely match the actual crack stress response, significantly enhancing the reliability of subsequent path modeling and discrimination.

[0095] In this embodiment, S3 specifically includes:

[0096] S31, inputting the image block into an image feature extraction branch, and using a direction-sensitive convolution operator to extract edge and texture features of the crack in the image block, wherein the direction-sensitive convolution operator includes multiple oriented gradient convolution kernels arranged along 0°, 45°, 90°, and 135° directions, and is used to enhance the response of the slender structure in the image block in different directions, thereby generating an image saliency feature map;

[0097] S32. Inputting the pseudo stress map of the image block into a stress feature extraction branch, and using a multi-scale dilated convolution operator to extract the morphology and range features of the stress concentration area. The multi-scale dilated convolution operator includes parallel convolution paths with different dilation rates. The dilation rate refers to the number of pixel intervals between consecutive sampling points in the convolution kernel, and is used to expand the receptive field without changing the convolution kernel size, and output a stress significance feature map.

[0098] S33, the fusion branch introduces a cross-domain attention mechanism, and generates a pixel-level cross-attention weight matrix by calculating the similarity between the image saliency feature map and the stress saliency feature map in the spatial dimension;

[0099] S34, based on the pixel-level cross attention weight matrix, performing pixel-by-pixel weighted fusion on the image saliency feature map and the stress saliency feature map to obtain a fused feature map;

[0100] S35. Perform a flattening operation on the fused feature map, wherein the flattening operation includes expanding the three-dimensional tensor structure with height, width and number of channels in the fused feature map into a one-dimensional vector in sequence according to channel priority, forming a fused feature vector and outputting it.

[0101] Using dual-channel inputs of images and pseudo-stress maps, the method enhances feature expression capabilities through a decoupled extraction and fusion mechanism. The image feature branch utilizes direction-sensitive convolution to effectively identify elongated crack textures, while the stress feature branch utilizes multi-scale dilated convolution to expand the receptive field, helping to capture stress distribution structures at different scales. A pixel-level cross-attention mechanism is introduced into the fusion process, enabling precise spatial alignment and fusion of the two feature types. The resulting fused feature vector not only retains texture information but also incorporates mechanical response properties, providing stronger support for the joint physical and visual expression of path modeling.

[0102] In this embodiment, the S4 specifically includes:

[0103] S41, through the improved fox algorithm, the fused feature vector is 2D projected and coded to construct the initial population set of candidate crack paths , each candidate crack path individual Represented as a sequence of continuous pixel coordinates ,in Indicates the The coordinates of pixels are used to describe the crack propagation path within the image block;

[0104] S42. Define a fitness function for each candidate crack path individual:

[0105] ;

[0106] in, Represents the candidate crack path individuals The fitness value of 、 and represents the weight coefficient, In the pseudo stress map of the image block corresponding to the fused feature vector, the pixel The stress intensity value at the coordinate, Represents the candidate crack path individuals The path length, Represents the candidate crack path individuals The set of angles between the paths, represents the variance of the path direction change;

[0107] S43, arrange the fitness values ​​of all candidate crack path individuals in descending order, and select the first The candidate crack path individuals with fitness values ​​constitute the elite path individual set ;

[0108] S44, execute the elite path individual Perform crossover operations on the parent path individuals and generate new path individuals by alternately extracting path segments:

[0109] ;

[0110] in, represents the new path individual, is the parent path individual, express From the first point to the The coordinate sequence of the points, express No. Point to The coordinate sequence of the points, Indicates path segment concatenation, Indicates the total number of waypoints;

[0111] S45, adding Gaussian noise perturbation to each coordinate point in the new path individual to achieve path variation;

[0112] S46, repeating steps S42 to S45 until the number of evolutionary rounds reaches a preset maximum value, and finally selecting the candidate crack path individual with the largest fitness as the optimal crack path individual;

[0113] S47, drawing path pixels on a mask of the same size as the original image block according to the continuous pixel coordinate sequence of the optimal crack path individual, to construct a crack path guidance map;

[0114] S48. Calculate the direction vectors of adjacent coordinate points of the optimal crack path individual, normalize them into direction unit vectors, map each direction unit vector to a corresponding pixel position, and construct a direction vector field.

