An electronic component defect detection method and system

Abstract

The present application relates to the technical field of electronic component detection, and specifically relates to an electronic component defect detection method and system, comprising: obtaining a visual detection image of the internal structure of an electronic component to be detected; and performing multi-scale feature enhancement preprocessing on the visual detection image of the electronic component to be detected.The present application constructs a dual-feature recognition model that fuses space-time correlation, synchronously extracts the category, location information of the internal defect ontology of the electronic component, and the texture features and degree level of the surrounding microstructure deformation area, and accurately captures the defect concomitant deformation information that is ignored by traditional detection methods.On this basis, according to the formation process rules and spatial geometric relationship of the defect ontology and the microstructure deformation area, the time sequence and spatial position correlation features are quantitatively extracted, and a defect-deformation correlation feature matrix is constructed, so that the mathematical representation of the internal formation correlation between the two is realized in the detection process.

Description

Technical Field

[0001] The present invention relates to the technical field of electronic component detection, and specifically provides a method and system for detecting defects in electronic components. Background Art

[0002] Defect detection of electronic components is a core process in the electronic manufacturing industry chain to ensure product reliability and usage safety. With the continuous increase in the packaging density of electronic components, their internal structures have become increasingly refined, and the formation and manifestation forms of defects have also become more complex. When various defects form inside components, they do not exist only in the form of independent defect bodies. The generation of some defects is accompanied by physical deformations of the surrounding microstructures. These deformations are directly related to the formation of the defect bodies and will directly affect the actual usage performance of the components. Currently, all automatic detection methods in the industry are based on obtaining the detection images of components, completing defect recognition through deep learning models, then quantitatively analyzing the defect-related features, and finally completing the qualification determination of components. Such methods have gradually replaced manual inspection and become the mainstream detection method in the industry.

[0003] The invention patent with the publication number CN112950560A: Method, Device and System for Detecting Defects in Electronic Components proposes to complete the training of a defect detection model by constructing a good product image library and a defect image library, identify the defect types of electronic components by the trained model, then match the corresponding area or length calculation method according to different defect types to complete the quantification of the defect bodies, and finally complete the qualification determination of components by comparing the quantification results with preset thresholds. At the same time, a continuous update mechanism for the model is designed to enable the model to adapt to more diverse defect image features. This solution effectively improves the efficiency and accuracy of single defect detection.

[0004] Existing various automatic detection solutions for electronic components only focus on the identification and quantitative analysis of defect bodies, and do not pay attention to the deformations of the surrounding microstructures accompanying the formation of some defects. Although these deformations are not defined as independent defect types, they are directly related to the formation of defect bodies and will become the direct inducement for the subsequent expansion of defects. The influence range and manifestation degree of the deformations will also directly amplify the negative impact of the defect bodies on the performance of components. Since the existing solutions do not capture the correlation information between the defect bodies and the deformations of the surrounding microstructures, and only make qualification determinations based on the quantification results of the defect bodies, they cannot accurately judge the actual impact degree of the defects on the components, resulting in some components with potential performance risks being determined as qualified. Moreover, the industry has not proposed effective technical means for detecting the correlation between defect bodies and the deformations of the surrounding microstructures, nor established a quantitative correlation determination logic between the two, making it impossible to accurately judge the actual impact degree of component defects. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for detecting defects in electronic components. This solves the problem that existing detection methods only focus on the defect itself and fail to capture the correlation information between the defect and the surrounding microstructure deformation, making it impossible to accurately determine the actual impact of the defect and easily misidentify components with potential performance risks.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for detecting defects in electronic components, comprising the following steps:

[0007] Acquire visual inspection images of the internal structure of the electronic components to be inspected;

[0008] Multi-scale feature enhancement preprocessing is performed on the visualized inspection image of the electronic components to be inspected. The basic features of the multi-scale image are extracted by Gaussian pyramid downsampling, and the upsampled image is reconstructed by Laplacian pyramid. The preprocessed image is output by combining the gray-level difference enhancement processing of neighboring pixels.

[0009] The preprocessed image is input into a dual-feature recognition model that integrates spatiotemporal correlation, and the category and location information of the defect body, as well as the texture features and degree level information of the surrounding microstructural deformation region are extracted simultaneously.

[0010] Based on the formation process rules and spatial geometric relationship between the defect body and the microstructure deformation region, spatiotemporal correlation features are extracted, and the spatiotemporal correlation features are quantized and encoded to construct a defect-deformation correlation feature matrix;

[0011] Dual-region quantization analysis based on the characteristics of the defect body and the microstructure deformation region is performed on the defect body and the microstructure deformation region respectively to obtain the quantization results of the defect body and the deformation region.

[0012] Based on the defect-deformation correlation feature matrix, the dynamic influence coefficient between the defect body and the microstructure deformation is generated by gradient descent iteration.

