Lightweight component surface defect quantitative evaluation method, system and device

Abstract

The invention relates to a lightweight component surface defect quantitative evaluation method, system and device, and the method comprises the steps: obtaining a multispectral surface image, laser three-dimensional data and ultrasonic detection data of a lightweight component, and carrying out the fusion registration based on a preset multi-sensor synchronization protocol, and obtaining a component joint data set; performing dynamic pruning analysis on the component joint data set to obtain geometric contour features, and performing internal feature extraction on the component joint data set to obtain internal abnormal response features; stress gradient distribution data and volume ratio parameters are carried out on the geometric contour features and the internal anomaly features; and obtaining an original working condition data set of the lightweight component, carrying out service working condition coupling analysis on the multi-dimensional prediction matrix, and carrying out multi-dimensional quality l parameter identification to obtain component comprehensive quality information. According to the method, the morphology mechanical coupling feature extraction precision of composite defects such as cracks and holes can be improved.

Description

Technical Field

[0001] This invention relates to the technical field of component inspection, and in particular to a method, system and device for quantitative assessment of surface defects in lightweight components. Background Technology

[0002] As core components in the manufacturing of equipment such as new energy vehicles, the accurate detection of surface and internal defects in lightweight components directly affects the safety and reliability of the overall structure. With the development of high-precision manufacturing technology, the geometric complexity and material diversity of components have significantly increased. Traditional detection methods based on single-modal sensing (such as single optical or ultrasonic testing) are insufficient to meet the quantitative assessment needs of multi-dimensional defects. Existing technologies typically employ staged independent analysis of surface morphology and internal anomalies, lacking multi-source data spatiotemporal synchronization and cross-modal feature correlation mechanisms, leading to increased rates of missed detection and false positives for minute defects. Furthermore, defect evolution prediction models do not fully consider the coupling effect of material property differences and dynamic service conditions, resulting in significant deviations between quality assessment results and actual failure behavior. Summary of the Invention

[0003] The main objective of this invention is to provide a method, system, and device for quantitative evaluation of surface defects in lightweight components, which can improve the accuracy of extracting the morphological and mechanical coupling features of complex defects such as cracks and holes.

[0004] To achieve the above objectives, the present invention provides a method for quantitatively evaluating surface defects in lightweight components, comprising: Multispectral surface images, laser 3D data, and ultrasonic test data of lightweight components are acquired and fused and registered based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components. Dynamic pruning analysis is performed on the joint dataset of the components to obtain geometric contour features, and internal feature extraction is performed on the joint dataset of the components to obtain internal abnormal response features; Stress gradient distribution data and volume ratio parameters are obtained for the geometric contour features and the internal anomaly features, and defect prediction is performed according to the preset component material type mapping rules to obtain a multi-dimensional prediction matrix. The original working condition dataset of the lightweight component is obtained, and the service condition coupling analysis is performed on the multi-dimensional prediction matrix. Multi-dimensional quality parameters are identified to obtain the comprehensive quality information of the component.

[0005] Furthermore, the acquisition of multispectral surface images, laser 3D data, and ultrasonic testing data of the lightweight component, and the fusion and registration based on a preset multi-sensor synchronization protocol, yields a joint dataset of the component, including: The lightweight component is subjected to multi-band combined irradiation on its surface, and the reflected beam is separated by a preset beam-splitting prism array to obtain a multi-channel spectral image; Subpixel-level feature point matching is performed on the multi-channel spectral image to obtain the multispectral surface image; The original laser data of the lightweight component is obtained, and the surface curvature gradient value is dynamically adjusted to obtain the three-dimensional laser data. The original coded pulse and original echo signals of the lightweight component are acquired and phase-aligned fusion is performed to obtain the ultrasonic detection data. The multispectral surface image and the laser three-dimensional data are projected and registered to obtain initial component assembly data. Based on the ultrasonic testing data, the initial component assembly data is spatially aligned to obtain the component joint dataset.

[0006] Further, the dynamic pruning parsing of the joint component dataset to obtain geometric contour features, and the internal feature extraction of the joint component dataset to obtain internal anomaly response features, include: The joint dataset of the components is subjected to spectral multi-channel band decomposition to obtain a texture feature map; A lightweight convolutional network channel is constructed on the texture feature map according to a preset convolutional kernel activation dynamic pruning algorithm to obtain a defect edge detection network; The geometric contour features are obtained by performing layer-by-layer feature mapping on the texture feature map through the defect edge detection network. The joint dataset of the components is subjected to ultrasonic time-frequency separation processing to obtain the defect reflection wave signal; Based on the joint dataset of the components, the acoustic path difference phase calibration of the defect reflected wave signal is performed to obtain an acoustic impedance anomaly distribution map; The coordinate matching relationship between the geometric contour features and the acoustic impedance anomaly distribution map is calculated to obtain the surface internal feature correlation matrix. The internal correlation matrix of the surface is fused using feature channels to obtain internal abnormal response features.

[0007] Further, the stress gradient distribution data and volume percentage parameters of the geometric contour features and the internal anomaly features are processed, and defect prediction is performed according to a preset component material type mapping rule to obtain a multi-dimensional prediction matrix, including: Multi-scale Gaussian curvature calculation is performed on the geometric contour features, and the surface deformation region is extracted to obtain the local curvature distribution features; The local curvature distribution features and the internal anomaly features are superimposed with anti-mutation strength to obtain a curvature gradient map of the defect region. Based on the curvature gradient map of the defect region and the laser three-dimensional data, spatial topological relationships are reconstructed to obtain a three-dimensional stress field. The stress gradient direction is traced in the three-dimensional stress field to obtain the stress gradient distribution data; Based on the laser 3D data, the curvature gradient map of the defect region is segmented into an octree space, and the density distribution is calculated to obtain the volume percentage parameter. Based on the component material type mapping rule, multimodal failure prediction is constructed using the stress gradient distribution data and the volume ratio parameter to obtain the multidimensional prediction matrix.

[0008] Further, the step of performing multi-scale Gaussian curvature calculation on the geometric contour features and extracting the surface deformation region to obtain local curvature distribution features includes: Multi-scale curvature calculation is performed on the geometric contour features to obtain an initial curvature distribution map; The initial curvature distribution map is subjected to curvature gradient consistency verification to obtain a verified curvature distribution map; The verified curvature distribution map is superimposed with feature channels to obtain an enhanced curvature distribution map; Based on a preset curvature gradient magnitude threshold, abrupt curvature gradient extraction is performed on the enhanced curvature distribution map to obtain curvature gradient abrupt regions; Connectivity analysis is performed on the curvature gradient abrupt change region, and region growth and boundary optimization are carried out to obtain a surface deformation region mask; Based on the surface deformation region mask, the deformation space of the enhanced curvature distribution map is truncated to obtain the curvature distribution extreme points; Based on the extreme points of the curvature distribution, feature extraction is performed on the enhanced curvature distribution map to obtain the local curvature distribution features.

[0009] Further, the step of performing octree spatial segmentation on the curvature gradient map of the defect region based on the laser three-dimensional data and calculating the density distribution to obtain the volume proportion parameter includes: The curvature gradient threshold of the curvature gradient map of the defect region is determined to obtain the region level depth; Based on the region hierarchy depth, the curvature gradient map of the defect region is recursively segmented into an octree space to obtain a segmentation structure set; Based on the laser 3D data, subspace density statistics are performed on the segmented structure set to obtain statistical information; Density anomaly neighborhood interpolation compensation is performed on the statistical information to obtain a compensation subspace set; The equivalent volume weight coefficient is obtained by calculating the subspace coefficients based on the compensation subspace set. The equivalent volume weight coefficients are weighted and superimposed in a subspace to obtain the volume percentage parameter.

[0010] Further, the process of obtaining the original operating condition dataset of the lightweight component, performing service condition coupling analysis on the multi-dimensional prediction matrix, and identifying multi-dimensional quality parameters to obtain comprehensive component quality information includes: The temperature distribution data, dynamic load spectrum data, and vibration spectrum data of the lightweight component are acquired and integrated to obtain the original working condition dataset. The original working condition dataset is timestamped based on the multi-dimensional prediction matrix to obtain the working condition feature set. The working condition feature set and the multi-dimensional prediction matrix are superimposed as tensors, and cross-modal attention mechanism is applied to obtain the enhanced prediction matrix. Stress coupling calculations are performed on the enhanced prediction matrix to obtain stress interaction feature vectors; Based on a preset failure mode database, the stress interaction feature vector is matched for defect similarity to obtain failure mode codes and hazard labels. The comprehensive quality information is obtained by performing quality analysis on the failure mode coding and the hazard characterization labels based on the enhanced prediction matrix.