[0115] An evolutionary strategy is introduced to construct a crack path search mechanism. A fitness function is constructed by combining pseudo-stress response, path length, and directional consistency, making path evolution more goal-oriented. Through individual path crossover and Gaussian perturbation mutation operations, the model is capable of maintaining diversity and optimizing local structures. Ultimately, an optimal path sequence and corresponding direction field are generated, capturing the true crack extension trend while providing clear path guidance information.

[0116] The improved fox algorithm employed is an evolutionary optimization method inspired by the foraging behavior of foxes in nature and specifically improved in this invention. Its core concept is to simulate the behavioral strategy of foxes in searching for crack paths in different regions, balancing exploration and convergence through a fitness-driven path evolution mechanism. The algorithm has the ability to encode high-dimensional path structures and can represent and optimize continuous crack pixel trajectories in image blocks. During the evolutionary process, a constraint term based on the variance of path direction changes is introduced, making the generated crack paths more directional stable and physically reasonable, making them particularly suitable for modeling crack structures with slender morphologies and well-defined directional extensions.

[0117] Compared with traditional crack extraction methods based on heat map backtracking or probability graph tracking, the improvement of the Fox algorithm in the present invention not only achieves the traceability and continuity of the path structure, but also provides basic information with clear structure, clear direction, and explainable behavior for subsequent mask generation and topological analysis through the joint output of the path and direction field, significantly improving the recognition ability and credibility of the crack extension state.

[0118] In this embodiment, the S5 specifically includes:

[0119] S51, the image block pseudo stress map is input to the edge detection submodule, using the Canny edge detection algorithm to extract the edge of the image block pseudo stress map to obtain an edge response map;

[0120] S52: Input the crack path guidance map and the direction vector field into the crack path direction submodule, define the effective range of the edge response map using the crack path guidance map, perform a direction consistency check on each pixel within the effective range using the direction vector field, retain only edge pixels that are consistent with the path direction, and output a path direction enhancement map;

[0121] S53: Input the fused feature vector to a pseudo-crack suppression submodule, calculate the Euclidean distance between the feature value of each pixel in the fused feature vector and the edge pixel value in the corresponding path direction enhancement image, use the Euclidean distance as a judgment basis, filter out pixel areas whose Euclidean distance is greater than a set distance threshold, and output a pseudo-crack suppression image;

[0122] S54, fusing the edge response map, the path direction enhancement map, and the pseudo crack suppression map into a preliminary crack mask map by using a pixel-by-pixel weighted summation method;

[0123] S55 , performing binarization processing on the preliminary crack mask image, eliminating small holes and noise areas, and outputting an image block crack mask.

[0124] Crack mask generation no longer relies on a single image channel. Instead, it fuses pseudo-stress maps, crack path guidance maps, directional vector fields, and image-stress fusion features for joint modeling, resulting in stronger structural constraints. During the generation process, a three-stage, layer-by-layer filtering process involving edge detection, directional guidance, and pseudo-crack removal significantly enhances the accuracy and physical consistency of the results. In particular, the use of directional vector fields for consistency verification of edge pixels and the application of fused features for Euclidean distance constraints effectively avoid the misidentification of pseudo-cracks such as scratches and texture faults, improving mask quality and practicality.

