A method for detecting flaws in an LED encapsulant material

Abstract

The application provides an LED packaging material flaw detection method, comprising: for each layer image after separation, performing image enhancement processing to obtain clear material images of each layer; for the identified cross flaw area, performing local fine scanning to obtain micron-level three-dimensional structure data, and reconstructing fine topographic features of the cross flaw; according to the reconstructed cross flaw structure, combining various material performance parameters stored in a material database in advance, simulating and calculating the influence of the flaw on the performance of the LED, including optical, thermal and mechanical performance; judging the position, size, type and influence degree of the flaw on the performance, quantitatively scoring the overall quality of the LED packaging, and obtaining the final detection result and quality grade judgment.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for detecting defects in LED packaging materials. Background Technology

[0002] In the field of defect detection for LED packaging materials, a severe challenge arises from a multi-layered, interconnected problem. The complexity of LED packaging structures means that defects may exist at different interfaces and layers, such as between the substrate and the chip, or between the phosphor layer and the epoxy resin layer, forming intersecting defects. Detecting these defects is far more difficult than conventional methods. Single-layer detection methods struggle to comprehensively capture all potential defects, and traditional detection techniques, limited by their independent assessment of specific layers or materials, are ill-suited for effectively identifying complex defects across layers. For example, the interaction between substrate microcracks and poor chip adhesion, and the interaction between uneven phosphor distribution and epoxy resin layer bubble defects, not only severely impact the optical performance and reliability of LEDs but may also lead to misjudgments and missed detections. Furthermore, the significant differences in the physical and optical properties of different material layers make adopting unified detection standards and methods extremely difficult; transparent epoxy resin layers and opaque phosphor layers require drastically different imaging techniques and analytical algorithms. Adding to the complexity, some defects may not be significant at a single layer, but once multi-layer interactions are involved, they can significantly affect the overall performance of the LED. In summary, designing a comprehensive detection method that can take into account multiple levels and interfaces to accurately identify and evaluate cross-defects in LED packaging materials requires not only breaking through the traditional single detection approach, but also deeply considering the interaction mechanism between materials and developing new multimodal sensing technologies and data fusion algorithms to address this technical challenge. This is a technical problem that urgently needs to be solved. Summary of the Invention

[0003] This invention provides a method for detecting defects in LED packaging materials, mainly including:

[0004] Multispectral imaging was used to scan the LED package and acquire image data at different wavelengths. Based on the spectral response characteristics of each material layer in different bands, independent image information of the substrate, chip, phosphor layer and epoxy resin layer was separated.

[0005] Image enhancement processing is performed on the separated images of each layer to obtain clearer images of each material layer;

[0006] For each enhanced image layer, a deep learning model is used for feature extraction and defect detection to identify potential defect areas, including cracks, bubbles, and impurities, and to mark the location and type of defects.

[0007] Based on the detected defect information of each layer, a three-dimensional spatial model is constructed, and defects of different layers are mapped to the same coordinate system. The spatial distribution relationship of defects is analyzed, and cross-layer defect areas are identified.

[0008] For the identified cross-defect areas, a local fine scan is performed to obtain micron-level three-dimensional structural data and reconstruct the fine morphological features of the cross-defects.

[0009] Based on the reconstructed cross-defect structure, combined with various material performance parameters pre-stored in the material database, the impact of defects on LED performance, including optical, thermal and mechanical properties, is simulated and calculated.

[0010] The location, size, type, and impact on performance of the defects are determined, and the overall quality of the LED package is quantitatively scored to obtain the final test results and quality grade determination.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] This invention discloses a method for detecting defects in LED packaging materials, overcoming the limitations of traditional single detection methods. It achieves comprehensive capture and accurate evaluation of potential defects in the packaging structure, effectively solving the problem of detecting cross-layer defects. Its advantage lies in its ability to consider the multi-layered characteristics of the packaging materials. Through multispectral imaging technology, it performs a comprehensive scan of the LED package, acquiring high-resolution image data of different material layers at specific wavelengths. Advanced image enhancement processing clearly reveals the details of each material layer, ensuring the accuracy of subsequent defect detection. Subsequently, a deep learning model is applied to feature extraction and defect recognition. Its powerful pattern recognition capabilities can identify and label various defects, such as microcracks, bubbles, and impurities, without overlooking any, even under multi-layer interactions. More importantly, the constructed three-dimensional spatial model can map defects at different levels to the same coordinate system, intuitively presenting the spatial distribution of defects, with particular attention to important cross-layer defect areas that only appear in multi-layer interactions. Localized fine scanning and 3D reconstruction of these areas revealed the detailed morphology of the cross-defects. Combined with a material property database, the comprehensive impact of these defects on the optical, thermal, and mechanical properties of LEDs was simulated and calculated, providing a quantitative assessment basis for the severity of the defects. Ultimately, by comprehensively considering the location, size, type of defects, and their impact on performance, a quantitative scoring and grading of the overall quality of LED packaging was achieved. This significantly improved the accuracy and efficiency of inspection, reduced errors caused by human factors, lowered production costs, and significantly improved the quality consistency and reliability of LED products. Attached Figure Description

[0013] Figure 1This is a flowchart of a method for detecting defects in LED packaging materials according to the present invention.

[0014] Figure 2 This is a schematic diagram of a method for detecting defects in LED packaging materials according to the present invention.

