LED packaging material flaw detection method

Abstract

The invention provides an LED packaging material defect detection method, which comprises the following steps: performing image enhancement processing on each layer of separated image to obtain each layer of clear material image; carrying out local fine scanning on the identified cross defect area to obtain micron-sized three-dimensional structure data, and reconstructing fine morphology characteristics of the cross defect; according to the reconstructed cross flaw structure, combining various material performance parameters pre-stored in a material database to simulate and calculate the influence of flaws on the LED performance, including optical, thermal and mechanical properties; and judging the position, the size, the type and the influence degree on the performance of the flaws, and carrying out quantitative scoring on the overall quality of LED packaging to obtain a final detection result and quality grade judgment.

Description

Technical Field

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

[0002] In the field of defect detection of LED packaging materials, there is a severe challenge of multi-level cross-problems. Due to the complexity of the LED packaging structure, defects may exist at different interfaces and levels such as between the substrate and the chip, between the phosphor layer and the epoxy resin layer, forming cross-defects. The detection difficulty of such defects is far beyond the norm. It is difficult to fully capture all potential defects with a single-level detection method. Traditional detection methods are limited by their independent evaluation of specific layers or materials, and it is difficult to effectively identify cross-level composite defects. For example, the mutual influence of substrate microcracks and poor chip bonding, as well as the interaction between uneven phosphor distribution and bubble defects in the epoxy resin layer, not only seriously affect the optical performance and reliability of the LED, but may also lead to misjudgment and missed detection results. In addition, the physical properties and optical attributes of different material layers are significantly different, making it extremely difficult to adopt unified detection standards and methods. The transparent epoxy resin layer and the opaque phosphor layer require completely different imaging technologies and analysis algorithms. To make it more complicated, some defects may not be significant on a single level, but once multi-layer interactions are involved, they will significantly affect the overall performance of the LED. In summary, how to design a comprehensive detection method that can take into account multiple levels and multiple interfaces to achieve accurate identification and evaluation of cross-defects in LED packaging materials requires not only breaking through the traditional single detection idea, but also in-depth consideration of the interaction mechanism between materials and developing new multimodal sensing technology and data fusion algorithms to cope with this technical difficulty. It is a technical problem that needs to be solved urgently. Summary of the invention

[0003] The present invention provides a method for detecting defects in LED packaging materials, which mainly includes:

[0004] The multi-spectral imaging method is used to scan the LED package to obtain image data at different wavelengths. According to the spectral response characteristics of each layer of material in different bands, the independent image information of the substrate, chip, phosphor layer and epoxy resin layer is separated.

[0005] For each layer of separated images, image enhancement processing is performed to obtain a clear image of each layer of material;

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

[0007] According to the detected defect information of each layer, a three-dimensional space model is constructed to map the defects at different levels into the same coordinate system, analyze the distribution relationship of the defects in space, and identify the defect areas that cross layers.

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

[0009] Based on the reconstructed cross-defect structure and the 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 of the defects on the performance 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 solution provided by the embodiment of the present invention may have the following beneficial effects:

[0012] The present invention discloses a defect detection method for LED packaging materials, which overcomes the limitations of traditional single detection methods, realizes the comprehensive capture and accurate evaluation of potential defects in the packaging structure, and effectively solves the problem of cross defect detection. The advantage is that it can take into account the multi-level characteristics of the packaging material, and perform an all-round scan of the LED package through multi-spectral imaging technology to obtain high-resolution image data of different material layers at specific wavelengths. With the help of advanced image enhancement processing, the details of each layer of material can be clearly displayed, ensuring the accuracy of subsequent defect detection. Subsequently, the deep learning model is applied to feature extraction and defect recognition, and its powerful pattern recognition ability can identify and mark various defects, such as microcracks, bubbles and impurities, 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 present the spatial distribution of defects, and pay special attention to those important cross defect areas that only appear in multi-layer interactions. The local fine scanning and three-dimensional reconstruction of these areas revealed the fine morphology of the cross-defects. Combined with the material performance database, the comprehensive impact of the defects on the optical, thermal and mechanical properties of the LED was simulated and calculated, providing a quantitative evaluation basis for the severity of the defects. Finally, through the comprehensive consideration of the location, size, type of defects and their impact on performance, the quantitative scoring and grade determination of the overall quality of LED packaging was achieved, which greatly improved the accuracy and efficiency of detection, reduced the errors caused by human factors, reduced production costs, and significantly improved the quality consistency and reliability of LED products. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1The present invention is a flow chart of a method for detecting defects in LED packaging materials.

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

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

[0016] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0017] like Figure 1 -3. In this embodiment, a method for detecting defects in LED packaging materials may specifically include:

[0018] Step S101, using a multi-spectral imaging method to scan the LED package, obtain image data at different wavelengths, and separate independent image information of the substrate, chip, phosphor layer and epoxy resin layer according to the spectral response characteristics of each layer of material in different bands.