[0013] By applying multi-level fusion judgment logic and combining the defect body quantification result, the deformation region quantification result, and the dynamic influence coefficient, the performance risk level and qualification conclusion of the electronic component under test are determined.

[0014] Furthermore, the specific steps of the multi-scale feature enhancement preprocessing are as follows:

[0015] The original visualized detection image is continuously downsampled through a Gaussian pyramid to obtain image pyramid levels at different resolutions to characterize the basic structural features.

[0016] In the process of Laplacian pyramid reconstruction, interpolation is performed on images of adjacent levels to extract high-frequency detail information, and interpolation algorithms are used to complete the reconstruction of the upsampled image.

[0017] Within a local window of the reconstructed image, the grayscale difference between the center pixel and its neighboring pixels is calculated. Based on the grayscale difference, the gradient features of the defect body edge and the subtle texture features of the surrounding microstructure deformation are linearly compensated to suppress random noise signals in the image background.

[0018] Furthermore, the dual-feature recognition model that integrates spatiotemporal correlation includes a basic feature extraction network, a defect recognition branch, and a deformation feature extraction branch;

[0019] The basic feature extraction network is used to extract the low-level shared features of the preprocessed image;

[0020] The defect identification branch locates and classifies the defect entity based on the underlying shared features;

[0021] The deformation feature extraction branch runs in parallel with the defect recognition branch. It captures the surrounding area affected by defect formation through an attention mechanism, identifies the location of the microstructure deformation area, extracts texture feature vectors from the microstructure deformation area, and classifies the deformation performance of the microstructure deformation area into different degree levels.

[0022] Furthermore, the specific process for extracting spatiotemporal correlation features includes:

[0023] In the time dimension, the process rules corresponding to the defect body are retrieved, and the sequential characteristics of defect generation and deformation occurrence in the manufacturing process are matched.

[0024] In the spatial dimension, the Euclidean distance characteristics from the edge of the defect body to the geometric center of the microstructure deformation region are calculated, the positional overlap characteristics of the defect body and the microstructure deformation region on the two-dimensional projection plane are analyzed, and the spatial envelopment ratio characteristics of the microstructure deformation region on the defect body are calculated.

[0025] Furthermore, the method for constructing the defect-deformation correlation feature matrix is ​​as follows:

[0026] The extracted temporal features, spatial distance features, spatial overlap features, and spatial wrapping features are mapped to a unified quantization space.

[0027] The spatiotemporal correlation features are converted into feature vectors using a numerical encoding method, and the feature vectors are arranged according to a preset topological structure to form a dedicated correlation feature matrix that represents the correlation between the defect body and the microstructure deformation region.

[0028] Furthermore, the specific implementation method of the dual-region quantitative analysis is as follows:

[0029] For the defect entity, a corresponding geometric shape fitting algorithm is selected according to its category to calculate the area, perimeter, and aspect ratio features of the defect entity, and generate the quantification result of the defect entity.

[0030] For the microstructure deformation region, the influence range involved in the deformation is calculated using the region contour closure fitting technique, and the texture pixel distribution density within the deformation region is statistically analyzed. Combined with the identified degree level, the quantification result of the deformation region is generated by weighted summarization.

[0031] Furthermore, the step of generating the dynamic influence coefficient includes:

[0032] Obtain the initial weight coefficients corresponding to the defect ontology, and establish an objective function for calculating the influence coefficients based on the feature vector distribution in the defect-deformation correlation feature matrix;

[0033] The gradient descent algorithm is used to perform multiple iterations for optimization. The coefficients are dynamically corrected based on the influence range and degree of deformation in the quantification results of the deformation region. When the objective function converges to the preset range, the dynamic influence coefficient representing the weight of the deformation on the defect performance is output.

[0034] Furthermore, the multi-level fusion judgment logic includes a basic judgment layer, a correlation quantization layer, and a performance risk layer:

[0035] The basic judgment layer compares the quantification result of the defect ontology with a preset single defect threshold to obtain a preliminary qualification criterion.

[0036] The correlation quantization layer performs multiplication and accumulation operations on the defect ontology quantization result, the deformation region quantization result, and the dynamic influence coefficient to calculate the defect-deformation fusion quantization value.

[0037] The performance risk layer matches a preset risk matrix based on the size range of the defect-deformation fusion quantization value to determine the performance risk level of the component, and outputs a final judgment conclusion in conjunction with the preliminary qualification criteria.

[0038] Furthermore, the method also includes a two-dimensional model update step:

[0039] The acquired original image, preprocessed image, dual-feature recognition result, associated feature matrix, quantitative analysis result, dynamic influence coefficient and judgment conclusion are integrated into a standardized data sample and added to the training set of the dual-feature recognition model for incremental learning.

[0040] Simultaneously, the actual performance test data of the components after testing is obtained, the matching degree between the test judgment result and the performance test data is calculated, and the judgment threshold and weight parameters in the multi-level fusion judgment logic are dynamically adjusted according to the deviation of the matching degree.