[0125] In this embodiment, S6 specifically includes:

[0126] S61, re-mapping the crack mask of each image block to the corresponding surface image position according to the initial division coordinates of the image block in the sliding window method, and aligning and splicing the image blocks according to the original coordinates in a blank matrix of the original image size;

[0127] S62: For the overlapping areas between the image blocks, multiple pixel values ​​from the crack masks of different image blocks at the same pixel position are fused, and a confidence-weighted average method is used to perform fusion calculation to obtain the final mask value of each overlapping pixel point:

[0128] ;

[0129] in, Represents pixel points The final mask value of represents the total number of image blocks, Indicates the Image blocks at position The crack mask value on Indicates the Image blocks at position Confidence weights for the crack mask values ​​on :

[0130] ;

[0131] in, represents the exponential function, Represents pixel points To The Euclidean distance between the centers of the image patches, represents the Gaussian distance attenuation coefficient;

[0132] S63. The final mask value of each pixel after all fusion is combined into a complete crack structure map.

[0133] When stitching block-level crack masks into a complete image, a spatial alignment and confidence-weighted overlapping region mechanism are introduced to avoid the edge faults and crack misalignment common in traditional block stitching. A Gaussian attenuation function based on Euclidean distance is used to dynamically assign fusion weights to overlapping regions. This gives pixels near the center of the block a higher weight in the fusion, while weights at the edges are moderately attenuated, thereby improving the consistency and structural continuity of the fused image. The resulting crack structure image is spatially coherent and morphologically realistic, demonstrating enhanced ability to reconstruct the entire crack landscape.

[0134] In this embodiment, the S7 specifically includes:

[0135] S71, performing a binarization operation on the complete crack structure image, marking pixels whose final mask values ​​are greater than a preset mask threshold as crack pixels, and marking the rest as background pixels, to form a binary crack mask image;

[0136] S72, extracting all connected regions in the binary crack mask image, and defining each connected region as a crack substructure;

[0137] S73, performing a skeleton extraction operation on each crack substructure, and extracting a crack path centerline with a single pixel width using a distance transform;

[0138] S74, based on the center line of the crack path, identifying the coordinates of the starting point and the ending point in the crack path, and calculating the direction vector of the crack path, defining the direction vector as a normalized vector from the starting point to the ending point of the path;

[0139] S75. If a single crack path has multiple branches, record the pixel positions of all branch points and represent the connection relationship between each branch and the main path in the form of a graph structure;

[0140] S76. Output the crack propagation result, where the crack propagation result includes the main crack path, branch structure, direction vector, starting point, and end point coordinates.

[0141] After reconstructing the crack structure graph, the crack region is converted into a graph structure through connected domain analysis and skeleton extraction, which then extracts the starting point, end point, direction vector, and branch topology information. This processing method not only preserves the crack morphology but also imparts quantifiable structural semantics, facilitating subsequent crack propagation trend analysis and growth modeling. Compared to simple mask detection or line segment regression methods, this method supports topological hierarchical information modeling, possesses the ability to identify main paths and describe branch connections, and provides reliable input data for structural health assessment and life prediction.

[0142] Example 1:

[0143] To verify the feasibility of this invention, it was applied to a periodic maintenance mission on a turbofan engine from China Aviation Industry Corporation. This engine had accumulated over 3,500 flight hours. Its critical hot-end structures—the high-pressure turbine guide vanes, the front casing, and the compressor rotor disk—were at risk of fatigue crack initiation under long-term thermal shock. Traditional crack detection, which primarily relies on magnetic particle and fluorescent penetrant testing, is not only time-consuming and highly dependent on operator experience, but also difficult to quantify crack continuity and growth.

[0144] In this example, a mobile inspection platform equipped with a high-resolution industrial camera (4096×3072 resolution, 30Hz sampling frequency) was used to capture images of the target component surface. Uniform illumination was maintained during the acquisition process, and a flexible light source was used to reduce reflection interference. The resulting image was normalized and then partitioned into image blocks using the sliding window mechanism described in this invention. The window size was set to 256×256 pixels with a step size of 128 pixels, ensuring 50% overlap between image blocks, thereby enhancing the robustness of continuous feature extraction of crack boundaries.