[0011] The present invention also provides a quantitative assessment system for surface defects of lightweight components, applied to the quantitative assessment method for surface defects of lightweight components described in any one of the above-mentioned methods, comprising: The acquisition module is used to acquire multispectral surface images, laser three-dimensional data and ultrasonic test data of lightweight components, and perform fusion registration based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components. The analysis module is used to perform dynamic pruning parsing on the joint dataset of components to obtain geometric contour features, and to extract internal features from the joint dataset of components to obtain internal abnormal response features. The association module is used to perform stress gradient distribution data and volume ratio parameters on the geometric contour features and the internal anomaly features, and to perform defect prediction according to the preset component material type mapping rules to obtain a multi-dimensional prediction matrix. The processing module is used to acquire the original working condition dataset of the lightweight component, perform service condition coupling analysis on the multi-dimensional prediction matrix, and identify multi-dimensional quality parameters to obtain the comprehensive quality information of the component.

[0012] The present invention also provides a device for quantitatively assessing surface defects in lightweight components, comprising: Memory, used to store programs; A processor is used to execute the program to implement the various steps of the quantitative assessment method for surface defects of lightweight components as described in any of the above-mentioned methods.

[0013] The method, system, and apparatus for quantitative assessment of surface defects in lightweight components provided by this invention have the following beneficial effects: By synchronously registering multispectral, laser 3D, and ultrasonic detection data in time and space, a precise mapping and correlation of surface texture geometry and acoustic impedance response is established, solving the problem of defect spatial positioning deviation caused by isolated processing of multi-source data in traditional methods, and realizing cross-modal joint detection of micron-level defects. Based on a dynamic pruning mechanism of convolution kernel activation, multispectral reflectance characteristics and ultrasonic time-frequency features are fused to construct an edge detection network for composite defects, overcoming the limitations of traditional single-sensor feature analysis and improving the accuracy of morphological and mechanical coupling feature extraction for composite defects such as cracks and pores. Through multidimensional parameter coupling analysis of service condition data (temperature gradient, dynamic load) and defect stress gradient distribution, a defect evolution model adaptive to material properties is constructed, reducing the prediction error of the correlation between volume proportion parameters and stress field distribution, and significantly improving the reliability of defect propagation behavior prediction under complex conditions. Based on a dynamic matching mechanism of yield strength threshold using a component material type mapping rule base, the defect sensitivity of heterogeneous materials such as carbon fiber composites and titanium alloys is quantitatively adapted. By constructing a structured output of maintenance priority codes and process compensation parameters, a closed-loop control link of defect assessment, process adjustment, and quality traceability is built, which reduces the defect rework rate of lightweight components and significantly improves the quality control efficiency of high-end equipment manufacturing processes. Attached Figure Description

[0014] Figure 1 This is a flowchart of a method for quantitatively evaluating surface defects in lightweight components provided by the present invention; Figure 2 This is a structural diagram of a lightweight component surface defect quantitative assessment system provided by the present invention; Figure 3 This is a structural diagram of a device for quantitatively evaluating surface defects in lightweight components, provided by the present invention.

[0015] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments.

[0018] Reference Figure 1As shown, the present invention provides a method for quantitatively evaluating surface defects in lightweight components, comprising: Step S1: Acquire multispectral surface images, laser 3D data and ultrasonic test data of lightweight components, and perform fusion and registration based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components; Step S2: Perform dynamic pruning analysis on the joint dataset of components to obtain geometric contour features, and extract internal features from the joint dataset of components to obtain internal anomaly response features; Step S3: Calculate stress gradient distribution data and volume ratio parameters for geometric contour features and internal anomaly features, and predict defects according to preset component material type mapping rules to obtain a multi-dimensional prediction matrix. Step S4: Obtain the original working condition dataset of lightweight components, perform service condition coupling analysis on the multi-dimensional prediction matrix, and identify multi-dimensional quality parameters to obtain the comprehensive quality information of the components.

[0019] Based on the steps described above, the detailed process is as follows: Step S1: Simultaneously acquire surface and internal data of the lightweight component using an embedded multispectral imaging module, a linear laser scanner, and an ultrasonic phased array probe. The multispectral imaging module splits light in the 400-1000nm wavelength range at 50nm intervals to generate a 12-channel spectral image, and achieves independent capture of reflected light in each band through a beam-splitting prism array. The laser scanner dynamically adjusts the scanning point spacing according to the Gaussian curvature gradient of the component surface, using a high-density scan of 0.02mm in areas of abrupt curvature change (such as crack edges) and a sparse scan of 0.1mm in flat areas to construct a three-dimensional point cloud model with millimeter-level accuracy. The ultrasonic phased array probe transmits coded pulses in full-matrix capture mode, and after receiving the echo signal, eliminates the phase distortion caused by the curved surface component through an acoustic path difference compensation algorithm to generate a 256×256 pixel acoustic impedance matrix. Multi-sensor data is synchronized at the microsecond level using a GPS-disciplined clock source. An improved ICP algorithm is used to align the spatial coordinates of the laser point cloud and the acoustic impedance matrix. Simultaneously, projection registration of the multispectral image and the 3D point cloud is achieved based on SURF feature points. The final output is a 3D joint dataset containing optical texture mapping, 3D topology, and internal acoustic features, providing a unified data base for subsequent feature analysis.

[0020] Step S2: Perform multi-channel band decomposition on the multispectral surface images in the 3D joint dataset to extract reflectance distribution feature maps for each band. Construct a lightweight convolutional network based on the dynamic pruning algorithm of convolutional kernel activation: Initialize a set of edge detection convolutional kernels covering the 0°, 45°, 90°, and 135° directions. Calculate the activation distribution histogram of each kernel on the multispectral feature map through sliding convolution. Remove redundant convolutional kernels with activation contribution coefficients lower than the dynamic threshold, and retain high-response kernels to construct a defect edge detection network. Apply this network to the multispectral feature map to output sub-pixel-level geometric contour feature vectors. Simultaneously perform short-time Fourier transform on the ultrasonic detection data to separate the direct wave and defect reflection wave signals. Perform path difference compensation based on the spatial topological relationship of the 3D joint dataset to generate an acoustic impedance anomaly distribution map aligned with the geometric contour space. Input the geometric contour feature vector and the acoustic impedance anomaly distribution map into the cross-modal attention module to calculate the surface-internal feature association weight matrix. Generate internal anomaly response features containing deformation-mechanical coupling characteristics through channel splicing and dimensionality reduction.

[0021] Step S3: Calculate Gaussian curvature at multiple scales based on geometric contour feature vectors, extracting local curvature gradient abrupt change regions as stress-sensitive areas. Construct a tetrahedral mesh model of the defect region using laser 3D point clouds. Set stress boundary conditions on material discontinuities extracted from the acoustic impedance anomaly distribution map, and reconstruct the 3D stress field distribution using a finite element interpolation algorithm. Use the Sobel operator to track the gradient direction of the 3D stress field, generating a principal stress direction vector set and a stress concentration coefficient matrix. Simultaneously, perform octree spatial segmentation on the defect region: dynamically control the segmentation depth based on the curvature gradient threshold, recursively subdivide high-gradient regions with a precision of 0.1 mm³, statistically analyze the point cloud density of each subspace, and compensate for missing data through neighborhood interpolation. Finally, calculate the equivalent volume percentage parameter using weighted averages. Call the material type mapping rule library, normalize the stress coefficients based on the yield strength threshold of the component material, and input the normalized stress, volume parameters, and principal stress directions into the LSTM time-series prediction model. Output a multi-dimensional prediction matrix containing defect propagation probability, critical failure load, and pattern encoding.