[0015] Figure 3 This is another schematic diagram of a method for detecting defects in LED packaging materials according to the present invention. Detailed Implementation

[0016] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] like Figure 1 -3, This embodiment of the method for detecting defects in LED packaging materials may specifically include:

[0018] Step S101: The LED package is scanned using a multispectral imaging method to obtain image data at different wavelengths. Based on the spectral response characteristics of each layer of material in different bands, independent image information of the substrate, chip, phosphor layer and epoxy resin layer is separated.

[0019] A multispectral imaging device is used to perform omnidirectional scanning of the LED package. Key bands are selected according to a preset wavelength range for scanning, resulting in multiple sets of image data. Principal component analysis is performed on the multispectral image data to identify the key bands that best distinguish different material layers. The spectral response characteristics of each material layer in the LED package are extracted in different bands, establishing a spectral response curve database containing characteristic curves of the substrate, chip, phosphor layer, and epoxy resin layer. The image set is classified pixel-level according to the spectral response curve database. A support vector machine algorithm is used to compare the spectral response of each pixel with the characteristic curves in the database to generate preliminary separation images of the substrate, chip, phosphor layer, and epoxy resin layer. Three-dimensional reconstruction is performed on the preliminary separation images, and structured light projection is used for estimation to generate the internal three-dimensional structure of the LED package containing spatial location and morphological information of the substrate, chip, phosphor layer, and epoxy resin layer.

[0020] Specifically, a multispectral imaging device is used to perform a 360-degree omnidirectional scan of the LED package. Based on a preset wavelength range of 300nm to 1000nm, key bands are selected for scanning to acquire multiple sets of image data. The image data is preprocessed to remove noise and distortion, resulting in a corrected multispectral image set. Principal component analysis is performed on the multispectral image data to identify the key bands that best distinguish different material layers, providing a basis for subsequent feature extraction. The spectral response characteristics of each layer of the LED package in different bands are extracted from the corrected multispectral image set, establishing a spectral response curve database containing feature curves for the substrate, chip, phosphor layer, and epoxy resin layer. A support vector machine algorithm is used to compare the spectral response of each pixel in the image set with the feature curves in the database. Based on the spectral matching results, the image set is classified at the pixel level, and pixels with similar spectral response characteristics are grouped to generate preliminary separation images of the substrate, chip, phosphor layer, and epoxy resin layer. Morphological processing using opening and closing operations is applied to remove noise and smooth boundaries. Then, a watershed algorithm is applied for region segmentation to improve the accuracy of the separated images. The optimized separated image is reconstructed in 3D. Combining the structural information and material properties of the LED package, the depth is estimated using structured light projection. The 3D structure is reconstructed by projecting a known pattern and analyzing its deformation. At the same time, photometric stereo vision technology is combined to improve the reconstruction accuracy by using images under different lighting conditions. A 3D model of the internal structure of the LED package is generated, including the spatial position and morphological information of the substrate, chip, phosphor layer and epoxy resin layer.

[0021] The multispectral imaging device employs a linear CCD sensor, coupled with a rotating platform to achieve 360-degree scanning, covering a wavelength range of 300nm to 1000nm. Seven key bands—340nm, 450nm, 550nm, 650nm, 750nm, 850nm, and 950nm—are selected for scanning. Image preprocessing uses median filtering to remove salt-and-pepper noise and perspective transformation to correct distortion. Principal component analysis, through eigenvalue decomposition, selects the top three principal components with a contribution rate exceeding 95% as key bands. The spectral response curve database is built based on pre-measured standard samples in the laboratory, containing characteristic curves of 50 common LED packaging materials. The support vector machine algorithm uses a radial basis function kernel with a C parameter set to 1.0 and a gamma parameter set to 0.1, with 10,000 training samples. The pixel-level classification threshold is set to 0.95, meaning that when the matching similarity is greater than 0.95, the pixel is classified into the corresponding material layer. In morphological processing, opening operations used 3x3 rectangular structuring elements, and closing operations used 5x5 rectangular structuring elements, with an iteration count of 2 for both. The watershed algorithm used distance transformation as input, marked interior and background points, and determined the watershed line using gradient descent. Structured light projection used a 9x9 Gray code pattern with a projection frequency of 60Hz, and the camera acquisition frequency was 120Hz, with a synchronization error of less than 1ms. Photometric stereo vision used five LED light sources in different directions with angles of 0°, 45°, 90°, 135°, and 180°, with an exposure time of 10ms. 3D reconstruction used a voxel mesh with a spatial resolution of 0.1mm, and calculated the material properties of each voxel using trilinear interpolation. The final generated 3D model contained the spatial location and morphological information of the substrate, chip, phosphor layer, and epoxy resin layer, achieving a model accuracy of 50μm.

[0022] Step S102: Perform image enhancement processing on the separated images of each layer to obtain clearer images of each material layer.

[0023] Image enhancement processing is performed on the separated images to obtain a first image, including processing the original image according to a preset block size and contrast limit threshold; a bilateral filter is applied to the first image to obtain a denoised second image, wherein the bilateral filter calculates weights based on preset spatial domain standard deviation, value domain standard deviation, and window size; a nonlinear sharpening method is applied to the second image to enhance edges, thereby obtaining an edge-enhanced third image, wherein the nonlinear sharpening method is implemented through Gaussian blur and edge image generation; the resolution of the third image is enhanced using a pre-trained ESPCN algorithm to obtain a resolution-enhanced fourth image; and a sharpened material layer image is reconstructed based on the fourth image.

