A wind power quality witness point-oriented component defect precision detection method

Abstract

The application provides a wind power quality witness point-oriented component defect accurate detection method, and relates to the technical field of nondestructive testing. The method comprises the following steps: acquiring multi-modal detection data of a transition arc area between an inner tooth ring and a raceway of a yaw bearing of a wind turbine generator at a quality witness point; performing spatial registration and three-dimensional topological surface unfolding processing on the multi-modal detection data, and constructing a spatial transformation matrix with the local reference surface of the transition arc area as the origin; solving the coordinate mapping relationship from a three-dimensional detection point to a two-dimensional unfolded plane through linear mapping operation and translation vector iteration of the spatial transformation matrix; projecting the variable curvature spatial topography of the transition arc area to a unified two-dimensional coordinate domain according to the coordinate mapping relationship, and obtaining alignment feature data of the transition arc area. The application improves manufacturing efficiency and product full life cycle operation reliability.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, and in particular to a method for precise detection of component defects guided by wind power quality witnessing points. Background Technology

[0002] The yaw bearing of a wind turbine is a core load-bearing component that enables precise adjustment of the nacelle's windward angle. Its operational reliability directly determines the service life and power generation efficiency of the entire wind turbine. The transition arc area at the junction of the internal gear tooth root and the raceway is a high-stress concentration zone and a primary location for the initiation of fatigue cracks and quenching spalling defects. In the yaw bearing manufacturing process, the period after raceway quenching and finishing, and before rolling element assembly, is designated as a critical quality monitoring point.

[0003] Surface defect detection in this transition arc region primarily relies on manual visual inspection and conventional machine vision inspection. However, due to the variable curvature surface geometry of this region and the physical obstruction at the tooth root, conventional visual light sources easily cause signal scattering in this area, resulting in low signal-to-noise ratios in the acquired image data. A more significant technical drawback is the presence of numerous periodic machining marks on the surface of this region. Existing detection algorithms often employ global image feature extraction or single threshold segmentation, failing to effectively decouple machining marks from micro-fatigue cracks or quenching spalling within the feature space. Under the strong interference of the tool mark texture, existing methods are prone to misclassifying tool marks as cracks or missing true micro-cracks hidden at the bottom of the tool marks, leading to persistently high false positive and false negative rates at quality verification points, which cannot support the stringent WHS release standards for core wind power components. Summary of the Invention

[0004] This invention provides a wind power quality witness point-oriented method for precise detection of component defects, which improves manufacturing efficiency and product lifecycle operational reliability.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows: Firstly, a method for precise detection of component defects guided by wind power quality witnessing points, the method comprising: Multimodal detection data of the transition arc region between the internal gear ring and raceway of the yaw bearing of a wind turbine generator were collected at the quality witness point; spatial registration and three-dimensional topological surface unfolding were performed on the multimodal detection data to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin; By solving the coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane through linear mapping operation of the spatial transformation matrix and iterative translation vector, the variable curvature spatial topography of the transition arc region is projected onto a unified two-dimensional coordinate domain according to the coordinate mapping relationship, and the alignment feature data of the transition arc region is obtained. Based on the alignment feature data of the transition arc region, the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain are extracted; the three-dimensional topological geometric normal distribution features and surface texture gradient features are input into a pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features. Based on the false defect suppression features and the real defect enhancement features, a defect topology map of the transition arc region is constructed; based on the defect topology map, the defect confidence of each suspected defect region in the transition arc region is calculated. The defect confidence level is compared with the dynamic release threshold corresponding to the quality witness point to obtain the accurate defect detection result of the transition arc area, and the status control command of the quality witness point is output.

[0006] Furthermore, multimodal testing data were acquired at quality witness points in the transition arc region between the internal gear ring and raceway of the yaw bearing of the wind turbine generator, including: A synchronous spatial scan was performed on the transition arc region to acquire the original detection sequence, which includes spatial point cloud coordinate sequence, surface optical grayscale texture map and ultrasonic echo time-frequency signal; Based on the preset physical reference markers in the original detection sequence, the initial spatial position vectors of visual feature nodes and acoustic feature nodes in the global coordinate system of the yaw bearing internal gear ring are extracted. By performing cross-modal coordinate mapping verification using the initial spatial position vector, the spatial dimensions of the surface optical grayscale texture map and the ultrasonic echo time-frequency signal are aligned to the corresponding grid nodes of the spatial point cloud coordinate sequence, thus completing pixel-level registration and data fusion of multimodal signals under a unified spatiotemporal reference to obtain multimodal detection data.

[0007] Furthermore, spatial registration and 3D topological surface unfolding are performed on the multimodal detection data to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin, including: Local sampling windows are divided along the curvature change trajectory of the three-dimensional topological point set. For the three-dimensional spatial coordinate set within each local sampling window, the sum of squared residuals of the spatial coordinates is minimized to solve the final local reference surface parameters corresponding to each local sampling window and determine the normal reference vector and the coordinates of the reference surface center. Using the normal reference vector as the reference rotation axis and the center coordinates of the reference plane as the origin of the new local coordinate system, a local spatial mapping parameter set containing rotational and translational transformation components is constructed. The local spatial mapping parameter set is serialized and spliced ​​along the curvature change trajectory and topological continuity constraint is iterated to eliminate spatial step misalignment and deformation gradient abrupt change between adjacent local windows, thus obtaining a spatial transformation matrix covering the global range of the transition arc region.

[0008] Furthermore, by performing linear mapping operations on the spatial transformation matrix and iteratively solving the coordinate mapping relationship from the 3D detection point to the 2D unfolded plane using translation vectors, the variable curvature spatial topography of the transition arc region is projected onto a unified 2D coordinate domain based on the coordinate mapping relationship, resulting in alignment feature data for the transition arc region, including: The spatial transformation matrix is ​​applied to the three-dimensional detection point set in the multimodal detection data to perform an initial linear coordinate mapping operation, transforming the three-dimensional spatial coordinates to the basic projected coordinates of the two-dimensional unfolded plane; The local curvature deviation between the basic projected coordinates and the nodes of the two-dimensional unfolded plane theoretical mesh is calculated to obtain the curvature compensation residual sequence; Based on the curvature compensation residual sequence, the translation vector iterative update operation is performed to dynamically adjust the displacement compensation components of each local sampling window in the two-dimensional unfolded plane, eliminate the projection stretching and compression distortion caused by the variable curvature spatial topography, and obtain the result of the updated displacement compensation components and the initial linear coordinate mapping operation. The updated displacement compensation components are superimposed and fused with the results of the initial linear coordinate mapping calculation to construct a complete coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane. Based on the coordinate mapping relationship, the variable curvature spatial topography of the transition arc region is mapped layer by layer to a unified two-dimensional coordinate domain, and topological continuity calibration is performed on the mapped multimodal feature layer to obtain the alignment feature data of the transition arc region.

[0009] Furthermore, based on the alignment feature data of the transition arc region, the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain are extracted, including: Local neighborhood topology analysis is performed on the alignment feature data of the transition arc region to extract the three-dimensional spatial pose information of each feature node in the unified two-dimensional coordinate domain, and a local differential mesh of the micro-undulation morphology of the transition arc region is constructed. Based on the local differential mesh, differential geometric operators are performed along the normal direction of the mesh nodes to calculate the deflection angle of the tangent plane of the spatial surface of each local region and the principal direction of curvature, and then aggregated to obtain the three-dimensional topological geometric normal distribution characteristics. The surface optical grayscale texture map and ultrasonic echo time-frequency signal mapping layer in the alignment feature data of the transition arc region are simultaneously extracted. Multi-scale directional filtering operation is performed in a unified two-dimensional coordinate domain to extract the grayscale jump rate and ultrasonic energy attenuation gradient in the neighborhood of each pixel. Cross-modal gradient vector synthesis is performed on the gray-level jump rate and ultrasonic energy attenuation gradient to eliminate interference components in the machining artifact direction and obtain surface texture gradient features.

[0010] Furthermore, the 3D topological geometric normal distribution features and surface texture gradient features are input into a pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features, including: The three-dimensional topological geometric normal distribution features and surface texture gradient features are spliced ​​and fused in the channel dimension to construct a multimodal joint feature matrix; The multimodal joint feature matrix is ​​input into the dual-stream feature encoding network of the tool mark-crack feature decoupling model to extract the geometric implicit features that characterize the macroscopic undulation of the surface and the texture frequency domain implicit features that characterize the frequency domain distribution of the surface micro-texture, respectively. By using the cross-attention decoupling layer in the tool mark-crack feature decoupling model, the cross-modal correlation weight between the implicit features of geometric morphology and the implicit features of texture frequency domain is calculated, and the tool mark interference component with periodic directional regularity and the crack response component with local random orientation are separated in the feature space. The feature mask suppression operation is performed on the tool mark interference component to filter out the pseudo-effect caused by machining tool marks and obtain the pseudo-defect suppression feature. Based on the feature map distribution of the pseudo-defect suppression feature, the main direction of the tool mark texture is determined. The crack response components are subjected to feature weight amplification and spatial reconstruction operations to highlight the local stress concentration response of micro fatigue cracks and obtain the true defect enhancement features.

[0011] Furthermore, based on the pseudo-defect suppression features and the real defect enhancement features, a defect topology map of the transition arc region is constructed, including: Based on the real defect enhancement features, the boundary contours and geometric center points of each suspected defect region within the transition arc area are extracted to construct an initial defect node set. By performing directional consistency filtering on the initial defect node set through the pseudo-defect suppression feature, pseudo-defect nodes with an angle less than a preset angle threshold with the main direction of the tool mark texture in the pseudo-defect suppression feature are removed, resulting in a pure defect node set. Calculate the spatial Euclidean distance and feature similarity between any two defect nodes in the pure defect node set. When the spatial Euclidean distance is less than a preset distance threshold and the feature similarity is greater than a preset similarity threshold, establish a topological connection edge between the corresponding two defect nodes. Using the set of pure defect nodes as graph vertices and the topological connecting edges as graph edges, and mapping the defect depth response values ​​of the corresponding nodes in the real defect enhancement features to the graph vertex weights, a defect topology graph of the transition arc region is assembled.

[0012] Furthermore, based on the defect topology map, the defect confidence level of each suspected defect region in the transition arc region is calculated, including: The defect topology graph is input into a graph convolutional neural network. The graph vertex weights and feature information of adjacent defect nodes are aggregated through topological connection edges to update the local topological feature representation of each defect node. Based on the updated local topological feature representation, the weighted fusion value of the topological connectivity coefficient and the defect depth response intensity of each suspected defect region is calculated. The weighted fusion value is input into a pre-constructed nonlinear confidence mapping function to calculate the comprehensive defect probability of each suspected defect region in the multidimensional feature space. By normalizing the overall defect probability, the defect confidence level of each suspected defect area in the transition arc region is obtained.

[0013] Furthermore, the defect confidence level is compared with the dynamic release threshold corresponding to the quality witness point to obtain the accurate defect detection result in the transition arc region, and the status control command of the quality witness point is output, including: Obtain the historical operating parameters of the yaw bearing internal gear ring and the process constraints of the current quality witness point. Calculate the dynamic release threshold corresponding to the quality witness point through a preset threshold adaptive adjustment model. The defect confidence level of each suspected defect area is compared with the dynamic release threshold point by point to screen out the target defect areas whose defect confidence level is greater than the dynamic release threshold. By statistically analyzing the number of target defect areas, their spatial distribution density, and the maximum defect confidence level, accurate defect detection results for the transition arc region are obtained. Based on the accurate defect detection results, a preset quality control strategy library is matched to obtain the corresponding status control instructions. The status control instructions include release instructions, interception and re-inspection instructions, or scrap isolation instructions, which are then sent to the execution terminal of the quality witness point.

[0014] Secondly, a wind power quality witness point-oriented precision detection system for component defects includes: The acquisition module is used to acquire multimodal detection data of the transition arc region between the inner gear ring and the raceway of the yaw bearing of the wind turbine generator at the quality witness point; the multimodal detection data is spatially registered and three-dimensional topological surface unfolded to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin; The mapping module is used to solve the coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane through linear mapping operation of the spatial transformation matrix and iterative translation vector. Based on the coordinate mapping relationship, the variable curvature spatial topography of the transition arc region is projected to a unified two-dimensional coordinate domain to obtain the alignment feature data of the transition arc region. The extraction module is used to extract the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain based on the alignment feature data of the transition arc region; the three-dimensional topological geometric normal distribution features and surface texture gradient features are input into a pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features. The calculation module is used to construct a defect topology map of the transition arc region based on the pseudo-defect suppression features and the real defect enhancement features; and to calculate the defect confidence of each suspected defect region in the transition arc region based on the defect topology map. The comparison module is used to compare the defect confidence level with the dynamic release threshold corresponding to the quality witness point, obtain the accurate defect detection result of the transition arc area, and output the status control command of the quality witness point.