[0022] Step S4: Real-time acquisition of component service condition data, including temperature gradient distribution, dynamic load spectrum, and vibration spectrum, is performed using embedded temperature sensors, strain gauges, and accelerometers. Wavelet denoising is applied to the raw service condition data to extract time-frequency features strongly correlated with defect evolution. These features are then spatiotemporally aligned with a multi-dimensional prediction matrix using a tensor superposition algorithm to generate a service condition-defect coupled feature set. The dynamic influence weights of temperature, load, and vibration on defect propagation are calculated using the entropy weight method, and the amplitude distribution of the stress concentration coefficient matrix is ​​corrected. The failure mode database is accessed, and the similarity between the corrected feature vector and historical cases is measured using Mahalanobis distance. Failure mode codes and risk level labels are output. The risk labels, corrected stress parameters, and volume percentage are input into the decision tree model. A low-dimensional quality assessment vector is generated through dimensionality reduction projection. After comparison with a preset material fatigue threshold, a comprehensive quality information matrix containing remaining life predictions, laser welding energy parameters, and inspection cycle suggestions is output, achieving closed-loop control of the entire process from defect assessment to process optimization and maintenance decision-making.

[0023] This invention provides a quantitative assessment method for surface defects in lightweight components. Through spatiotemporal synchronous registration of multispectral, laser 3D, and ultrasonic detection data, it establishes a precise mapping and correlation between surface texture geometry and acoustic impedance response, solving the spatial positioning deviation problem caused by isolated processing of multi-source data in traditional methods. This enables cross-modal joint detection of micron-level defects. A dynamic pruning mechanism based on convolution kernel activation integrates multispectral reflectance characteristics and ultrasonic time-frequency features to construct an edge detection network for composite defects, overcoming the limitations of traditional single-sensor feature analysis and improving the accuracy of morphological and mechanical coupling feature extraction for composite defects such as cracks and pores. Through multidimensional parameter coupling analysis of service condition data (temperature gradient, dynamic load) and defect stress gradient distribution, a material property-adaptive defect evolution model is constructed, reducing the prediction error of the correlation between volume proportion parameters and stress field distribution, and significantly improving the reliability of defect propagation behavior prediction under complex conditions. A dynamic matching mechanism based on the yield strength threshold of a component material type mapping rule library enables quantitative adaptation of defect sensitivity metrics for heterogeneous materials such as carbon fiber composites and titanium alloys. By constructing a structured output of maintenance priority codes and process compensation parameters, a closed-loop control link of defect assessment, process adjustment, and quality traceability is built, which reduces the defect rework rate of lightweight components and significantly improves the quality control efficiency of high-end equipment manufacturing processes.

[0024] In one embodiment, multispectral surface images, laser 3D data, and ultrasonic testing data of a lightweight component are acquired, and fused and registered based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the component, including: In the detection of surface defects in lightweight components, a multi-band combined illumination system optically covers the component surface using a broadband light source. The light source covers a wavelength range of 400-1000 nm, divided into 12 independent spectral channels according to a preset wavelength spacing rule. The beam-splitting prism array consists of a group of high-precision optical prisms, each corresponding to a specific wavelength range. The reflected beam is separated by wavelength to individual sensors using the principle of physical beam splitting. The design rules of the beam-splitting prism array include the matching relationship between the prism tilt angle and the wavelength diffraction angle, ensuring that beams from different wavelengths do not overlap. After processing by the beam-splitting prism array, a multi-channel spectral image containing 12 independent spectral channels is generated. Each channel image records the surface reflection characteristics differences at the corresponding wavelength.

[0025] Multi-channel spectral images require the elimination of image shifts caused by mechanical vibration or assembly errors. Subpixel-level feature point matching employs the Accelerated Robust Feature (SURF) algorithm, which detects key feature points in each channel image, calculates feature descriptors, and establishes cross-channel matching relationships. Matching rules are based on the consistency of feature point geometric position and grayscale distribution, eliminating outlier matching points. A homography matrix transformation model is used to align the image coordinates of different channels, achieving subpixel-level registration by minimizing feature point projection errors. The registered multi-channel spectral images are then pixel-level superimposed to generate a multispectral surface image with a geometric resolution error of less than 0.1 pixels, fully preserving surface texture and spectral reflectance characteristics.

[0026] Laser 3D data is acquired using a linear laser scanner, and the scanning parameters are dynamically adjusted based on the surface curvature gradient of the component. The surface curvature gradient value is obtained by calculating the second derivative of the Gaussian curvature in real time and is used to determine the spacing between scanning points. The preset adjustment rule is: when the absolute value of the curvature gradient is less than 0.05mm... -1 At that time, a sparse scan of 0.1 mm was used; the curvature gradient was between 0.05 and 0.2 mm. -1 The time was adjusted to a medium density scan of 0.05mm; the curvature gradient exceeded 0.2mm. -1 High-density scanning of 0.02mm is activated. The dynamically adjusted scanning process generates high-fidelity laser 3D data with point cloud density error controlled within ±3%, accurately characterizing the micro-geometry of the component surface.

[0027] Ultrasonic testing data was acquired using a 128-element ultrasonic phased array probe in full matrix capture (FMC) mode. The raw coded pulse signal was a modulated pulse with 80% bandwidth and a center frequency of 10MHz. After receiving the raw echo signal, a path difference phase compensation algorithm was used to eliminate the signal propagation path differences caused by the curved components. The compensation rule was based on the three-dimensional topological relationship of the component surface, calculating the path difference between the ultrasonic transmitting and receiving elements to the defect point, and performing phase correction on the echo signal. The phase-aligned data was then used to generate a 256×256 pixel acoustic impedance distribution matrix using an inverse synthetic aperture algorithm, accurately characterizing the acoustic properties of internal defects in the material.

[0028] Registration of multispectral surface images and laser 3D data was achieved using an improved Iterative Closest Point (ICP) algorithm. This algorithm introduces a curvature constraint term into the standard ICP, improving registration accuracy by matching the geometric curvature features of the point clouds. The registration rule minimizes the weighted sum of Euclidean distance and curvature differences between point clouds, ensuring accurate mapping between optical texture and 3D geometry. The registered data generates an initial component assembly dataset containing the correlation between surface texture and 3D point cloud mappings, with a spatial alignment error of less than 15 μm.

[0029] The initial component assembly data needs to be spatially aligned with the ultrasonic testing data. The alignment process is based on the defect coordinate information in the ultrasonic acoustic impedance distribution matrix, mapping the ultrasonic data to a laser three-dimensional coordinate system through rigid body transformation. The alignment rule minimizes the spatial positional deviation of defect feature points in the multi-source data, achieved using a combination of feature point matching and least-squares optimization. The final generated component joint dataset is a three-dimensional voxel structure, with each voxel containing multispectral reflectance, three-dimensional coordinates, and acoustic impedance values. This achieves multimodal data fusion of surface morphology, internal defects, and material properties, providing a unified data foundation for subsequent quantitative defect assessment.

[0030] This embodiment achieves high-precision acquisition of multispectral surface images through multi-band combined illumination and wavelength separation mechanism using a beam-splitting prism array. This eliminates ambient light interference while preserving the differences in reflectivity across different bands, improving the detection sensitivity of surface textures and minute defects. A sub-pixel-level feature point matching algorithm aligns multi-channel spectral images via homography matrix transformation, resolving image shift issues caused by mechanical vibration and ensuring geometric consistency of multispectral surface images. A dynamic curvature gradient adjustment mechanism for laser 3D data adaptively adjusts the scanning density based on the degree of surface deformation, achieving high-resolution point cloud acquisition in complex curved surface regions, avoiding redundant data storage and improving 3D reconstruction efficiency. Phase alignment and fusion technology for ultrasonic testing data eliminates the influence of curved components on the sound wave propagation path through a path difference compensation algorithm, enhancing the acoustic feature extraction capability of internal defects and achieving high-contrast imaging of internal material anomalies. The final generated component joint dataset integrates multi-source information, supporting comprehensive evaluation of defect morphology, mechanical properties, and evolutionary behavior, significantly improving the comprehensiveness and reliability of lightweight component quality inspection.

[0031] In one embodiment, dynamic pruning parsing is performed on the joint dataset of components to obtain geometric contour features, and internal feature extraction is performed on the joint dataset of components to obtain internal anomaly response features, including: In the surface defect detection of lightweight components, the joint dataset includes multispectral surface images, laser 3D data, and ultrasonic testing data. The spectral multi-channel band decomposition module, based on the physical spectral dispersion results of a beam-splitting prism array, decomposes the multispectral surface image into texture feature maps of 12 independent spectral channels at wavelength intervals. Predefined rules for the beam-splitting prism array require each prism to correspond to a specific wavelength's diffraction angle, ensuring independent imaging of the spectral reflectance characteristics of different bands. The decomposed texture feature maps contain the differences in surface reflectance distribution at each wavelength, used to characterize the material's surface texture and micro-deformation features.