[0024] Specifically, for each layer of the separated image, a contrast-limited adaptive histogram equalization algorithm is used to enhance the image contrast. The image is divided into 8x8 blocks, and histogram equalization is performed on each block with a contrast limit threshold of 0.01 to prevent excessive noise enhancement, resulting in an image with improved contrast. A bilateral filter is then used to denoise the enhanced image, with a spatial domain standard deviation of 3, a value domain standard deviation of 0.1, and a window size of 5x5. Weights are calculated based on the spatial distance between pixels and the difference in grayscale values, eliminating noise while preserving edge information to obtain a smooth image with preserved edges. The denoised image is then non-linearly sharpened using the unsharpening masking method. First, Gaussian blur is applied with a standard deviation of 1.5 to obtain a blurred image. Then, the difference between the original image and the blurred image is calculated to obtain the edge image. The edge image is multiplied by a factor of 0.7 and added back to the original image to improve edge sharpness and detail, generating an edge-enhanced image. The ESPCN (Efficient Sub-Pixel Convolutional Neural Network) algorithm is used to upscale the edge-enhanced image. The training dataset contains high- and low-resolution image pairs of various LED packaging materials. The network structure consists of three convolutional layers, with the last layer using sub-pixel convolutional upsampling and a magnification factor of 2. By learning the mapping relationship from low-resolution to high-resolution images, high-resolution material images of each layer are reconstructed, resulting in a final image with higher clarity. A contrast-limited adaptive histogram equalization algorithm first divides the 1024x1024 pixel image into 8x8 blocks, each 128x128 pixels. A grayscale histogram is calculated for each block, with a contrast limit threshold of 0.01. Components exceeding the threshold are cropped and evenly distributed across the entire histogram. Histogram equalization is performed on each block, and adjacent blocks are merged using bilinear interpolation to obtain an image with enhanced contrast. A bilateral filter is applied to the image with a spatial domain standard deviation of 3 pixels, a value domain standard deviation of 0.1, and a window size of 5x5 pixels. For the substrate layer, due to its relatively coarse texture, the spatial domain standard deviation is adjusted to 5 pixels; for the chip layer, due to its rich detail, the value domain standard deviation is reduced to 0.05. The Unsharpmasking technique uses a Gaussian blur kernel (standard deviation of 1.5 pixels) to obtain a blurred image. The difference between the original image and the blurred image is used to obtain the edge image, which is then multiplied by a factor of 0.7 and added back to the original image. For the phosphor layer, due to its strong graininess, the factor is adjusted to 0.5 to avoid over-sharpening. The ESPCN network structure contains three convolutional layers: the first layer uses 64 5x5 convolutional kernels, the second layer uses 32 3x3 convolutional kernels, the third layer uses 9 3x3 convolutional kernels, and the last layer uses subpixel convolutional upsampling with a magnification factor of 2.The training dataset contains 10,000 pairs of high- and low-resolution images of LED packaging materials. The low-resolution images are 128x128 pixels in size, and the high-resolution images are 256x256 pixels in size. The network was trained using the Adam optimizer with a learning rate of 0.001, a batch size of 64, and for 100 epochs. After processing, a high-resolution 256x256 pixel image with twice the sharpness, significantly reduced noise, and sharp edges was obtained.

[0025] Step S103: For the enhanced images of each layer, use a deep learning model to perform feature extraction and defect detection, identify potential defect areas from the images, including cracks, bubbles, and impurities, and mark the location and type information of the defects.

[0026] A ResNet-50 network is used to extract features from each layer of the enhanced image. The first four convolutional blocks of the network are retained, and the fully connected layers are removed to obtain feature maps containing rich semantic information. A region proposal network is used to slide convolutional kernels of a preset size on the feature map to generate anchor boxes of multiple scales and ratios, filter and optimize candidate regions, and obtain the location information and bounding box coordinates of defect regions. For the selected candidate regions, a fixed-size feature map is extracted, and feature transformation is performed through two fully connected layers. A softmax classifier is used to classify the feature map into three categories: cracks, bubbles, and impurities. At the same time, a corresponding category label and precise location are assigned to each defect region. Based on the location information and type label of the defect region, it is marked on the original image, bounding boxes are drawn and type labels are added to generate a labeled image containing defect location and type information.

