A method for identifying surface cracks of single crystal turbine blades based on multispectral imaging

Abstract

The application discloses a single crystal turbine blade surface crack identification method based on multispectral imaging, and belongs to the field of nondestructive testing of an aero-engine. The method first collects ultraviolet, visible light and short-wave infrared multi-narrow-band original spectral images, generates a multispectral data cube through sub-pixel level spatial registration and reflectivity normalization; then extracts a pixel spectral response curve, calculates a spectral index and combines a spatial gradient feature to construct a multi-dimensional fusion feature map; a deep learning model embedded with a double attention mechanism is used to identify a crack candidate area, false defects are removed by using cooling hole coordinates and crystal orientation information extracted from a three-dimensional numerical model of the blade, and finally, accurate identification results are output, and crack geometric measurement and three-dimensional positioning report generation can be optionally realized. The method effectively suppresses high light reflection and structural interference, significantly improves the detection rate of micron-level hidden cracks, reduces the false detection rate, and adapts to the needs of automatic batch detection.

Description

Technical Field

[0001] This invention belongs to the field of non-destructive testing and intelligent identification technology for aero-engines, specifically a method for identifying surface cracks in single-crystal turbine blades based on multispectral imaging. Background Technology

[0002] Turbine blades are the most important hot-end components of aero engines, especially high-pressure turbine blades, which operate in extremely harsh environments, including high temperatures, high speeds, and high aerodynamic loads. Damage to high-pressure turbine blades will directly affect the safe and reliable operation of the aero engine, and may even lead to serious engine failure.

[0003] To improve the high-cycle fatigue resistance and high-temperature resistance of high-pressure turbine blades, a single-crystal material has been proposed for manufacturing engine blades in existing technologies, with single-crystal DD6 being one of the most representative materials. A significant characteristic of single-crystal blades is that the entire blade is a single grain, thus eliminating the boundary slip and dislocation phenomena found in polycrystalline materials. This significantly improves the fatigue strength of engine blades, and the high-temperature resistance of this single-crystal material is also significantly superior to that of nickel-based alloys.

[0004] However, the early identification of surface cracks in single-crystal DD6 material used in high-pressure turbine blades for aero-engines faces certain challenges. Traditional optical inspection methods suffer from insufficient sensitivity to latent cracks and susceptibility to interference from structures such as cooling holes when dealing with the highly reflective surfaces, complex geometries, and micron-level crack characteristics of single-crystal blades. Existing imaging methods based on a single visible light band have limitations in distinguishing real cracks from crystal textures or manufacturing traces, affecting the accuracy and reliability of the detection results.

[0005] Therefore, there is an urgent need to design a method that can comprehensively evaluate whether single-crystal DD6 material can be used for high-pressure turbine blades. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying surface cracks on single-crystal turbine blades based on multispectral imaging, in order to solve or mitigate the problems in the background art where traditional optical detection methods are insufficiently sensitive to hidden cracks, easily affected by structural interference such as cooling holes, and difficult to distinguish between real cracks and crystal textures or manufacturing traces under conditions of high reflectivity, complex geometry, and micron-level crack characteristics of single-crystal high-pressure turbine blades.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following solution:

[0008] A method for identifying surface cracks in single-crystal turbine blades based on multispectral imaging includes the following steps:

[0009] Step S1: Acquire the original spectral images of the single-crystal turbine blade to be tested in multiple preset narrow bands, and construct a multispectral image sequence;

[0010] Step S2: Perform sub-pixel-level spatial registration on the multispectral image sequence to generate a multispectral data cube;

[0011] Step S3: Extract the spectral response curve of each pixel in the multispectral data cube, calculate the spectral index characterizing the crack feature using the reflectance of the key band in the spectral response curve, and generate the corresponding spectral index map; at the same time, perform spatial gradient operation on the multispectral data cube to obtain the spatial gradient feature map; and concatenate the spectral index map and the spatial gradient feature map through channels to construct a multidimensional fusion feature map.

[0012] Step S4: Input the multi-dimensional fused feature map into a preset deep learning recognition model, and identify the crack candidate region through multi-scale feature extraction and attention mechanism;

[0013] Step S5: Based on the surface metallographic structure characteristics of the single-crystal turbine blade, perform false defect elimination on the crack candidate region and output the final surface crack identification result.

[0014] The core innovation of this method lies in addressing the industry pain points of detecting high reflectivity, complex structure, and micron-level cracks in single-crystal turbine blades. It constructs an innovative full-process system of "multi-dimensional data acquisition - precise feature fusion - intelligent recognition - targeted interference removal", breaking through the limitations of traditional single-band detection and manual interpretation. Its innovation is mainly reflected in three aspects: First, it innovatively adopts a multi-narrow band collaborative acquisition mode of ultraviolet, visible light, and short-wave infrared, combined with sub-pixel-level spatial registration and reflectivity normalization processing, which not only captures the weak scattering signal of crack openings, but also penetrates the covering layer to detect hidden cracks, solving the problem of insufficient sensitivity of traditional optical detection to hidden cracks; Second, it is the first to construct a multi-dimensional fusion feature map by channel splicing of spectral indices (based on the difference in reflectivity between crack and background bands) and spatial gradient features (Sobel gradient operation), and paired with an FPN deep learning model with embedded dual attention mechanism, it achieves precise coupling of crack spectral features and spatial edge features, which greatly improves the recognition sensitivity of micron-level cracks; Third, it innovatively designs a dual pseudo-defect removal mechanism based on the three-dimensional digital model of the blade, which accurately distinguishes non-crack structural interference from crystal texture noise through air film cooling hole spatial matching and crystal orientation vector determination, completely solving the pain point of traditional methods that are difficult to distinguish between real cracks and manufacturing traces and structural interference. At the same time, the optional three-dimensional geometric measurement and report generation links realize "identification-measurement-tracing". This engineering closed-loop system provides a new, automated, and high-precision solution for non-destructive testing of aero-engine blades.