[0019] A multi-spectral imaging device is used to perform an all-round scan of the LED package, and key bands are selected for scanning according to a preset wavelength range to obtain multiple sets of image data. The multi-spectral image data is subjected to principal component analysis to identify the key bands that can best distinguish different material layers. The spectral response characteristics of each layer of the LED package material in different bands are extracted, and a spectral response curve database containing characteristic curves of the substrate, chip, phosphor layer and epoxy resin layer is established. The image set is pixel-level classified according to the spectral response curve database, and the spectral response of each pixel point is compared with the characteristic curve in the database through the support vector machine algorithm to generate a preliminary separation image of the substrate, chip, phosphor layer and epoxy resin layer. 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.

[0020] Specifically, a multispectral imaging device is used to perform a 360-degree full-scale scan of the LED package. According to the preset wavelength range from 300nm to 1000nm, key bands are selected for scanning to obtain multiple sets of image data. The image data is preprocessed to remove noise and distortion, and a corrected multispectral image set is obtained. The multispectral image data is subjected to principal component analysis to identify the key bands that can best distinguish different material layers, providing a basis for subsequent feature extraction. The spectral response characteristics of each layer of the LED package material in different bands are extracted from the corrected multispectral image set, and a spectral response curve database is established, including the characteristic curves of the substrate, chip, phosphor layer and epoxy resin layer. The spectral response of each pixel in the image set is compared with the characteristic curve in the database through the support vector machine algorithm. The image set is pixel-level classified according to the spectral matching results, 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 is performed using open and closed operations to remove noise and smooth the boundaries, and then the watershed algorithm is applied to perform regional segmentation to improve the accuracy of the separated image. The optimized separated image is reconstructed into three dimensions. Combined with the structural information and material properties of the LED package, the structured light projection method is used for depth estimation. The three-dimensional structure is reconstructed by projecting a known pattern and analyzing its deformation. At the same time, combined with photometric stereo vision technology, images under different lighting conditions are used to improve the reconstruction accuracy, and a three-dimensional model of the internal structure of the LED package is generated, which includes the spatial position and morphological information of the substrate, chip, phosphor layer and epoxy resin layer.

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

[0022] Step S102: performing image enhancement processing on the separated layers of images to obtain clear images of the layers of materials.

[0023] The separated layers of images are subjected to image enhancement processing to obtain a first image, including processing the original image according to a preset block size and a contrast limiting threshold; the first image is processed by a bilateral filter to obtain a denoised second image, the bilateral filter calculates weights according to preset spatial domain standard deviation, range standard deviation and window size; the second image is subjected to edge enhancement by a nonlinear sharpening method to obtain an edge-enhanced third image, the nonlinear sharpening method is achieved by Gaussian blurring and edge image generation; the third image is subjected to resolution enhancement by using a pre-trained ESPCN algorithm to obtain a fourth image with enhanced resolution; and a clear material layer image is reconstructed based on the fourth image.

[0024] Specifically, for each layer of separated images, the contrast-limited adaptive histogram equalization algorithm is used to enhance the contrast of the image. The image is divided into 8x8 blocks, and each block is histogram equalized. The contrast limit threshold is set to 0.01 to prevent excessive noise enhancement, and an image with improved contrast is obtained. The enhanced image is denoised using a bilateral filter, and the spatial domain standard deviation is set to 3, the value domain standard deviation is set to 0.1, and the window size is 5x5. The weight is calculated based on the spatial distance and gray value difference between pixels. The noise is eliminated while maintaining the edge information to obtain a smooth and edge-retained image. The unsharpmasking method is applied to the denoised image for nonlinear sharpening. First, Gaussian blur is used with a standard deviation of 1.5 to obtain a blurred image. Then, the original image is subtracted from the blurred image to obtain an edge image. The edge image is multiplied by a factor of 0.7 and added back to the original image to improve the edge clarity and detail performance of the image, and an edge-enhanced image is generated. The ESPCN (Efficient Sub-Pixel Convolutional Neural Network) algorithm is used to improve the resolution of edge-enhanced images. The training data set contains high-resolution and low-resolution image pairs of various LED packaging materials. The network structure contains three convolutional layers. The last layer uses sub-pixel convolution upsampling, and the magnification factor is set to 2. By learning the mapping relationship from low-resolution images to high-resolution images, high-resolution images of each layer of materials are reconstructed to obtain a final image with higher clarity. The contrast-limited adaptive histogram equalization algorithm first divides the 1024x1024 pixel image into 8x8 small blocks, each with 128x128 pixels. The grayscale histogram is calculated for each small block, and the contrast limit threshold is set to 0.01. The part exceeding the threshold is cropped and evenly distributed to the entire histogram. Histogram equalization is performed on each small block, and adjacent blocks are merged by bilinear interpolation to obtain a contrast-enhanced image. The bilateral filter sets the spatial domain standard deviation to 3 pixels, the value domain standard deviation to 0.1, and the window size to 5x5 pixels to filter the image. For the substrate layer, due to the rough texture, the spatial domain standard deviation is adjusted to 5 pixels; for the chip layer, due to the rich details, the value domain standard deviation is reduced to 0.05. The unsharpmasking technology uses a Gaussian blur kernel (standard deviation of 1.5 pixels) to obtain a blurred image. The original image is subtracted from the blurred image to obtain an edge image, which is multiplied by a factor of 0.7 and added back to the original image. For the phosphor layer, due to the strong granularity, the factor is adjusted to 0.5 to avoid over-sharpening. The ESPCN network structure contains 3 convolutional layers. The first layer uses 64 5x5 convolution kernels, the second layer uses 32 3x3 convolution kernels, the third layer uses 9 3x3 convolution kernels, and the last layer uses sub-pixel convolution upsampling with an amplification factor of 2.The training data set contains 10,000 pairs of high-resolution and low-resolution LED packaging material images. The low-resolution image size is 128x128 pixels, and the high-resolution image size is 256x256 pixels. The network training uses the Adam optimizer, the learning rate is set to 0.001, the batch size is 64, and the training is 100 epochs. After processing, a 256x256 pixel high-resolution image with a 2x improvement in clarity, significantly reduced noise, and sharp edges is obtained.