[0015] The above-described solution of the present invention has at least the following beneficial effects: By employing multimodal detection data synchronous acquisition and pixel-level spatial registration, combined with 3D topological surface unfolding and curvature compensation iterative projection techniques, a spatial transformation matrix covering the entire transition arc region is constructed. This maps the variable curvature spatial topography to a unified two-dimensional coordinate domain, thus overcoming the technical problems of signal acquisition blind spots, low data signal-to-noise ratio, and projection distortion caused by the variable curvature geometric features of the transition arc region and physical occlusion of the tooth root. Furthermore, by using a dual-stream cross-attention tool mark-crack feature decoupling model to separate periodic tool mark interference from the random-path real defect response, and based on the defect topology map... By combining graph convolutional neural networks to calculate the comprehensive defect confidence level and generating dynamic release thresholds based on historical operating conditions and process constraints, this technology overcomes the technical problems of high false positive and false negative rates caused by the coupling of machining tool marks and real defect features, and the inability of static thresholds to match the release requirements of different quality witness points. This enables precise detection of minute defects in the entire transition arc area of ​​the inner gear ring of wind turbine yaw bearings without blind spots, reduces the false positive and false negative rates of quality witness points, meets the stringent WHS release standards for core wind power components, and avoids premature bearing failure caused by non-conforming products entering the assembly process. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a wind power quality witness point-oriented method for precise detection of component defects, provided by an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a wind power quality witness point-oriented precision detection system for component defects, provided by an embodiment of the present invention.

[0018] Figure 3 This is a schematic diagram of the original surface texture gradient features.

[0019] Figure 4This is a schematic diagram of the defect topology construction. Detailed Implementation

[0020] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art.

[0021] like Figure 1 As shown, an embodiment of the present invention proposes a method for accurate detection of component defects guided by wind power quality witnessing points. The method includes the following steps: Step 1: Obtain multimodal detection data of the transition arc region between the inner gear ring and raceway of the yaw bearing of the wind turbine generator at the quality witness point; perform spatial registration and three-dimensional topological surface unfolding processing on the multimodal detection data to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin; Step 2: Solve the coordinate mapping relationship from the 3D detection point to the 2D unfolded plane through linear mapping operation of the spatial transformation matrix and iterative translation vector. Based on the coordinate mapping relationship, project the variable curvature spatial topography of the transition arc region onto a unified 2D coordinate domain to obtain the alignment feature data of the transition arc region. Step 3: Based on the alignment feature data of the transition arc region, extract the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain; input the three-dimensional topological geometric normal distribution features and surface texture gradient features into the pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features. Step 4: Construct a defect topology map of the transition arc region based on the pseudo-defect suppression features and the real defect enhancement features; calculate the defect confidence of each suspected defect region in the transition arc region based on the defect topology map; Step 5: Compare the defect confidence level with the dynamic release threshold corresponding to the quality witness point to obtain the accurate defect detection result of the transition arc area, and output the status control command of the quality witness point.

[0022] In this embodiment of the invention, the multimodal detection data spatial registration and three-dimensional topological surface unfolding, combined with the spatial transformation matrix and translation vector iterative solution of the variable curvature spatial topography two-dimensional projection technique, overcomes the technical problems of signal acquisition blind spots, low data signal-to-noise ratio, and projection stretching and extrusion distortion caused by the geometric features of the variable curvature surface in the transition arc region and the physical occlusion of the tooth root; at the same time, because the three-dimensional topological geometric normal distribution and surface texture gradient feature extraction, and the tool mark-crack feature decoupling model are used to achieve false defect suppression and real defect enhancement, the technical problem of high false detection and missed detection rate caused by the coupling of machining tool mark false defects and real fatigue cracks and quenching spalling features is overcome; and because the missing The technology of constructing a topology graph and using a graph convolutional neural network to calculate defect confidence, combined with the dynamic release threshold comparison of quality witness points to output state control commands, overcomes the technical problem that static thresholds cannot adapt to the differentiated release requirements of quality witness points under different operating conditions and process constraints. This enables precise detection of minute defects in the entire range of the transition arc area of ​​the inner gear ring of wind turbine yaw bearings without blind spots, reduces the missed detection rate and false alarm rate of quality witness points, meets the stringent WHS release standards of core wind power components, avoids premature bearing failure and unplanned downtime of wind turbine units caused by non-conforming products entering the assembly process, reduces unnecessary rework and scrap, and improves manufacturing efficiency and product life cycle operational reliability.

[0023] In a preferred embodiment of the present invention, step 1 above may include: Step 1.1: Perform synchronous spatial scanning on the transition arc region to acquire the original detection sequence, which includes a spatial point cloud coordinate sequence, a surface optical grayscale texture map, and ultrasonic echo time-frequency signals. Specifically, the acquisition device integrates a line laser profile sensor, a high-resolution area array camera, and an ultrasonic phased array probe on a rigid support. The fields of view of these three components are pre-calibrated spatially using a calibration target, unified to a sensor coordinate system with the optical center of the line laser sensor as the origin, and a conversion relationship is established between the camera image pixel coordinates, the ultrasonic probe array element coordinates, and the line laser point cloud spatial coordinates. A synchronous trigger controller sends synchronous pulses of the same frequency to the three sensors to achieve synchronous scanning of the transition arc region between the yaw bearing's inner gear ring and raceway at equal spatial intervals. The line laser profile sensor scans the transition arc surface along the circumference of the inner gear ring at a sampling interval of 0.1 mm, acquiring the three-dimensional spatial coordinates of discrete points on each sampling profile line to form a spatial point cloud coordinate sequence. ; A phased array camera synchronously acquires surface optical grayscale texture maps. Each frame corresponds to a transition arc surface with a known arc length. After flat-field correction and distortion elimination, the image pixel grayscale values ​​are used to establish a frame index relationship based on the acquisition timestamp and point cloud coordinate sequence. An ultrasonic phased array probe emits and receives ultrasonic longitudinal waves near the transition arc surface in a fan-shaped scanning manner, acquiring the A-scan signal received by each array element. The echo amplitude data at different depths at each sampling position is obtained through sound field reconstruction, and the peak time of the signal envelope is used to characterize the ultrasonic echo time-frequency signal. .

[0024] Step 1.2: Based on the preset physical reference markers in the original detection sequence, extract the initial spatial position vectors of visual feature nodes and acoustic feature nodes in the global coordinate system of the yaw bearing inner gear ring. Specifically, this includes: pre-processing three hemispherical physical reference markers that are not on the same straight line on the outer circular end face of the yaw bearing inner gear ring. The surface roughness of the markers differs from that of the substrate, and can form distinguishable responses in both optical images and ultrasonic reflection signals. Based on surface optical grayscale texture maps, the image is first preprocessed with grayscale equalization and noise filtering. A circular symmetry detection algorithm is then used to traverse circular templates of different radii within the local region where the marker points are located. An edge gradient direction voting mechanism is used to initially locate the candidate circle centers of the marker points. The edge contours of the candidate regions are then refined by interpolation, and the image coordinates of the imaging circle centers of each marker point are extracted. Combined with the camera's intrinsic and extrinsic parameters and the corresponding region's point cloud depth information provided by the line laser contour sensor, the three-dimensional visual feature node coordinates of each marker point in the global coordinate system of the yaw bearing's internal gear ring are calculated. Simultaneously, from the ultrasonic echo time-frequency signal, the abrupt change characteristics of the echo amplitude when the ultrasonic beam is perpendicularly incident on the bottom surface of the marker point are extracted for the arc position of each marker point. The sound path and spatial position of the phased array probe corresponding to this abrupt change point are taken. Based on the ultrasonic velocity constant and probe installation calibration parameters, the image coordinates of the candidate circle centers are calculated. The acoustic feature node coordinates of each marker point were calculated, where, It is the first The three-dimensional coordinates of the acoustic feature nodes on the bottom surface of each physical reference marker. It is the three-dimensional position of the phased array probe in the global coordinate system of the yaw bearing. It is the sound velocity constant of ultrasound in the bearing matrix material. It is the two-way propagation time of the ultrasonic wave corresponding to the echo from the bottom surface of the marked point. It is the unit vector of the main direction of the ultrasonic beam, and the coordinates of the visual feature nodes and the acoustic feature nodes are ultimately used as the initial spatial position vector of the multimodal system.

[0025] Step 1.3 involves performing cross-modal coordinate mapping verification using the initial spatial position vector. This aligns the spatial dimensions of the surface optical grayscale texture map and the ultrasonic echo time-frequency signal to the corresponding grid nodes of the spatial point cloud coordinate sequence. This completes pixel-level registration and data fusion of multimodal signals under a unified spatiotemporal reference, yielding multimodal detection data. Specifically, this includes: constructing a cross-modal coordinate mapping verification relationship from acoustic feature nodes to visual feature nodes, using the visual feature node coordinates as a reference; defining a 3×3 orthogonal rotation matrix R and a 3×1 translation column vector T as rigid body transformation parameters; establishing an overdetermined system of equations using three marker points; and solving for the final R and T by minimizing the sum of weighted spatial deviations. The weights are then set... Inversely proportional to the acoustic positioning uncertainty of each marker point, These are the index numbers of three physical reference points. The objective is to solve for: ; Solve using singular value decomposition. First, calculate the weighted cross-covariance matrix: ; in, , .

[0026] Perform singular value decomposition on matrix H: ; The estimated value of the rotation matrix is: ; Then, after performing determinant corrections, the effective rotation matrix R is obtained. The estimated value of the translation vector is: ; The calculated rotation matrix R and translation vector T are applied to the spatial coordinates corresponding to all ultrasonic echo time-frequency signals, mapping the ultrasonic data to the global coordinate system of the point cloud. The surface optical grayscale texture map is then mapped pixel-by-pixel to the point cloud coordinate grid through the calibrated homography between the camera and the line laser sensor. For each grid node in the point cloud, if there are valid mapped grayscale values ​​and ultrasonic echo values, they are assigned to the attributes of that node; if a grid node has no valid mapped values, it is filled by inverse distance weighted interpolation of adjacent valid nodes, resulting in multimodal detection data with a one-to-one correspondence between spatial point cloud coordinates, grayscale values, and ultrasonic echo amplitudes.

[0027] In this embodiment of the invention, the technical means of synchronously scanning and acquiring spatial point cloud coordinate sequences, surface optical grayscale texture maps and ultrasonic echo time-frequency signals, extracting initial spatial position vectors of visual and acoustic feature nodes based on preset physical reference markers, and aligning multimodal signals to corresponding grid nodes of spatial point clouds through cross-modal coordinate mapping verification overcomes the technical problems of existing single-modal detection that cannot simultaneously acquire surface morphology and internal defect information, coordinate misalignment and feature correspondence deviation caused by inconsistent spatial references of different modal data, and insufficient cross-modal data fusion accuracy. Thus, it achieves the synchronous and complete acquisition of surface microcracks and internal quenching defect information in the transition arc region, and completes pixel-level accurate registration and fusion of multimodal signals under a unified spatiotemporal reference.

[0028] In a preferred embodiment of the present invention, step 1 above may include: Step 1.4: Divide the local sampling window along the curvature change trajectory of the three-dimensional topological point set. For the three-dimensional spatial coordinate set within each local sampling window, perform the minimum operation of the sum of squared spatial coordinate residuals to solve the final local reference surface parameters corresponding to each local sampling window, and determine the normal reference vector and the coordinates of the reference surface center. Specifically, in the multimodal detection data that has been registered and fused, along the curvature change trajectory of the transition arc region of the yaw bearing inner gear ring, divide the continuous surface of this region into multiple overlapping local cylindrical neighborhood windows according to the principle of equal arc length. For the k-th local sampling window...

[0029] Within each local window, a local tangent plane is fitted using the moving least squares method as a temporary reference plane for that region. The specific solution process is as follows: finding a plane equation. ,in, For the 1st The unit normal vector to be found within a local window. Let k be the in-plane reference point to be determined within the k-th local window. Let be the three-dimensional spatial coordinates of any point on the plane, which is defined by its unit normal vector. and the coordinates of any point in the plane The defined unit normal vector perpendicular to the plane is the required normal reference vector, and the spatial coordinates of the geometric center of the plane are the coordinates of the reference plane center.

[0030] In order to obtain the final and Construct a point set All data points The objective functional is the sum of squared residuals of the perpendicular distances to the plane to be determined, which can be expressed as: ; By solving for the minimum value of this objective functional, we can obtain an optimal fitting plane that spans the window set and minimizes the sum of the squared distances from each point to the plane. It can be taken as the centroid of the window point set, that is The calculation minimizes the sum of squared spatial coordinate residuals within each local sampling window, thus determining their respective independent normal reference vectors. coordinates of the center of the reference plane .

[0031] Step 1.5: Using the normal reference vector as the reference rotation axis and the center coordinates of the reference plane as the origin of the new local coordinate system, construct a set of local spatial mapping parameters containing rotation and translation transformation components. Specifically, this includes: based on the solved reference plane geometric parameters of each local window, transforming them into a set of rigid body motion transformation parameters that locally flatten the 3D surface to a 2D plane. For the k-th local sampling window, using the center coordinates of the reference plane as the reference axis... As the origin of spatial translation, the translation transformation component is defined as a vector. The calculation formula is as follows: ; This translation vector moves the origin of the local coordinate system to the center of the reference plane.