[0015] Preferably, the preset narrow band in step S1 includes at least: an ultraviolet band of 320nm-400nm, used to capture weak scattering signals at the crack opening; a visible light band of 450nm-700nm, used to obtain the macroscopic morphology and color characteristics of the turbine blade surface; and a short-wave infrared band of 900nm-1700nm, used to detect the tendency of hidden cracks under the coating layer on the turbine blade surface.

[0016] The core innovation of this design lies in breaking through the functional limitations of traditional single-band detection. Addressing the pain points of multi-scenario crack detection in single-crystal turbine blades, it achieves precise selection and synergistic empowerment design across multiple narrow bands: ultraviolet, visible light, and short-wave infrared. Traditional detection methods often rely on a single visible light band, which cannot capture the weak signals at the opening of micron-level cracks, nor can it penetrate surface coatings to detect latent cracks, resulting in incomplete detection range and low sensitivity. This invention creatively matches dedicated narrow bands for different detection needs: the 320nm-400nm ultraviolet band precisely targets the weak scattered signals at the crack opening, solving the industry pain point of difficulty in capturing early micron-level crack signals; the 450nm-700nm visible light band ensures complete acquisition of the macroscopic morphology and color characteristics of the blade surface, providing a basic reference for subsequent feature comparison; and the 900nm-1700nm short-wave infrared band overcomes the obstruction limitations of surface coatings, enabling early detection of latent crack trends and filling the gap in traditional technology for subsurface defect detection. The three bands are not simply superimposed, but form a complementary and synergistic system of "weak signal capture - macroscopic feature acquisition - hidden defect detection". This system builds a full-dimensional and multi-layered original data foundation from the source, completely changing the situation of traditional single-band detection that "misses out on one thing". It provides a key prerequisite for subsequent multi-dimensional feature fusion and accurate identification, reflecting innovative thinking on the adaptation relationship between band functions and detection needs.

[0017] Preferably, step S2 further includes: setting a standard reflection reference plate, first obtaining the standard brightness value of the standard reflection reference plate in each narrow band, and then extracting the gray value of each pixel in the original spectral image of the single-crystal turbine blade under the narrow band; by calculating the ratio between the gray value of the pixel in the original image of the single-crystal turbine blade in the narrow band and the standard brightness value of the reference plate in the corresponding narrow band, the reflectivity of the original image of the single-crystal turbine blade is normalized, and finally the interference of light fluctuation and high reflectivity of the blade is eliminated, generating a multispectral data cube with a unified light reference.

[0018] The core innovation of this design lies in overcoming the technical limitations of traditional multispectral data calibration, which suffers from "uniform processing and poor band adaptability." It constructs an active standardization system of "precise band-specific calibration - pixel-level ratio calibration," fundamentally solving the data distortion problem caused by illumination fluctuations and high leaf reflectivity. Traditional testing often relies on a single benchmark for general calibration of all band images, failing to adapt to the differences in optical characteristics across different narrow bands, leading to crack feature signals being masked by environmental interference. This invention creatively employs a "one-to-one band" calibration logic: first, a dedicated standard is established for each preset narrow band; then, the standard brightness value of the corresponding band is obtained through a standard reflection reference plate; finally, the pixel grayscale value of the original spectral image of the leaf under test in that band is precisely extracted; and the reflectivity is normalized through the ratio calculation. This design is not a simple numerical adjustment but fully considers the differences in illumination response across different narrow bands, allowing each band's leaf image to eliminate interference with its own dedicated standard. The resulting multispectral data cube not only possesses a unified illumination benchmark but also restores the reflectivity data to its true physical meaning for each band. Its innovative value lies in upgrading the calibration logic from "passively adapting to the environment" to "actively defining standards." This not only solves the problem of insufficient multi-band adaptation in traditional calibration methods, but also provides a highly consistent and reliable data foundation for subsequent key steps such as spectral response curve extraction and spectral index calculation. It ensures the accuracy of crack identification from the source and reflects the deep optimization and innovation of the calibration logic for multispectral detection data.

[0019] Preferably, step S3 specifically includes:

[0020] The reflectance values ​​of each pixel in different spectral bands are extracted from the multispectral data cube to form the spectral response curve;

[0021] Selecting the crack-sensitive band λ1 and the background reference band λ2, the difference in reflectance between the cracked and normal regions is greatest under the crack-sensitive band λ1, while the turbine blade surface reflectance is stable and least affected by crystal orientation under the background reference band λ2. A spectral index SI is defined, which satisfies: SI = (R... λ1 - R λ2 ) / (R λ1 + R λ2 ), where R λ1 R represents the reflectivity of the crack-sensitive band. λ2 The reflectance of the background reference band is used; the corresponding values ​​of each pixel on the surface of the turbine blade under test are calculated according to the spectral index SI; the spectral index values ​​of all pixels are presented according to their spatial positions to generate a spectral index map, so as to enhance the difference in spectral energy between the cracked area and the normal area and achieve the initial enhancement of the crack signal.