[0025] Step S103, using a deep learning model to perform feature extraction and defect detection on the enhanced layers of the image, identifying potential defect areas from the image, including cracks, bubbles, and impurities, and marking the location and type information of the defects.

[0026] The ResNet-50 network is used to extract features from each layer of the enhanced image, the first four convolution blocks of the network are retained, and the fully connected layer of the network is removed to obtain a feature map containing rich semantic information. A region proposal network is used to slide a convolution kernel of a preset size on the feature map to generate a multi-scale and multi-ratio anchor box, screen and optimize the candidate area, and obtain the location information and bounding box coordinates of the defect area. For the screened candidate area, a fixed-size feature map is extracted and feature transformation is performed through two fully connected layers. The softmax classifier is used to perform multi-category classification on the feature map, and the defects are divided into three categories: cracks, bubbles and impurities. At the same time, the corresponding category label and precise location are assigned to each defect area. According to the location information and type label of the defect area, it is marked on the original image, the bounding box is drawn and the type label is added to generate a marked image containing the defect location and type information.

[0027] Specifically, the ResNet-50 network is used to extract features from each layer of enhanced images, the first four convolution blocks are retained, the fully connected layer is removed, and each layer of LED packaging material images is input into the network for processing. The low-level, medium-level and high-level features of the image are extracted through multi-layer convolution and pooling operations, and a 1 / 16 size feature map containing rich semantic information is obtained. The feature map is processed using a region proposal network, and a 3x3 convolution kernel is slid on the feature map to generate multi-scale and multi-ratio anchor boxes. The candidate regions are screened and optimized through binary classification and bounding box regression. Finally, the non-maximum suppression algorithm is applied to retain the top 300 candidate regions with the highest scores, and the location information and bounding box coordinates of the defect area are obtained. For the screened candidate regions, the ROIAlign method is used to extract a fixed-size 7x7 feature map, and then the feature transformation is performed through two fully connected layers. The softmax classifier is used to perform multi-category classification of the defect type, and the defects are divided into three categories: cracks, bubbles and impurities. At the same time, the smooth L1 loss is used to fine-tune the bounding box, and the corresponding category label and precise position are assigned to each defect area. According to the location information and type label of the defect area, the original image is marked, the bounding box is drawn and the type label is added. Different colors are used to represent different types of defects. Red represents cracks, green represents bubbles, and blue represents impurities. The type label and confidence score are added in the upper left corner of the bounding box. The label format is "type: score". A labeled image containing the defect location and type information is generated for subsequent quality assessment and defect analysis. The ResNet-50 network processes 512x512 pixel LED packaging material images, and after 49 convolutional layers and pooling layers, a 32x32x2048 feature map is output. The region proposal network uses a 3x3 convolution kernel on the feature map with a step size of 1 to generate 9 anchor boxes of different scales and ratios, generating a total of 32x32x9=9216 initial candidate regions. Through bounding box regression and binary classification, the top 1000 candidate regions are screened out, and then non-maximum suppression is performed, with an IoU threshold of 0.7, and 300 final candidate regions are retained. The ROIAlign method uniformly transforms candidate regions of different sizes into a 7x7 feature map to maintain the continuity of spatial information. Two fully connected layers, with 1024 and 512 neurons respectively, transform the features. The softmax classifier outputs 4 category probabilities (background, cracks, bubbles, impurities), and 4 regressors fine-tune the bounding box coordinates. The classification threshold is set to 0.5. If it exceeds the threshold, it is judged as a defect of the corresponding category. On the original 512x512 image, a 2-pixel wide bounding box is used to mark the defects. Cracks are marked in red (255, 0, 0), bubbles are marked in green (0, 255, 0), and impurities are marked in blue (0, 0, 255). In the upper left corner of the bounding box, "Type: 0.XX" is marked in 10-point white font, where 0.XX is the confidence score.The entire processing process runs on the GPU, and the processing time for a single image is about 0.2 seconds, which can achieve real-time detection. The average precision (mAP) of 1,000 test images reached 0.85, among which crack detection accuracy was the highest (0.90), followed by bubbles (0.85), and impurities were the lowest (0.80).