[0032] The normal reference vector obtained by the solution Construct rotational transformation components using the reference rotation axis. Rotation transformation It is a 3x3 orthogonal matrix whose geometric meaning is to rotate the local surface of the current window until its normal is aligned with a standard axis of the world coordinate system, such as the Z-axis. The construction depends on The spatial geometric relationship between the alignment axis and the target can be calculated using the Rodriguez rotation formula, which describes the coordinate transformation between any two spatial vectors.

[0033] Rotation transformation components With translation components Combine them to form the local space mapping parameter set of the k-th local window. This parameter set fully describes a three-dimensional point. The process of transforming from the original 3D detection coordinate system to a 2D unfolded plane with a local reference plane is as follows: ; in, This is a coordinate vector that identifies the local planar geometric information after mapping.

[0034] Step 1.6 involves serializing and iterating the local spatial mapping parameter set along the curvature change trajectory, and applying topological continuity constraints to eliminate spatial step misalignment and abrupt deformation gradient changes between adjacent local windows, resulting in a spatial transformation matrix covering the global range of the transition arc region. Specifically, this includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] The curvature change trajectory along the transition arc region is stitched together. Since the reference plane of each window is calculated independently, directly applying their mapping relationship will cause coordinate misalignment in the overlapping area of ​​adjacent windows, which manifests as spatial step misalignment and abrupt changes in deformation gradient.

[0035] To eliminate this discontinuity, a topological continuity constraint iterative algorithm is used for global optimization. The goal of this optimization process is to find a new set of optimized transformation parameters. This minimizes the mapping error and deformation energy between adjacent windows. The specific iterative calculation method is as follows: Define a global energy function, which includes a data term and a smoothing term. The transformed parameters after data term constraint optimization are compared with the independently calculated original parameters. The smoothing term keeps the windows as close as possible; for example, the mapping coordinates of the same spatial point in the overlapping region of the k-th and k+1-th windows remain continuous.

[0036] During the iteration process, the rotation and translation parameters of each window are continuously adjusted to minimize the global energy function. The process of minimizing the energy function forces the transformation matrices of adjacent windows to generate almost identical two-dimensional coordinates in the overlapping area, thereby smoothly transitioning the deformation differences caused by curvature changes and eliminating splicing cracks.

[0037] The series of continuous, stepless misaligned local spatial mapping parameter sets obtained after global optimization are integrated into a unified data structure according to the spatial distribution sequence of the window, which is the spatial transformation matrix covering the global range of the transition arc region.

[0038] In this embodiment of the invention, the technical means of dividing local sampling windows along the curvature change trajectory of the three-dimensional topological point set, solving the final local reference surface parameters of each window by minimizing the sum of squared spatial coordinate residuals, constructing a local spatial mapping parameter set containing rotation and translation transformation components, and serializing and splicing the local parameter set along the curvature trajectory and applying topological continuity constraints for iteration overcomes the technical problems that a single global reference surface cannot adapt to the variable curvature geometric features of the transition arc region, that global projection is prone to stretching and squeezing distortion, and that there are spatial step misalignments and deformation gradient abrupt changes in the splicing of adjacent local windows. Thus, a high-precision spatial transformation matrix covering the global range of the transition arc region is constructed, ensuring the topological continuity and geometric accuracy of the mapping from the three-dimensional detection point to the two-dimensional unfolded plane, and eliminating the projection distortion of the variable curvature surface.

[0039] In a preferred embodiment of the present invention, step 2 above may include: Step 2.1 involves applying the spatial transformation matrix to the set of 3D detection points in the multimodal detection data, performing an initial linear coordinate mapping operation to transform the 3D spatial coordinates to the basic projected coordinates of the 2D unfolded plane. Specifically, this includes: obtaining the constructed spatial transformation matrix covering the global range of the transition arc region. This matrix stores the local coordinate transformation relationship corresponding to each local sampling window in the form of a piecewise mapping parameter set. The 3D detection point set contained in the multimodal detection data is then processed according to each 3D detection point. Given its spatial location, index the local sampling window k to which it belongs, and retrieve the local spatial mapping parameter set corresponding to that window. Perform initial linear coordinate mapping operations to convert the three-dimensional spatial coordinates... Basic projection coordinates for conversion to a two-dimensional unfolded plane The mapping operation is expressed as:

[0040] in, This means extracting the first two orthogonal components of the 3D coordinate vector after rotation and translation transformation, projecting them onto a 2D unfolded plane, and obtaining a 2D coordinate vector. The above mapping operation is performed on all data points in the 3D detection point set to generate a basic projected coordinate sequence that corresponds one-to-one with the original 3D points.

[0041] Step 2.2 calculates the local curvature deviation between the base projection coordinates and the theoretical mesh nodes in the 2D unfolded plane, obtaining the curvature compensation residual sequence. Specifically, this includes: pre-constructing a set of regularly distributed theoretical mesh nodes on the 2D unfolded plane. The generation rule for these theoretical mesh nodes is to pre-construct a set of regularly distributed theoretical mesh nodes on the 2D unfolded plane. This mesh is a rectangular mesh covering the entire projection range of the transition arc region. Its mesh spacing is determined jointly based on the average sampling spacing of the 3D point cloud and the average curvature radius of the corresponding window, ensuring that the mesh density can capture the deviation caused by curvature changes. The coordinates of the mesh nodes are... Where m is the grid node index, and the grid spacing is set according to the average sampling resolution of the spatial point cloud in the multimodal detection data. For each basic projected coordinate obtained through the initial linear mapping... Search for the nearest theoretical grid node in the local neighborhood of its own grid cell. Calculate the local curvature deviation between the base projected coordinates and the corresponding theoretical mesh nodes. : ; The deviation arises from the inability of the variable curvature surface to perfectly match the planar mesh under global linear mapping. It reflects the residual nonlinear deformation component after linear expansion. The local curvature deviations corresponding to all detection points are arranged in spatial order on the transition arc surface to form a curvature compensation residual sequence.

[0042] Step 2.3 involves performing an iterative update operation on the translation vector based on the curvature compensation residual sequence. This dynamically adjusts the displacement compensation components of each local sampling window within the two-dimensional unfolded plane, eliminating projection stretching and compression distortions caused by the variable curvature spatial topography. The result is obtained by mapping the updated displacement compensation components to the initial linear coordinates. Specifically, this includes: using the curvature compensation residual sequence as the observation signal, constructing an iterative update algorithm for the translation vector for each local sampling window; aggregating and analyzing the curvature compensation residuals of all detection points within the k-th local sampling window; extracting the overall displacement correction requirement for the window region; and defining the displacement compensation component of the k-th window in the t-th iteration as... initial value 0, the iterative update rule is: ; in, This represents the set of detection point indices belonging to the k-th local sampling window. This represents the number of detection points within the window. This is the iteration step size control coefficient, with a value between 0 and 1. This represents the residual value recalculated in the t-th iteration. After each iteration, the updated displacement compensation component is superimposed onto the base projection coordinates, and the curvature compensation residual is recalculated. The iteration termination condition is set to the mean square value change of the residual sequence between two adjacent iterations being lower than a preset convergence threshold. When the iteration terminates, the projection stretching and compression distortion caused by the variable curvature spatial topography is eliminated, and the updated displacement compensation component is obtained. And the basic topology determined by the initial linear coordinate mapping operation.

[0043] Step 2.4 involves superimposing and fusing the updated displacement compensation components with the results of the initial linear coordinate mapping calculation to construct a complete coordinate mapping relationship from the 3D detection point to the 2D unfolded plane. Specifically, this includes: integrating the final displacement compensation components of each local sampling window obtained through iterative solving. The result of the mapping operation with the initial linear coordinates Perform overlay and fusion. For detection points belonging to the k-th local sampling window... Its final mapped coordinates The calculation method is as follows: ; The basic projected coordinates obtained from the initial linear mapping are added point-by-point in the vector space to the nonlinear displacement compensation components obtained from iterative optimization, thus constructing a complete coordinate mapping relationship from the 3D detection point to the 2D unfolded plane. This mapping relationship is stored in the form of a lookup table or interpolation function, recording each 3D detection point. The index in the original space and its precise coordinates in the two-dimensional unfolded plane A two-way correspondence between them.

[0044] Step 2.5 involves mapping the variable curvature spatial topography of the transition arc region layer by layer to a unified two-dimensional coordinate domain based on the coordinate mapping relationship, and performing topological continuity calibration on the mapped multimodal feature layers to obtain aligned feature data for the transition arc region. Specifically, this includes: mapping the multimodal data layers acquired sequentially within the transition arc region to a unified two-dimensional coordinate domain based on the constructed complete coordinate mapping relationship; mapping the geometric height information of the spatial point cloud to the distance values ​​from the surface of each pixel on the two-dimensional plane to generate a height topography layer; projecting the surface optical grayscale texture map to the two-dimensional coordinate domain according to the same mapping relationship to generate a grayscale texture layer; and mapping the energy attenuation features in the ultrasonic echo time-frequency signal to the corresponding acoustic response layer at that coordinate point. During the mapping process, the coordinate mapping relationship ensures that the data layers of different modes are aligned pixel by pixel on the two-dimensional plane.

[0045] Topological continuity calibration is performed on the mapped multimodal feature layers. The adjacency relationships of adjacent pixels on the 2D unfolded plane are checked to ensure continuity in the original 3D space. For local mesh breaks or overlaps caused by interpolation during mapping, Laplacian smoothing correction based on the original 3D topological adjacency relationships is applied. After topological continuity calibration, the height topology layer, grayscale texture layer, and acoustic response layer are integrated into a unified data structure to obtain the transition arc region alignment feature data.

[0046] In this embodiment of the invention, the technical means of performing initial linear coordinate mapping through spatial transformation matrix, calculating local curvature deviation to generate curvature compensation residual sequence, dynamically adjusting displacement compensation component by iteratively updating translation vector based on residual sequence, superimposing and fusing compensation component with initial mapping result to construct complete coordinate mapping relationship, mapping variable curvature spatial morphology layer by layer and performing multimodal feature layer topology continuity calibration, thus overcoming the technical problems of simple linear mapping being unable to adapt to the variable curvature geometric features of transition arc region causing projection stretching and compression distortion, misalignment of projection coordinates and theoretical grid nodes caused by local curvature differences, and feature correspondence deviation caused by topological discontinuity after multimodal feature layer mapping, thereby achieving distortion-free projection of variable curvature spatial morphology of transition arc region to unified two-dimensional coordinate domain, ensuring spatial alignment accuracy and topological continuity of multimodal feature layer, and obtaining transition arc region alignment feature data.

[0047] In a preferred embodiment of the present invention, step 3 above may include: Step 3.1 involves performing local neighborhood topology analysis on the alignment feature data of the transition arc region to extract the 3D spatial pose information of each feature node within a unified 2D coordinate domain, and constructing a local differential mesh for the microscopic undulation morphology of the transition arc region. Specifically, this includes: in the alignment feature data of the transition arc region, taking each feature node within the unified 2D coordinate domain as the center, quickly extracting the set of K nearest neighbors for each feature node according to the spatial index structure. The value of K is determined based on the sampling resolution of the spatial point cloud in the multimodal detection data and the scale of the defect target, and is usually set to an integer value between 8 and 24.

[0048] For each feature node and its K nearest neighbor set, the original 3D Cartesian coordinates corresponding to each 2D coordinate are back-indexed using the established coordinate mapping relationship from the complete 3D detection point to the 2D unfolded plane. Since the mapping relationship is a bidirectional, exact correspondence and has been calibrated for topological continuity, the true coordinate values ​​of each feature node in the original 3D space of the transition arc region of the yaw bearing's inner gear ring can be recovered without error. This inverse indexing operation also extracts the 3D spatial pose information of the feature node, including its spatial location and the distribution of its local neighborhood along the three coordinate axes.

[0049] With feature nodes Centered on the K nearest neighbors, a local covariance matrix is ​​constructed using their 3D coordinates. : ; in, Let K be the mean vector of the 3D coordinates of its K nearest neighbors. Perform eigenvalue decomposition on this 3×3 covariance matrix to obtain three eigenvalues. and the corresponding feature vectors Minimum eigenvalue corresponding feature vector This is an approximate value of the principal normal direction of the local point set, used to characterize the microscopic normal information at the feature nodes.

[0050] Using each feature node as a topological vertex, adjacent nodes are connected in three-dimensional space according to the K-nearest neighbor relationship to generate triangular or polygonal patches, thus constructing a local differential mesh that characterizes the microscopic undulations of the transition arc region.