[0032] Texture feature maps are input into a lightweight convolutional network architecture. The initial set of convolutional kernels includes multi-directional edge detection kernels (0°, 45°, 90°, and 135° directions). A dynamic kernel activation pruning algorithm calculates the activation contribution coefficient based on the activation response distribution of each convolutional kernel on the texture feature map. The rule for determining the activation contribution coefficient is: when a convolutional kernel's activation value is below a dynamic threshold in more than 80% of the image region, it is considered a redundant kernel and is removed. The pruned network retains high-response convolutional kernels, forming a defect edge detection network whose channel count is reduced to 40% of the initial network, significantly reducing computational complexity.

[0033] In the process of constructing a defect edge detection network through dynamic pruning, a set of convolutional kernels covering multi-directional edge detection is first initialized based on the channel characteristics of the multispectral texture feature map. These convolutional kernels are designed as gradient operators at different angles, including horizontal, vertical, and diagonal directions, to capture the edge features of potential defects in the surface texture. Each convolutional kernel performs a sliding window convolution operation on the input texture feature map, generating a corresponding activation response map that reflects the distribution intensity of edges in different directions in the image.

[0034] In the dynamic activation evaluation phase, the activation contribution coefficient of each convolutional kernel is calculated by statistically analyzing its activation response distribution across the entire texture feature map. The contribution coefficient is determined by the ratio of the cumulative activation intensity of the convolutional kernel within the effective detection region to the maximum possible activation intensity. The evaluation rule is set as follows: if the activation value of a convolutional kernel remains consistently below a dynamically adjusted threshold in most areas of the image, its contribution to edge features is deemed insufficient. The dynamic threshold is adaptively adjusted based on the median of the activation distribution of all convolutional kernels to ensure that the selection criteria match the data characteristics.

[0035] The pruning operation is performed based on the contribution coefficient ranking, removing redundant convolutional kernels with contribution coefficients below a dynamic threshold, and retaining only high-response kernels sensitive to defect edges. The retained convolutional kernels are then weighted according to activation priority, forming a lightweight network channel structure. This network gradually fuses multi-scale edge information by stacking the retained convolutional kernels layer by layer, ultimately outputting the geometric contour features of the defect edges. The pruned network reduces the number of parameters while enhancing the directional detection capability of defect edges such as cracks and dents, ensuring a balance between computational efficiency and feature parsing accuracy.

[0036] In this process, the initialization direction coverage of the convolution kernel, the dynamic evaluation mechanism of the activation response, and the adaptive threshold screening strategy together constitute the core optimization logic of the defect edge detection network, achieving accurate feature extraction from multispectral textures to geometric deformations.

[0037] The defect edge detection network performs layer-by-layer feature mapping on the texture feature map, extracting edge gradient information of the multispectral texture through the retained convolutional kernels. The network output is processed by a nonlinear activation function, fusing the edge responses of each spectral channel to generate geometric contour features with sub-pixel accuracy. The geometric contour features represent the geometric deformation regions on the component surface in the form of a 3D point cloud, including the boundary coordinates and gradient magnitude information of defects such as cracks and pits, with a spatial resolution error of less than 0.05 mm.

[0038] The ultrasonic testing data in the component joint dataset includes raw echo signals in full-matrix capture mode. The ultrasonic time-frequency separation module uses a short-time Fourier transform algorithm to convert the time-domain echo signals into a joint time-frequency distribution map. The separation rule is based on frequency-energy distribution characteristics, extracting the defect reflection wave components in the high-frequency band (8-12MHz) and filtering out low-frequency direct waves and environmental noise. The separated defect reflection wave signals retain the acoustic scattering characteristics caused by internal material defects, providing a high signal-to-noise ratio input for subsequent acoustic impedance calculations.

[0039] Phase calibration of the defect-reflected wave signal is based on the laser 3D data in the joint component dataset. The influence of curved components on the ultrasonic propagation path is eliminated through an acoustic path difference compensation algorithm. The compensation rule requires calculating the theoretical time delay of the acoustic wave propagation path based on the 3D coordinates of the defect point and the position of the probe array elements, and then aligning the phase of the echo signal. The calibrated signal is then used to reconstruct the acoustic impedance distribution using an inverse synthetic aperture imaging algorithm, generating a 256×256 pixel acoustic impedance anomaly distribution map. Highlighted areas in the map represent internal material defects or density anomalies.

[0040] Spatial alignment between geometric contour features and acoustic impedance anomaly distribution maps relies on a unified coordinate system within the joint dataset of components. The coordinate matching module uses a feature point matching algorithm to spatially correlate defect boundary points in the geometric contour features with the center points of the acoustic impedance anomaly regions. The matching rule requires feature point pairs with an Euclidean distance of less than 0.1 mm and consistent curvature gradient directions to be considered valid correlations. The generated surface internal feature correlation matrix is ​​a two-dimensional matrix, where matrix element values ​​represent the correlation strength between surface defects and internal anomalies, ranging from 0 to 1, with 1 representing a complete spatial correspondence.

[0041] The surface internal feature correlation matrix is ​​input to the feature channel fusion module, which includes a channel attention mechanism and a weighted superposition operation. The channel attention mechanism assigns fusion weights to different channels based on the numerical distribution of the correlation matrix, with channels having a correlation value higher than 0.7 having their weights doubled. The internal anomaly response features generated after weighted superposition include the coupling response strength of surface deformation and internal defects, defect depth estimation, and material failure risk level, providing multi-dimensional input for subsequent defect prediction.

[0042] This embodiment utilizes spectral multi-channel band decomposition technology to decompose multispectral surface images into independent texture feature maps based on wavelength characteristics. This preserves surface reflection differences across different bands, enhancing the ability to identify surface textures and micro-defects and addressing the issue of single-spectral imaging being susceptible to environmental interference. The convolutional kernel activation dynamic pruning algorithm adaptively selects high-contribution convolutional kernels based on the edge response distribution of the texture feature maps to construct a lightweight network. This reduces computational complexity while improving the directional sensitivity of defect edge detection and avoiding feature confusion caused by redundant kernels. Ultrasonic time-frequency separation processing extracts defect reflection wave signals through frequency domain energy distribution, effectively suppressing environmental noise and direct wave interference, and enhancing the accuracy of acoustic feature extraction for internal defects. Sound path difference phase calibration, combined with the component's three-dimensional morphology data, corrects ultrasonic propagation path errors, eliminates acoustic impedance imaging distortion caused by surface effects, and achieves precise spatial localization of internal defects.

[0043] In one embodiment, stress gradient distribution data and volume percentage parameters are obtained from geometric contour features and internal anomaly features, and defect prediction is performed according to a preset component material type mapping rule to obtain a multi-dimensional prediction matrix, including: Based on 3D point cloud data with geometric contour features, the geometric deformation characteristics of component surfaces are analyzed using a multi-scale Gaussian curvature calculation module. The multi-scale Gaussian curvature calculation employs filter kernels of different sizes to cover curvature analysis needs from micro to macro levels. Pre-defined scale rules require that the filter kernel size be inversely proportional to the component surface resolution, ensuring the ability to capture details in high-curvature gradient regions. The calculated local curvature distribution features are stored as point cloud attributes, with each point containing a curvature value and gradient direction. High-curvature regions are marked as surface deformation regions, representing the geometric boundaries of defects such as cracks and pits.

[0044] The local curvature distribution features are derived from 3D point cloud data output by the multi-scale Gaussian curvature calculation module. A 2D curvature gradient map is generated through rasterization, with each raster cell recording the curvature gradient amplitude and direction. The acoustic impedance anomaly distribution map is generated from ultrasonic testing data after phase calibration; the matrix values ​​reflect the intensity of anomalies in the material's internal acoustic properties. Both types of data are unified to the same spatial coordinate system through a coordinate mapping module. Raster resolution errors are corrected to within 0.02 mm using bilinear interpolation to ensure spatial alignment accuracy.