[0022] The Sobel gradients of the multispectral data cube are calculated along the x and y directions in the spatial dimension to obtain the gradient magnitude and gradient direction, forming a spatial gradient feature map.

[0023] The spectral index map and the spatial gradient feature map are concatenated by channel stitching to construct a multidimensional fusion feature map.

[0024] The core innovation of this step lies in breaking through the technical bottleneck of traditional crack detection, which suffers from "single feature extraction and low signal recognition." It constructs an innovative logical chain of "precise spectral feature enhancement - refined spatial structure capture - deep fusion of two-dimensional features," achieving comprehensive and highly sensitive extraction of crack features. Traditional methods often rely on isolated spectral information or spatial morphology, making it difficult to consider both the spectral differences and edge structural features of cracks. Furthermore, crack signals are easily masked by crystal orientation and background noise. The innovation of this invention is reflected in three aspects: First, it innovatively designs a "targeted spectral feature extraction" mechanism to meet the needs of crack detection. This involves screening crack-sensitive bands (maximizing the difference in reflectivity between the crack and normal areas) and background reference bands (minimizing crystal orientation interference), and using SI=(R... λ1 -R λ2 ) / (R λ1 +R λ2 The design employs a multi-dimensional approach: 1) **Precisely amplifying the spectral energy difference between the cracked and normal regions using a ratio formula:** This effectively addresses the weakness and background noise of micrometer-level crack signals. 2) **Simultaneously introducing Sobel gradient operations:** This captures the gradient magnitude and direction information along the x and y directions, accurately extracting the crack's edge contour features and overcoming the limitation of single spectral features in depicting crack spatial morphology. 3) **Deep fusion of the spectral index map and spatial gradient feature map using channel splicing, rather than simple superposition, ensures the multi-dimensional fused feature map simultaneously covers both the "spectral distinction between crack and background" and the "spatial structure of the crack itself." This eliminates interference from crystal orientation and surface reflection while preserving the core features of the crack.** This design, combining "spectral targeted enhancement + fine spatial characterization + dual-dimensional deep fusion," achieves precise mining and integration of crack information at the feature extraction level. It completely overcomes the shortcomings of traditional single feature extraction, which suffers from "information gaps and weak anti-interference," providing high-quality, comprehensive feature support for accurate identification by subsequent deep learning models. This demonstrates a profound understanding and innovative transformation of crack detection feature requirements.

[0025] Preferably, in step S4, the deep learning recognition model adopts a feature pyramid network (FPN) structure, with the bottom layer composed of ResNet series residual blocks, and a channel attention module and a spatial attention module are introduced in the residual connections. The channel attention module is used to weight and highlight the crack-sensitive feature channels, and the spatial attention module is used to enhance the spatial saliency of the crack edge region, so as to achieve accurate recognition of crack candidate regions.

[0026] The core innovation of this design lies in overcoming the technical bottleneck of traditional deep learning models in crack identification, which suffers from "strong generalization but weak targeting" in feature extraction. Through a customized structural design combining "backbone network optimization + precise empowerment via dual attention mechanisms," it achieves high sensitivity and accuracy in identifying crack candidate regions. Traditional deep learning models often employ general network architectures for blade detection tasks. These models either struggle to capture micron-level cracks due to insufficient multi-scale feature fusion or suffer from crack signals being masked by background interference due to a lack of targeted feature enhancement mechanisms. Consequently, they are ill-suited to the characteristics of single-crystal turbine blade cracks: "weak, dispersed, and easily confused with structural noise."

[0027] The innovation of this invention lies in its triple collaborative design: First, it innovatively selects Feature Pyramid Network (FPN) as the core architecture. Its inherent multi-scale feature fusion capability can accurately cover cracks of different sizes (from micron-sized microcracks to more obvious cracks), solving the contradiction of traditional single-scale networks that "either miss small cracks or misjudge large-sized interferences." Second, it uses ResNet series residual blocks as the bottom backbone, and overcomes the gradient vanishing problem in deep networks through residual connections, ensuring that weak crack features can still be effectively transmitted in deep networks, laying a deep feature foundation for accurate identification. Third, it creatively introduces a "channel-space" dual attention module into the residual connections, rather than a simple superposition attention mechanism. The channel attention module can intelligently weight and highlight the sensitive feature channels that are key to crack identification, filtering out interference from irrelevant background channels. The spatial attention module specifically strengthens the spatial saliency of the crack edge region, solving the pain point of blurred crack edges and difficulty in distinguishing them from crystal texture boundaries.

[0028] This integrated design of "multi-scale fusion + deep feature preservation + dual-dimensional attention-targeted enhancement" is not a simple patchwork of existing technologies, but a customized innovation specifically for crack detection in single-crystal turbine blades. It upgrades the model from "passively extracting all features" to "actively focusing on core crack features," solving the problem of insufficient sensitivity of traditional models to weak crack features and significantly improving its resistance to interference from complex backgrounds. Ultimately, it achieves precise locking of crack candidate regions, providing high-quality candidate samples for subsequent false defect removal, demonstrating the deep adaptation and innovative integration of deep learning models with the specific needs of detection scenarios.