[0051] Step 3.2: Based on the local differential mesh, perform differential geometric operator operations along the normal direction of the mesh nodes to calculate the deflection angle of the tangent plane of the spatial surface in each local region and the principal direction of curvature, and aggregate them to obtain the three-dimensional topological geometric normal distribution characteristics. Specifically, this includes: traversing all mesh vertices one by one on the local differential mesh. For each vertex... The set of one or two rings of triangular facets formed by its local neighborhood is taken as the local surface approximation region. The vertex normal vectors are then refined by weighted averaging of the normals of all facets within this region. The weighting function is usually taken as the area of ​​the patch or the centroid coordinate coefficient, so that the normal estimation remains robust under non-uniform grids.

[0052] With vertices Refined normal vector at the location For reference normal, Projecting the coordinates of the vertices in a ring neighborhood onto the tangent plane defined by the normal yields a two-dimensional projection vector sequence. On the tangent plane, the two-dimensional coordinates are used to fit a local quadratic surface, or the coordinate transformations on the tangent plane are used to construct a discrete approximation of the shape operator. Specifically, this can be achieved by solving the Weingarten equation. ,in, Unit normal vector of the surface The differential change The differential of the tangent vector of the surface, Let be a shape operator, be a self-adjoint linear operator from tangent space to tangent space, be a 2×2 symmetric matrix, and its eigenvalues ​​are... and Let be the two principal curvatures of the surface, and let be the corresponding eigenvectors in the directions of the principal curvatures.

[0053] Based on principal curvature and Calculate the tangent plane deflection angle and principal curvature direction at each vertex. The tangent plane deflection angle is defined as the spatial deflection angle of the local tangent plane at that vertex relative to the theoretically designed surface of the transition arc region, and can be obtained through the normal vector. The angle between the curve and the theoretical reference normal vector is used. The principal direction of curvature is directly taken as the direction of the eigenvector corresponding to the maximum principal curvature in three-dimensional space. Simultaneously, to characterize the category of local surface morphology, Gaussian curvature can be calculated. and mean curvature These two values ​​respectively reflect whether the local surface is elliptical, hyperbolic, or parabolic, as well as the degree of local curvature.

[0054] Based on the above calculations, a local geometric feature vector containing multi-dimensional information is aggregated for each mesh vertex, specifically including: normal vector. The three components, tangential plane deflection angle The three components of the direction vector of maximum principal curvature, and the value of maximum principal curvature. Minimum principal curvature value Gaussian curvature Mean curvature The local geometric feature vectors of all vertices are reorganized according to the spatial order of the two-dimensional coordinate domain to form a feature layer set aligned with the original two-dimensional unfolded plane, thus aggregating the three-dimensional topological geometric normal distribution features.

[0055] Step 3.3 involves simultaneously extracting the surface optical grayscale texture map and the ultrasonic echo time-frequency signal mapping layer from the alignment feature data of the transition arc region. Multi-scale directional filtering is then performed within a unified two-dimensional coordinate domain to extract the grayscale jump rate and ultrasonic energy attenuation gradient within the neighborhood of each pixel. Specifically, this includes simultaneously extracting the surface optical grayscale texture map layer and the ultrasonic echo time-frequency signal mapping layer from the alignment feature data of the transition arc region obtained in Step 2, based on a unified two-dimensional coordinate domain pixel index. The two layers are spatially registered at the pixel level. Let the pixel grayscale value of the surface optical grayscale texture map be I(u, v), and the energy response value of the ultrasonic echo time-frequency signal mapping layer be U(u, v), where (u, v) are discrete pixel coordinates in the two-dimensional coordinate domain.

[0056] Within a unified two-dimensional coordinate domain, a set of multi-scale, multi-directional bandpass filters is constructed. The filter design needs to match the known directional characteristics of machining tool marks and the scale range of the crack target. Specifically, a two-dimensional Gabor filter bank is selected, and its kernel function is defined as: ; in, , , For wavelength, It is the direction angle. Let the standard deviation be the Gaussian envelope. The aspect ratio is specified. The filter bank covers at least 4–8 directions and 3 different scales. The filter bank is then subjected to two-dimensional convolution operations with the optical grayscale layer I(u,v) and the ultrasonic energy layer U(u,v), respectively.

[0057] For an optical grayscale texture map, at each pixel (u, v), a K×K neighborhood window is taken, and the amplitude of the filtering response in each direction within the window is analyzed to extract the grayscale jump rate. The grayscale jump rate is defined as the difference between the amplitude of the filtering response in the maximum direction and the amplitude of the filtering response in the minimum direction within the window, i.e.: ; This calculation method highlights the directional abrupt changes in local texture. Cracks typically manifest as sudden changes in direction inconsistent with the surrounding regular tool marks, and records the orientation angle corresponding to the maximum response. , representing the main direction of local texture.

[0058] For the ultrasonic echo time-frequency signal mapping layer, the ultrasonic energy attenuation gradient is extracted based on the filtering response in each direction. Due to the scattering and energy absorption effects of ultrasound on surface microcracks, the ultrasonic energy in the crack region exhibits a steep attenuation. The gradient magnitude of the ultrasonic energy filtering response within the pixel neighborhood is calculated: ; in, and These are differential approximations of the ultrasonic energy layer in the horizontal and vertical directions, respectively. To enhance directivity, gradient calculation is fused with directional filtering in each direction. First, directional filtering is performed, then the first-order directional derivative is calculated along that direction, and the envelope of the magnitudes of each directional derivative is taken as the final ultrasonic energy attenuation gradient. .

[0059] Step 3.4 involves performing cross-modal gradient vector synthesis on the gray-level jump rate and the ultrasonic energy attenuation gradient to eliminate interference components in the machining artifact direction and obtain surface texture gradient features. Specifically, this includes: based on the obtained gray-level jump rate at each pixel (u, v)... and its main direction and ultrasonic energy attenuation gradient and its direction Construct a two-dimensional gradient vector representation.

[0060] For optical modes, the gradient vector is defined as: ; For acoustic modes, the gradient vector is defined as: ; Cross-modal synthesis of gradient vectors from two modes yields the original texture gradient vector field. The synthesis can be achieved using weighted summation. ; in, and To normalize the weights, they are automatically optimized and determined based on the signal-to-noise ratio and defect contrast of the two modes on the training set, ensuring that the synthesized vectors can reflect both the gray-scale abrupt changes in surface texture and capture the physical characteristics of subsurface acoustic attenuation.

[0061] To eliminate interference components in the machining artifact direction, it is necessary to estimate the dominant direction field of the background tool mark texture. By calculating the direction distribution histogram of the entire transition arc region on the two-dimensional unfolded plane, the dominant direction of the periodic tool marks is determined using the density peak detection method. In the defect-free region, The direction will be concentrated towards Design a directional suppression mask function for the region and its vicinity. Angle less than preset threshold Within the range, it produces a suppression coefficient close to 0, while in other directions the coefficient is close to 1.

[0062] Apply the orientation suppression mask to the original texture gradient vector field. For each pixel, its gradient vector is projected onto the component near the main direction of the knife mark and attenuated. The specific operation is as follows: Calculate Unit vector of the principal direction of the tool mark The inner product of , if the absolute value of the inner product is greater than . Then from Subtract from the middle The directional components yield the gradient vector after artifact suppression. : ; in, This is for fine-tuning the threshold. This operation ensures that smooth gradients consistent with the main direction of the tool mark texture are significantly weakened, while gradients caused by random directions due to real defects such as cracks or those intersecting with the tool mark texture are preserved and enhanced.

[0063] Gradient vector field after eliminating machining artifact interference Perform scalarization and spatial aggregation. Calculate the texture gradient magnitude at each pixel: ; And its direction: ; Simultaneously, by using locally connected regions as units, scattered noise is suppressed to obtain continuous and physically meaningful surface texture gradient feature maps. For example, Figure 3 The image shows the original surface texture gradient magnitude map extracted directly from the alignment feature data, without any tool mark suppression processing. The horizontal and vertical axes represent the length and width of the two-dimensional unfolded plane, respectively, within a range of 0-100 pixels. Bright, diagonal lines arranged densely and periodically at 45° are clearly visible in the image.

[0064] In this embodiment of the invention, the following technical means are employed: local neighborhood topological analysis is performed on the alignment feature data of the transition arc region to construct a local differential mesh of micro-undulation morphology; the deflection angle of the tangent plane of the spatial surface and the main curvature direction are calculated based on differential geometric operators to obtain the three-dimensional topological geometric normal distribution features; multi-scale directional filtering is performed on the multi-modal feature layer to extract the gray-level jump rate and ultrasonic energy attenuation gradient; and cross-modal gradient vectors are synthesized to eliminate the directional interference components of machining artifacts. Therefore, this overcomes the technical problems of existing feature extraction methods, such as the inability to simultaneously and accurately capture surface micro-geometric deformation and multi-modal texture defect information, the susceptibility of single-modal texture features to strong interference from machining tool marks, and the insufficient dimension of defect representation due to the separation of geometric and texture features. This achieves the coordinated and accurate extraction of geometric features and multi-modal texture features, effectively suppressing the directional interference of machining artifacts, and obtaining the three-dimensional topological geometric normal distribution features and surface texture gradient features.

[0065] In a preferred embodiment of the present invention, step 3 above may include: Step 3.5 involves concatenating and fusing the 3D topological geometric normal distribution features and surface texture gradient features along the channel dimension to construct a multimodal joint feature matrix. Specifically, this includes obtaining the 3D topological geometric normal distribution features. This feature is represented as a multidimensional matrix in a unified two-dimensional coordinate domain, denoted as: ; in, and These represent the number of pixels in the height and width directions of the two-dimensional unfolded plane, respectively. This refers to the number of geometric feature channels. Geometric feature channels specifically include: normal vectors. The three component channels , Tangent plane deflection angle Channel, maximum principal curvature Channel, minimum principal curvature Channel, Gaussian curvature Channel, mean curvature Channel, and the three component channels of the maximum principal curvature direction vector. Each channel is strictly aligned in space with the pixel coordinates of the two-dimensional unfolded plane.

[0066] Simultaneously, surface texture gradient features are acquired, which are also represented as a multidimensional matrix in a unified two-dimensional coordinate domain, denoted as: ; in, This refers to the number of texture feature channels. Specifically, texture feature channels include: grayscale jump rate amplitude channels. Optical texture main direction angular channel Ultrasonic energy attenuation gradient amplitude channel The strongest acoustic response direction angle channel Cross-modal synthesis gradient magnitude channel Cross-modal synthesis gradient direction angle channel And statistical feature channels of directional filtering response at each scale, each channel is also pixel-level aligned with the pixel coordinates of the two-dimensional unfolded plane.

[0067] The two feature matrices are preprocessed with channel-dimensional standardization. For each channel, calculate its mean and standard deviation over the entire transition arc region. Use Z-score normalization to independently normalize each channel, ensuring uniformity of dimensions for all geometric features. The standardized procedure is as follows: ; in, and Geometric features The mean and standard deviation of the channel.

[0068] Similarly, regarding it Each channel in the process performs the same standardized operation: ; Perform a spatial resolution alignment check on the two standardized feature matrices. Since geometric and texture features may originate from local neighborhood operations at different scales, it is necessary to ensure the spatial dimensions of both. Completely identical. If resolution differences exist, the lower-resolution feature matrix is ​​sampled to a higher-resolution size using bilinear interpolation.

[0069] The standardized geometric feature matrix and texture feature matrix The concatenation is performed along the channel dimension, specifically by joining the two tensors along the channel axis: ; Total number of channels after splicing Each pixel position A length of is formed at the location The joint feature vector encodes both the three-dimensional surface geometry and multimodal texture anomaly information at that location.

[0070] For the constructed multimodal joint feature matrix Boundary processing is performed. Since the transition arc region is not a complete rectangular region on the 2D unfolded plane, its boundary is determined by the edge contour of the original 3D surface. For invalid pixel positions outside the boundary, their joint feature vector is set to zero, and the binary mask matrix of the valid region is recorded. .

[0071] Step 3.6 involves inputting the multimodal joint feature matrix into the dual-stream feature encoding network of the tool mark-crack feature decoupling model. This extracts implicit geometric features characterizing the macroscopic undulations of the surface and implicit texture features characterizing the frequency domain distribution of surface micro-textures. Specifically, this includes constructing the dual-stream feature encoding network structure for the tool mark-crack feature decoupling model. This network consists of two parallel, structurally similar but parameter-independent deep convolutional neural network branches, referred to as the geometric encoding stream and the texture encoding stream, respectively. Each encoding stream uses a residual convolutional network as its backbone structure, containing multiple downsampling stages and feature mapping layers.

[0072] For a geometrically encoded stream, the input is a multimodal joint feature matrix. The subset that belongs to the original geometric feature channels is denoted as The design goal of Geometric Encoding Flow is to extract implicit geometric features that characterize the macroscopic undulations of a surface. Since tool marks and cracks differ fundamentally in their geometric morphology—tool marks exhibit periodic, consistent-direction regular grooves and undulations, while cracks manifest as locally random, irregular micro-indentations and surface distortions caused by stress concentration—Geometric Encoding Flow needs a sufficient receptive field to capture multiple cycles of tool mark textures and a wider range of surface trend changes.