[0145] During image block processing, a pseudo-stress estimation network is first constructed to predict the pseudo-stress field for each image block. This network takes a standardized image block as input and outputs a normalized stress response heat map. It employs an improved local differential encoding and a Gaussian stress enhancement kernel to effectively amplify the stress response in the suspected crack tip region. During network training, the pseudo-stress network is fine-tuned on publicly available material fatigue crack image datasets (such as CFD-Crack500) and a dataset of field-measured images from our laboratory. The training set consists of approximately 4,800 image blocks, and the test set consists of 1,200. The network ultimately achieves an average pixel error of less than 0.04 on the validation set.

[0146] Next, the image block and the corresponding pseudo-stress map are fed into an image feature decoupling encoding module, which contains two sub-paths: the image channel and the stress channel. A high-dimensional fused feature vector is generated through a cross-domain attention mechanism and used as input for crack path modeling. The improved Fox algorithm of the present invention is used to perform path individual evolutionary search, with the initial population size set to 100 and the maximum number of iterations set to 50. The fitness function comprehensively considers the total amount of pseudo-stress response, path length, and directional consistency. In each round of evolution, new individuals are generated using path segment intersection and Gaussian perturbation, and elite individuals are retained based on the path score. The final output is a path guidance map and a direction vector field, which serve as a key reference for guiding mask generation.

[0147] During crack mask generation, a strategy combining edge detection and directional consistency constraints is used to construct a mask image. False crack regions are then removed using the Euclidean distance between the fused feature vector and the directional map. Gaussian-weighted fusion and splicing are used to restore the complete crack structure image for each image block mask in overlapping areas. Finally, skeleton extraction and topological mapping are used to extract topological information, including the main crack path, branches, starting and ending points, and directional vectors, as the output.

[0148] In this mission, the proposed method processed three engine components, capturing 1,920 images and generating approximately 13,600 image blocks. Ultimately, 38 high-risk cracks were identified, each containing an average of 156 path pixels and possessing two to four natural branches. Compared to manual inspection, this method achieved an average crack integrity rate of 96.3%, with a missed detection rate below 3.2%. This method improves recognition accuracy by approximately 12.5% ​​compared to traditional image semantic segmentation models.

[0149] This embodiment fully verifies the feasibility and effectiveness of the proposed machine vision-based fatigue crack propagation identification method for aircraft engine parts in actual engineering environments. By integrating pseudo-stress estimation, image-stress feature decoupling coding, improved Fox algorithm path modeling, and crack topology structure recognition, this method not only achieves high-precision and high-robustness crack detection, but also can output structured topological information including main paths, branches, direction vectors, and start and end points. Compared with traditional crack detection methods based on image semantic segmentation or heat map backtracking, this embodiment shows significant advantages in crack recognition completeness, missed detection rate, and interpretability. It is suitable for fatigue monitoring tasks such as aircraft engines that have extremely high safety and precision requirements, and provides a solid data foundation for subsequent crack evolution modeling and life prediction.

[0150] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for identifying fatigue crack growth in aircraft engine parts based on machine vision, characterized in that: The steps include: S1. Standardizing a surface image of an aircraft engine component and dividing the surface image into a plurality of fixed-size image blocks with overlapping areas using a sliding window method to obtain an image block sequence; S2. constructing a pseudo stress estimation network, and inputting each image block in the image block sequence into the pseudo stress estimation network to generate a corresponding pseudo stress map of the image block; S3, inputting the image block and the corresponding image block pseudo stress map into an image feature decoupling coding module, wherein the image feature decoupling coding module includes an image feature extraction branch, a stress feature extraction branch, and a fusion branch, and outputs a fusion feature vector; S4. Inputting the fused feature vector into an improved Fox algorithm, wherein the improved Fox algorithm selects an optimal crack path individual from a plurality of candidate crack path individuals based on an evolutionary strategy optimization method, and generates a crack path guidance graph and a direction vector field; S5. Inputting the image block pseudo stress map, crack path guidance map, direction vector field, and fused feature vector into a crack mask generation module, wherein the crack mask generation module includes an edge detection submodule, a crack path direction submodule, and a pseudo crack suppression submodule, and outputs an image block crack mask; S6. splicing the crack masks of the image blocks according to the spatial positions of the image blocks in the surface image, and fusing the overlapping areas using a confidence-weighted method to generate a complete crack structure map; S7. Extracting the crack structure topology relationship based on the complete crack structure diagram and outputting the crack propagation result.

2. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The sliding window method sets the window step size to be smaller than the window size, so that there is an overlapping area between adjacent image blocks, and the overlapping area is used to enhance the feature continuity between the image blocks.

3. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The S2 specifically includes: S21, constructing a pseudo stress estimation network, wherein the pseudo stress estimation network includes a local feature differential encoding module, a stress concentration area enhancement module, and a simulated stress field decoding module; S22, inputting the image block into a local feature differential encoding module, calculating the grayscale gradient difference between each pixel in the image block and its neighboring pixels, and forming a differential feature map reflecting local structural changes; S23, the stress concentration area enhancement module designs a stress concentration filter kernel based on the prior knowledge of stress concentration that the stress field gradient is the largest at the crack tip and decreases toward the surrounding area. The stress concentration filter kernel is a two-dimensional Gaussian dynamic convolution kernel with the largest center weight and gradually decreases outward. The module performs nonlinear enhancement on the differential feature map to generate a stress enhancement feature map. S24, inputting the stress enhancement feature map into the simulated stress field decoding module, and using the Fourier spectrum domain inverse transform method to obtain an initial simulated stress field distribution map with the same size as the image block; S25, performing pixel-by-pixel amplitude normalization processing on the initial simulated stress field distribution map to generate an image block pseudo stress map with a value range between 0 and 1; S26: Obtain an image block pseudo stress map corresponding to each image block in the image block sequence.

4. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The S3 specifically includes: S31, inputting the image block into an image feature extraction branch, and using a direction-sensitive convolution operator to extract edge and texture features of the crack in the image block, wherein the direction-sensitive convolution operator includes multiple oriented gradient convolution kernels arranged along 0°, 45°, 90°, and 135° directions, and is used to enhance the response of the slender structure in the image block in different directions, thereby generating an image saliency feature map; S32. Inputting the pseudo stress map of the image block into a stress feature extraction branch, and using a multi-scale dilated convolution operator to extract the morphology and range features of the stress concentration area. The multi-scale dilated convolution operator includes parallel convolution paths with different dilation rates. The dilation rate refers to the number of pixel intervals between consecutive sampling points in the convolution kernel, and is used to expand the receptive field without changing the convolution kernel size, and output a stress significance feature map. S33, the fusion branch introduces a cross-domain attention mechanism, and generates a pixel-level cross-attention weight matrix by calculating the similarity between the image saliency feature map and the stress saliency feature map in the spatial dimension; S34, based on the pixel-level cross attention weight matrix, performing pixel-by-pixel weighted fusion on the image saliency feature map and the stress saliency feature map to obtain a fused feature map; S35. Perform a flattening operation on the fused feature map, wherein the flattening operation includes expanding the three-dimensional tensor structure with height, width and number of channels in the fused feature map into a one-dimensional vector in sequence according to channel priority, forming a fused feature vector and outputting it.

5. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The S4 specifically includes: S41, through the improved fox algorithm, the fused feature vector is 2D projected and coded to construct the initial population set of candidate crack paths , each candidate crack path individual Represented as a sequence of continuous pixel coordinates ,in Indicates the The coordinates of pixels are used to describe the crack propagation path within the image block; S42. Define a fitness function for each candidate crack path individual: ； in, Represents the candidate crack path individuals The fitness value of 、 and represents the weight coefficient, In the pseudo stress map of the image block corresponding to the fused feature vector, the pixel The stress intensity value at the coordinate, Represents the candidate crack path individuals The path length, Represents the candidate crack path individuals The set of angles between the paths, represents the variance of the path direction change; S43, arrange the fitness values ​​of all candidate crack path individuals in descending order, and select the first The candidate crack path individuals with fitness values ​​constitute the elite path individual set ; S44, execute the elite path individual Perform crossover operations on the parent path individuals and generate new path individuals by alternately extracting path segments: ； in, represents the new path individual, is the parent path individual, express From the first point to the The coordinate sequence of the points, express No. point to The coordinate sequence of the points, Indicates path segment concatenation, Indicates the total number of waypoints; S45, adding Gaussian noise perturbation to each coordinate point in the new path individual to achieve path variation; S46, repeating steps S42 to S45 until the number of evolutionary rounds reaches a preset maximum value, and finally selecting the candidate crack path individual with the largest fitness as the optimal crack path individual; S47, drawing path pixels on a mask of the same size as the original image block according to the continuous pixel coordinate sequence of the optimal crack path individual, to construct a crack path guidance map; S48. Calculate the direction vectors of adjacent coordinate points of the optimal crack path individual, normalize them into direction unit vectors, map each direction unit vector to a corresponding pixel position, and construct a direction vector field.

6. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The S5 specifically includes: S51, the image block pseudo stress map is input to the edge detection submodule, using the Canny edge detection algorithm to extract the edge of the image block pseudo stress map to obtain an edge response map; S52: Input the crack path guidance map and the direction vector field into the crack path direction submodule, define the effective range of the edge response map using the crack path guidance map, perform a direction consistency check on each pixel within the effective range using the direction vector field, retain only edge pixels that are consistent with the path direction, and output a path direction enhancement map; S53: Input the fused feature vector to a pseudo-crack suppression submodule, calculate the Euclidean distance between the feature value of each pixel in the fused feature vector and the edge pixel value in the corresponding path direction enhancement image, use the Euclidean distance as a judgment basis, filter out pixel areas whose Euclidean distance is greater than a set distance threshold, and output a pseudo-crack suppression image; S54, fusing the edge response map, the path direction enhancement map, and the pseudo crack suppression map into a preliminary crack mask map by using a pixel-by-pixel weighted summation method; S55 , performing binarization processing on the preliminary crack mask image, eliminating holes and noise areas, and outputting an image block crack mask.

7. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The S6 specifically includes: S61, re-mapping the crack mask of each image block to the corresponding surface image position according to the initial division coordinates of the image block in the sliding window method, and aligning and splicing the image blocks according to the original coordinates in a blank matrix of the original image size; S62: For the overlapping areas between the image blocks, multiple pixel values ​​from the crack masks of different image blocks at the same pixel position are fused, and a confidence-weighted average method is used to perform fusion calculation to obtain the final mask value of each overlapping pixel point: ； in, Represents pixel points The final mask value of represents the total number of image blocks, Indicates the Image blocks at position The crack mask value on Indicates the Image blocks at position Confidence weights for the crack mask values ​​on : ； in, represents the exponential function, Represents pixel points To The Euclidean distance between the centers of the image patches, represents the Gaussian distance attenuation coefficient; S63. The final mask value of each pixel after all fusion is combined into a complete crack structure map.

8. The method for identifying fatigue crack growth of aircraft engine parts based on machine vision according to claim 1, characterized in that: The S7 specifically includes: S71, performing a binarization operation on the complete crack structure image, marking pixels whose final mask values ​​are greater than a preset mask threshold as crack pixels, and marking the rest as background pixels, to form a binary crack mask image; S72, extracting all connected regions in the binary crack mask image, and defining each connected region as a crack substructure; S73, performing a skeleton extraction operation on each crack substructure, and extracting a crack path centerline with a single pixel width using a distance transform; S74, based on the center line of the crack path, identifying the coordinates of the starting point and the ending point in the crack path, and calculating the direction vector of the crack path, defining the direction vector as a normalized vector from the starting point to the ending point of the path; S75. If a single crack path has multiple branches, record the pixel positions of all branch points and represent the connection relationship between each branch and the main path in the form of a graph structure; S76. Output the crack propagation result, where the crack propagation result includes the main crack path, branch structure, direction vector, starting point, and end point coordinates.

Images