[0041] The present invention also provides an electronic component defect detection system, comprising:

[0042] The image acquisition module is used to acquire visual inspection images of the internal structure of the electronic components to be inspected;

[0043] The multi-scale preprocessing module, connected to the image acquisition module, is used to perform multi-scale feature enhancement preprocessing on the visualized detection image of the electronic components to be detected. It extracts the basic features of the multi-scale image through Gaussian pyramid downsampling, reconstructs the upsampled image using Laplacian pyramid, and outputs the preprocessed image by combining neighborhood pixel gray-level difference enhancement processing.

[0044] The dual-feature recognition module is used to input the preprocessed image into a dual-feature recognition model that integrates spatiotemporal correlation, and simultaneously extract the category and location information of the defect body, as well as the texture features and degree level information of the surrounding microstructural deformation area;

[0045] The associated feature extraction module is used to extract spatiotemporal associated features based on the formation process rules and spatial geometric relationship between the defect body and the microstructure deformation region, and to quantize and encode the spatiotemporal associated features to construct a defect-deformation associated feature matrix.

[0046] The dual-region quantization module is used to perform dual-region quantization analysis based on the characteristics of the defect body and the microstructure deformation region respectively, to obtain the quantization results of the defect body and the deformation region.

[0047] The dynamic coefficient generation module is used to generate the dynamic influence coefficient between the defect body and the microstructure deformation by gradient descent iterative calculation based on the defect-deformation correlation feature matrix.

[0048] The fusion judgment module is used to apply multi-level fusion judgment logic, combine the defect body quantification result, the deformation region quantification result and the dynamic influence coefficient, and determine the performance risk level and qualification conclusion of the electronic component to be tested.

[0049] The dual-dimensional model update module is used to receive data samples generated during system operation and simultaneously optimize and adjust the model parameters of the dual-feature recognition module and the judgment parameters of the fusion judgment module.

[0050] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0051] This invention constructs a dual-feature recognition model that integrates spatiotemporal correlation, simultaneously extracting the category and location information of the internal defect of electronic components, as well as the texture features and degree levels of the surrounding microstructural deformation regions. This accurately captures the defect-associated deformation information that is ignored by traditional detection methods. Based on this, according to the formation process rules and spatial geometric relationships of the defect and the microstructural deformation regions, its temporal and spatial location correlation features are quantitatively extracted, and a defect-deformation correlation feature matrix is ​​constructed. This enables a mathematical representation of the inherent formation correlation between the two during the detection process.

[0052] Furthermore, by performing independent quantitative analysis on both the defect itself and the deformed region, and combining this with the dynamic influence coefficient generated by the gradient descent algorithm, the quantitative results of both are fused with the correlation weights at multiple levels for a multi-level assessment. Finally, the performance risk level and qualification of the component are accurately evaluated based on the defect-deformation fused quantitative value. This solves the core problem of existing technologies that fail to assess the correlation between the defect itself and surrounding deformations due to isolated defect analysis, leading to inaccurate judgments of the actual performance risk of components. By introducing correlation analysis and fusion judgment mechanisms, the ability to identify and intercept components with potential performance risks is effectively improved. This represents a non-obvious technological leap from isolated defect detection to comprehensive correlation impact assessment, ensuring a high degree of consistency between the detection conclusions and the actual reliability of the component. Attached Figure Description

[0053] Figure 1 This is an overall flowchart of the electronic component defect detection method of the present invention;

[0054] Figure 2 This is a flowchart illustrating the specific process of multi-scale feature enhancement preprocessing in this invention.

[0055] Figure 3 This is a schematic diagram of the dual-feature recognition model that integrates spatiotemporal correlation according to the present invention;

[0056] Figure 4 This is a flowchart illustrating the multi-level fusion determination logic of the present invention.

[0057] Figure 5 This is a schematic diagram of the module architecture of the electronic component defect detection system of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0059] Example 1

[0060] Please see Figure 1 This invention provides a method for detecting defects in electronic components, comprising the following steps:

[0061] Acquire visual inspection images of the internal structure of the electronic components to be inspected;

[0062] Multi-scale feature enhancement preprocessing is performed on the visualized inspection image of the electronic components to be inspected. The basic features of the multi-scale image are extracted by Gaussian pyramid downsampling, and the upsampled image is reconstructed by Laplacian pyramid. The preprocessed image is output by combining the gray-level difference enhancement processing of neighboring pixels.

[0063] The preprocessed image is input into a dual-feature recognition model that integrates spatiotemporal correlation, and the category and location information of the defect body, as well as the texture features and degree level information of the surrounding microstructural deformation region are extracted simultaneously.