[0045] The anti-abrupt change strength superposition module fuses curvature gradient and acoustic impedance anomaly features based on a dynamic weight allocation rule. The curvature gradient weight is dynamically set according to the normalized gradient amplitude; the higher the gradient amplitude (e.g., in crack edge regions), the greater the weight. The acoustic impedance anomaly weight is adjusted according to the anomaly intensity; the weight doubles when the anomaly value exceeds a certain proportion of the material's average acoustic impedance. The total weight coefficient is the product of the two weights, significantly enhancing the response only in regions where both curvature and acoustic properties are anomaly-prone. The weighted superposition formula is the sum of the weighted products of the curvature gradient value and the acoustic impedance anomaly value, generating the initial fusion matrix.

[0046] The initial fusion matrix undergoes a neighborhood consistency check. If the curvature or acoustic impedance anomaly of a certain raster cell exceeds three times the neighborhood mean, it is determined to be noise interference, triggering Gaussian smoothing filtering to suppress abrupt changes. Morphological closing operations are performed on the fused raster data to perform dilation-erosion operations, filling small holes and smoothing defect boundaries. The final generated defect region curvature gradient map is a two-dimensional matrix with values ​​ranging from 0 to 255. High-value regions represent the strong coupling effect between surface deformation and internal defects.

[0047] The formula for superimposing anti-mutation strength is as follows: ; F(x, y): The grid cell value at coordinates (x, y) in the fused defect region curvature gradient map. C(x, y): The local curvature gradient value at coordinates (x, y), output by the multi-scale Gaussian curvature calculation module.

[0048] (x, y): Curvature gradient weighting coefficient. A(x, y): Acoustic impedance anomaly at coordinates (x, y), generated from ultrasonic testing data. (x, y): Acoustic impedance anomaly weighting coefficient. (x, y): Two-dimensional grid coordinates, corresponding to the spatial position on the surface of the component.

[0049] Local curvature distribution features and internal anomaly features (acoustic impedance anomaly distribution map) are input into the anti-mutation strength superposition module. The anti-mutation strength superposition rule requires weighted fusion of curvature gradient and acoustic impedance abrupt change regions. The weight allocation is dynamically adjusted based on the amplitude of the curvature gradient and the intensity of the acoustic impedance anomaly. The fused defect region curvature gradient map is stored in the form of a two-dimensional matrix. The matrix element values ​​represent the deformation-mechanical coupling strength of the defect region, avoiding misjudgments caused by single feature analysis.

[0050] The curvature gradient map of the defect region and the laser 3D data are input into the finite element analysis module to construct a 3D tetrahedral mesh model. Spatial topology reconstruction rules are set based on the high-stress-sensitive areas marked in the curvature gradient map, setting mesh node constraints, and the stress distribution of the mesh nodes is calculated using an interpolation algorithm. The reconstructed 3D stress field is stored in a voxel structure, with each voxel containing the stress amplitude, principal stress direction, and shear stress component, characterizing the stress concentration and distribution characteristics within the material.

[0051] The three-dimensional stress field is input to the gradient direction tracking module, which analyzes the stress variation trend at each node using gradient operators. The tracking rules require extending the path along the principal stress direction to identify gradient abrupt change points in stress concentration regions and record their coordinates and gradient magnitudes. The generated stress gradient distribution data includes the stress concentration factor, gradient direction vector set, and maximum shear stress distribution in the defect region, providing mechanical behavior input for failure prediction.

[0052] Laser 3D data and the curvature gradient map of the defect area are input into the octree segmentation module, which dynamically adjusts the segmentation depth based on the curvature gradient threshold. The segmentation rules require recursive subdivision of regions with curvature gradients exceeding a preset threshold, with a minimum subspace size of 0.1 mm³. After segmentation, the point cloud density distribution within each subspace is statistically analyzed, and point cloud loss in low-density regions is compensated through neighborhood interpolation. The density distribution calculation results are weighted and superimposed to generate a volume proportion parameter, quantifying the spatial proportion of the defect area within the overall component.

[0053] Stress gradient distribution data and volume percentage parameters are input into a multimodal failure prediction model. The model calls upon yield strength thresholds, fatigue limit parameters, and fracture toughness indices from the component material type mapping rule library. Prediction rules match corresponding failure criteria based on material type; for example, the maximum stress criterion is used for carbon fiber composites, and the Von Mises yield criterion is used for titanium alloys. The model outputs a multi-dimensional prediction matrix containing defect propagation probability, critical load threshold, failure mode coding, and remaining life prediction values, providing a quantitative basis for quality assessment.

[0054] This embodiment utilizes multi-scale curvature analysis, anti-mutation feature fusion, and stress field reconstruction techniques to achieve a quantitative assessment of the mechanical behavior of defects in lightweight components. Local curvature distribution characteristics are fused with acoustic impedance anomaly data to generate a gradient map of the defect region, driving accurate reconstruction of the three-dimensional stress field. Octree segmentation and density statistics quantify the defect volume ratio, and combined with material property mapping rules, a multi-modal failure prediction model is constructed. This model outputs a multi-dimensional prediction matrix containing parameters such as propagation probability and critical load, providing high-confidence data support for component life prediction and maintenance decisions.

[0055] In one embodiment, multi-scale Gaussian curvature calculation is performed on the geometric contour features, and the surface deformation region is extracted to obtain local curvature distribution features, including: Based on 3D point cloud data with geometric contour features, the geometric properties of component surfaces are analyzed using a multi-scale Gaussian curvature calculation module. The multi-scale curvature calculation employs filter kernels of different sizes to cover curvature analysis needs from micro to macro scales. The preset filter kernel size rule requires that it be inversely proportional to the component surface resolution, ensuring the ability to capture details in high curvature gradient regions (such as crack edges). The filter kernel captures localized minute deformations at the microscale (0.1 mm) and identifies overall warping trends at the macroscale (1.0 mm). The calculated initial curvature distribution map is stored in the form of a two-dimensional raster, with each raster cell recording a normalized curvature value (range 0-1), and high-value regions representing the severity of surface deformation.

[0056] The initial curvature distribution map is input to the gradient consistency verification module, which removes outliers through statistical analysis of neighborhood gradient directions. The verification rule requires that the curvature gradient direction of each grid cell be consistent with the majority of directions within its neighborhood (3×3 window); if the direction deviation exceeds 30 degrees, it is marked as an outlier. Outliers are corrected based on valid neighborhood data using a bilinear interpolation algorithm to ensure the spatial continuity of the curvature gradient. The verified curvature distribution map retains the geometric features of surface deformation, eliminates isolated outliers caused by noise or data acquisition errors, and generates a verified curvature distribution map.

[0057] The curvature distribution map is input to the feature channel overlay module, which then overlays the multi-scale curvature calculation results according to their weights. The preset weighting rule requires a 0.6 weight for microscale curvature and a 0.4 weight for macroscale curvature, enhancing the sensitivity for detecting microscopic defects. The overlaid enhanced curvature distribution map integrates deformation features from different scales through pixel-level fusion, forming high-contrast surface deformation characterization data. The enhanced curvature distribution map is stored as a grayscale image; higher grayscale values ​​indicate more significant deformation.

[0058] The enhanced curvature distribution map is input into the gradient threshold segmentation module. A preset dynamic threshold rule automatically adjusts based on the global distribution of curvature gradient magnitudes. The threshold is set to twice the standard deviation of the mean curvature gradient; grids exceeding the threshold are marked as abrupt curvature gradient regions. In the segmented binary image, bright areas correspond to the geometric boundaries of surface defects such as cracks and pits, while low-value areas represent normal surfaces. The spatial coordinates and gradient magnitudes of abrupt curvature gradient regions are stored as a point cloud dataset for subsequent connectivity analysis.

[0059] Point cloud data of abrupt curvature gradient regions are input into the connectivity analysis module, which identifies independent defect regions based on the eight-neighbor connectivity criterion. The region growth rule requires that seed points start from the point with the highest gradient magnitude and gradually expand to adjacent points, with the expansion condition being that the gradient value of adjacent points is no less than 80% of that of the seed point. The grown regions are then smoothed at their boundaries using morphological closing operations (dilation followed by erosion), filling in small holes caused by noise and generating a surface deformation region mask. The mask is a binary image that marks the precise contours of the defect regions.

[0060] The surface deformation region mask and the enhanced curvature distribution map are input to the spatial cropping module. The enhanced curvature distribution map is cropped using the mask, retaining only the curvature data within the defect area. The cropped data then extracts the curvature distribution extrema. Extrema are determined by the rule of maximum curvature value and consistent gradient direction within a local neighborhood (5×5 window). The set of extrema records the coordinates and curvature gradient direction of the core defect region, characterizing the mechanical behavior of the deformation.