[0027] Specifically, a ResNet-50 network is used to extract features from each layer of the enhanced image. The first four convolutional blocks are retained, and the fully connected layers are removed. Each layer of LED packaging material image is input into the network for processing. Low-level, mid-level, and high-level features of the image are extracted through multi-layer convolution and pooling operations to obtain a feature map of 1 / 16 size containing rich semantic information. A region proposal network is used to process the feature map. A 3x3 convolutional kernel is slid across the feature map to generate anchor boxes of multiple scales and proportions. Candidate regions are filtered and optimized through binary classification and bounding box regression. Finally, a non-maximum suppression algorithm is applied to retain the top 300 candidate regions with the highest scores to obtain the location information and bounding box coordinates of the defect regions. For the selected candidate regions, a fixed-size 7x7 feature map is extracted using the ROIAlign method. Then, feature transformation is performed through two fully connected layers. A softmax classifier is used to classify the defect types into three categories: cracks, bubbles, and impurities. At the same time, a smooth L1 loss is used to fine-tune the bounding boxes, assigning a corresponding category label and precise location to each defect region. Based on the location information and type labels of the defect areas, the original image is marked, bounding boxes are drawn, and type labels are added. Different colors are used to represent different types of defects: red for cracks, green for bubbles, and blue for impurities. A type label and confidence score are added to the upper left corner of the bounding box, with the label format "type: score". This generates a labeled image containing defect location and type information for subsequent quality assessment and defect analysis. The ResNet-50 network processes a 512x512 pixel LED packaging material image, passing it through 49 convolutional and pooling layers to output a 32x32x2048 feature map. The region proposal network uses 3x3 convolutional kernels with a stride of 1 on the feature map to generate anchor boxes of 9 different scales and proportions, producing a total of 32x32x9 = 9216 initial candidate regions. Through bounding box regression and binary classification, the top 1000 candidate regions are selected, and then subjected to non-maximum suppression with an IoU threshold of 0.7, retaining 300 final candidate regions. The ROIAlign method transforms candidate regions of different sizes into a uniform 7x7 feature map, preserving the continuity of spatial information. Two fully connected layers, with 1024 and 512 neurons respectively, transform the features. A softmax classifier outputs probabilities for four categories (background, crack, bubble, and impurity), while four regressors fine-tune the bounding box coordinates. A classification threshold of 0.5 is set; if the value exceeds the threshold, the image is classified as a defect of the corresponding category. On the original 512x512 image, defects are marked with 2-pixel-wide bounding boxes: cracks are marked in red (255, 0, 0), bubbles in green (0, 255, 0), and impurities in blue (0, 0, 255). At the top left corner of the bounding box, "Type: 0.XX" is labeled in 10-point white font, where 0.XX represents the confidence score.The entire processing runs on the GPU, with a single image processing time of approximately 0.2 seconds, enabling real-time detection. The average accuracy (mAP) for 1000 test images reached 0.85, with the highest accuracy for crack detection (0.90), followed by bubbles (0.85), and the lowest accuracy for impurities (0.80).

[0028] Step S104: Based on the detected defect information of each layer, construct a three-dimensional spatial model, map the defects of different layers to the same coordinate system, analyze the spatial distribution relationship of defects, and identify defect areas that cross layers.

[0029] Based on the detected defect information from each layer, the two-dimensional defect coordinates are converted into a three-dimensional spatial point set using a direct coordinate mapping method. By setting the thickness parameter of each material layer, the depth value of the defect in the Z-axis direction is calculated, generating three-dimensional point cloud data containing XYZ coordinates. The three-dimensional point cloud data is then voxelized, dividing the continuous space into discrete cubic meshes. For each point cloud data, its voxel index is calculated, and the attribute information of the point is recorded in that voxel, resulting in a discretized three-dimensional spatial model. In the discretized three-dimensional spatial model, each non-empty voxel is marked. Adjacent voxels are traversed; if a voxel is found to be connected across layers, it is identified as a potential cross-layer defect region. The marked voxels are grouped, aggregating voxels that are spatially close and belong to the same defect type into continuous defect regions. The clustering results are post-processed, merging clusters of the same type of defect with small distances. Based on the clustered defect regions, the volume, surface area, and number of layers spanned by each defect region are calculated, ultimately generating a three-dimensional defect distribution report.

[0030] Specifically, based on the detected defect information of each layer, the two-dimensional defect coordinates are converted into a three-dimensional spatial point set using the direct coordinate mapping method. For each two-dimensional defect coordinate (x, y), the z-coordinate is calculated according to the index i of its layer and the preset layer thickness d, z = i * d. By setting the thickness parameter of each layer of material, the depth value of the defect in the Z-axis direction is calculated, generating three-dimensional point cloud data containing XYZ coordinates. The point cloud density is consistent with the resolution of the original image. The generated three-dimensional point cloud data is voxelized, dividing the continuous space into discrete cubic meshes. The voxel size is set to 0.1mm x 0.1mm x 0.1mm, and an octree data structure is used for spatial partitioning to improve query efficiency. For each point cloud data, its voxel index is calculated, and the attribute information of the point is recorded in the voxel. The point cloud data is mapped to the corresponding voxel to obtain a discretized three-dimensional spatial model. In the voxelized 3D spatial model, each non-empty voxel is labeled, recording the material layer and defect type to which it belongs. The 26-connectivity rule is used to traverse adjacent voxels. For each non-empty voxel, it is checked whether its adjacent voxels belong to different material layers. If voxels connected across layers are found, they are marked as potential cross-layer defect regions. Continuous defect voxels in the Z-axis direction are identified to determine whether cross-layer defect regions exist. A density-based spatial clustering algorithm is used to group the labeled voxels, setting the neighborhood radius to 0.5 mm and the minimum number of points to 5. Voxels that are spatially close and belong to the same defect type are aggregated into continuous defect regions. The clustering results are post-processed, merging defect clusters of the same type with a distance of less than 1 mm. The volume, surface area, and number of layers spanned by each defect region are calculated to generate a detailed 3D defect distribution report. For a 512x512 pixel LED package image, where each pixel represents 0.01 mm, three material layers (substrate, chip, and phosphor layer) were detected, with thicknesses of 0.2 mm, 0.1 mm, and 0.3 mm, respectively. A direct coordinate mapping method generated a 3D point cloud of 262,144 points, containing 1,532 defect points. Voxelization divided the space into 800x800x60 voxels, each 0.1 mm x 0.1 mm x 0.1 mm in size. The maximum depth of the octree data structure was set to 6, with leaf nodes containing a maximum of 8 points. During 26-connectivity rule traversal, each non-empty voxel examined its 26 neighboring voxels, identifying 73 potential cross-layer defect regions. A density-based spatial clustering algorithm, with a neighborhood radius ε = 0.5 mm and a minimum number of points MinPts = 5, identified 124 independent defect clusters. Post-processing merged 28 pairs of the same type of defect clusters with a distance of less than 1 mm, resulting in 96 defect areas. Among them, there were 43 cracks (35 in a single layer and 8 across layers), 31 bubbles (27 in a single layer and 4 across layers), and 22 impurities (20 in a single layer and 2 across layers).The largest defect area is a crack spanning the substrate and chip layers, with a volume of 0.015 mm³ and a surface area of ​​0.42 mm³. The generated 3D defect distribution report includes the type, location, volume, surface area, number of layers spanned, and distribution ratio of each defect area, providing accurate spatial information for quality control and process improvement.