[0064] Based on the formation process rules and spatial geometric relationship between the defect body and the microstructure deformation region, spatiotemporal correlation features are extracted, and the spatiotemporal correlation features are quantized and encoded to construct a defect-deformation correlation feature matrix;

[0065] Dual-region quantization analysis based on the characteristics of the defect body and the microstructure deformation region is performed on the defect body and the microstructure deformation region respectively to obtain the quantization results of the defect body and the deformation region.

[0066] Based on the defect-deformation correlation feature matrix, the dynamic influence coefficient between the defect body and the microstructure deformation is generated by gradient descent iteration.

[0067] By applying multi-level fusion judgment logic and combining the defect body quantification result, the deformation region quantification result, and the dynamic influence coefficient, the performance risk level and qualification conclusion of the electronic component under test are determined.

[0068] First, visualize the internal structure of the electronic components to be inspected. Mainstream visualization inspection methods such as industrial CT and high-magnification microscopy can be used to obtain visualized inspection images of the internal structure of the components with appropriate resolution. For example, for conventional electronic components such as chip ceramic capacitors and miniature resistors, an imaging result with a resolution of 2048×2048 pixels can be selected. This resolution can clearly present the fine structure and potential defects inside the components, providing basic image data for subsequent inspection steps.

[0069] In one embodiment, see Figure 2 After image acquisition, multi-scale feature enhancement preprocessing is performed on the visualized detection image. The specific steps of multi-scale feature enhancement preprocessing are as follows:

[0070] The original visualized detection image is continuously downsampled through a Gaussian pyramid to obtain image pyramid levels at different resolutions to characterize the basic structural features.

[0071] In the process of Laplacian pyramid reconstruction, interpolation is performed on images of adjacent levels to extract high-frequency detail information, and interpolation algorithms are used to complete the reconstruction of the upsampled image.

[0072] Within a local window of the reconstructed image, the grayscale difference between the center pixel and its neighboring pixels is calculated. Based on the grayscale difference, the gradient features of the defect body edge and the subtle texture features of the surrounding microstructure deformation are linearly compensated to suppress random noise signals in the image background.

[0073] Specifically, in the Gaussian pyramid downsampling process, Gaussian filtering and downsampling operations are performed sequentially on the original visualized detection image. The Gaussian filtering uses a two-dimensional Gaussian kernel for convolution calculation, and the formula for the two-dimensional Gaussian kernel is: ,in The standard deviation of the Gaussian kernel, for example, taking This value is the conventional optimal value for image filtering, which can effectively smooth the image while preserving structural features. The downsampling adopts a 2x downsampling method, that is, every other pixel in the filtered image is selected to complete 4 levels of downsampling in sequence. For example, based on the original image of 2048×2048, image pyramid levels of resolution of 1024×1024, 512×512, 256×256, and 128×128 are obtained in sequence. The multi-level images can comprehensively characterize the basic structural features inside the components, taking into account both macroscopic structure and microscopic details.

[0074] The Laplacian pyramid reconstruction is based on the images of each level of the Gaussian pyramid mentioned above. Interpolation is performed on images of adjacent levels to extract high-frequency detail information. The interpolation formula is as follows: ,in For the first of Laplace's pyramids High-frequency detail images of the layers, For the first Gaussian pyramid Layer image, For the upsampling operation, a bilinear interpolation algorithm is used. For convolution operations, Using the same Gaussian kernel as during downsampling, the interpolation algorithm ensures pixel continuity in the upsampled image, avoiding step-like distortion. After high-frequency detail extraction, the overall upsampled image is reconstructed by layer-by-layer superposition of the high-frequency detail image and the upsampled image, restoring the resolution to the same level as the original image. Based on the reconstructed image, neighborhood pixel gray-level difference enhancement is performed. A 3×3 local window is selected to traverse the reconstructed image pixel by pixel. For example, considering the fineness of the microstructure of electronic components, a 3×3 window can accurately capture edge gradient features and avoid noise interference. The mean gray-level difference between the center pixel and its 8 neighboring pixels within the window is calculated. ,in The grayscale value of the center pixel. For the first The grayscale values ​​of a neighboring pixel are used to linearly compensate for the gradient features of the defect body edge and the subtle texture features of the surrounding microstructural deformation based on the average of these grayscale differences. The compensation formula is as follows: ,in The grayscale value of the center pixel after compensation. For example, a compensation coefficient is taken based on the overall grayscale distribution of the image. For areas with grayscale values ​​below 128, take the larger value. For areas with grayscale values ​​higher than 128, take the smaller value. This linear compensation method can enhance the feature contrast of the defect and deformation areas, while effectively suppressing random noise signals in the image background. Because the pixel grayscale difference in the background area is extremely small, the grayscale value does not change significantly after compensation, while the pixel grayscale difference in the defect and deformation areas is large, and the features are more prominent after compensation. Finally, the preprocessed image with multi-scale feature enhancement is output.