[0029] Preferably, the specific process of spurious defect removal in step S5 includes:

[0030] Verification A: Crack-free structure elimination based on spatial matching

[0031] Obtain the design three-dimensional digital model of the single-crystal turbine blade, and extract the coordinate distribution map of the film cooling holes and crystal orientation information from it;

[0032] Spatial matching is performed between the candidate crack region and the coordinate distribution map of the film cooling hole. If the candidate crack region is located within 1.5 times the hole diameter at the edge of the film cooling hole, and its circularity after Hough circle transformation is ≥0.85, it is marked as a non-crack structure and removed.

[0033] Verification B: Texture noise removal based on crystal orientation

[0034] The extension direction vector of the crack candidate region is calculated based on the crystal orientation information. If the angle between the extension direction vector and the crystal growth principal axis is less than 15°, it is determined to be crystal texture noise and is removed.

[0035] Finally, the cracked area that has been verified by both verification A and verification B is retained as the identification result.

[0036] The core innovation of this system lies in overcoming the limitations of traditional false defect removal methods, which rely on "single features and lack targeting." It constructs a "dual-target verification system based on a three-dimensional digital model of blade design," precisely addressing two major pain points in single-crystal turbine blade inspection: "interference from film cooling hole structure" and "confusion caused by crystal texture noise." Traditional inspection methods often rely on manual interpretation or screening for false defects using single image features. This approach is neither suitable for the complex geometry of blades nor can it distinguish the morphological differences between crystal textures and real cracks, resulting in a high false detection rate.

[0037] The innovation of this invention lies in its triple-targeted design: First, it innovatively uses the three-dimensional digital model of the blade design as the core basis for eliminating false defects, obtaining precise coordinate distribution and crystal orientation information of the film cooling holes from the source. This gives the elimination logic clear physical structural support, rather than simply relying on image pixel features, completely changing the traditional passive screening mode of "image-based analysis." Second, for interference from film cooling holes, it designs a dual judgment rule of "spatial position constraint + morphological feature verification." By locking the suspected interference area within a spatial range of 1.5 times the hole diameter, and then using the circularity of ≥0.85 detected by Hough circle transform as morphological evidence, it achieves accurate identification of false defects at the edge of the cooling holes, solving the problem that traditional methods struggle to distinguish between "cooling hole edge traces and short cracks." Third, for crystal texture noise, it creatively introduces "orientation vector angle judgment." The mechanism calculates the angle between the extension direction of the crack candidate region and the crystal growth axis, and identifies regions with an angle <15° as texture noise. It accurately grasps the core difference that "crystal texture is distributed along the growth axis and the actual crack extension direction is random", filling the gap that traditional technology cannot quantitatively distinguish between crystal texture and crack.

[0038] This integrated solution of "design data empowerment + dual-targeted verification" is not a simple combination of existing screening methods, but an innovative logic tailored to the structural characteristics and metallographic features of single-crystal turbine blades. It upgrades the elimination of false defects from "fuzzy screening" to "precise judgment," achieving comprehensive coverage of the two main sources of interference while ensuring that genuine crack areas are not mistakenly deleted. Ultimately, it controls the false detection rate to within 2.5%, providing crucial assurance for the accuracy and reliability of the test results. This embodies the innovative approach of "deep integration of testing technology with product design and material properties."

[0039] Preferably, the method further includes step S6: based on the mapping relationship between pixels and physical dimensions in the multispectral image, and combined with the surface parameters in the three-dimensional digital model of the single-crystal turbine blade design, geometric measurements are performed on the identified crack region to calculate its length, opening width, and area. The crack location is then mapped to the three-dimensional coordinate system of the single-crystal turbine blade profile to generate a crack report with spatial positioning information. The geometric measurement adopts the principle of binocular vision triangulation, combined with the intrinsic and extrinsic parameter matrices of the multispectral camera, to convert the two-dimensional pixel coordinates into three-dimensional spatial coordinates. The crack opening width is determined jointly by local curvature fitting and edge gradient extreme value localization. The generated crack report can be directly imported into the product lifecycle management system.

[0040] The core innovation of this step lies in breaking through the technical limitations of traditional crack detection, which is "only identification, difficult to quantify, and lacks traceability." It constructs a complete innovative system encompassing "precise geometric measurement - three-dimensional spatial positioning - engineering report output," upgrading the detection results from "qualitative identification" to "quantitative usability," and completely breaking down the key barrier between "detection data and engineering applications." Even if traditional detection methods can identify cracks, they often face problems such as large errors in geometric parameter measurement, inability to accurately map crack locations to the three-dimensional structure of the blade, and the need for manual secondary processing of reports, making it difficult to directly provide effective support for blade life assessment and maintenance decisions.