[0031] Step S105: For the identified cross-defect areas, perform local fine scanning to obtain micron-level three-dimensional structural data and reconstruct the fine morphological features of the cross-defects.

[0032] A laser scanning microscope is used to acquire three-dimensional structural data of the cross-defect region. The three-dimensional structural data includes information about the defect region and its adjacent boundaries. The acquired three-dimensional structural data is preprocessed, including noise removal, to obtain denoised three-dimensional structural data. Moving least squares is used to smooth the point cloud data, resulting in smoothed point cloud data. Three-dimensional surface fitting is performed on the smoothed point cloud data to generate a mesh model representing the fine morphology of the cross-defect. Local geometric features are extracted from the mesh model, and local geometric feature parameters are calculated and determined. Based on the local geometric feature parameters, a morphological skeleton extraction algorithm is applied to identify the crack direction and accurately locate the defect edges, obtaining three-dimensional defect model data containing fine morphological features.

[0033] Specifically, based on the coordinates of the identified cross-fault areas, a confocal laser scanning microscope was used for detailed local scanning. An adaptive scanning strategy was employed, combining grid scanning and contour tracking scanning, with the scanning range dynamically adjusted to ensure complete coverage of the cross-fault areas and their surrounding 1mm boundaries. The scanning step size was set to 0.5μm to acquire high-density 3D structural data. The acquired high-density 3D structural data was preprocessed, using a statistical outlier filtering algorithm to remove noise, followed by smoothing the point cloud using moving least squares to preserve edge features while reducing measurement noise, resulting in clear microstructural point cloud data. A Poisson surface reconstruction algorithm was used to fit a 3D surface to the processed point cloud data, setting the octree depth to 10 and the sampling point count to at least 100 per square unit. Adaptive density control was introduced to increase reconstruction accuracy in the fault edge areas, generating a mesh model representing the fine morphology of the cross-faults. The mesh resolution was set to 0.1μm to ensure accurate description of minute structural features. Local geometric features are extracted from the reconstructed 3D mesh model. Principal curvature analysis and shape index calculation are used to extract local geometric features. Curvature, normal vector and depth information are calculated. Morphological skeleton extraction algorithm is applied to identify crack direction. The level set method is combined to accurately locate the defect edge and generate 3D defect model data containing fine morphological features.

[0034] The confocal laser scanning microscope used a 488nm laser source with a scanning range of 2mm x 2mm and a Z-axis resolution of 0.1μm. An adaptive scanning strategy first performed a coarse scan with 100μm intervals to identify defect boundaries, then a fine scan with 10μm intervals in the boundary regions, finally achieving a scanning step size of 0.5μm in the core region. The acquired raw point cloud data contained approximately 16 million points. A statistical outlier filtering algorithm was used with a neighborhood of 50 points and a standard deviation factor of 2.5, removing approximately 2% of noise. The moving least squares method used a quadratic polynomial fitting with a search radius of 5μm and 5 iterations, reducing the number of point clouds to 15 million after smoothing. In Poisson surface reconstruction, an octree depth of 10 corresponds to a minimum voxel size of approximately 0.05μm, and the sampling density was set to 400 points / μm³, with adaptive density control increasing to 800 points in the edge regions.

[0035] / μm3. The reconstructed 3D mesh model contains approximately 5 million triangular facets. Principal curvature analysis uses a spherical neighborhood with a radius of 3μm to calculate Gaussian curvature and mean curvature, with a shape index threshold set to 0.7 to distinguish between concave and convex features. Morphological skeleton extraction employs a 3D thinning algorithm with a minimum branch length of 1μm. The level set method uses a fast marching algorithm with an evolution time step of 0.1 and 100 iterations to accurately locate defect edges. The final generated 3D defect model data includes a geometric mesh, curvature map, shape index map, and skeleton structure, with a total data size of approximately 200MB.

[0036] Step S106: Based on the reconstructed cross-defect structure and combined with various material performance parameters pre-stored in the material database, simulate and calculate the impact of defects on LED performance, including optical, thermal, and mechanical properties.

[0037] The process involves: acquiring the reconstructed cross-flaw structure; retrieving corresponding material performance parameters from the material database based on the cross-flaw structure; processing the properties of the transition region using a material interpolation function; setting the minimum mesh element size to construct a refined LED structure model containing the flaws; simulating the photon propagation path in the LED structure model using the Monte Carlo ray tracing algorithm, setting the photon quantity and wavelength range; calculating the scattering coefficient based on the phosphor's reabsorption and scattering effects to obtain data on the impact of flaws on optical performance; discretizing the steady-state heat conduction equation using the finite volume method, including solving the discrete equation set using the conjugate gradient method; obtaining information on thermal performance changes caused by flaws by calculating thermal resistance and junction temperature; performing stress-strain analysis using linear elasticity theory; implementing adaptive mesh refinement, including setting refinement criteria based on stress gradient; if the stress gradient is greater than a preset threshold, reducing the mesh size, iterating the refinement process until the stress difference between two consecutive iterations is less than the preset threshold, obtaining prediction results for stress concentration and potential failure locations.