[0061] The extreme points of curvature distribution are input to the feature extraction module, which uses a density clustering algorithm (such as DBSCAN) to divide the extreme points into independent defect clusters. The preset clustering rules require that the distance between adjacent points be less than 0.1 mm and the deviation of the curvature gradient direction be less than 15 degrees. After clustering, the area, perimeter, average curvature, and gradient direction distribution of the deformed region are calculated for each defect cluster, generating local curvature distribution features. These features are stored in a structured data format, containing the geometric, mechanical properties, and spatial distribution information of the defects, providing input for subsequent defect classification and prediction.

[0062] This embodiment utilizes a multi-scale Gaussian curvature calculation module, employing filter kernels of different sizes to simultaneously capture microscopic dents and macroscopic warping features. This overcomes the limitations of traditional methods that miss small or large-scale defects due to single-scale detection, achieving full-dimensional analysis of surface deformation. Curvature gradient consistency verification eliminates outliers through neighborhood direction statistical analysis and corrects isolated noise points using bilinear interpolation, eliminating curvature distribution distortion caused by acquisition errors and improving data spatial continuity. Feature channel overlay and fusion of multi-scale curvature data enhances crack edge sensitivity with high weighting at the microscale, generating a high-contrast enhanced curvature distribution map and improving the accuracy of distinguishing the boundary between defects and normal areas. Dynamic threshold segmentation adaptively sets the segmentation threshold based on the global distribution of curvature gradients, accurately locating the geometric boundaries of defects such as cracks and dents, avoiding over-segmentation or missed detection problems caused by fixed thresholds. Deformation space truncation and density clustering extract curvature extrema points in the core defect region, associating deformation gradient directions with mechanical properties to construct defect geometry-mechanical coupling features. This provides high-confidence quantitative input for life prediction and process compensation, achieving end-to-end closed-loop optimization from deformation detection to failure assessment.

[0063] In one embodiment, the curvature gradient map of the defect region is spatially segmented using an octree based on laser 3D data, and the density distribution is calculated to obtain the volume percentage parameters, including: Based on the gradient magnitude distribution of the curvature gradient map of the defect area, an octree segmentation level depth for different regions is determined through a dynamic threshold determination module. The curvature gradient threshold determination rule is set as follows: when the local curvature gradient value exceeds twice the global average gradient value, the region needs to be subdivided into depths, and the level depth is increased by one level; regions with gradient values ​​lower than the average gradient value maintain the basic level depth. The determination result generates a region level depth map, where each pixel records the corresponding octree segmentation depth value. High gradient regions (such as crack edges) are marked as depths of 3-5, and low gradient regions are marked as depths of 1-2.

[0064] Based on the regional hierarchical depth map, a recursive octree spatial segmentation is performed on the curvature gradient map of the defect region. The segmentation rule requires starting from the initial bounding box and subdividing the space level by level according to the hierarchical depth value: if the level depth corresponding to the curvature gradient value of the current subspace has not reached the termination condition, then the subspace is divided into eight secondary subspaces. The termination condition is set when the subspace size reaches the minimum resolution (0.1 mm³) or the curvature gradient value is lower than 50% of the dynamic threshold. After segmentation, a segmentation structure set containing multi-level subspaces is generated. Each subspace in the structure set records its spatial boundary, parent node index, and mean curvature gradient value.

[0065] Based on the spatial distribution characteristics of the curvature gradient map of the defect region, recursive octree segmentation is performed starting from the initial bounding box. The initial bounding box covers the extreme range of the laser 3D data of the defect region, defined as a cubic space containing all curvature gradient values. The segmentation rules require dynamic adjustment of the segmentation level according to the region's hierarchical depth map: if the mean curvature gradient in the subspace exceeds twice the global mean, a high target level depth (e.g., level 3) is assigned, requiring further subdivision; regions with a gradient mean close to the global mean are assigned a medium level depth (e.g., level 2); regions with a gradient mean lower than the global mean retain the basic level depth (e.g., level 1). The subspace that triggers segmentation is divided into eight subspaces along the x, y, and z axes. Each subspace inherits some attributes from its parent space, including spatial boundaries, parent node indices, and the mean curvature gradient. The termination condition is set when the subspace size reaches the minimum resolution (0.1 mm³) or the mean curvature gradient is lower than 50% of the dynamic threshold. After the segmentation is completed, a segmentation structure set containing multi-level subspaces is generated. Each subspace records its spatial coordinate range, parent node identifier and gradient statistics, forming a tree-like hierarchical structure.

[0066] The formula for calculating the coordinate range of each subspace is as follows: ; , , : The minimum coordinates of the parent and child spaces on the x, y, and z axes. L is the side length of the parent and child spaces on a certain axis (all axes are of equal length).

[0067] The dynamic threshold is generated based on the global statistical characteristics of the curvature gradient map of the defect region, and the segmentation sensitivity is dynamically adjusted by combining the gradient mean and standard deviation. The global gradient mean is calculated by the arithmetic mean of all pixels in the curvature gradient map, and the standard deviation reflects the dispersion of the gradient distribution. The dynamic threshold is set as a weighted sum of the mean and standard deviation to ensure that high-gradient regions are subdivided first, while low-gradient regions are retained as coarse segments. During the segmentation process, gradient mean filtering is applied to each subspace to suppress the interference of isolated noise points on the segmentation decision. The final generated segmentation structure set is stored in a tree data structure, supporting fast querying and spatial backtracking, and providing multi-resolution input for subsequent density statistics.

[0068] The segmented structure set is input into the density statistics module, which calculates the point cloud density within each subspace using laser 3D point cloud data. The density statistics rule is: the ratio of the number of laser point clouds in a subspace to the volume of the subspace is used to generate an initial density distribution map. A preset baseline density is dynamically set based on the overall point cloud density of the component. If the density of a subspace is less than 30% of the baseline density, it is marked as a low-density anomaly region. The statistical information includes the location, volume, number of point clouds, and density label of each subspace, and is stored as a structured dataset.

[0069] Neighborhood interpolation compensation is performed on subspaces marked as low-density anomalies. The compensation rule requires selecting density data from neighboring subspaces based on the K-nearest neighbor algorithm (K=6), and reconstructing the point cloud distribution of the missing region through cubic spline interpolation. After interpolation, the density values ​​of the subspaces are updated to the neighborhood weighted average, with weights inversely proportional to spatial distance. The compensated subspace set retains the original subspace structure, updating only the number of point clouds and density values ​​in low-density regions, thus eliminating density distortion caused by blind spots in data acquisition.

[0070] The compensation subspace set is input to the weight calculation module, which calculates the equivalent volume weight coefficient based on the ratio of the subspace density to the baseline density. The calculation rules are as follows: if the subspace density is higher than the baseline density, the weight coefficient is 1.0; if the density is between 30% and 100% of the baseline density, the weight coefficient is scaled linearly; the weight of a subspace with a density below 30% is set to zero. The weight coefficient is multiplied by the subspace volume to generate an equivalent volume weight distribution map, quantifying the true contribution of each subspace to the volume of the defect region.

[0071] The equivalent volume weight distribution map is input into the overlay module, which sums the equivalent volumes of all subspaces to generate a total volume percentage parameter for the defect area. The overlay rules require excluding subspaces with zero weights and only summing the product of the volume and weight of the effective subspaces. The final volume percentage parameter is expressed as a percentage, reflecting the spatial proportion of the defect area within the overall component, providing a quantitative basis for defect severity assessment.

[0072] This embodiment dynamically adjusts the octree segmentation level depth through a curvature gradient threshold determination mechanism, achieving refined segmentation of high-gradient defect regions and efficient processing of low-gradient regions. This solves the problems of wasted computational resources or loss of detail caused by traditional fixed-depth segmentation, significantly improving spatial segmentation efficiency and accuracy. Recursive octree spatial segmentation generates multi-resolution subspace structures based on gradient characteristics, accurately adapting to the geometric morphology of complex curved surface defects and avoiding insufficient analysis of micro-cracks or large-scale warping by single-scale segmentation. Laser 3D data-driven subspace density statistics, combined with neighborhood interpolation compensation technology, reconstruct the density distribution of point cloud missing regions caused by occlusion or reflection, eliminating the interference of blind spots in volume calculation and enhancing data integrity. The equivalent volume weight coefficient dynamically quantifies the subspace contribution through density ratios, solving the volume assessment bias problem caused by neglecting density unevenness in traditional methods and improving the quantitative accuracy of defect region spatial proportion. The weighted superposition algorithm integrates the compensated subspace data to generate high-confidence volume proportion parameters, providing a reliable basis for defect severity classification and process optimization, achieving closed-loop optimization from data acquisition to quantitative decision-making.