[0075] In one embodiment, see Figure 3 After obtaining the preprocessed image, it is input into a dual-feature recognition model that integrates spatiotemporal correlation for feature extraction. The specific structure and operating logic of this model are as follows:

[0076] The dual-feature recognition model that integrates spatiotemporal correlation includes a basic feature extraction network, a defect recognition branch, and a deformation feature extraction branch;

[0077] The basic feature extraction network is used to extract the low-level shared features of the preprocessed image;

[0078] The defect identification branch locates and classifies the defect entity based on the underlying shared features;

[0079] The deformation feature extraction branch runs in parallel with the defect recognition branch. It captures the surrounding area affected by defect formation through an attention mechanism, identifies the location of the microstructure deformation area, extracts texture feature vectors from the microstructure deformation area, and classifies the deformation performance of the microstructure deformation area into different degree levels.

[0080] Specifically, the basic feature extraction network uses the ResNet50 network model. For example, considering the feature extraction requirements of the detected image, ResNet50 can effectively extract the low-level shared features and avoid the gradient vanishing problem. This network performs convolution and pooling operations on the preprocessed image through the residual block structure, and sequentially extracts the low-level texture, mid-level contour, high-level semantic and other shared features of the image. The output feature map provides the basis for the feature extraction of the subsequent two branches.

[0081] The defect identification branch and the deformation feature extraction branch operate in parallel based on the output feature map of the basic feature extraction network. The defect identification branch adopts the Faster R-CNN target detection algorithm, which generates candidate regions of the defect ontology through the region proposal network based on the underlying shared features. Then, feature extraction and classification regression are performed on the candidate regions to achieve accurate localization and category identification of the defect ontology. The categories of identifiable defect ontology include common internal defects of electronic components such as cracks, pinholes, cavities, and impurities. The localization results are output in the form of pixel coordinates, which clearly indicates the specific location of the defect ontology in the image.

[0082] The deformation feature extraction branch integrates the CBAM attention mechanism, which assigns weights to the feature map from both channel and spatial dimensions. This mechanism can accurately capture the surrounding microstructural deformation regions affected by defect formation and effectively eliminate interference from irrelevant regions. After identifying the location of the surrounding microstructural deformation regions, a convolutional neural network is used to extract features from these regions, generating a 256-dimensional texture feature vector. Simultaneously, based on the degree and extent of texture distortion and influence of the surrounding microstructural deformation regions, the microstructural deformation is classified into three levels: slight, moderate, and severe. The classification is based on the texture pixel distortion rate of the surrounding microstructural deformation regions. For example, a distortion rate <10% indicates slight deformation, 10% ≤ distortion rate <30% indicates moderate deformation, and a distortion rate ≥30% indicates severe deformation. This classification quantifies the degree of microstructural deformation, providing a basis for subsequent analysis. This dual-feature recognition model can simultaneously extract the category and location information of the defect itself, as well as the texture features and degree level information of the surrounding microstructural deformation area. Compared with the traditional single defect recognition model, it can simultaneously capture the correlation features between defects and deformations, laying the foundation for subsequent correlation analysis.

[0083] After extracting the defect and deformation-related features, spatiotemporal correlation features are extracted based on the formation process rules and spatial geometric relationships of the defect body and the surrounding microstructure deformation regions. Specifically, the implementation is as follows:

[0084] In the time dimension, the process rules corresponding to the defect body are retrieved, and the sequential characteristics of defect generation and deformation occurrence in the manufacturing process are matched.

[0085] In the spatial dimension, the Euclidean distance characteristics from the edge of the defect body to the geometric center point of the surrounding microstructure deformation region are calculated, the positional overlap characteristics of the defect body and the surrounding microstructure deformation region on the two-dimensional projection plane are analyzed, and the spatial enclosure ratio characteristics of the microstructure deformation region to the defect body are calculated.

[0086] Specifically, the extraction of temporal features in the time dimension needs to be combined with the manufacturing process library of electronic components. This library pre-stores the complete manufacturing process of various electronic components and the types of defects and deformations that are likely to occur in each process. It retrieves the type of the component to be tested and the category of the identified defect body, matches the corresponding manufacturing process, and determines the order in which defects occur and microstructure deformations occur. For example, the internal crack defect of a chip multilayer ceramic capacitor occurs in the sintering process, while the warping deformation of the surrounding dielectric layer occurs in the cooling process after sintering. The crack defect occurs before the warping deformation. This temporal feature is quantified using a 0-1 encoding method. If the defect occurs before the deformation, the temporal feature value is 1; if the deformation occurs before the defect, the temporal feature value is 0; if both occur simultaneously, the temporal feature value is 0.5.