[0028] Step S104, constructing a three-dimensional space model based on the detected defect information of each layer, mapping the defects at different levels into the same coordinate system, analyzing the distribution relationship of the defects in space, and identifying the defect areas that cross layers.

[0029] According to the detected defect information of each layer, the direct coordinate mapping method is used to convert the two-dimensional defect coordinates into a three-dimensional space point set. By setting the thickness parameters of each layer of material, the depth value of the defect in the Z-axis direction is calculated to generate three-dimensional point cloud data containing XYZ coordinates. The three-dimensional point cloud data is voxelized, and the continuous space is divided into discrete cubic grids. For each point cloud data, the voxel index to which it belongs is calculated, and the attribute information of the point is recorded in the voxel to obtain a discretized three-dimensional space model. In the discretized three-dimensional space model, each non-empty voxel is marked, and the adjacent voxels are traversed. If a voxel connected across layers is found, it is judged as a potential cross-layer defect area. The marked voxels are grouped, and the voxels with similar spatial positions and belonging to the same defect type are aggregated into a continuous defect area. The clustering results are post-processed to merge the same type of defect clusters with small distances. According to the defect area after clustering, the volume, surface area and number of layers spanned by each defect area are calculated, and finally a three-dimensional defect distribution report is generated.

[0030] Specifically, according to the defect information of each layer detected, the direct coordinate mapping method is used to convert the two-dimensional defect coordinates into a three-dimensional space point set. For each two-dimensional defect coordinate (x, y), the z coordinate is calculated according to the index i of the layer and the preset layer thickness d, z = i * d, and the depth value of the defect in the Z-axis direction is calculated by setting the thickness parameter of each layer of material, and the three-dimensional point cloud data containing XYZ coordinates is generated. The point cloud density is consistent with the original image resolution. The generated three-dimensional point cloud data is voxelized, and the continuous space is divided into discrete cubic grids. The voxel size is set to 0.1mmx0.1mmx0.1mm, and the octree data structure is used for space division to improve query efficiency. For each point cloud data, the voxel index to which it belongs 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 space model. In the voxelized three-dimensional spatial model, each non-empty voxel is marked, and the material layer and defect type to which the voxel belongs are recorded. The adjacent voxels are traversed using the 26-connectivity rule. For each non-empty voxel, check whether the upper and lower adjacent voxels belong to different material layers. If a cross-layer connected voxel is found, it is marked as a potential cross-layer defect area, and the continuous defect voxels in the Z-axis direction are identified to determine whether there is a cross-layer defect area. The marked voxels are grouped using a density-based spatial clustering algorithm, and the neighborhood radius is set to 0.5mm and the minimum number of points is 5. Voxels with similar spatial positions and belonging to the same defect type are aggregated into continuous defect areas. The clustering results are post-processed to merge the same type of defect clusters with a distance of less than 1mm, calculate the volume, surface area and number of layers spanned by each defect area, and generate a detailed three-dimensional defect distribution report. For a 512x512 pixel LED package image, each pixel represents 0.01mm, and a total of 3 layers of materials (substrate, chip, phosphor layer) are detected, with thickness of each layer of 0.2mm, 0.1mm and 0.3mm respectively. The direct coordinate mapping method generates a 3D point cloud of 262,144 points, including 1,532 defect points. The voxelization process divides the space into 800x800x60 voxels, each voxel size is 0.1mmx0.1mmx0.1mm. The maximum depth of the octree data structure is set to 6, and the leaf node contains a maximum of 8 points. When traversing the 26-connectivity rule, each non-empty voxel is checked for its surrounding 26 adjacent voxels, and 73 potential cross-layer defect areas are found. The density-based spatial clustering algorithm sets the neighborhood radius ε = 0.5mm, the minimum number of points MinPts = 5, and identifies 124 independent defect clusters. Post-processing merged 28 pairs of defect clusters of the same type with a distance less than 1 mm, and finally obtained 96 defect areas, including 43 cracks (35 single-layer and 8 cross-layer), 31 bubbles (27 single-layer and 4 cross-layer), and 22 impurities (20 single-layer and 2 cross-layer).The largest defect area is a crack that spans the substrate and chip layers, with a volume of 0.015mm3 and a surface area of ​​0.42mm3. The generated 3D defect distribution report contains the type, location, volume, surface area, number of layers spanned, and distribution ratio in each layer of each defect area, which can provide accurate spatial information for quality control and process improvement.

[0031] Step S105 , performing a local fine scan on the identified cross defect area, acquiring micron-level three-dimensional structure data, and reconstructing fine morphological features of the cross defect.

[0032] A laser scanning microscope is used to obtain three-dimensional structural data of a cross defect area, wherein the three-dimensional structural data includes information about the defect area and its adjacent boundaries; the collected three-dimensional structural data is preprocessed, including removing noise points from the three-dimensional structural data to obtain denoised three-dimensional structural data; a moving least square method is used to perform point cloud smoothing on the denoised three-dimensional structural data to obtain smoothed point cloud data; three-dimensional surface fitting is performed on the smoothed point cloud data to generate a grid model representing the fine morphology of the cross defect; local geometric features are extracted on the grid model, and local geometric feature parameters are calculated and determined; based on the local geometric feature parameters, a morphological skeleton extraction algorithm is used to identify the direction of the crack and then accurately locate the defect edge to obtain three-dimensional defect model data containing fine morphological features.