[0041] The innovation of this invention lies in its triple synergistic breakthrough: First, it innovatively achieves precise end-to-end conversion of "pixel-physical size-3D coordinates." By combining the binocular vision triangulation principle with the intrinsic and extrinsic parameter matrices of a multispectral camera, the crack pixel coordinates in the 2D image are transformed into 3D spatial coordinates. Simultaneously, considering the complex curved surface characteristics of the blade, a combined scheme of "local curvature fitting + edge gradient extreme value localization" is adopted, solving the pain point of traditional measurements' difficulty in accurately quantifying crack opening width on curved blades. This ensures that the measurement errors for length and opening width are controlled within ±0.01mm, meeting engineering-grade accuracy requirements. Second, it creatively integrates crack detection with the 3D digital model of blade design. It not only corrects measurement deviations through surface parameters but also accurately maps the crack location to the 3D coordinate system of the blade profile, achieving "spatial source tracing" of the crack and solving the problem of traditional detection methods that "only know there is a crack, but don't know where it is." The limitations of the previous system have been eliminated, providing a precise basis for subsequent maintenance and positioning. Thirdly, an innovative engineering closed-loop output mechanism has been designed, and the generated crack report is directly compatible with the Product Lifecycle Management (PLM) system. No manual transcription or format conversion is required, which completely changes the status quo of traditional test reports being "isolated and difficult to reuse". It realizes the seamless integration of test data with the entire lifecycle management of blade design, production and maintenance.

[0042] This integrated design, combining "quantitative precision, three-dimensional positioning, and engineering application," is not merely a simple aggregation of functions, but rather a profound expansion and innovation of the value of inspection, specifically addressing the practical engineering needs of aero-engine blade inspection. It elevates crack detection from "discovering problems" to "providing crucial data support for solving problems," filling the gaps in traditional inspection regarding quantitative precision and three-dimensional traceability, while significantly improving the practicality and efficiency of inspection results. It provides a comprehensive, high-precision solution for aero-engine blade quality control, life assessment, and maintenance decisions, embodying the innovative concept of "deep integration of inspection technology and engineering applications."

[0043] The beneficial effects of this invention are as follows:

[0044] 1. Significantly improved detection rate of micron-level hidden cracks: It integrates information from multiple bands, including ultraviolet, visible light, and short-wave infrared. The ultraviolet band enhances the scattering signal of micron-level cracks, and the short-wave infrared penetrates the surface coating to detect subsurface cracks. Combined with the spectral index enhancement algorithm, the detection rate of surface and hidden cracks ≥5μm is ≥99%, solving the problem of insufficient sensitivity of traditional optical detection to hidden cracks.

[0045] 2. Significantly reduced false defect detection rate: Based on the dual false defect elimination mechanism of the three-dimensional digital model of single-crystal turbine blade, it accurately distinguishes between air film cooling hole interference (judged within the range of circularity ≥ 0.85 and 1.5 times the hole diameter) and crystal texture noise (judged within the range of angle with growth main axis < 15°), with a false detection rate of ≤ 2.5%, overcoming the limitation of single imaging technology in distinguishing between cracks and structural traces.

[0046] 3. Positioning and measurement accuracy meets engineering requirements: Sub-pixel level spatial registration (error ≤ 0.1 pixel) combined with binocular vision triangulation principle, crack length and opening width measurement errors ≤ ±0.01mm, three-dimensional positioning accuracy ≤ ±0.05mm, which can directly provide accurate data support for blade life assessment and maintenance decision-making.

[0047] 4. Automated adaptation to batch inspection scenarios: From image acquisition and feature processing to recognition report generation, no manual intervention is required. The inspection time for a single blade is ≤40 minutes. It is compatible with production line automation transformation and solves the problems of low efficiency and reliance on experience in traditional manual inspection.

[0048] 5. Strong anti-interference capability: By suppressing high metal reflectivity through polarization filters and eliminating light fluctuation interference through reflectivity normalization, it maintains stable recognition performance in blade inspection scenarios with complex curved surfaces and dense cooling holes, adapting to the stringent inspection requirements of aero-engine blades. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0050] Figure 2 A schematic diagram of the generation of multidimensional fusion feature maps.

[0051] Figure 3 This is a schematic diagram of dual verification and three-dimensional localization of the crack candidate region. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0053] like Figure 1 As shown, Figure 1This is a schematic diagram of the overall process of the method of the present invention. A method for identifying surface cracks in a single-crystal turbine blade based on multispectral imaging includes the following steps:

[0054] Step S1: Acquire the original spectral images of the single-crystal turbine blade to be tested in multiple preset narrow bands, and construct a multispectral image sequence;

[0055] Step S2: Perform sub-pixel-level spatial registration on the multispectral image sequence to generate a multispectral data cube;

[0056] Step S3: Extract the spectral response curve of each pixel in the multispectral data cube, calculate the spectral index characterizing the crack feature using the reflectance of the key band in the spectral response curve, and generate the corresponding spectral index map; at the same time, perform spatial gradient operation on the multispectral data cube to obtain the spatial gradient feature map; and concatenate the spectral index map and the spatial gradient feature map through channels to construct a multidimensional fusion feature map.

[0057] Step S4: Input the multi-dimensional fused feature map into a preset deep learning recognition model, and identify the crack candidate region through multi-scale feature extraction and attention mechanism;

[0058] Step S5: Based on the surface metallographic structure characteristics of the single-crystal turbine blade, perform false defect elimination on the crack candidate region and output the final surface crack identification result.

[0059] Preferably, the preset narrow band in step S1 includes at least: an ultraviolet band of 320nm-400nm, used to capture weak scattering signals at the crack opening; a visible light band of 450nm-700nm, used to obtain the macroscopic morphology and color characteristics of the turbine blade surface; and a short-wave infrared band of 900nm-1700nm, used to detect the tendency of hidden cracks under the coating layer on the turbine blade surface.