[0073] The geometric coding stream employs a five-stage residual coding structure. The first stage uses a 7×7 convolutional kernel with a stride of 2 to initially downsample the input features, expanding the output channel count to 64. This is followed by batch normalization and ReLU activation. The second to fourth stages each consist of two residual modules, each containing two 3×3 convolutional layers and introducing skip connections. The output channel counts for each stage are 128, 256, and 512, respectively, with the spatial resolution gradually decreasing to 1 / 4, 1 / 8, and 1 / 16 of the original size. The fifth stage uses dilated convolutions with dilation rates of 2, 4, and 8, expanding the receptive field without further reducing spatial resolution. This allows each element in the output feature map to perceive geometric undulations within a sufficiently large spatial neighborhood, thus distinguishing the macroscopic geometric differences between periodic knife marks and local cracks. The final output of the geometric coding stream is denoted as... ,in and The spatial dimensions after downsampling. The channel dimension is the implicit geometric feature.

[0074] For texture-coded streams, the input is a multimodal joint feature matrix. The subset that belongs to the original texture feature channels is denoted as The design goal of texture-encoded streams is to extract implicit frequency-domain features that characterize the frequency-domain distribution of surface micro-textures. Since optical textures and ultrasonic echo signals contain rich frequency information about surface microstructures, and the texture responses of cracks and tool marks are separable in the frequency domain, with tool mark textures exhibiting narrow-band directional spectra and crack textures exhibiting wide-band, non-directional spectral distributions, texture-encoded streams need to possess sophisticated analytical capabilities in the frequency domain dimension.

[0075] The texture coding stream employs a five-stage residual structure similar to the geometric coding stream, but with some adjustments in detail: A multi-scale frequency domain preprocessing module is added before the first stage. This module performs Gaussian blur and Laplacian pyramid decomposition on the input texture features at different scales, generating texture feature representations at three scales, corresponding to high-frequency details, mid-frequency texture, and low-frequency background, respectively. These three feature layers are concatenated along the channel dimension and fed into the coding network. A frequency domain attention module is inserted between the third and fourth stages of the coding network. This module performs a Fast Fourier Transform on the feature map, learns a frequency domain weight matrix, and initially suppresses narrowband frequency components containing periodic knife-mark information while enhancing broadband frequency components containing random crack textures. The final output of the texture coding stream is denoted as... Among them, space size Maintain consistency with the spatial dimensions of the geometrically encoded stream output. The channel dimension of the implicit texture features.

[0076] The geometric encoding stream and the texture encoding stream are kept strictly synchronized in the spatial resolution dimension through a uniform step size setting, ensuring that the implicit feature maps output by the two branches correspond element-wise in spatial location.

[0077] Step 3.7: Through the cross-attention decoupling layer in the tool mark-crack feature decoupling model, calculate the cross-modal correlation weights between the implicit features of geometric morphology and the implicit features of texture frequency domain. Separate the tool mark interference component with periodic directional regularity and the crack response component with local random orientation within the feature space. Specifically, this includes: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] and This serves as the input to the cross-attention decoupling layer. For ease of computation, the spatial dimensions are expanded, and the two feature maps are reshaped into matrix form: and ,in This represents the total number of spatial locations. Each row vector in the vector represents the implicit geometric features of a spatial location. The corresponding row vectors represent the implicit features of the texture in the frequency domain at the same spatial location.

[0078] A query, key, and value projection mechanism for cross-attention is constructed. To enhance the expressive power of features, three learnable linear projection matrices are used to generate the query, key, and value, respectively, for the cross-attention path from the geometric modality to the texture modality. Using geometric features as queries and texture features as keys and values: ; ; ; in, For the projection dimensions of the query and the key, The projection dimension of the value.

[0079] Similarly, for cross-attention paths from texture modality to geometry modality Using texture features as queries and geometric features as keys and values: ; ; ; Computational geometry to texture ( The cross-attention weight matrix for each spatial location in the geometric modality. Calculate its query vector With all spatial locations in the texture modality key vector The similarity between the two components is normalized using Softmax to obtain the attention weights. Since cracks and tool marks differ in their cross-modal correlation patterns of geometry and texture, the geometric fluctuations of periodic tool marks exhibit strong and regular cross-modal coupling with oriented textures, while the coupling pattern of geometric abrupt changes in random cracks and texture anomalies shows a non-periodic scattering characteristic. By analyzing the statistical properties of the attention weight matrix, key signals for separating the two types of components can be obtained.

[0080] use Attention weights are used to compute texture modal output guided by geometric features: ; Similarly, calculating texture to geometry ( The cross-attention weight matrix: ; And geometric modal output guided by texture features: ; The separation of the tool mark interference component and the crack response component is performed in the feature space. This separation operation is based on the following key understanding: the tool mark feature exhibits a strong periodic directional regularity in both modes, and its cross-modal attention weights show a highly structured band-like or grid-like distribution; while the crack feature exhibits a local random orientation in both modes, and its attention weights show a sparse and unstructured scattered distribution.

[0081] To quantify this difference, for each spatial location Calculate the local orientation consistency index of its attention weights. Specifically, based on position... Centered on, analysis Attention weight matrix The Middle The weights of the rows are distributed in the autocorrelation function within a two-dimensional neighborhood. If the autocorrelation function exhibits a significant periodic peak along a certain direction, it is determined that the position exhibits a periodic directional pattern and belongs to the characteristics of tool mark interference; if the autocorrelation function has no obvious directionality and the energy is mainly concentrated near zero delay, it is determined that the position exhibits a local random orientation and belongs to the characteristics of crack response.

[0082] Based on the above discrimination, a spatially analytical binary decoupling mask is generated. The location with a value of 1 represents the region dominated by the tool mark interference component, and the location with a value of 0 represents the region dominated by the crack response component. Alternatively, two soft-weighted continuous masks can also be generated. and ,satisfy This is used to model the coexistence and transition of the two types of components in the feature space more precisely.

[0083] By using a decoupling mask to perform weighted separation on the two cross-modal guided features, the tool mark interference component and the crack response component are obtained: ; ; in, This represents element-wise multiplication. A 1-row vector is used for broadcasting to the feature channel dimension. The tool mark interference component contains cross-modal coupling features that are highly correlated with the periodic tool mark texture; The crack response component contains cross-modal coupling features related to the random crack orientation.

[0084] Step 3.8: Perform feature masking suppression operation on the tool mark interference components to filter out the spurious effects caused by machining tool marks, obtaining spurious defect suppression features. Based on the feature map distribution of the spurious defect suppression features, determine the main direction of the tool mark texture. Specifically, this includes: using the separated tool mark interference components... To process the object, Reshaping back to a two-dimensional spatial form, we obtain The feature map is analyzed in the spatial dimension to identify the characteristic distribution pattern of the tool mark texture. Because the tool marks have a consistent direction in the transition arc region, their... The response vector in the 3D feature space should exhibit high correlation and amplitude stationarity among adjacent spatial positions along the direction of the tool mark.

[0085] To construct a spatial self-enhancing suppression mechanism, a learnable gated convolutional layer is designed, with its kernel aligned along the principal direction of the knife mark. (The principal direction of the knife mark is...) This has already been calculated in step 3.4 during the gradient vector synthesis stage. The unit direction vector in the tool mark direction is then... Sampling is performed on a spatial grid to generate a set of spatial offsets for the tool mark orientation. A gated convolutional layer performs convolution operations along a one-dimensional profile of the main tool mark direction, learning a position-dependent suppression gating value, generating a suppression coefficient vector for each spatial location in the feature map. : ; in, For the Sigmoid activation function, the gate value is limited to... In regions where the knife mark feature response is strong, the gating network will learn to output a gating value close to 1 for strong suppression; in non-knife mark regions, it will output a gating value close to 0, preserving the original features.

[0086] Using gated value pairs Perform feature mask suppression operation: ; in, The gating matrix is ​​a row-ordered arrangement of the gating vectors for all spatial locations. To suppress masking. After the tool mark interference component is suppressed, what remains are mainly small-amplitude scattered responses that cannot be interpreted as tool mark features by the gating network. These scattered responses are essentially background noise in the feature space and have no substantial correlation with the subsequent identification of crack features.

[0087] To further ensure the removal of spurious defect features, a minimum feature magnitude threshold filtering step is introduced after gating suppression. For each spatial location... Calculate the L2 norm of its suppressed eigenvectors. If the norm is lower than the preset threshold If the threshold value is zero, then the feature vector at that position is set to zero. The value is adaptively determined based on the statistical distribution of background noise in the feature space, and is usually set as the 95th percentile of the residual norm after stab mark suppression in the training set.

[0088] The tool mark component after gating suppression and threshold filtering is the pseudo-defect suppression feature, denoted as... The high-response areas in the feature map have been greatly reduced, and the pseudo-defect features that originally appeared in the tool mark texture area have been effectively filtered out, avoiding the situation where tool marks are misidentified as crack nodes when constructing the defect topology map.

[0089] Step 3.9 involves performing feature weight amplification and spatial reconstruction operations on the crack response components to highlight the local stress concentration response of micro-fatigue cracks and obtain the true defect enhancement characteristics. Specifically, this includes using the separated crack response components... For processing objects. Reshaping back to a two-dimensional spatial form, we obtain The crack response component includes the responses of real defects such as micro-fatigue cracks and quench spalling in the cross-modal characteristic space. However, these responses may have low amplitudes due to the small crack size and weak signal, and require special enhancement processing to be effectively identified in subsequent defect detection.

[0090] An adaptive amplification operation is performed on the feature weights of the crack response components. Since different types of defects exhibit different response patterns in the feature space, a channel attention mechanism is employed to learn the importance weights of each feature channel, selectively amplifying channels highly correlated with the crack. Specifically, for... Perform global average pooling to obtain a 3D channel description vector : ; The channel weight vector is calculated through a bottleneck structure consisting of two fully connected layers. : ; in , , The dimension after dimensionality reduction (usually) ), Activated for ReLU Activate Sigmoid. and Channel-by-channel multiplication achieves adaptive weight amplification along the channel dimension: ; During training, the channel attention network automatically learns that channels associated with crack features receive larger weights because they contribute more to defect classification, while channels coupled with background noise receive smaller weights.

[0091] Spatial reconstruction is performed on the weighted crack features. Considering that micro-fatigue cracks often extend linearly or arc-like in space, exhibiting spatial continuity, while random noise is spatially isolated and scattered, the spatial reconstruction operation aims to enhance the spatial continuity of the crack while suppressing isolated noise points. This operation is implemented through a spatial self-attention module: Calculate the weighted feature map Feature similarity between any two spatial locations: ; in, and For two independent 1×1 convolution mappings, This represents the mapped feature dimension. along The orientation is Softmax normalized to obtain the spatial self-attention weights. If the location and location Regions extending from the same crack should have highly similar feature vectors. A larger value is chosen; naturally, the attention weight between the crack area and the noise area will be lower.

[0092] Spatial aggregation and reconstruction of feature values ​​using self-attention weights: ; in, This is the third 1×1 convolution mapping. This operation allows the enhanced features at each location to aggregate the crack response of the surrounding associated region. For continuously extending small cracks, homogeneous features from adjacent locations converge, and the crack signal is significantly amplified. For isolated pseudo-response points, due to the lack of sufficient neighborhood support, their aggregated weights are evenly distributed, and the amplitude of the reconstructed response decreases rather than increases.

[0093] The spatially reconstructed crack features are fused with the channel-weighted features, and independent information from the original crack response is preserved through a residual connection: ; in The fusion coefficient is a learnable coefficient, initially set to 1.0 and automatically adjusted during training. The fused coefficient... This is a feature that enhances the true defects.

[0094] Spatial resolution reconstruction is performed on the augmented features of the real defects to restore them to the same spatial size as the original input. A method combining bilinear interpolation upsampling with skip connections is employed to... The feature map is progressively upsampled to the target resolution and then concatenated and fused with intermediate layer features of the corresponding scale from a dual-stream coding network. A lightweight convolutional decoding unit refines the features, ultimately outputting a true defect enhancement feature map. .

[0095] In this embodiment of the invention, a multimodal joint feature matrix is ​​constructed by splicing and fusing the three-dimensional topological geometric normal distribution features and surface texture gradient features in the channel dimension. Implicit geometric features and implicit texture frequency domain features are extracted separately through a dual-stream feature coding network. Cross-modal correlation weights are calculated using a cross-attention decoupling layer to separate periodic tool mark interference components and locally random crack response components. Feature masking suppression is performed on the tool mark interference components, and feature weight amplification and spatial reconstruction are performed on the crack response components. Therefore, this overcomes the technical problems of existing detection algorithms, such as the inability of global feature extraction or single threshold segmentation to effectively decouple pseudo-defects and real defects of machining tool marks in the feature space, insufficient decoupling capability of single-modal features, and high false positive and false negative rates caused by strong interference from tool mark textures masking micro-fatigue crack features. This achieves accurate separation of tool mark and crack features in the feature space, effectively filtering out pseudo-effects caused by machining tool marks, highlighting the local stress concentration response of micro-fatigue cracks and quenching spalling, and generating pseudo-defect suppression features and real defect enhancement features.