[0073] In one embodiment, the original operating condition dataset of the lightweight component is obtained, and a service condition coupling analysis is performed on the multi-dimensional prediction matrix, followed by multi-dimensional quality parameter identification, to obtain the comprehensive quality information of the component, including: In the service condition monitoring of lightweight components, operating condition data are simultaneously acquired using an embedded temperature sensor array, a dynamic strain gauge, and a triaxial vibration accelerometer. The temperature sensor array acquires surface temperature distribution data of the component at a sampling rate of 1 Hz, generating a temperature gradient thermogram. The dynamic strain gauge records load spectrum data at a sampling rate of 10 kHz, including the temporal amplitude changes of tensile, compressive, and shear loads. The triaxial vibration accelerometer captures vibration spectrum data of the component in the x, y, and z directions at a sampling rate of 20 kHz, generating a frequency domain energy distribution map through fast Fourier transform. The three types of data are synchronized at the microsecond level using a GPS-disciplined clock source and spatially aligned according to the component's three-dimensional coordinates to generate a raw operating condition dataset containing the spatiotemporal distribution characteristics of temperature, load, and vibration.

[0074] The multi-dimensional prediction matrix (including parameters such as defect stress gradient and volume percentage) and the original operating condition dataset are input into the timestamp alignment module. The alignment rules require data with different sampling rates to be unified to the highest sampling rate (20kHz), and low-frequency data (such as temperature) are completed through linear interpolation to complete the time series. The aligned operating condition feature set is a four-dimensional tensor structure, with dimensions including timestamp, spatial coordinates, data type (temperature / load / vibration), and eigenvalues, supporting cross-modal correlation analysis.

[0075] The operating condition feature set and multi-dimensional prediction matrix are fused through a tensor overlay module, expanding the operating condition data (temperature, load, vibration) and defect parameters (stress gradient, volume percentage) into a five-dimensional tensor. The cross-modal attention mechanism calculates the dynamic weights of each modal data based on the entropy weighting method: the temperature gradient has a weight of 0.4 for the thermal stress coupling effect, the load amplitude has a weight of 0.5 for the mechanical stress effect, and the vibration spectrum has a weight of 0.1 for the fatigue damage contribution. The weighted data are then overlaid through channel weighting to generate an enhanced prediction matrix, where the matrix element values ​​reflect the comprehensive risk coefficient of defect evolution and operating condition coupling.

[0076] The enhanced prediction matrix is ​​input into the stress coupling calculation module, which analyzes the interaction effects between thermal expansion stress caused by temperature gradient, mechanical stress caused by dynamic load, and fatigue stress accumulated from vibration energy based on the material constitutive relation. The coupling rule requires that when the thermal and mechanical stresses are in the same direction, the equivalent stress amplitude is calculated by superposition; when they are in opposite directions, the net stress is calculated by vector synthesis. The calculation results generate a stress interaction characteristic vector, including the peak equivalent stress, the number of stress cycles, and the cumulative fatigue damage value.

[0077] The stress interaction feature vector is input to the failure mode matching module, which then calls a pre-set failure mode database (containing typical failure cases such as crack propagation, plastic deformation, and fatigue fracture). The matching rule uses Mahalanobis distance to measure the similarity between the feature vector and the database samples; the sample with the smallest distance is determined as the most likely failure mode. The matching result outputs a failure mode code (e.g., F001 represents crack propagation) and a hazard label (e.g., high-risk, medium-risk, low-risk). The label is automatically classified according to a distance threshold: less than 1.0 is high-risk, 1.0-2.0 is medium-risk, and greater than 2.0 is low-risk.

[0078] Failure mode codes, hazard labels, and enhanced prediction matrices are input into the quality analysis module. The module calculates the predicted remaining life based on the defect propagation rate model and matches laser welding energy parameters and inspection cycle recommendations with the maintenance process database. Quality analysis rules require high-risk defects to trigger immediate maintenance alerts, medium-risk defects to have their inspection cycles shortened to 50% of the original plan, and low-risk defects to maintain routine monitoring. The final generated comprehensive component quality information is a structured data table containing remaining life, maintenance priority codes, process parameters, and inspection strategies, supporting full lifecycle management decisions.

[0079] This embodiment integrates temperature, load, and vibration spectrum data through multi-source synchronous data acquisition and spatiotemporal alignment technology to achieve comprehensive analysis of multi-physics coupling effects in the service environment, overcoming the limitations of single-field analysis in traditional methods. The timestamp alignment module unifies the time reference of heterogeneous data, and interpolation compensation ensures cross-modal data consistency, improving the accuracy of defect correlation analysis. A cross-modal attention mechanism dynamically allocates temperature, load, and vibration weights, enhancing the interactive characterization of mechanical and thermal stresses and constructing a quantitative assessment model for defect risk. Stress coupling calculation analyzes the multi-stress vector synthesis effect, quantifying the equivalent stress peak value and fatigue damage, breaking through the accuracy bottleneck of single-field models. A failure mode database drives Mahalanobis distance matching to objectively determine failure codes and hazard levels, eliminating human experience bias. The comprehensive quality module correlates remaining service life, maintenance priority, and process parameters to generate inspection cycles and maintenance strategies, achieving closed-loop management of the entire component lifecycle and supporting precise operation and maintenance decisions for high-end equipment.

[0080] Reference Figure 2 The present invention also provides a quantitative assessment system for surface defects of lightweight components, applicable to the quantitative assessment method for surface defects of lightweight components according to any one of the above-mentioned methods, comprising: The acquisition module is used to acquire multispectral surface images, laser 3D data and ultrasonic test data of lightweight components, and perform fusion registration based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components. The analysis module is used to perform dynamic pruning analysis on the joint dataset of components to obtain geometric contour features, and to extract internal features from the joint dataset of components to obtain internal anomaly response features. The correlation module is used to obtain stress gradient distribution data and volume ratio parameters for geometric contour features and internal anomaly features, and to predict defects according to the preset component material type mapping rules, so as to obtain a multi-dimensional prediction matrix. The processing module is used to acquire the original working condition dataset of lightweight components, perform service condition coupling analysis on the multi-dimensional prediction matrix, and identify multi-dimensional quality l-parameters to obtain the comprehensive quality information of the components.

[0081] This invention provides a quantitative assessment system for surface defects in lightweight components. Through spatiotemporal synchronous registration of multispectral, laser 3D, and ultrasonic detection data, it establishes a precise mapping and correlation between surface texture geometry and acoustic impedance response, solving the spatial positioning deviation problem caused by isolated processing of multi-source data in traditional methods. This enables cross-modal joint detection of micron-level defects. A dynamic pruning mechanism based on convolution kernel activation integrates multispectral reflectance characteristics and ultrasonic time-frequency features to construct an edge detection network for composite defects. This overcomes the limitations of traditional single-sensor feature analysis, improving the accuracy of morphological and mechanical coupling feature extraction for composite defects such as cracks and pores. Through multidimensional parameter coupling analysis of service condition data (temperature gradient, dynamic load) and defect stress gradient distribution, a material property-adaptive defect evolution model is constructed. This reduces the prediction error of the correlation between volume proportion parameters and stress field distribution, significantly improving the reliability of defect propagation behavior prediction under complex conditions. A dynamic matching mechanism based on the yield strength threshold of a component material type mapping rule library enables quantitative adaptation of defect sensitivity metrics for heterogeneous materials such as carbon fiber composites and titanium alloys. By constructing a structured output of maintenance priority codes and process compensation parameters, a closed-loop control link of defect assessment, process adjustment, and quality traceability is built, which reduces the defect rework rate of lightweight components and significantly improves the quality control efficiency of high-end equipment manufacturing processes.

[0082] Reference Figure 3 The present invention also provides a device for quantitatively assessing surface defects in lightweight components, comprising: Memory, used to store programs; A processor is used to execute a program to implement the various steps of a method for quantitatively assessing surface defects in lightweight components, as described above.