[0060] Preferably, step S2 further includes: setting a standard reflection reference plate, first obtaining the standard brightness value of the standard reflection reference plate in each narrow band, and then extracting the gray value of each pixel in the original spectral image of the single-crystal turbine blade under the narrow band; by calculating the ratio between the gray value of the pixel in the original image of the single-crystal turbine blade in the narrow band and the standard brightness value of the reference plate in the corresponding narrow band, the reflectivity of the original image of the single-crystal turbine blade is normalized, and finally interference such as light fluctuation and high reflectivity of the blade is eliminated, generating a multispectral data cube with a unified illumination reference.

[0061] like Figure 2 As shown, Figure 2This is a schematic diagram of the multidimensional fusion feature map generation. Step S3 specifically includes: extracting the reflectance values ​​of each pixel in different spectral bands from the multispectral data cube to form the spectral response curve;

[0062] Selecting the crack-sensitive band λ1 and the background reference band λ2, the difference in reflectance between the cracked and normal regions is greatest under the crack-sensitive band λ1, while the turbine blade surface reflectance is stable and least affected by crystal orientation under the background reference band λ2. A spectral index SI is defined, which satisfies: SI = (R... λ1 - R λ2 ) / (R λ1 + R λ2 ), where R λ1 R represents the reflectivity of the crack-sensitive band. λ2 The reflectance of the background reference band is used; the corresponding values ​​of each pixel on the surface of the turbine blade under test are calculated according to the spectral index SI; the spectral index values ​​of all pixels are presented according to their spatial positions to generate a spectral index map, so as to enhance the difference in spectral energy between the cracked area and the normal area and achieve the initial enhancement of the crack signal.

[0063] The Sobel gradients of the multispectral data cube are calculated along the x and y directions in the spatial dimension to obtain the gradient magnitude and gradient direction, forming a spatial gradient feature map.

[0064] The spectral index map and the spatial gradient feature map are concatenated by channel stitching to construct a multidimensional fusion feature map.

[0065] Preferably, in step S4, the deep learning recognition model adopts a feature pyramid network (FPN) structure, with the bottom layer composed of ResNet series residual blocks, and a channel attention module and a spatial attention module are introduced in the residual connections. The channel attention module is used to weight and highlight the crack-sensitive feature channels, and the spatial attention module is used to enhance the spatial saliency of the crack edge region, so as to achieve accurate recognition of crack candidate regions.

[0066] like Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the dual verification and three-dimensional localization of the crack candidate region. The specific process of false defect removal in step S5 includes:

[0067] Verification A: Crack-free structure elimination based on spatial matching

[0068] Obtain the design three-dimensional digital model of the single-crystal turbine blade, and extract the coordinate distribution map of the film cooling holes and crystal orientation information from it;

[0069] Spatial matching is performed between the candidate crack region and the coordinate distribution map of the film cooling hole. If the candidate crack region is located within 1.5 times the hole diameter at the edge of the film cooling hole, and its circularity after Hough circle transformation is ≥0.85, it is marked as a non-crack structure and removed.

[0070] Verification B: Texture noise removal based on crystal orientation

[0071] The extension direction vector of the crack candidate region is calculated based on the crystal orientation information. If the angle between the extension direction vector and the crystal growth principal axis is less than 15°, it is determined to be crystal texture noise and is removed.

[0072] Finally, the cracked area that has been verified by both verification A and verification B is retained as the identification result.

[0073] Preferably, the method further includes step S6: based on the mapping relationship between pixels and physical dimensions in the multispectral image, and combined with the surface parameters in the three-dimensional digital model of the single-crystal turbine blade design, geometric measurements are performed on the identified crack region to calculate its length, opening width, and area. The crack location is then mapped to the three-dimensional coordinate system of the single-crystal turbine blade profile to generate a crack report with spatial positioning information. The geometric measurement adopts the principle of binocular vision triangulation, combined with the intrinsic and extrinsic parameter matrices of the multispectral camera, to convert the two-dimensional pixel coordinates into three-dimensional spatial coordinates. The crack opening width is determined jointly by local curvature fitting and edge gradient extreme value positioning. The generated crack report can be directly imported into the product lifecycle management system.

[0074] I. Application Scenarios

[0075] An aero-engine maintenance plant needs to inspect the surface cracks of single-crystal DD6 turbine blades after they have been put into service. These blades are core hot-end components of the engine and are subjected to high temperatures and high speeds during operation. Micron-level hidden cracks are easily generated on their surfaces. In addition, the blades have dense film cooling holes and special crystal textures. Traditional optical inspection often results in problems such as "missing hidden cracks and misjudging interference". A precise and efficient inspection solution is needed.

[0076] II. Specific Implementation Steps

[0077] Step 1: Take "multi-band inspection photos" of the leaves.

[0078] We used a multi-band dedicated camera to take pictures of the blade under test in three specific bands:

[0079] Ultraviolet band (320nm-400nm): Specifically captures scattered signals at the crack opening, which are like "weak reflections", and can even initially capture tiny cracks as small as 5μm;

[0080] Visible light band (450nm-700nm): Photograph the macroscopic appearance and color of the leaf surface, such as whether the leaf has obvious scratches or deformation;

[0081] Short-wave infrared band (900nm-1700nm): can penetrate the protective coating on the blade surface and detect the tendency of hidden cracks "under the coating".