[0096] In a preferred embodiment of the present invention, step 4 above may include: Step 4.1: Based on the real defect enhancement features, extract the boundary contours and geometric center points of each suspected defect region within the transition arc region to construct an initial defect node set. Specifically, the real defect enhancement features are multi-channel feature maps after channel weight amplification and spatial reconstruction. Their spatial dimensions are completely consistent with the two-dimensional unfolded plane of the transition arc region, with dimensions of [dimension not specified]. ,in and These are the height and width pixels of the two-dimensional unfolded plane, respectively. This represents the number of output channels for the augmented feature map of the real defects. Channel-wise maximum response fusion is performed on the augmented feature map of the real defects to generate a single-channel defect response heatmap. The fusion method involves applying maximum response fusion to all channels at each pixel location. The maximum value of each channel feature is taken to obtain the defect response heatmap. Each pixel value represents the intensity of the defect response at that location; the larger the value, the higher the probability of a real defect.

[0097] Adaptive threshold segmentation is performed on the defect response heatmap, and the global final segmentation threshold is automatically calculated using Otsu's method. Pixel values ​​greater than in the heatmap The area is marked as a suspected defect area, with a pixel value less than or equal to... The region is marked as the background region. Morphological opening and closing operations are performed sequentially on the segmented binary image: first, a 3×3 rectangular structuring element is used for opening to eliminate isolated noise points and small edge burrs in the image; then, a 5×5 rectangular structuring element is used for closing to fill the tiny holes inside the suspected defect region, resulting in a complete and continuous binary mask of the suspected defect region.

[0098] An 8-neighborhood contour tracking algorithm was used to extract the closed boundary contours of all suspected defect regions, and the sequence of all pixel coordinates contained in each contour was recorded. For each extracted closed contour, the coordinates of its geometric center point were calculated; the geometric center point is the arithmetic mean of all pixel coordinates within that contour. For the ... Given a suspected defect region whose outline contains a set of pixel coordinates, find the coordinates of the geometric center point of that region. The calculation formula is: , ; Simultaneously, a corresponding feature vector is extracted for each suspected defect region. The feature vector consists of the average feature value of the enhanced feature map of the actual defect within that region, i.e., the average feature value of all pixel locations within that region. Taking the arithmetic mean of the eigenvectors of dimension 1, we obtain the eigenvector of dimension 2. Feature vectors of suspected defective regions The geometric center coordinates, feature vector, contour pixel coordinate sequence, and maximum defect response intensity of each suspected defect region are used as the core attributes of the node to construct an initial set of defect nodes. ,in This represents the total number of suspected defective areas extracted.

[0099] Step 4.2: Perform directional consistency filtering on the initial defect node set using the pseudo-defect suppression feature, removing pseudo-defect nodes whose angle with the main direction of the tool mark texture in the pseudo-defect suppression feature is less than a preset angle threshold, thus obtaining a pure defect node set. Specifically, this includes: extracting the main direction of the tool mark texture from the pseudo-defect suppression feature. The pseudo-defect suppression feature is a feature map after suppressing the tool mark interference component feature, which fully preserves the global directional distribution information of the machining tool mark texture. Oriented gradient histogram statistics are performed on the pseudo-defect suppression feature map. On the two-dimensional unfolded plane of the entire transition arc region, directional intervals from 0° to 180° are divided at 15° intervals. The sum of gradient magnitudes within each directional interval is calculated, and the center angle of the directional interval with the largest sum of gradient magnitudes is taken as the main direction of the tool mark texture. The unit is degrees.

[0100] For each node in the initial defect node set Calculate the principal direction of the corresponding suspected defect region. The principal direction of the suspected defect region is obtained by performing principal component analysis on the set of contour pixel coordinates of the region: first, the covariance matrix of the contour pixel coordinates is calculated, then eigenvalue decomposition is performed on the covariance matrix, and the angle between the eigenvector corresponding to the largest eigenvalue and the horizontal axis is taken as the principal direction of the region. The unit is degrees.

[0101] Calculate the principal direction of each initial defect node. With the main direction of the knife mark texture The minimum included angle between The formula for calculating the included angle is: ; Preset angle threshold The threshold is 20°, determined based on the statistical characteristics of the directional consistency of machining tool marks in the transition arc region of the wind turbine yaw bearing's inner gear ring. When the suspected defect area corresponding to the node is determined to be highly consistent with the direction of the tool mark texture, it belongs to a pseudo-defect node caused by machining tool marks, and is removed from the initial defect node set; when If the node is not found, retain it. After directional consistency filtering, a set of pure defective nodes is obtained. ,in The number of suspected defective areas to be retained, and .

[0102] Step 4.3: Calculate the spatial Euclidean distance and feature similarity between any two defective nodes in the pure defective node set. When the spatial Euclidean distance is less than a preset distance threshold and the feature similarity is greater than a preset similarity threshold, establish a topological connection edge between the corresponding two defective nodes. Specifically, this includes: traversing all unordered node pairs in the pure defective node set and calculating the spatial Euclidean distance and feature similarity between any two different nodes. and ( Spatial Euclidean distance between ) and feature similarity .

[0103] Feature similarity The cosine similarity between the feature vectors of two nodes is used to characterize the degree of feature similarity between two suspected defective regions.

[0104] Preset distance threshold The similarity threshold is set to 5 pixels. It is 0.7. When both conditions are met... and When two conditions are met, determine the node. and For different components belonging to the same continuously extending crack, establish an undirected topological connection edge between the two nodes; otherwise, do not establish a connection edge. After traversing all node pairs, obtain the set of all topological connection edges.

[0105] Step 4.4: Using the set of pure defect nodes as graph vertices and topological connecting edges as graph edges, and mapping the defect depth response values ​​of corresponding nodes in the real defect enhancement features to graph vertex weights, the defect topology graph of the transition arc region is assembled. Specifically, this includes: using the set of pure defect nodes... As the set of vertices of a graph , with topologically connected edge set As the set of edges in a graph, a defective topology graph is constructed to form an undirected graph structure. ,in This is the vertex weight matrix.

[0106] Vertex weights are obtained by mapping the defect depth response value of the corresponding node in the real defect enhancement feature map. The defect depth response value comes from the depth response channel feature value at the geometric center point of the node in the real defect enhancement feature map, denoted as . Its value range is directly related to the energy attenuation characteristics of the ultrasonic echo signal, normalizing the defect depth response values ​​of different orders of magnitude to a unified weight range. Vertex weights are stored as core attributes of the graph vertices. Vertex weight matrix In this process, the geometric center coordinates, feature vector, contour pixel coordinate sequence, principal direction, and maximum defect response intensity of each vertex are stored as additional attributes of the graph vertices.

[0107] The final assembled defect topology map fully represents the spatial distribution, geometric shape, feature similarity, topological connectivity, and defect depth information of all suspected real defects within the transition arc region. Each connected component in the map corresponds to a continuously extending crack or a cluster of defects. Vertex weights reflect the severity of the corresponding defect location, and topological connecting edges reflect the spatial association and feature similarity between different defect regions. For example... Figure 4The background shown is a grayscale 2D unfolded plane, with the horizontal and vertical axes representing the length and width of the 2D unfolded plane, respectively, within a range of 0-100 pixels. Three clean defect nodes, N1, N2, and N3, are drawn in the figure after directional filtering. Each node is represented by a circle of different sizes, with the circle size representing the defect depth response intensity. The calculated defect confidence score is labeled next to each node; for example, N1 has a confidence score C=0.92, N2 has a confidence score C=0.88, and N3 has a confidence score C=0.65. A bold line connects N1 and N2, representing a topological connection established based on spatial distance and feature similarity, indicating that N1 and N2 are likely different parts of the same extending crack, while N3 is an isolated node.

[0108] In this embodiment of the invention, an initial defect node set is constructed by extracting the boundary contours and geometric center points of each suspected defect region within the transition arc region based on real defect enhancement features. A pseudo-defect node set is then filtered for directional consistency using pseudo-defect suppression features to remove pseudo-defect nodes whose angle with the main direction of the tool mark texture is less than a preset threshold. A topological connection edge is established when the spatial Euclidean distance and feature similarity between pure defect nodes meet the double threshold condition. Finally, a defect topology graph is assembled using pure defect nodes as graph vertices and topological connection edges as graph edges, and mapping the defect depth response values ​​of corresponding nodes in the real defect enhancement features to graph vertex weights. This overcomes the technical problems of existing defect detection methods, such as the inability to utilize pseudo-defect suppression features for secondary precise filtering leading to a large number of residual pseudo-defect nodes, the inability to effectively characterize the spatial topological association and feature similarity between defect regions, the tendency to split continuously extending cracks into multiple independent defects, and the single defect representation dimension failing to reflect defect depth information. This achieves efficient purification of initial defect nodes, constructing a structured defect topology graph that fully reflects the spatial distribution, topological connectivity, and depth attributes of defects, clearly demonstrating the extension trend and local aggregation characteristics of micro-fatigue cracks.

[0109] In a preferred embodiment of the present invention, step 4 above may include: Step 4.5: Input the defect topology graph into a graph convolutional neural network. Aggregate the graph vertex weights and feature information of adjacent defect nodes through topological connection edges, and update the local topological feature representation of each defect node. Specifically, this includes: processing the obtained defect topology graph... Input a pre-trained two-layer graph convolutional neural network, and aggregate the graph vertex weights and original feature information of adjacent defect nodes through topological connection edges to generate a local topological feature representation that integrates spatial topological associations.

[0110] Before inputting the defective topology graph into the convolutional neural network, the graph is first converted to a standard graph data format. A vertex feature matrix is ​​then constructed. ,in This represents the total number of clean defect nodes. The number of output channels for enhancing features of real defects. Each row of the matrix corresponds to the initial feature vector of a defect node, the first... Dimension is the original feature vector of this node. The last dimension represents the graph vertex weight of that node. Simultaneously construct the adjacency matrix. If node With nodes If there are topologically connected edges, then ,otherwise Furthermore, all diagonal elements of the adjacency matrix are set to 1, and self-loop connections are added to each node to ensure that the node's own feature information is preserved during the aggregation process.

[0111] Perform symmetric normalization on the adjacency matrix and calculate the degree matrix. The degree matrix is ​​a diagonal matrix, and its diagonal elements are... For nodes The degree of adjacency is the total number of topological edges directly connected to that node. The normalized adjacency matrix... The calculation formula is: ； The normalized adjacency matrix With vertex feature matrix The first graph convolutional layer is input, and the first feature aggregation operation is performed. The learnable weight matrix of the first graph convolutional layer is: ,in The feature dimension of the first hidden layer is set to 64. The output feature matrix of the first graph convolution layer. The calculation formula is: ； in, It is the ReLU activation function, used to introduce nonlinear transformations and enhance the feature representation ability of the network.

[0112] The feature matrix output from the first layer The second graph convolutional layer is input to perform a second feature aggregation operation, further fusing topological association information from longer distances. The learnable weight matrix of the second graph convolutional layer is: ,in The feature dimension of the second hidden layer is set to 32. The output feature matrix of the second graph convolution is... The calculation formula is: ; Output feature matrix This represents the updated local topological features of each defect node, and the matrix's first... row vector Corresponding node The local topological features. This feature integrates the original features of the node itself, vertex weights, and feature and weight information of neighboring nodes, comprehensively characterizing the defective topological structure characteristics of the local region where the node is located.

[0113] Step 4.6, based on the updated local topological feature representation, calculate the weighted fusion value of the topological connectivity coefficient and the defect depth response intensity of each suspected defect region, specifically including: based on the updated local topological feature representation Calculate the topological connectivity coefficient for each suspected defect region. Response strength with defect depth The two values ​​are then weighted and fused to obtain a weighted fusion value. .

[0114] Topological connectivity coefficients characterize the degree of connectivity of a defect node in the defect topology graph, reflecting the likelihood that it belongs to a continuously extending crack. For a node... Topological connectivity coefficient The calculation method is as follows: first extract the set of all nodes in the connected component to which the node belongs. Calculate the total number of nodes in the connected component. and the sum of the vertex weights of all nodes within that connected component. Topological connectivity coefficients The vertex weight of this node The ratio of the weight of the connected components to the total weight of the connected components is calculated using the following formula: ; The range of values ​​for the topological connectivity coefficient is: The larger the value, the more important the node is in its connected component, and the higher the probability that it belongs to the core region of the crack.

[0115] Defect depth response strength Directly use the graph vertex weight of this node. This value has been normalized to a linear mapping. The range accurately reflects the depth and severity of the defect.