[0033] Specifically, according to the identified coordinates of the cross defect area, a confocal laser scanning microscope is used for local fine scanning. An adaptive scanning strategy is adopted, combined with raster scanning and contour tracking scanning, and the scanning range is dynamically adjusted to ensure that the cross defect area and its surrounding 1mm boundary are fully covered. The scanning step is set to 0.5μm to obtain high-density three-dimensional structural data. The collected high-density three-dimensional structural data are preprocessed, and the statistical outlier filtering algorithm is used to remove noise points. Then, the moving least squares method is used to smooth the point cloud, retaining the edge features while reducing the measurement noise, and obtaining clear microstructure point cloud data. The processed point cloud data is fitted with a three-dimensional surface using the Poisson surface reconstruction algorithm. The octree depth is set to 10, and the number of sampling points per square unit is not less than 100. At the same time, adaptive density control is introduced to increase the reconstruction accuracy in the defect edge area, and a grid model representing the fine morphology of the cross defect is generated. The grid resolution is set to 0.1μm to ensure accurate description of the tiny structural features. Local geometric features are extracted from the reconstructed three-dimensional mesh model using principal curvature analysis and shape index calculation. The curvature, normal vector and depth information are calculated. The morphological skeleton extraction algorithm is used to identify the direction of the crack. The level set method is used to accurately locate the edge of the defect and generate three-dimensional defect model data containing fine morphological features.

[0034] The confocal laser scanning microscope uses a 488nm laser source, with a scanning range of 2mm x 2mm and a Z-axis resolution of 0.1μm. The adaptive scanning strategy first performs a coarse scan at a spacing of 100μm to identify the defect boundary, and then performs a fine scan at a spacing of 10μm in the boundary area, and finally achieves a scanning step of 0.5μm in the core area. The acquired raw point cloud data contains approximately 16 million points. The statistical outlier removal algorithm sets the number of neighborhood points to 50 and the standard deviation multiple to 2.5, removing a total of approximately 2% of noise points. The moving least squares method uses quadratic polynomial fitting, with a search radius of 5μm and 5 iterations. The number of point clouds after smoothing is reduced to 15 million. In Poisson surface reconstruction, the octree depth of 10 corresponds to a minimum voxel size of approximately 0.05μm, the sampling density is set to 400 points / μm3, and the adaptive density control is increased to 800 points in the edge area.

[0035] / μm3. The reconstructed 3D mesh model contains about 5 million triangular facets. The principal curvature analysis uses a spherical neighborhood with a radius of 3μm to calculate the Gaussian curvature and the mean curvature. The shape index threshold is set to 0.7 to distinguish concave and convex features. The morphological skeleton extraction uses a 3D refinement algorithm, and the minimum branch length is set to 1μm. The level set method uses a fast marching algorithm with an evolution time step of 0.1 and 100 iterations to accurately locate the defect edge. The final generated 3D defect model data contains geometric meshes, curvature maps, shape index maps, and skeleton structures, with a total data volume of about 200MB.

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

[0037] Obtain a reconstructed cross-defect structure; call the corresponding material performance parameters from the material database according to the cross-defect structure; use the material interpolation function to process the properties of the transition area; set the minimum unit size of the grid, and build a fine LED structure model containing defects. Use the Monte Carlo ray tracing algorithm to simulate the propagation path of photons in the LED structure model, set the number of photons and the wavelength range; calculate the scattering coefficient based on the reabsorption and scattering effects of the phosphor, and obtain the data on the impact of defects on optical performance. Use the finite volume method to discretize the steady-state heat conduction equation, including using the conjugate gradient method to solve the discrete equation group; obtain the thermal performance change information caused by the defect by calculating the thermal resistance and junction temperature. Use the linear elastic mechanics theory to perform stress-strain analysis; implement adaptive mesh refinement, including setting the refinement criterion based on the stress gradient; if the stress gradient is greater than the preset threshold, reduce the mesh size, iterate the refinement process until the stress difference between two consecutive iterations is less than the preset threshold, and obtain the prediction results of stress concentration and potential failure locations.