[0082] These three sets of photos, when combined, form a "multi-band image sequence," much like giving the leaf a comprehensive "spectral check-up."

[0083] 2. Step 2: Calibrate and align to generate a "3D data package".

[0084] First, place a "standard reflection reference board" (similar to a "grayscale card" when taking pictures) in the shooting environment, and take pictures of its standard brightness values ​​in three bands as "calibration benchmarks";

[0085] The three sets of spectral images of the blades were precisely stitched together using "sub-pixel alignment technology" (alignment error ≤ 0.1 pixels, equivalent to millimeter-level precision alignment).

[0086] The brightness value of each pixel in the leaf photo is divided by the standard brightness value corresponding to the reference board to eliminate the interference caused by the light intensity and reflection on the leaf surface during the photo capture. Finally, a "multi-band stereo data package" (multispectral data cube) is generated to ensure that the spectral information at each location is accurate and reliable.

[0087] 3. Step 3: Extract features and fuse "crack clues"

[0088] Extract the spectral response curve of each pixel in three bands from the 3D data package;

[0089] Select two key bands: one is the "crack-sensitive band" (where the spectral difference between the cracked and normal regions is greatest), and the other is the "background reference band" (where the blade surface spectrum is stable and unaffected by crystal texture). Use the formula SI = (R λ1 - R λ2 ) / (R λ1 + R λ2 Calculate the "spectral index," which amplifies the difference between the cracked and normal areas, generating a "spectral index map," in which the crack will be more prominent.

[0090] The Sobel gradient algorithm is used to extract the contour features of the blade surface from the 3D data and generate a spatial gradient feature map, which clearly shows the edge of the crack.

[0091] The "spectral index map" and the "spatial gradient feature map" are merged into a "multidimensional fusion feature map". This map contains both the "spectral signal" and "contour morphology" of the crack, which is equivalent to integrating all the "clues" of the crack.

[0092] 4. Step 4: AI model "finds candidate crack regions"

[0093] Input the "multi-dimensional fused feature map" into a specially trained AI model:

[0094] The "skeleton" of this AI model is the "multi-scale recognition structure" (FPN), which can simultaneously identify large cracks and tiny cracks as small as 5μm.

[0095] The model also includes two "intelligent filtering modules": one that specifically highlights features related to cracks (channel attention module), and the other that specifically strengthens crack edges (spatial attention module).

[0096] After a quick scan, the AI ​​will mark all areas suspected of having cracks, forming a "crack candidate area list".

[0097] 5. Step 5: Precisely "eliminate interference" and locate the real crack.

[0098] Retrieve the 3D design model of the blade and extract two key pieces of information: the precise location of the film cooling holes (coordinate distribution map) and the growth direction of the crystals (crystal orientation information).

[0099] First, eliminate interference from cooling holes: If the candidate crack area is within 1.5 times the diameter of the cooling hole edge and the shape is close to circular (circularity ≥ 0.85), it is determined to be a "cooling hole edge trace" and is directly rejected;

[0100] Next, exclude crystal texture interference: if the angle between the extension direction of the candidate region and the crystal growth principal axis is less than 15°, it is determined to be "crystal texture noise" (similar to the texture of wood, not cracks) and is also removed.

[0101] The remaining crack candidate areas are the "real crack areas".

[0102] 6. Step 6: Measurement + Generation of "Traceability Report"

[0103] By combining the surface parameters of the blade 3D model, the "binocular vision measurement technology" is used to convert the crack pixel coordinates in the image into real three-dimensional spatial coordinates.

[0104] The length, opening width (error ≤ ±0.01mm) and area of ​​the crack are accurately measured. The crack width is determined by the method of "surface fitting + edge positioning", which does not affect the accuracy even if the blade has a complex curved surface.

[0105] Generate a "crack report with spatial location" which marks the specific location of the crack (corresponding to the coordinates of the blade 3D model) and size parameters. This report can be directly imported into the engine's "full life cycle management system" for convenient subsequent maintenance and life assessment.

[0106] III. Performance Verification: To demonstrate the advantages of this method, we selected 100 single-crystal DD6 turbine blades with different crack sizes (5μm-1mm) and compared them using the "traditional single visible light detection method" and the "multispectral imaging detection method of this invention." The results are as follows:

[0107] Table 1: Performance Comparison of the Invention and Existing Turbine Blade Detection Methods

[0108]