[0116] For topological connectivity coefficients Response strength with defect depth Perform weighted fusion, weighted fusion value The calculation formula is: ； in and For the weighting coefficients, satisfying Based on the hazardous characteristics of defects in the transition arc region of the internal gear ring of a wind turbine yaw bearing, the defect depth has a greater impact on the bearing's operational reliability than topological connectivity; therefore, a [details omitted]. , The weighting coefficient can be optimized and adjusted through cross-validation of a large number of labeled defect samples.

[0117] Step 4.7: Input the weighted fusion value into the pre-constructed nonlinear confidence mapping function to calculate the comprehensive defect probability of each suspected defect region in the multidimensional feature space. Specifically, this includes: inputting the weighted fusion value of each suspected defect region... Input a pre-constructed nonlinear confidence mapping function to calculate the comprehensive defect probability of the region in the multidimensional feature space. The nonlinear confidence mapping function is constructed using a three-layer fully connected neural network, trained on a dataset containing 10,000 labeled defect samples. The input layer of this network has a dimension of 1, corresponding to the weighted fusion value. The first hidden layer has a dimension of 16 and uses the ReLU activation function; the second hidden layer has a dimension of 8 and uses the ReLU activation function; the output layer has a dimension of 1 and uses the Sigmoid activation function to ensure the overall defect probability of the output. The value range is (0, 1). For the input weighted fusion value... Overall defect probability The calculation process is as follows: ; in, The learnable weight matrix for each layer, These are the learnable bias vectors for each layer, and all parameters are optimized during training using the backpropagation algorithm. (Comprehensive defect probability) By comprehensively considering the topological connectivity, depth response intensity, and defect feature patterns in a multi-dimensional feature space, this method more accurately reflects the probability that a region is a real defect. The closer the value is to 1, the higher the probability that the region is a real defect; the closer the value is to 0, the higher the probability that the region is a false defect or noise.

[0118] Step 4.8: By performing normalization processing on the comprehensive defect probability, the defect confidence of each suspected defect region in the transition arc region is obtained. Specifically, this includes: performing global minimum-maximum normalization processing on the comprehensive defect probability of all suspected defect regions, mapping these probability values ​​uniformly to the interval between 0 and 1, and finally obtaining the final defect confidence of each suspected defect region.

[0119] Iterate through all clean defect nodes, comparing the comprehensive defect probability values ​​corresponding to each node, and then filter out the maximum and minimum values. If the maximum and minimum values ​​are exactly equal, it means that there is no difference in the comprehensive defect probability of all suspected defect areas. In this case, the defect confidence score of all nodes is uniformly set to 0.5. If the maximum value is greater than the minimum value, calculate the defect confidence score for each suspected defect area separately. The defect confidence score obtained after the above normalization process is fixed between 0 and 1, which intuitively reflects the defect severity of each suspected defect area and the probability of it being a real defect. A defect confidence score of 0 means that there is no defect risk in the area, while a defect confidence score of 1 means that the area is the real defect with the highest severity. The final defect confidence score is obtained.

[0120] In this embodiment of the invention, the defect topology graph is input into a graph convolutional neural network. The network aggregates the graph vertex weights and feature information of adjacent defect nodes through topological connection edges to update the local topological feature representation. It calculates the weighted fusion value of the topological connectivity coefficient and the defect depth response intensity. It calculates the comprehensive defect probability in the multidimensional feature space through a nonlinear confidence mapping function. It then performs normalization processing on the comprehensive defect probability to obtain the defect confidence. Therefore, this method overcomes the technical problems of existing defect confidence calculation methods that rely only on the independent features of a single defect region, cannot utilize the spatial topological association information between defects, ignore the key attributes of defect depth and connectivity, and whose linear statistical methods cannot accurately reflect the true defect probability under multidimensional features, resulting in low confidence calculation accuracy and poor defect judgment reliability. This method fully integrates the spatial topological structure, depth response intensity, and local feature information of defects, realizes the accurate quantification of the comprehensive defect probability of each suspected defect region, and generates an objective and accurate defect confidence.

[0121] In a preferred embodiment of the present invention, step 5 above may include: Step 5.1: Obtain the historical operating condition parameters of the yaw bearing internal gear ring and the process constraints of the current quality witness point. Calculate the dynamic release threshold corresponding to the quality witness point using a preset threshold adaptive adjustment model. Specifically, this includes: obtaining the historical operating condition parameters of the yaw bearing internal gear ring for this model and the process constraints of the current quality witness point. The historical operating condition parameters cover the cumulative operating data of this bearing model in wind farms with different climate types across the country, specifically including the rated load factor, annual average alternating load cycle count, minimum temperature in extreme low-temperature environments, design service life, critical defect expansion size in past failure cases, and remaining life data corresponding to different defect levels. The process constraints of the current quality witness point include the material grade of the internal gear ring in the current production batch, the measured average value of the quenched layer depth, the surface roughness grade of the finished product, the process stage corresponding to this witness point, the special quality level requirements agreed upon in the customer contract, and the process capability index of the current production line.

[0122] The threshold adaptive adjustment model is constructed using the XGBoost gradient boosting tree framework. The training dataset contains 5000 sets of labeled samples of qualified bearings throughout their entire life cycle under different working conditions and processes. The input is a 12-dimensional feature vector, where 6 dimensions represent historical operating condition parameters such as rated load coefficient and annual average alternating load cycle count, and 6 dimensions represent current process constraints such as quenching layer depth and surface roughness. The data preprocessing stage sequentially performs median imputation of missing values, 3σ principle outlier removal, Z-score standardization of numerical features, and target encoding of categorical features. The model output label is the final release threshold under the corresponding parameter combination. The maximum permissible defect confidence level is calculated through finite element stress analysis and fatigue crack propagation theory, and is combined with feedback from actual engineering operations and review by senior quality engineers. The model was calibrated and confirmed. The core architecture uses a CART decision tree as the base learner, with 100 trees, a maximum depth of 6, a learning rate of 0.1, and both subsampling and column sampling rates of 0.8. L1 regularization coefficients of 0.1 and L2 regularization coefficients of 1.0, along with a minimum leaf node sample size of 5, are used to control model complexity. The loss function is a weighted mean square error adapted to the priority principle of wind power quality and safety. High-risk samples that underestimate the release threshold are given 3 times the weight, while samples that overestimate or accurately predict the threshold are given 1 times the weight. During training, the training, validation, and test sets are divided using stratified sampling in a 7:2:1 ratio. An early stopping mechanism and 5-fold cross-validation are introduced to optimize the model. The final model achieves a false release rate ≤0.1% and a false interception rate ≤5% on the test set, meeting the accuracy and safety requirements of industrial inspection.

[0123] Standardization is performed on all input features to map features of different dimensions to a standard normal distribution interval with a mean of 0 and a variance of 1, thereby eliminating the impact of dimensional differences on the model's prediction results.

[0124] The preprocessed 12-dimensional feature vector is input into the gradient boosting tree regression model. The model uses ensemble prediction from 100 decision trees to output the initial prediction threshold for the current quality witness point. To ensure the safety and robustness of the threshold, a safety factor is introduced. The initial prediction threshold is adjusted, and the safety factor is increased. The value ranges from 0.85 to 0.95, and is dynamically adjusted based on the processing quality stability of the first 10 products in the current batch. The higher the processing quality stability, the better. The closer the value is to 0.95, the better. This represents the final dynamic release threshold after correction. The calculation formula is: ; The dynamic release threshold is forcibly limited to a range of 0.6 to 0.9 to ensure that a large number of qualified products are not mistakenly intercepted due to an excessively low threshold, nor that unqualified products with potential failure risks flow into subsequent assembly processes due to an excessively high threshold.

[0125] Step 5.2 involves comparing the defect confidence score of each suspected defect area with the dynamic release threshold point by point, and filtering out target defect areas whose defect confidence score is greater than the dynamic release threshold. Specifically, this includes obtaining all calculated... The defect confidence level corresponding to each pure defect node, and the calculated dynamic release threshold for the current quality witness point. .

[0126] Perform a point-by-point comparison of the defect confidence score for each suspected defect area: sequentially compare the defect confidence score of each node. With dynamic release threshold If a comparison is made, If the area is identified as a target defect area with quality risk, its corresponding node information, geometric contour, location coordinates, and estimated defect depth are all added to the target defect area set. ;like If the defect is identified as a pseudo-defect with no quality risk or an acceptable minor process defect, it will not be included in the target defect region set. After traversing all pure defect nodes, the final target defect region set is obtained.

[0127] Step 5.3 involves statistically analyzing the number, spatial distribution density, and maximum defect confidence of target defect regions to obtain accurate defect detection results for the transition arc region. Specifically, this includes: based on the target defect region set, statistically analyzing three core quality assessment indicators to generate complete accurate defect detection results for the transition arc region; and directly counting the total number of target defect regions included in the set. This indicator directly reflects the total number of defects with quality risks within the transition arc region. The distribution density of the target defect area within the effective detection area of ​​the transition arc region is calculated to characterize the degree of local defect aggregation. The effective detection area of ​​the transition arc region of the internal gear ring of this type of yaw bearing is obtained. The unit is square centimeters. This area is pre-calculated based on the nominal diameter of the internal gear ring, the design width of the transition arc, and the scanning coverage of the detection equipment. Spatial distribution density. The calculation formula is: ; The unit of spatial distribution density is defects per square centimeter. The higher the value, the higher the degree of concentration of defects in a local area, and the greater the risk of stress concentration during operation leading to rapid crack propagation.

[0128] like If the maximum defect confidence level is 0, then the maximum defect confidence level is directly set to 0. The maximum defect confidence level reflects the risk level of the most serious defect in the transition arc region. The closer the value is to 1, the higher the probability that the defect will cause early failure of the bearing during service. The above three core statistical indicators are integrated with the detailed information of each target defect region to generate a complete and accurate defect detection result report. The report also includes the geometric center point coordinates, the size of the outline bounding rectangle, the estimated defect depth, and the position marking map on the two-dimensional unfolded plane of the transition arc region for each target defect region.

[0129] Step 5.4: Based on the accurate defect detection results, match the pre-set quality control strategy library to obtain the corresponding status control instructions. These instructions include release instructions, interception and re-inspection instructions, or scrap and isolation instructions, and are then sent to the execution terminal at the quality witness point. Specifically, this involves: pre-constructing a standardized quality control strategy library, which stores standard quality control processes and status control instructions corresponding to different defect detection results. The matching rules of the strategy library are based on the number of target defect areas. Spatial distribution density and maximum defect confidence Three core indicators were established, and the specific matching rules are as follows: Clearance instruction: When When the transition arc area is cleared, it indicates that there are no target defects requiring treatment, and the quality status fully meets the release requirements of the current witness point, generating a release pass instruction. The instruction includes the unique number of the internal gear ring in the current batch, the quality witness point number, the test result conclusion of being qualified, the release time, the inspector number, and the automatically generated electronic quality seal.

[0130] Intercepting Re-inspection Command: An intercepting re-inspection command is generated when any of the following conditions are met: and per square centimeter and ; If the longest side of the bounding rectangle of a single target defect area is less than 0.5 mm and the estimated defect depth is less than 0.1 mm, the interception re-inspection instruction includes the specific coordinate range of the area to be re-inspected, the re-inspection technical requirements, the re-inspection time limit, and the information of the full-time inspection personnel responsible for the re-inspection.

[0131] Decommission isolation command: A decommission isolation command is generated when any of the following conditions are met: or Units per square centimeter; ; If the longest side of the circumscribed rectangle of a single target defect area is greater than 1 mm or the estimated defect depth is greater than 0.3 mm, the scrap isolation instruction includes the unique number of the non-conforming internal gear ring, a detailed description of the defect location and severity, physical isolation requirements, non-conforming product identification method, and information on quality traceability and cause analysis procedures.

[0132] After generating the corresponding status control instructions, the instructions are sent in real time to all execution terminals at the quality witness point via industrial Ethernet, including workshop manufacturing execution system terminals, operation displays of testing equipment, and audible and visual alarm devices on the production line. The accurate defect detection results report and status control instructions are synchronously encrypted and stored in the enterprise's product quality traceability system, realizing traceability, queryability, and auditability of quality data throughout the entire life cycle of wind power core components.

[0133] In this embodiment of the invention, by combining historical operating parameters of the yaw bearing internal gear ring with current process constraints at the quality witness point, calculating the dynamic release threshold through a preset threshold adaptive adjustment model, comparing the defect confidence level with the dynamic threshold point by point to screen target defect areas, statistically analyzing the spatial distribution density of target defects and the maximum defect confidence level to obtain detection results, and generating and issuing control instructions for interception, re-inspection, or scrapping isolation based on the detection results and a preset quality control strategy library, this invention overcomes the technical problems of existing technologies that use static fixed release thresholds, cannot adapt to the differentiated release requirements of quality witness points under different operating conditions and process constraints, cannot comprehensively assess the overall quality status of the transition arc area based solely on the judgment result of a single defect, and suffer from low efficiency due to the lack of automated hierarchical management of the quality control process. This achieves dynamic adaptive adjustment of the release threshold at the quality witness point, comprehensively and accurately assesses the severity of defects and the overall quality level of the transition arc area, automatically generates and issues precise quality status control instructions, meets the stringent hierarchical quality control requirements of wind power core components, improves the detection efficiency and control accuracy of the quality witness point, and ensures that non-conforming products are intercepted and processed in a timely manner.