[0083] In this embodiment, the processor and memory can be connected via a bus or other means. The memory may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as read-only memory, flash memory, hard disk, or solid-state drive. The processor may be a general-purpose processor, such as a central processing unit, digital signal processor, application-specific integrated circuit, or one or more integrated circuits configured to implement embodiments of the present invention.

[0084] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the system and each module described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0085] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for quantitatively assessing surface defects in lightweight components, characterized in that, include: Multispectral surface images, laser 3D data, and ultrasonic test data of lightweight components are acquired and fused and registered based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components. Dynamic pruning analysis is performed on the joint dataset of the components to obtain geometric contour features, and internal feature extraction is performed on the joint dataset of the components to obtain internal abnormal response features; Stress gradient distribution data and volume ratio parameters are obtained for the geometric contour features and the internal anomaly features, and defect prediction is performed according to the preset component material type mapping rules to obtain a multi-dimensional prediction matrix. The original working condition dataset of the lightweight component is obtained, and the service condition coupling analysis is performed on the multi-dimensional prediction matrix. Multi-dimensional quality parameters are identified to obtain the comprehensive quality information of the component.

2. The method for quantitative evaluation of surface defects in lightweight components according to claim 1, characterized in that, The process involves acquiring multispectral surface images, laser 3D data, and ultrasonic testing data of lightweight components, and fusing and registering them based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components, including: The lightweight component is subjected to multi-band combined irradiation on its surface, and the reflected beam is separated by a preset beam-splitting prism array to obtain a multi-channel spectral image; Subpixel-level feature point matching is performed on the multi-channel spectral image to obtain the multispectral surface image; The original laser data of the lightweight component is obtained, and the surface curvature gradient value is dynamically adjusted to obtain the three-dimensional laser data. The original coded pulse and original echo signals of the lightweight component are acquired and phase-aligned fusion is performed to obtain the ultrasonic detection data. The multispectral surface image and the laser three-dimensional data are projected and registered to obtain initial component assembly data. Based on the ultrasonic testing data, the initial component assembly data is spatially aligned to obtain the component joint dataset.

3. The method for quantitatively assessing surface defects of lightweight components according to claim 1, characterized in that, The dynamic pruning parsing of the joint dataset of components yields geometric contour features, and the internal feature extraction of the joint dataset of components yields internal anomaly response features, including: The joint dataset of the components is subjected to spectral multi-channel band decomposition to obtain a texture feature map; A lightweight convolutional network channel is constructed on the texture feature map according to a preset convolutional kernel activation dynamic pruning algorithm to obtain a defect edge detection network; The geometric contour features are obtained by performing layer-by-layer feature mapping on the texture feature map through the defect edge detection network. The joint dataset of the components is subjected to ultrasonic time-frequency separation processing to obtain the defect reflection wave signal; Based on the joint dataset of the components, the acoustic path difference phase calibration of the defect reflected wave signal is performed to obtain an acoustic impedance anomaly distribution map; The coordinate matching relationship between the geometric contour features and the acoustic impedance anomaly distribution map is calculated to obtain the surface internal feature correlation matrix. The internal correlation matrix of the surface is fused using feature channels to obtain internal abnormal response features.

4. The method for quantitative evaluation of surface defects in lightweight components according to claim 1, characterized in that, The process involves analyzing stress gradient distribution data and volume percentage parameters of the geometric contour features and internal anomaly features, and then performing defect prediction based on preset component material type mapping rules to obtain a multi-dimensional prediction matrix, including: Multi-scale Gaussian curvature calculation is performed on the geometric contour features, and the surface deformation region is extracted to obtain the local curvature distribution features; The local curvature distribution features and the internal anomaly features are superimposed with anti-mutation strength to obtain a curvature gradient map of the defect region. Based on the curvature gradient map of the defect region and the laser three-dimensional data, spatial topological relationships are reconstructed to obtain a three-dimensional stress field. The stress gradient direction is traced in the three-dimensional stress field to obtain the stress gradient distribution data; Based on the laser 3D data, the curvature gradient map of the defect region is segmented into an octree space, and the density distribution is calculated to obtain the volume percentage parameter. Based on the component material type mapping rule, multimodal failure prediction is constructed using the stress gradient distribution data and the volume ratio parameter to obtain the multidimensional prediction matrix.

5. The method for quantitative evaluation of surface defects in lightweight components according to claim 4, characterized in that, The step of performing multi-scale Gaussian curvature calculation on the geometric contour features and extracting the surface deformation region to obtain local curvature distribution features includes: Multi-scale curvature calculation is performed on the geometric contour features to obtain an initial curvature distribution map; The initial curvature distribution map is subjected to curvature gradient consistency verification to obtain a verified curvature distribution map; The verified curvature distribution map is superimposed with feature channels to obtain an enhanced curvature distribution map; Based on a preset curvature gradient magnitude threshold, abrupt curvature gradient extraction is performed on the enhanced curvature distribution map to obtain curvature gradient abrupt regions; Connectivity analysis is performed on the curvature gradient abrupt change region, and region growth and boundary optimization are carried out to obtain a surface deformation region mask; Based on the surface deformation region mask, the deformation space of the enhanced curvature distribution map is truncated to obtain the curvature distribution extreme points; Based on the extreme points of the curvature distribution, feature extraction is performed on the enhanced curvature distribution map to obtain the local curvature distribution features.

6. The method for quantitatively evaluating surface defects of lightweight components according to claim 4, characterized in that, The step of performing octree spatial segmentation of the curvature gradient map of the defect region based on the laser three-dimensional data and calculating the density distribution to obtain the volume proportion parameter includes: The curvature gradient threshold of the curvature gradient map of the defect region is determined to obtain the region level depth; Based on the region hierarchy depth, the curvature gradient map of the defect region is recursively segmented into an octree space to obtain a segmentation structure set; Based on the laser 3D data, subspace density statistics are performed on the segmented structure set to obtain statistical information; Density anomaly neighborhood interpolation compensation is performed on the statistical information to obtain a compensation subspace set; The equivalent volume weight coefficient is obtained by calculating the subspace coefficients based on the compensation subspace set. The equivalent volume weight coefficients are weighted and superimposed in a subspace to obtain the volume percentage parameter.

7. The method for quantitative evaluation of surface defects in lightweight components according to claim 1, characterized in that, The process involves acquiring the original operating condition dataset of the lightweight component, performing service condition coupling analysis on the multi-dimensional prediction matrix, and identifying multi-dimensional quality parameters to obtain the comprehensive quality information of the component, including: The temperature distribution data, dynamic load spectrum data, and vibration spectrum data of the lightweight component are acquired and integrated to obtain the original working condition dataset. The original working condition dataset is timestamped based on the multi-dimensional prediction matrix to obtain the working condition feature set. The working condition feature set and the multi-dimensional prediction matrix are superimposed as tensors, and cross-modal attention mechanism is applied to obtain the enhanced prediction matrix. Stress coupling calculations are performed on the enhanced prediction matrix to obtain stress interaction feature vectors; Based on a preset failure mode database, the stress interaction feature vector is matched for defect similarity to obtain failure mode codes and hazard labels. The comprehensive quality information is obtained by performing quality analysis on the failure mode coding and the hazard characterization labels based on the enhanced prediction matrix.

8. A quantitative assessment system for surface defects in lightweight components, characterized in that, The method for quantitative assessment of surface defects in lightweight components according to any one of claims 1-7 includes: The acquisition module is used to acquire multispectral surface images, laser three-dimensional data and ultrasonic test data of lightweight components, and perform fusion registration based on a preset multi-sensor synchronization protocol to obtain a joint dataset of the components. The analysis module is used to perform dynamic pruning parsing on the joint dataset of components to obtain geometric contour features, and to extract internal features from the joint dataset of components to obtain internal abnormal response features. The association module is used to perform stress gradient distribution data and volume ratio parameters on the geometric contour features and the internal anomaly features, and to perform defect prediction according to the preset component material type mapping rules to obtain a multi-dimensional prediction matrix. The processing module is used to acquire the original working condition dataset of the lightweight component, perform service condition coupling analysis on the multi-dimensional prediction matrix, and identify multi-dimensional quality parameters to obtain the comprehensive quality information of the component.

9. A device for quantitatively assessing surface defects in lightweight components, characterized in that, include: Memory, used to store programs; A processor is configured to execute the program to implement the steps of the quantitative assessment method for surface defects of lightweight components as described in any one of claims 1-7.

Images