[0038] Specifically, the reconstructed 3D structure of the cross-flaw is imported into COMSOL Multiphysics software. Based on the location and shape of the flaw, corresponding material property parameters, including refractive index, thermal conductivity, and elastic modulus, are retrieved from the material database. For flaws spanning multiple materials, material interpolation functions are used to handle the properties of the transition region. The minimum mesh element size is set to 0.1 μm to ensure accurate capture of flaw details, thus constructing a detailed LED structure model containing the flaws. For optical performance, the Monte Carlo ray tracing algorithm is used to simulate the propagation path of photons in the LED structure. The photon count is set to 10^6, the wavelength range is 300-800 nm, and 20 equally spaced sampling points are used. Considering the reabsorption and scattering effects of the phosphor, the Mie scattering theory is used to calculate the scattering coefficient, light extraction efficiency, and luminous uniformity, obtaining data on the impact of flaws on optical performance. For thermal properties, the steady-state heat conduction equation is discretized using the finite volume method, and the discrete equation system is solved using the conjugate gradient method. The convergence criterion is set to a residual of less than 10⁻⁶. Considering temperature-dependent material properties, such as the change in thermal conductivity with temperature, information on thermal performance changes caused by defects is obtained by calculating thermal resistance and junction temperature. For mechanical properties, stress-strain analysis is performed using linear elasticity theory, and h-adaptive mesh refinement is implemented. A refinement criterion is set based on the stress gradient. In regions where the stress gradient is greater than a preset threshold, the mesh size is reduced by 50%. The refinement process is iterated until the stress difference between two consecutive iterations is less than 1%, improving the calculation accuracy of defective regions and obtaining prediction results for stress concentration and potential failure locations.

[0039] The 3D cross-scratch structure imported into COMSOL Multiphysics software has a volume of 5μm x 5μm x 10μm and includes cracks spanning the chip and phosphor layers. The material interpolation function used is the hyperbolic tangent function, with a transition region width of 0.5μm. Mesh generation used tetrahedral elements with a minimum size of 0.1μm, generating approximately 1 million mesh elements. In the optical simulation, Monte Carlo ray tracing was performed with a photon count of 10^6 and a wavelength range of 300–800nm ​​divided into 20 sampling points. Mie scattering theory calculated the phosphor scattering coefficient to be 1000 cm^-1 at 450nm. The light extraction efficiency decreased from 85% to 82%, and the luminescence uniformity decreased by 5%. In the thermal analysis, the finite volume method discretized approximately 1.5 million equations, and the conjugate gradient method required 50 iterations to converge. The rate of change of thermal conductivity with temperature was set to -0.2% / K. Calculations showed an increase in thermal resistance of 0.5 K / W and a junction temperature increase of 3 °C. In the mechanical property analysis, the h-adaptive mesh refinement was set with a stress gradient threshold of 5 MPa / μm. After three iterations of refinement, the number of mesh elements increased to 1.5 million. The final calculations showed that the maximum stress concentration region was located at the crack tip, with a stress value reaching 350 MPa, 40% higher than the flawless case. The predicted potential failure location was within 0.5 μm of the crack tip. The entire multiphysics coupling analysis was completed in 4 hours on a 64-core high-performance computer.

[0040] Step S107: Determine the location, size, type, and impact on performance of the defects, quantify and score the overall quality of the LED package, and obtain the final test results and quality grade determination.

[0041] The spatial coordinates, volume, surface area, and impact on optical, thermal, and mechanical properties of defects are obtained to construct a multi-dimensional feature vector. This multi-dimensional feature vector includes the x, y, and z coordinates, volume, surface area, and impact on optical, thermal, and mechanical properties of the defect. Based on this multi-dimensional feature vector, a scoring model is established using a random forest algorithm. This scoring model uses the location, size, type of defect, and its impact on various performance aspects as evaluation indicators. A weighted score for the defect is calculated using the scoring model. Based on the weighted score, each evaluation indicator is divided into several levels, and the membership degree is calculated using a Gaussian membership function to construct a fuzzy relation matrix. A fuzzy synthesis operation is performed on the fuzzy relation matrix using a weighted average method to obtain the overall quality score of the LED package. If the overall quality score is less than a preset threshold, the quality level of the LED package is determined to be unqualified; if the overall quality score is greater than or equal to the preset threshold, the quality level of the LED package is determined to be qualified.