[0087] Spatial dimension feature extraction is based on the pixel coordinates of the defect body and the surrounding microstructural deformation region output by the dual-feature recognition model. First, the Euclidean distance feature is calculated, using the formula: ,in The pixel coordinates of the geometric center point of the defect body edge. The pixel coordinates of the geometric center point of the surrounding microstructural deformation region. The Euclidean distance between the two points is used to characterize the spatial distance relationship between the defect and the deformation region. Secondly, the overlapping characteristics are analyzed, and the overlapping area of ​​the defect body and the deformation region on the two-dimensional projection plane is calculated. Total area of ​​the defect body ratio , This is the positional overlap feature value, which can characterize the spatial overlap relationship between the two; finally, the spatial envelopment ratio feature is calculated to extract the edge perimeter of the defect body. and the perimeter of the defect body edge enclosed by the microstructural deformation region. The ratio of the two This refers to the spatial envelopment ratio feature value, which characterizes the spatial envelopment relationship between the deformed region and the defect body. By extracting features in both time and space dimensions, the intrinsic relationship between the defect body and the microstructural deformation region can be comprehensively captured, overcoming the limitation of traditional detection techniques that only focus on a single defect feature.

[0088] In one embodiment, after extracting spatiotemporal correlation features, they are quantized and encoded to construct a defect-deformation correlation feature matrix. The method for constructing the defect-deformation correlation feature matrix is ​​as follows:

[0089] The extracted temporal features, spatial distance features, spatial overlap features, and spatial wrapping features are mapped to a unified quantization space.

[0090] The spatiotemporal correlation features are converted into feature vectors using a numerical encoding method, and the feature vectors are arranged according to a preset topological structure to form a dedicated correlation feature matrix that represents the correlation between the defect body and the microstructure deformation region.

[0091] Specifically, the time-series feature values, spatial distance feature values, spatial overlap feature values, and spatial wrapping feature values ​​are first normalized and mapped to a unified quantization space of [0,1]. The normalization formula is as follows: ,in These are the original eigenvalues. These are the normalized eigenvalues. This is the minimum value of this type of feature. To find the maximum value of this type of feature, normalization eliminates the influence of differences in the dimensions of different features, ensuring the accuracy of subsequent matrix operations. The four normalized eigenvalues ​​are then converted into 4-dimensional feature vectors using numerical encoding. ,in These are the normalized time series feature values. These are the normalized distance feature values. These are the normalized overlapping eigenvalues. These are the normalized package feature values.

[0092] For electronic components undergoing batch inspection, the 4D feature vectors corresponding to each component are arranged sequentially according to the component's inspection number as the topological structure, forming a defect-deformation correlation feature matrix. ,in The matrix represents the number of components to be tested in a batch. The matrix consists of 4 rows and 4 columns. Each row corresponds to a spatiotemporal correlation feature of defect-deformation of a component under test, and each column corresponds to a normalized spatiotemporal correlation feature. This matrix can intuitively and quantitatively characterize the relationship between the defect body and the microstructure deformation region, providing a quantitative basis for the subsequent calculation of dynamic influence coefficients.

[0093] In one embodiment, after constructing the defect-deformation correlation feature matrix, dual-region quantization analysis is performed on both the defect body and the microstructure deformation region. The specific implementation of the dual-region quantization analysis is as follows:

[0094] For the defect body, a corresponding geometric shape fitting algorithm is selected according to its category to calculate the area, perimeter, and aspect ratio features of the defect body, and generate the quantization result of the defect body; for the microstructure deformation region, the influence range involved in the deformation is calculated using the region contour closure fitting technique, and the texture pixel distribution density in the deformation region is statistically analyzed. Combined with the identified degree level, the quantization result of the deformation region is generated by weighted summarization.

[0095] Specifically, for the quantitative analysis of the defect itself, a corresponding geometric shape fitting algorithm is selected based on the defect category. For example, near-circular defects such as pinholes and cavities are fitted using the least squares method, near-linear defects such as cracks and scratches are fitted using polygons, and irregular defects such as impurities and missing corners are fitted using contours. The geometric features of the defect are extracted using the corresponding fitting algorithm, and the actual area of ​​the defect is calculated. ,perimeter and aspect ratio The aspect ratio is the ratio of the length to the width of the circumscribed rectangle of the defect body. These three geometric feature values ​​are normalized to form a three-dimensional feature vector. This feature vector is the result of defect ontology quantification, which can comprehensively quantify the geometric features of defect ontology and reflect the actual size and shape features of defect.