[0038] Specifically, the reconstructed three-dimensional structure of cross-defects is imported into COMSOL Multiphysics software. According to the location and shape of the defects, the corresponding material performance parameters, including refractive index, thermal conductivity, elastic modulus, etc., are called from the material database. For defects that span multiple materials, the material interpolation function is used to process the properties of the transition area, and the minimum grid unit size is set to 0.1μm to ensure accurate capture of defect details and build a fine LED structure model containing defects. For optical performance, the Monte Carlo ray tracing algorithm is used to simulate the propagation path of photons in the LED structure, the number of photons is set to 10^6, the wavelength range is 300-800nm, and 20 equally spaced sampling points are divided. Considering the reabsorption and scattering effects of the phosphor, the Mie scattering theory is used to calculate the scattering coefficient, the light extraction efficiency and luminous uniformity are calculated, and the impact data of defects on optical performance are obtained. For thermal performance, the finite volume method is used to discretize the steady-state heat conduction equation, and the conjugate gradient method is used to solve the discrete equations. The convergence criterion is set to a residual less than 10^-6. Considering the temperature-dependent material properties, such as the change of thermal conductivity with temperature, the thermal resistance and junction temperature are calculated to obtain the information on the change of thermal performance caused by defects. For mechanical properties, stress-strain analysis is performed using the theory of linear elastic mechanics, and h-adaptive mesh refinement is implemented. The refinement criterion is set based on the stress gradient. In the area where the stress gradient is greater than the 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%, which improves the calculation accuracy of the defect area and obtains the prediction results of stress concentration and potential failure locations.

[0039] The three-dimensional structure of cross-defects imported into COMSOL Multiphysics software is a volume of 5μmx5μmx10μm, including cracks across the chip and phosphor layer. The material interpolation function uses the hyperbolic tangent function, and the width of the transition zone is set to 0.5μm. The mesh generation uses tetrahedral units with a minimum size of 0.1μm, and a total of about 1 million mesh units are generated. In the optical simulation, the Monte Carlo ray tracing sets the number of photons to 10^6, and the wavelength range of 300-800nm ​​is divided into 20 sampling points. The Mie scattering theory calculates that the scattering coefficient of the phosphor is 1000cm^-1 at 450nm. The light extraction efficiency dropped from the original 85% to 82%, and the luminous uniformity decreased by 5%. In the thermal analysis, about 1.5 million equations were obtained after discretization by the finite volume method, and 50 iterations were required for the conjugate gradient method to converge. The rate of change of thermal conductivity with temperature is set to -0.2% / K. The calculation shows that the thermal resistance increases by 0.5K / W and the junction temperature increases by 3°C. In the mechanical performance analysis, the stress gradient threshold is set to 5MPa / μm by h-adaptive mesh refinement. After three iterations of refinement, the number of mesh elements increases to 1.5 million. The final calculation shows that the maximum stress concentration area is located at the crack tip, and the stress value reaches 350MPa, which is 40% higher than the flawless case. The predicted potential failure position is within 0.5μm of the crack tip. The entire multi-physics field coupling analysis process runs on a 64-core high-performance computer for 4 hours.

[0040] Step S107, determining the position, size, type and impact of the defect on the performance, quantifying and scoring the overall quality of the LED package, and obtaining the final test result and quality grade determination.

[0041] The spatial coordinates, volume, surface area of ​​the defect and the degree of influence on the optical, thermal and mechanical properties are obtained, and a multidimensional feature vector is constructed, wherein the multidimensional feature vector includes the x, y, z coordinates, volume, surface area and the degree of influence on the optical, thermal and mechanical properties of the defect; a scoring model is established based on the multidimensional feature vector using a random forest algorithm, wherein the scoring model uses the location, size, type of the defect and the degree of influence on each performance as evaluation indicators; a weighted score of the defect is calculated using the scoring model; each evaluation indicator is divided into several levels based on the weighted score of the defect, and the membership is calculated using a Gaussian membership function to construct a fuzzy relationship matrix; a fuzzy synthesis operation is performed on the fuzzy relationship matrix using a weighted average method to obtain an overall quality score of the LED package; if the overall quality score is less than a preset threshold, the quality grade 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 grade of the LED package is determined to be qualified.

[0042] Specifically, according to the three-dimensional reconstruction results and performance simulation data, the geometric features of the defects such as spatial coordinates, volume, surface area, etc. are extracted, and the influence of optical, thermal and mechanical properties is combined to construct a 10-dimensional feature vector, including the x, y, z coordinates, volume, surface area, and the influence of optical, thermal and mechanical properties (one dimension each) of the defects. The simulation results are normalized to the range of 0-1 to characterize the comprehensive attributes of each defect. The scoring model is established using the random forest algorithm, and the location, size, type of defects and their influence on various properties are used as evaluation indicators. The model is trained using historical data, and the importance of each indicator is automatically learned as a weight. The number of trees is set to 100 and the maximum depth is 10 to balance the complexity and generalization ability of the model, and the weighted score of each defect is calculated. Using the fuzzy comprehensive evaluation method, each evaluation index is divided into five levels (very good, good, average, poor, and very poor) based on the defect weighted score. The membership is calculated using the Gaussian membership function, and the fuzzy relationship matrix is ​​constructed. The fuzzy synthesis operation is performed by the weighted average method to obtain the overall quality score of the LED package, and the score result is mapped to the predefined quality grade interval. According to the quality score and the preset threshold, the final inspection result and quality grade of the LED package are judged. The preset threshold is determined based on the statistical analysis of a large amount of historical data. The k-means clustering algorithm is used to divide the historical data into five categories corresponding to five quality grades. The cluster center is used as the representative value of each grade to generate a test report containing defect details, performance impact analysis, and quality grade determination, which provides a basis for subsequent quality control and process optimization. For a specific LED package sample, five defects are obtained after three-dimensional reconstruction and performance simulation. A 10-dimensional feature vector was constructed, in which the spatial coordinate range was (0-5000μm, 0-5000μm, 0-1000μm), the volume range was 0-1000μm3, the surface area range was 0-5000μm3, and the performance impact was normalized to 0-1. The random forest model was trained based on 10,000 historical samples, 100 decision trees, and a maximum depth of 10. The feature importance was 30% for position, 25% for size, 20% for optical influence, 15% for thermal influence, and 10% for mechanical influence. In the fuzzy evaluation, the mean of the Gaussian membership function was set to the central value of each level, and the standard deviation was 1 / 6 of the interval length. In the weighted average method, the weight was consistent with the feature importance obtained by the random forest. The k-means clustering was based on 50,000 historical samples, iterated 100 times, and obtained 5 cluster centers as quality level thresholds, which were 95 points, 85 points, 75 points, 65 points, and 55 points respectively. The LED package sample was evaluated for 5 defects, with scores of 92, 88,