[0042] Specifically, based on the 3D reconstruction results and performance simulation data, geometric features such as spatial coordinates, volume, and surface area of ​​defects are extracted. Combined with the degree of influence on optical, thermal, and mechanical properties, a 10-dimensional feature vector is constructed, including the defect's x, y, and z coordinates, volume, surface area, and the degree of influence on optical, thermal, and mechanical properties (each occupying one dimension). The simulation results are normalized to the range of 0-1 to characterize the comprehensive attributes of each defect. A random forest algorithm is used to establish a scoring model, using the defect's location, size, type, and its influence on various performance aspects as evaluation indicators. Historical data is used to train the model, automatically learning the importance of each indicator as weights. The number of trees is set to 100, and the maximum depth to 10, to balance the model's complexity and generalization ability. A weighted score for each defect is then calculated. Using a fuzzy comprehensive evaluation method, based on defect weighted scores, each evaluation indicator is divided into 5 levels (excellent, good, average, poor, very poor). Gaussian membership functions are used to calculate membership degrees, constructing a fuzzy relation matrix. Fuzzy synthesis is then performed using a weighted average method to obtain the overall quality score of the LED package. The score results are mapped to predefined quality level ranges. Based on the quality score and preset thresholds, the final inspection result and quality level of the LED package are determined. The preset thresholds are determined based on statistical analysis of a large amount of historical data. A k-means clustering algorithm is used to divide the historical data into 5 categories, corresponding to 5 quality levels. The cluster centers serve as representative values ​​for each level, generating an inspection report containing detailed defect information, performance impact analysis, and quality level determination, providing a basis for subsequent quality control and process optimization. For a specific LED package sample, 5 defects are obtained after 3D reconstruction and performance simulation. A 10-dimensional feature vector was constructed, with spatial coordinates ranging from 0 to 5000 μm, 0 to 5000 μm, and 0 to 1000 μm, volume ranging from 0 to 1000 μm³, and surface area ranging from 0 to 5000 μm³. Performance influence was normalized to 0-1. A random forest model was trained on 10,000 historical samples using 100 decision trees at a maximum depth of 10, yielding feature importance as follows: position 30%, size 25%, optical influence 20%, thermal influence 15%, and mechanical influence 10%. In the fuzzy evaluation, the mean of the Gaussian membership function was set as the center value for each level, and the standard deviation was 1 / 6 of the interval length. In the weighted average method, the weights were consistent with the feature importance obtained from the random forest. K-means clustering, based on 50,000 historical samples and iterated 100 times, yielded five cluster centers as quality level thresholds: 95, 85, 75, 65, and 55. Five defects in the LED package sample were evaluated, and the scores were 92, 88, and...

[0043] Scores of 79, 76, and 71, with an overall score of 83.5, correspond to a "Good" rating. The inspection report includes 10-dimensional feature values, individual scores, and an overall score for each defect, as well as the final quality level determination. The entire evaluation process, accelerated by GPU, was completed in 0.5 seconds.

[0044] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for detecting defects in LED packaging materials, characterized in that, The method includes: scanning the LED package using a multispectral imaging method to acquire image data at different wavelengths; separating independent image information of the substrate, chip, phosphor layer, and epoxy resin layer based on the spectral response characteristics of each material layer in different bands; performing image enhancement processing on the separated layer images to obtain clearer images of each material layer; using a deep learning model to perform feature extraction and defect detection on the enhanced layer images, identifying potential defect areas from the images, including cracks, bubbles, and impurities, and marking the location and type of defects; and constructing a three-dimensional spatial model based on the detected defect information of each layer, and then... Layer defects are mapped onto the same coordinate system to analyze their spatial distribution and identify cross-layer defect areas. For these identified cross-layer defect areas, a fine local scan is performed to acquire micron-level three-dimensional structural data, reconstructing the fine morphological features of the cross-layer defects. Based on the reconstructed cross-layer defect structure and combined with various material performance parameters pre-stored in the material database, the impact of defects on LED performance, including optical, thermal, and mechanical properties, is simulated and calculated. The location, size, type, and degree of impact on performance of the defects are determined, and the overall quality of the LED package is quantitatively scored to obtain the final inspection results and quality grade determination.

2. The method according to claim 1, wherein, The method employs multispectral imaging to scan the LED package, acquiring image data at different wavelengths. Based on the spectral response characteristics of each material layer in different bands, independent image information of the substrate, chip, phosphor layer, and epoxy resin layer is separated. This includes: performing a 360-degree scan of the LED package using a multispectral imaging device; selecting key bands within a preset wavelength range for scanning to obtain multiple sets of image data; performing principal component analysis on the multispectral image data to identify the key bands that best distinguish different material layers; extracting the spectral response characteristics of each material layer in the LED package at different bands to establish a spectral response curve database containing feature curves of the substrate, chip, phosphor layer, and epoxy resin layer; and performing pixel-level classification of the image set based on the spectral response curve database, comparing the spectral response of each pixel with the feature curves in the database using a support vector machine algorithm. A preliminary separation image of the substrate, chip, phosphor layer, and epoxy resin layer is generated; the preliminary separation image is reconstructed in three dimensions, and the structured light projection method is used for estimation to generate the internal three-dimensional structure of the LED package containing the spatial position and morphological information of the substrate, chip, phosphor layer, and epoxy resin layer.

3. The method according to claim 1, wherein, The step of performing image enhancement processing on the separated layer images to obtain a clearer image of each material layer includes: performing image enhancement processing on the separated layer images to obtain a first image, including processing the original image according to a preset block size and contrast limit threshold; processing the first image using a bilateral filter to obtain a denoised second image, wherein the bilateral filter calculates weights based on preset spatial domain standard deviation, value domain standard deviation, and window size; applying a nonlinear sharpening method to the second image for edge enhancement to obtain an edge-enhanced third image, wherein the nonlinear sharpening method is implemented through Gaussian blur and edge image generation; using a pre-trained ESPCN algorithm to enhance the resolution of the third image to obtain a resolution-enhanced fourth image; and reconstructing a clearer material layer image based on the fourth image.