[0096] For the quantitative analysis of microstructural deformation regions, the Canny edge detection algorithm is first used to extract the contours of the deformation regions. Then, the extracted contours are fitted using a region contour closure fitting technique, and the area of ​​the circumcircle of the fitted closed region is calculated. This value represents the area affected by the deformation; secondly, the number of texture distortion pixels within the deformation area is counted. Calculate the texture pixel distribution density This value characterizes the degree of texture distortion within the deformed area. Finally, combined with the deformation level identified by the dual-feature recognition model, corresponding weight coefficients are assigned to different levels. For example, based on the impact of deformation on defect performance, slight deformation is given a higher weight. Moderate deformation weight Severe deformation weight Convert the degree of deformation into a quantitative value. (Slight = 1, Moderate = 2, Severe = 3), multiply by the corresponding weighting coefficient to obtain the level weighted value. The range of deformation influence Texture pixel distribution density Rank weighted value After normalization, a three-dimensional feature vector is formed. This feature vector represents the quantification result of the deformation region, which can comprehensively quantify the influence range, texture features, and degree level of the microstructural deformation region. Dual-region quantification analysis can accurately quantify and characterize both the defect body and the microstructural deformation region, making up for the shortcomings of traditional detection techniques that only perform simple quantification of the defect body, and providing a more comprehensive quantitative basis for subsequent fusion judgment.

[0097] In one embodiment, after obtaining the defect-deformation correlation feature matrix, the defect ontology quantization result, and the deformation region quantization result, the dynamic influence coefficient between the defect ontology and microstructure deformation is generated through gradient descent iteration based on the defect-deformation correlation feature matrix. The specific steps are as follows:

[0098] Obtain the initial weight coefficients corresponding to the defect ontology, and establish an objective function for calculating the influence coefficients based on the feature vector distribution in the defect-deformation correlation feature matrix;

[0099] The gradient descent algorithm is used to perform multiple iterations for optimization. The coefficients are dynamically corrected based on the influence range and degree of deformation in the quantification results of the deformation region. When the objective function converges to the preset range, the dynamic influence coefficient representing the weight of the deformation on the defect performance is output.

[0100] Specifically, the initial weight coefficients are first obtained based on the category of the defect ontology. The initial weighting coefficients are preset based on process experience in the field of electronic component manufacturing, such as the initial weighting coefficients for minor defects like pinholes and voids. Initial weighting coefficients for severe defects such as cracks and fractures The initial coefficients are set based on the inherent risk level of defects to component performance, ensuring their rationality. Based on the eigenvector distribution in the defect-deformation correlation feature matrix, and combining the defect body quantization results and deformation region quantization results, an objective function for calculating the influence coefficients is established. The objective function is: ,in The dynamic influence coefficient to be solved is... For the first Defect quantification results of individual components For the first Quantification results of the deformation region of individual components It is a 2-norm. For example, take the regularization coefficient. This value can effectively avoid overfitting during the iteration process and ensure the generalization ability of the coefficient solution.

[0101] After establishing the objective function, the gradient descent algorithm is used for multiple iterations to find the optimal function. The iterative formula for gradient descent is as follows: ,in For the first The coefficient values ​​of the next iteration. For the first The coefficient values ​​of the next iteration. For example, take the learning rate. This value can ensure the speed of iterative convergence while avoiding oscillations. For the objective function in The gradient at the location. During the iteration process, the influence range is determined based on the quantization results of the deformation region. Weighted by the degree of deformation The coefficients are dynamically adjusted; the larger the influence range and the higher the degree of deformation, the greater the adjustment magnitude, ensuring that the dynamic influence coefficients accurately reflect the actual weight of the deformation region on the performance of the defect itself. A preset convergence threshold is set. When the difference between the objective function values ​​of two adjacent iterations is < When the objective function is considered to have converged, the output is... This refers to the dynamic influence coefficient between the defect itself and the microstructure deformation. The generation method of this dynamic influence coefficient can achieve adaptive solution of the coefficient based on the correlation characteristics between defects and deformation, which solves the limitation of traditional detection technology using fixed weights, and enables the coefficient to accurately match the actual correlation between different defects and deformations.

[0102] In one embodiment, see Figure 4 After generating the dynamic influence coefficient, a multi-level fusion judgment logic is applied, combining the defect body quantification results, deformation region quantification results, and dynamic influence coefficient, to determine the performance risk level and pass / fail conclusion of the electronic component under test. The multi-level fusion judgment logic is as follows:

[0103] The multi-level fusion judgment logic includes a basic judgment layer, a correlation quantization layer, and a performance risk layer:

[0104] The basic judgment layer compares the quantification result of the defect ontology with a preset single defect threshold to obtain a preliminary qualification criterion.

[0105] The correlation quantization layer performs multiplication and accumulation operations on the defect ontology quantization result, the deformation region quantization result, and the dynamic influence coefficient to calculate the defect-deformation fusion quantization value.

[0106] The performance risk layer matches a preset risk matrix based on the size range of the defect-deformation fusion quantization value to determine the performance risk level of the component, and outputs a final judgment conclusion in conjunction with the preliminary qualification criteria.