[0134] like Figure 2 As shown, embodiments of the present invention also provide a wind power quality witness point-oriented precision detection system for component defects, comprising: The acquisition module is used to acquire multimodal detection data of the transition arc region between the inner gear ring and the raceway of the yaw bearing of the wind turbine generator at the quality witness point; the multimodal detection data is spatially registered and three-dimensional topological surface unfolded to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin; The mapping module is used to solve the coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane through linear mapping operation of the spatial transformation matrix and iterative translation vector. Based on the coordinate mapping relationship, the variable curvature spatial topography of the transition arc region is projected to a unified two-dimensional coordinate domain to obtain the alignment feature data of the transition arc region. The extraction module is used to extract the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain based on the alignment feature data of the transition arc region; the three-dimensional topological geometric normal distribution features and surface texture gradient features are input into a pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features. The calculation module is used to construct a defect topology map of the transition arc region based on the pseudo-defect suppression features and the real defect enhancement features; and to calculate the defect confidence of each suspected defect region in the transition arc region based on the defect topology map. The comparison module is used to compare the defect confidence level with the dynamic release threshold corresponding to the quality witness point, obtain the accurate defect detection result of the transition arc area, and output the status control command of the quality witness point.

[0135] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0136] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for precise detection of component defects guided by wind power quality witnessing points, characterized in that, The method includes: Multimodal detection data of the transition arc region between the internal gear ring and raceway of the yaw bearing of a wind turbine generator were collected at the quality witness point; spatial registration and three-dimensional topological surface unfolding were performed on the multimodal detection data to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin; By solving the coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane through linear mapping operation of the spatial transformation matrix and iterative translation vector, the variable curvature spatial topography of the transition arc region is projected onto a unified two-dimensional coordinate domain according to the coordinate mapping relationship, and the alignment feature data of the transition arc region is obtained. Based on the alignment feature data of the transition arc region, the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain are extracted; the three-dimensional topological geometric normal distribution features and surface texture gradient features are input into a pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features. Based on the false defect suppression features and the real defect enhancement features, a defect topology map of the transition arc region is constructed; based on the defect topology map, the defect confidence of each suspected defect region in the transition arc region is calculated. The defect confidence level is compared with the dynamic release threshold corresponding to the quality witness point to obtain the accurate defect detection result of the transition arc area, and the status control command of the quality witness point is output.

2. The wind power quality witness point-guided precise component defect detection method according to claim 1, characterized in that, Acquire multimodal test data of the transition arc region between the internal gear ring and raceway of the yaw bearing of a wind turbine generator at quality witness points, including: A synchronous spatial scan was performed on the transition arc region to acquire the original detection sequence, which includes spatial point cloud coordinate sequence, surface optical grayscale texture map and ultrasonic echo time-frequency signal; Based on the preset physical reference markers in the original detection sequence, the initial spatial position vectors of visual feature nodes and acoustic feature nodes in the global coordinate system of the yaw bearing internal gear ring are extracted. By performing cross-modal coordinate mapping verification using the initial spatial position vector, the spatial dimensions of the surface optical grayscale texture map and the ultrasonic echo time-frequency signal are aligned to the corresponding grid nodes of the spatial point cloud coordinate sequence, thus completing pixel-level registration and data fusion of multimodal signals under a unified spatiotemporal reference to obtain multimodal detection data.

3. The wind power quality witness point-guided precise component defect detection method according to claim 2, characterized in that, Spatial registration and 3D topological surface unfolding processing are performed on the multimodal detection data to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin, including: Local sampling windows are divided along the curvature change trajectory of the three-dimensional topological point set. For the three-dimensional spatial coordinate set within each local sampling window, the sum of squared residuals of the spatial coordinates is minimized to solve the final local reference surface parameters corresponding to each local sampling window and determine the normal reference vector and the coordinates of the reference surface center. Using the normal reference vector as the reference rotation axis and the center coordinates of the reference plane as the origin of the new local coordinate system, a local spatial mapping parameter set containing rotational and translational transformation components is constructed. The local spatial mapping parameter set is serialized and spliced ​​along the curvature change trajectory and topological continuity constraint is iterated to eliminate spatial step misalignment and deformation gradient abrupt change between adjacent local windows, thus obtaining a spatial transformation matrix covering the global range of the transition arc region.

4. The wind power quality witness point-guided precise component defect detection method according to claim 3, characterized in that, The coordinate mapping relationship from the 3D detection point to the 2D unfolded plane is solved by linear mapping operations of the spatial transformation matrix and iterative translation vector calculation. Based on the coordinate mapping relationship, the variable curvature spatial topography of the transition arc region is projected onto a unified 2D coordinate domain to obtain the alignment feature data of the transition arc region, including: The spatial transformation matrix is ​​applied to the three-dimensional detection point set in the multimodal detection data to perform an initial linear coordinate mapping operation, transforming the three-dimensional spatial coordinates to the basic projected coordinates of the two-dimensional unfolded plane; The local curvature deviation between the basic projected coordinates and the nodes of the two-dimensional unfolded plane theoretical mesh is calculated to obtain the curvature compensation residual sequence; Based on the curvature compensation residual sequence, the translation vector iterative update operation is performed to dynamically adjust the displacement compensation components of each local sampling window in the two-dimensional unfolded plane, eliminate the projection stretching and compression distortion caused by the variable curvature spatial topography, and obtain the result of the updated displacement compensation components and the initial linear coordinate mapping operation. The updated displacement compensation components are superimposed and fused with the results of the initial linear coordinate mapping calculation to construct a complete coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane. Based on the coordinate mapping relationship, the variable curvature spatial topography of the transition arc region is mapped layer by layer to a unified two-dimensional coordinate domain, and topological continuity calibration is performed on the mapped multimodal feature layer to obtain the alignment feature data of the transition arc region.

5. The wind power quality witness point-guided precise component defect detection method according to claim 4, characterized in that, Based on the alignment feature data of the transition arc region, the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain are extracted, including: Local neighborhood topology analysis is performed on the alignment feature data of the transition arc region to extract the three-dimensional spatial pose information of each feature node in the unified two-dimensional coordinate domain, and a local differential mesh of the micro-undulation morphology of the transition arc region is constructed. Based on the local differential mesh, differential geometric operators are performed along the normal direction of the mesh nodes to calculate the deflection angle of the tangent plane of the spatial surface of each local region and the principal direction of curvature, and then aggregated to obtain the three-dimensional topological geometric normal distribution characteristics. The surface optical grayscale texture map and ultrasonic echo time-frequency signal mapping layer in the alignment feature data of the transition arc region are simultaneously extracted. Multi-scale directional filtering operation is performed in a unified two-dimensional coordinate domain to extract the grayscale jump rate and ultrasonic energy attenuation gradient in the neighborhood of each pixel. Cross-modal gradient vector synthesis is performed on the gray-level jump rate and ultrasonic energy attenuation gradient to eliminate interference components in the machining artifact direction and obtain surface texture gradient features.

6. The wind power quality witness point-guided precise component defect detection method according to claim 5, characterized in that, By inputting the 3D topological geometric normal distribution features and surface texture gradient features into a pre-trained tool mark-crack feature decoupling model, pseudo-defect suppression features and real defect enhancement features are obtained, including: The three-dimensional topological geometric normal distribution features and surface texture gradient features are spliced ​​and fused in the channel dimension to construct a multimodal joint feature matrix; The multimodal joint feature matrix is ​​input into the dual-stream feature encoding network of the tool mark-crack feature decoupling model to extract the geometric implicit features that characterize the macroscopic undulation of the surface and the texture frequency domain implicit features that characterize the frequency domain distribution of the surface micro-texture, respectively. By using the cross-attention decoupling layer in the tool mark-crack feature decoupling model, the cross-modal correlation weight between the implicit features of geometric morphology and the implicit features of texture frequency domain is calculated, and the tool mark interference component with periodic directional regularity and the crack response component with local random orientation are separated in the feature space. The feature mask suppression operation is performed on the tool mark interference component to filter out the pseudo-effect caused by machining tool marks and obtain the pseudo-defect suppression feature. Based on the feature map distribution of the pseudo-defect suppression feature, the main direction of the tool mark texture is determined. The crack response components are subjected to feature weight amplification and spatial reconstruction operations to highlight the local stress concentration response of micro fatigue cracks and obtain the true defect enhancement features.

7. The wind power quality witness point-guided precise component defect detection method according to claim 6, characterized in that, Based on the pseudo-defect suppression features and the real defect enhancement features, a defect topology map of the transition arc region is constructed, including: Based on the real defect enhancement features, the boundary contours and geometric center points of each suspected defect region within the transition arc area are extracted to construct an initial defect node set. By performing directional consistency filtering on the initial defect node set through the pseudo-defect suppression feature, pseudo-defect nodes with an angle less than a preset angle threshold with the main direction of the tool mark texture in the pseudo-defect suppression feature are removed, resulting in a pure defect node set. Calculate the spatial Euclidean distance and feature similarity between any two defect nodes in the pure defect node set. When the spatial Euclidean distance is less than a preset distance threshold and the feature similarity is greater than a preset similarity threshold, establish a topological connection edge between the corresponding two defect nodes. Using the set of pure defect nodes as graph vertices and the topological connecting edges as graph edges, and mapping the defect depth response values ​​of the corresponding nodes in the real defect enhancement features to the graph vertex weights, a defect topology graph of the transition arc region is assembled.

8. The wind power quality witness point-guided precise component defect detection method according to claim 7, characterized in that, Based on the defect topology map, the defect confidence level of each suspected defect region in the transition arc region is calculated, including: The defect topology graph is input into a graph convolutional neural network. The graph vertex weights and feature information of adjacent defect nodes are aggregated through topological connection edges to update the local topological feature representation of each defect node. Based on the updated local topological feature representation, the weighted fusion value of the topological connectivity coefficient and the defect depth response intensity of each suspected defect region is calculated. The weighted fusion value is input into a pre-constructed nonlinear confidence mapping function to calculate the comprehensive defect probability of each suspected defect region in the multidimensional feature space. By normalizing the overall defect probability, the defect confidence level of each suspected defect area in the transition arc region is obtained.

9. The wind power quality witness point-guided precise component defect detection method according to claim 8, characterized in that, The defect confidence level is compared with the dynamic release threshold corresponding to the quality witness point to obtain the accurate defect detection result in the transition arc region, and the status control command of the quality witness point is output, including: Obtain the historical operating parameters of the yaw bearing internal gear ring and the process constraints of the current quality witness point. Calculate the dynamic release threshold corresponding to the quality witness point through a preset threshold adaptive adjustment model. The defect confidence level of each suspected defect area is compared with the dynamic release threshold point by point to screen out the target defect areas whose defect confidence level is greater than the dynamic release threshold. By statistically analyzing the number of target defect areas, their spatial distribution density, and the maximum defect confidence level, accurate defect detection results for the transition arc region are obtained. Based on the accurate defect detection results, a preset quality control strategy library is matched to obtain the corresponding status control instructions. The status control instructions include release instructions, interception and re-inspection instructions, or scrap isolation instructions, which are then sent to the execution terminal of the quality witness point.

10. A wind power quality witness point-oriented precision detection system for component defects, the system implementing the method as described in any one of claims 1 to 9, characterized in that, include: The acquisition module is used to acquire multimodal detection data of the transition arc region between the inner gear ring and the raceway of the yaw bearing of the wind turbine generator at the quality witness point; the multimodal detection data is spatially registered and three-dimensional topological surface unfolded to construct a spatial transformation matrix with the local reference plane of the transition arc region as the origin; The mapping module is used to solve the coordinate mapping relationship from the three-dimensional detection point to the two-dimensional unfolded plane through linear mapping operation of the spatial transformation matrix and iterative translation vector. Based on the coordinate mapping relationship, the variable curvature spatial topography of the transition arc region is projected to a unified two-dimensional coordinate domain to obtain the alignment feature data of the transition arc region. The extraction module is used to extract the three-dimensional topological geometric normal distribution features and surface texture gradient features of the transition arc region in a unified two-dimensional coordinate domain based on the alignment feature data of the transition arc region; the three-dimensional topological geometric normal distribution features and surface texture gradient features are input into a pre-trained tool mark-crack feature decoupling model to obtain pseudo-defect suppression features and real defect enhancement features. The calculation module is used to construct a defect topology map of the transition arc region based on the pseudo-defect suppression features and the real defect enhancement features; Calculate the defect confidence level of each suspected defect region in the transition arc region based on the defect topology map; The comparison module is used to compare the defect confidence level with the dynamic release threshold corresponding to the quality witness point, obtain the accurate defect detection result of the transition arc area, and output the status control command of the quality witness point.

Images