[0109] IV. Conclusion

[0110] This method, through a complete workflow design of "multi-band data acquisition - precise calibration - feature fusion - AI recognition - targeted interference removal - quantitative reporting," perfectly solves the core pain points of traditional optical inspection: "low detection rate of hidden cracks, high false detection rate, insufficient accuracy, and low efficiency." Data shows that it improves the detection rate of hidden cracks by 17.5 percentage points, reduces the false detection rate by 9.2 percentage points, and increases detection efficiency by 2.7 times. It fully meets the high-precision, automated inspection requirements of single-crystal turbine blades for aero-engines, demonstrating significant practical value and technological advantages compared to existing technologies.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for identifying surface cracks in a single-crystal turbine blade based on multispectral imaging, characterized by comprising the following steps: Step S1: Acquire raw spectral images of the single-crystal turbine blade to be tested in multiple preset narrow bands to construct a multispectral image sequence; the preset narrow bands include at least: ultraviolet band 320nm-400nm, used to capture weak scattering signals at the crack opening; visible light band 450nm-700nm, used to obtain the macroscopic morphology and color characteristics of the turbine blade surface; short-wave infrared band 900nm-1700nm, used to detect the tendency of hidden cracks under the coating layer on the turbine blade surface; Step S2: Perform sub-pixel-level spatial registration on the multispectral image sequence to generate a multispectral data cube; Step S3: Extract the spectral response curve of each pixel in the multispectral data cube, calculate the spectral index characterizing the crack feature using the reflectance of the key band in the spectral response curve, and generate the corresponding spectral index map; at the same time, perform spatial gradient operation on the multispectral data cube to obtain the spatial gradient feature map; and concatenate the spectral index map and the spatial gradient feature map through channels to construct a multidimensional fusion feature map. The reflectance values ​​of each pixel in different spectral bands are extracted from the multispectral data cube to form the spectral response curve; Selecting the crack-sensitive band λ1 and the background reference band λ2, the difference in reflectance between the cracked and normal regions is greatest under the crack-sensitive band λ1, while the turbine blade surface reflectance is stable and least affected by crystal orientation under the background reference band λ2. A spectral index SI is defined, which satisfies: SI = (R... λ1 - R λ2 ) / (R λ1 + R λ2 ), where R λ1 R represents the reflectivity of the crack-sensitive band. λ2 The reflectance of the background reference band is used; the corresponding values ​​of each pixel on the surface of the turbine blade under test are calculated according to the spectral index SI; the spectral index values ​​of all pixels are presented according to their spatial positions to generate a spectral index map, so as to enhance the difference in spectral energy between the cracked area and the normal area and achieve the initial enhancement of the crack signal. The Sobel gradients of the multispectral data cube are calculated along the x and y directions in the spatial dimension to obtain the gradient magnitude and gradient direction, forming a spatial gradient feature map. The spectral index map and the spatial gradient feature map are concatenated by channels to construct a multi-dimensional fusion feature map. Step S4: Input the multi-dimensional fused feature map into a preset deep learning recognition model, and identify the crack candidate region through multi-scale feature extraction and attention mechanism; Step S5: Based on the surface metallographic structure characteristics of the single-crystal turbine blade, perform false defect elimination on the crack candidate region and output the final surface crack identification result; The specific process of pseudo-defect removal in step S5 includes: Verification A: Crack-free structure elimination based on spatial matching Obtain the design three-dimensional digital model of the single-crystal turbine blade, and extract the coordinate distribution map of the film cooling holes and crystal orientation information from it; Spatial matching is performed between the candidate crack region and the coordinate distribution map of the film cooling hole. If the candidate crack region is located at the edge of the film cooling hole... If the diameter is within 1.5 times the aperture and the circularity of the outline after Hough circle transformation is ≥0.85, then it is marked as a non-crack structure and discarded. Verification B: Texture noise removal based on crystal orientation The extension direction vector of the crack candidate region is calculated based on the crystal orientation information. If the angle between the extension direction vector and the crystal growth principal axis is less than 15°, it is determined to be crystal texture noise and is removed. Finally, the cracked area that has been verified by both verification A and verification B is retained as the identification result.

2. The method for identifying surface cracks in single-crystal turbine blades based on multispectral imaging according to claim 1, characterized in that, Step S2 further includes: setting a standard reflection reference plate, first obtaining the standard brightness value of the standard reflection reference plate in each narrow band, and then extracting the gray value of each pixel in the original spectral image of the single-crystal turbine blade under the narrow band; by calculating the ratio between the gray value of the pixel in the original image of the single-crystal turbine blade in the narrow band and the standard brightness value of the reference plate in the corresponding narrow band, the reflectivity of the original image of the single-crystal turbine blade is normalized, and a multispectral data cube with a unified illumination reference is generated.

3. The method for identifying surface cracks in single-crystal turbine blades based on multispectral imaging according to claim 1, characterized in that, In step S4, the deep learning recognition model adopts a feature pyramid network structure, with the bottom layer consisting of ResNet series residual blocks. Channel attention modules and spatial attention modules are introduced into the residual connections. The channel attention module is used to weight and highlight the crack-sensitive feature channels, and the spatial attention module is used to enhance the spatial saliency of the crack edge region, thereby achieving accurate recognition of crack candidate regions.

4. The method for identifying surface cracks in single-crystal turbine blades based on multispectral imaging according to claim 1, characterized in that, The process also includes step S6: Based on the mapping relationship between pixels and physical dimensions in the multispectral image, and combined with the surface parameters in the three-dimensional digital model of the single-crystal turbine blade design, geometric measurements are performed on the identified crack region to calculate its length, opening width, and area. The crack location is then mapped to the three-dimensional coordinate system of the single-crystal turbine blade profile to generate a crack report with spatial positioning information. The geometric measurement adopts the principle of binocular vision triangulation, combined with the intrinsic and extrinsic parameter matrices of the multispectral camera, to convert the two-dimensional pixel coordinates into three-dimensional spatial coordinates. The crack opening width is determined jointly by local curvature fitting and edge gradient extreme value localization. The generated crack report can be directly imported into the product lifecycle management system.

Images