[0043] 79, 76, 71, the overall score is 83.5, corresponding to the "good" grade. The inspection report contains the 10-dimensional feature value of each defect, the individual score and the overall score, as well as the final quality grade determination. The entire evaluation process takes 0.5 seconds to complete under GPU acceleration.

[0044] What is disclosed above is only a preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. A person skilled in the art can understand that all or part of the processes of the above embodiments and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A method for detecting defects in LED packaging materials, characterized in that: The method comprises: The multi-spectral imaging method is used to scan the LED package to obtain image data at different wavelengths. According to the spectral response characteristics of each layer of material in different bands, the independent image information of the substrate, chip, phosphor layer and epoxy resin layer is separated. For each layer of separated images, image enhancement processing is performed to obtain a clear image of each layer of material; For the enhanced images of each layer, a deep learning model is used to perform feature extraction and defect detection, identify potential defect areas from the image, including cracks, bubbles, and impurities, and mark the location and type information of the defects; According to the detected defect information of each layer, a three-dimensional space model is constructed to map the defects at different levels into the same coordinate system, analyze the distribution relationship of the defects in space, and identify the defect areas that cross layers. For the identified cross defect area, a local fine scan is performed to obtain micron-level three-dimensional structural data and reconstruct fine morphological features of the cross defect; Based on the reconstructed cross-defect structure and the 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 impact of the defects on the performance are determined, and the overall quality of the LED package is quantitatively scored to obtain the final test results and quality grade determination.

2. The method according to claim 1, wherein: The multi-spectral imaging method is used to scan the LED package to obtain image data at different wavelengths, and separate independent image information of the substrate, chip, phosphor layer and epoxy resin layer according to the spectral response characteristics of each layer of material in different bands, including: Use multi-spectral imaging equipment to perform all-round scanning of LED packages, select key bands for scanning according to the preset wavelength range, and obtain multiple sets of image data; Performing principal component analysis on the multispectral image data to identify key bands that can best distinguish different material layers; Extract the spectral response characteristics of each layer of LED packaging materials in different bands, and establish a spectral response curve database including the characteristic curves of substrate, chip, phosphor layer and epoxy resin layer; Performing pixel-level classification on the image set according to the spectral response curve database, and comparing the spectral response of each pixel with the characteristic curve in the database by using a support vector machine algorithm to generate preliminary separation images of the substrate, chip, phosphor layer, and epoxy resin layer; The preliminary separated images are reconstructed in three dimensions and estimated using a structured light projection method to generate an internal three-dimensional structure of the LED package including spatial position and morphological information of a substrate, a chip, a phosphor layer, and an epoxy resin layer.

3. The method according to claim 1, wherein: The image enhancement process is performed on the separated layers of images to obtain clear images of the layers of materials, including: Performing image enhancement processing on the separated layers of images to obtain a first image, including processing the original image according to a preset block size and contrast limiting threshold; Processing the first image using a bilateral filter to obtain a denoised second image, wherein the bilateral filter calculates weights according to a preset spatial domain standard deviation, a range standard deviation and a window size; Applying a nonlinear sharpening method to the second image to perform edge enhancement to obtain an edge-enhanced third image, wherein the nonlinear sharpening method is implemented by Gaussian blurring and edge image generation; Using a pre-trained ESPCN algorithm to improve the resolution of the third image, to obtain a fourth image with improved resolution; Based on the fourth image, a sharpened material layer image is reconstructed.

4. The method according to claim 1, wherein: The enhanced images of each layer are subjected to feature extraction and defect detection using a deep learning model, and potential defect areas, including cracks, bubbles, and impurities, are identified from the images, and the location and type information of the defects are marked, including: Using a ResNet-50 network to extract features from each layer of the enhanced image, retaining the first four convolutional blocks of the network, removing the fully connected layer of the network, and obtaining a feature map containing rich semantic information; Use a region proposal network to slide a convolution kernel of a preset size on the feature map, generate multi-scale and multi-ratio anchor frames, screen and optimize candidate regions, and obtain location information and bounding box coordinates of the defect region; For the selected candidate regions, extract a feature map of a fixed size and perform feature transformation through two fully connected layers; Using a softmax classifier to perform multi-category classification on the feature map, the defects are classified into three categories: cracks, bubbles, and impurities, and a corresponding category label and precise position are assigned to each defect area; According to the position information and type label of the defect area, the original image is marked, a boundary box is drawn and a type label is added to generate a marked image containing the defect position and type information.