4. The method according to claim 1, wherein, The enhanced images at each layer are then processed using a deep learning model for feature extraction and defect detection. Potential defect areas, including cracks, bubbles, and impurities, are identified, and their location and type are marked. This process includes: using a ResNet-50 network to extract features from each enhanced image layer, retaining the first four convolutional blocks, and removing the fully connected layers to obtain a feature map containing rich semantic information; and then using a region proposal network to slide convolutional kernels of a preset size onto the feature map. Anchor boxes of multiple scales and proportions are generated to filter and optimize candidate regions, obtaining the location information and bounding box coordinates of defective regions. For the filtered candidate regions, a fixed-size feature map is extracted and transformed through two fully connected layers. A softmax classifier is used to classify the feature map into three categories: cracks, bubbles, and impurities. At the same time, a corresponding category label and precise location are assigned to each defective region. Based on the location information and type label of the defective region, it is marked on the original image, bounding boxes are drawn, and type labels are added to generate a labeled image containing defect location and type information.

5. The method according to claim 1, wherein, Based on the detected defect information from each layer, a three-dimensional spatial model is constructed, mapping defects from different layers onto the same coordinate system, and analyzing the spatial distribution of defects. Identifying cross-layer defect regions includes: based on the detected defect information from each layer, converting two-dimensional defect coordinates into a three-dimensional spatial point set using a direct coordinate mapping method; calculating the depth value of the defect in the Z-axis direction by setting the thickness parameter of each material layer, generating three-dimensional point cloud data containing XYZ coordinates; performing voxelization processing on the three-dimensional point cloud data, dividing the continuous space into discrete cubic meshes; calculating the voxel index of each point cloud data, and recording the attribute information of the points in that voxel, obtaining a discretized three-dimensional spatial model; marking each non-empty voxel in the discretized three-dimensional spatial model, traversing adjacent voxels, and identifying potential cross-layer defect regions if cross-layer connected voxels are found; grouping the marked voxels, aggregating voxels with similar spatial locations and belonging to the same defect type into continuous defect regions, post-processing the clustering results, and merging defect clusters of the same type with small distances; calculating the volume, surface area, and number of layers crossed for each defect region based on the clustered defect regions, and finally generating a three-dimensional defect distribution report.

6. The method according to claim 1, wherein, The process involves performing a detailed local scan of the identified cross-flaw areas to acquire micron-level three-dimensional structural data and reconstruct the fine morphological features of the cross-flaws. This includes: acquiring three-dimensional structural data of the cross-flaw areas using a laser scanning microscope, where the three-dimensional structural data includes information about the flaw areas and their adjacent boundaries; preprocessing the acquired three-dimensional structural data, including removing noise to obtain denoised three-dimensional structural data; smoothing the point cloud using the moving least squares method on the denoised three-dimensional structural data to obtain smoothed point cloud data; fitting a three-dimensional surface to the smoothed point cloud data to generate a mesh model representing the fine morphology of the cross-flaws; extracting local geometric features from the mesh model and calculating and determining local geometric feature parameters; and using a morphological skeleton extraction algorithm to identify crack directions and precisely locate the flaw edges based on the local geometric feature parameters, thereby obtaining three-dimensional flaw model data containing fine morphological features.

7. The method according to claim 1, wherein, The process involves simulating and calculating the impact of defects on LED performance, including optical, thermal, and mechanical properties, based on the reconstructed cross-defect structure and various pre-stored material performance parameters in the material database. This includes: acquiring the reconstructed cross-defect structure; retrieving corresponding material performance parameters from the material database based on the cross-defect structure; processing the properties of the transition region using a material interpolation function; setting the minimum grid cell size to construct a refined LED structure model containing defects; simulating the propagation path of photons in the LED structure model using a Monte Carlo ray tracing algorithm, setting the number of photons and wavelength range; calculating the scattering coefficient based on the phosphor's reabsorption and scattering effects to obtain data on the impact of defects on optical performance; discretizing the steady-state heat conduction equation using the finite volume method, including solving the discrete equation set using the conjugate gradient method; obtaining information on thermal performance changes caused by defects by calculating thermal resistance and junction temperature; performing stress-strain analysis using linear elasticity theory; implementing adaptive mesh refinement, including setting refinement criteria based on stress gradient; if the stress gradient is greater than a preset threshold, reducing the mesh size and iterating the refinement process until the stress difference between two consecutive iterations is less than the preset threshold, thus obtaining prediction results for stress concentration and potential failure locations.

8. The method according to claim 1, wherein, The process of determining the location, size, type, and impact on performance of defects to quantify and score the overall quality of LED packaging, resulting in a final inspection result and quality grade determination, includes: acquiring the spatial coordinates, volume, surface area, and impact on optical, thermal, and mechanical properties of defects; constructing a multi-dimensional feature vector, which includes the x, y, and z coordinates, volume, surface area, and impact on optical, thermal, and mechanical properties of defects; establishing a scoring model using a random forest algorithm based on the multi-dimensional feature vector, where the location, size, type, and impact on various performance aspects of defects are used as evaluation indicators; calculating a weighted score for defects using the scoring model; dividing each evaluation indicator into several levels based on the weighted score, calculating membership degrees using a Gaussian membership function, and constructing a fuzzy relation matrix; performing fuzzy synthesis on the fuzzy relation matrix using a weighted average method to obtain the overall quality score of the LED packaging; if the overall quality score is less than a preset threshold, the quality grade of the LED packaging is determined to be unqualified; if the overall quality score is greater than or equal to the preset threshold, the quality grade of the LED packaging is determined to be qualified.

Images