5. The method according to claim 1, wherein: The method constructs a three-dimensional spatial model based on the detected defect information of each layer, maps the defects at different layers to the same coordinate system, analyzes the distribution relationship of the defects in space, and identifies the defect areas that cross layers, including: According to the detected defect information of each layer, the direct coordinate mapping method is used to convert the two-dimensional defect coordinates into a three-dimensional space point set. By setting the thickness parameters of each layer of material, the depth value of the defect in the Z-axis direction is calculated to generate three-dimensional point cloud data containing XYZ coordinates; voxel processing is performed on the three-dimensional point cloud data to divide the continuous space into discrete cubic grids, and for each point cloud data, the voxel index to which it belongs is calculated, and the attribute information of the point is recorded in the voxel to obtain a discretized three-dimensional space model; In the discretized three-dimensional space model, each non-empty voxel is marked, and the adjacent voxels are traversed. If a voxel connected across layers is found, it is determined to be a potential cross-layer defect area; The marked voxels are grouped, and the voxels with similar spatial positions and belonging to the same defect type are aggregated into continuous defect areas. The clustering results are post-processed to merge the defect clusters of the same type with small distances. According to the defect areas after clustering processing, the volume, surface area and number of layers spanned by each defect area are calculated, and finally a three-dimensional defect distribution report is generated.

6. The method according to claim 1, wherein: The method of performing a local fine scan on the identified cross defect area to obtain micrometer-level three-dimensional structure data and reconstruct fine morphological features of the cross defect includes: Using a laser scanning microscope to obtain three-dimensional structural data of the cross defect area, the three-dimensional structural data includes information about the defect area and its adjacent boundaries; Preprocessing the collected three-dimensional structure data, including removing noise points from the three-dimensional structure data to obtain denoised three-dimensional structure data; For the denoised three-dimensional structure data, performing point cloud smoothing processing using a moving least squares method to obtain smoothed point cloud data; Performing three-dimensional surface fitting on the smoothed point cloud data to generate a grid model representing the fine morphology of the cross defect; Extracting local geometric features on the grid model, and calculating and determining local geometric feature parameters; According to the local geometric feature parameters, a morphological skeleton extraction algorithm is applied to identify the crack direction and then accurately locate the defect edge to obtain three-dimensional defect model data containing fine morphological features.

7. The method according to claim 1, wherein: The method of simulating and calculating the influence of defects on LED performance, including optical, thermal and mechanical properties, based on the reconstructed cross defect structure and in combination with various material performance parameters pre-stored in the material database includes: obtaining a reconstructed cross-defect structure; Retrieving corresponding material performance parameters from a material database according to the cross defect structure; Use material interpolation functions to handle the properties of transition areas; Set the minimum cell size of the grid to build a detailed LED structure model including defects; Using the Monte Carlo ray tracing algorithm to simulate the propagation path of photons in the LED structure model, setting the number of photons and the wavelength range; Based on the reabsorption and scattering effects of phosphors, the scattering coefficient is calculated to obtain the impact of defects on optical performance; Discretization of the steady-state heat conduction equations using the finite volume method, including solving the discretized system of equations using the conjugate gradient method; Obtain information on thermal performance changes caused by defects by calculating thermal resistance and junction temperature; Use linear elasticity theory to perform stress-strain analysis; Implement adaptive mesh refinement, including setting refinement criteria based on stress gradients; If the stress gradient is greater than a preset threshold, the mesh size is reduced, and the refinement process is iterated until the stress difference between two consecutive iterations is less than the preset threshold, thereby obtaining the prediction results of stress concentration and potential failure locations.

8. The method according to claim 1, wherein: The determination of the position, size, type and impact of the defects on the performance, quantitative scoring of the overall quality of the LED package, and obtaining the final test results and quality grade determination include: Obtaining the spatial coordinates, volume, surface area of ​​the defect and the degree of influence on the optical, thermal and mechanical properties, and constructing a multidimensional feature vector, wherein the multidimensional feature vector includes the x, y and z coordinates, volume, surface area, and the degree of influence on the optical, thermal and mechanical properties of the defect; A scoring model is established based on the multidimensional feature vector using a random forest algorithm, wherein the scoring model uses the location, size and type of the defect and the degree of influence on various performances as evaluation indicators; Calculating a defect weighted score using the scoring model; According to the defect weighted score, each evaluation index is divided into several levels, the membership is calculated using a Gaussian membership function, and a fuzzy relationship matrix is ​​constructed; Performing fuzzy synthesis operation on the fuzzy relationship matrix by 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 grade of the LED package is determined to be unqualified; If the overall quality score is greater than or equal to a preset threshold, the quality grade of the LED package is determined to be qualified.

Images