[0107] Specifically, the implementation of the basic judgment layer is based on a preset single defect threshold. This threshold is set based on industry standards and usage requirements for electronic components, and is a three-dimensional threshold vector corresponding to the defect quantification result. ,in The threshold for area features, The threshold for the perimeter feature. The threshold value for aspect ratio features is used to quantize the defect ontology. Each feature value is compared with the corresponding value of the threshold vector one by one. If all feature values ​​do not exceed the corresponding threshold, it is initially judged as qualified. If any feature value exceeds the corresponding threshold, it is initially judged as unqualified. This preliminary qualification criterion can make a basic judgment based on the inherent features of the defect ontology, ensuring the basic nature of the judgment.

[0108] The correlation quantization layer performs multiplication and summation operations on the defect ontology quantization results, deformation region quantization results, and dynamic influence coefficients to calculate the defect-deformation fusion quantization value. The calculation formula is as follows: ,in For dot product operation, To achieve a fusion quantification value, this formula organically integrates the quantification characteristics of the defect entity, the quantification characteristics of the deformation region, and the dynamic influence weights of the two. This allows for a comprehensive quantification of the combined impact of defects and deformation on component performance. Compared to traditional quantification methods based solely on the defect entity, this fusion quantification value better reflects the actual performance risk of the component.

[0109] The performance risk layer pre-defines a risk matrix corresponding to the fusion quantization value. This risk matrix is ​​divided into different numerical ranges based on the performance requirements of electronic components, for example, a fusion quantization value < For low risk, < Medium risk. For high-risk situations, a qualified threshold is set. , The value is determined by combining industry standards and usage scenarios, such as for consumer-grade electronic components. The value is lower than that of industrial-grade electronic components, and the calculated fusion quantization value is used. The system matches the numerical range and pass / fail threshold of the risk matrix to determine the performance risk level of the component. Then, it combines this with the preliminary pass / fail criteria obtained from the basic judgment layer to output the final pass / fail conclusion. If the preliminary judgment is pass / fail and the fused quantization value is < If it is initially deemed qualified but the fusion quantification value is not met, it will be deemed qualified. Regardless of the magnitude of the fusion quantification value, if a preliminary judgment is made that the value is initially deemed unqualified, it will ultimately be deemed unqualified. This multi-level fusion judgment logic can achieve layer-by-layer judgment from basic to related, and from single to comprehensive, ensuring the accuracy and comprehensiveness of the judgment results.

[0110] In one embodiment, the method further includes a two-dimensional model update step:

[0111] The acquired original image, preprocessed image, dual-feature recognition result, associated feature matrix, quantitative analysis result, dynamic influence coefficient and judgment conclusion are integrated into a standardized data sample and added to the training set of the dual-feature recognition model for incremental learning.

[0112] Simultaneously, the actual performance test data of the components after testing is obtained, the matching degree between the test judgment result and the performance test data is calculated, and the judgment threshold and weight parameters in the multi-level fusion judgment logic are dynamically adjusted according to the deviation of the matching degree.

[0113] Specifically, after completing the defect detection of each batch of electronic components, the original visualized detection images, preprocessed images after multi-scale feature enhancement preprocessing, defect and deformation feature recognition results output by the dual-feature recognition model, the constructed defect-deformation correlation feature matrix, the defect ontology and deformation region quantization results obtained from dual-region quantization analysis, the dynamic influence coefficients obtained from gradient descent iteration calculation, and the performance risk level and qualification conclusions obtained from fusion judgment are standardized and organized to form standardized data samples. These data samples are then added to the training set of the dual-feature recognition model. The training set is divided into a training set and a validation set at an 8:2 ratio. Incremental learning is performed on the dual-feature recognition model, and the parameters of the model are fine-tuned using stochastic gradient descent, enabling the model to continuously learn new defect and deformation features, adapt to more diverse electronic component defect detection scenarios, and improve the model's recognition accuracy and generalization ability.

[0114] Simultaneously, actual performance tests are conducted on the completed electronic components to obtain actual performance test data, including withstand voltage, conductivity, and stability. Based on the performance test data, the actual qualification and performance risk of the components are determined, and the actual results are compared with the test judgment results to calculate the matching degree. ,in To determine the number of components whose results match the actual results, Given the total number of components tested in the batch, with a preset matching threshold of 90%, if the calculated matching degree... If the matching degree is within a certain range, it indicates that the threshold and weight parameters in the multi-level fusion judgment logic are set reasonably and do not need to be adjusted; if the matching degree is less than a certain value, it indicates that the matching degree is within a certain range. Based on the deviation, the single defect threshold, risk matrix threshold, pass threshold, and various weight parameters in the multi-level fusion judgment logic are dynamically adjusted. The larger the deviation, the greater the adjustment, ensuring that the judgment parameters continuously adapt to actual detection needs and guaranteeing the accuracy of the judgment results. This dual-dimensional model update step enables dual optimization of the model and judgment parameters, allowing the entire detection method to continuously iterate and upgrade, adapting to the development of electronic component manufacturing technology and changes in defect types. It effectively solves the problems of overfitting and fixed judgment parameters that are common in traditional detection models, ensuring the long-term effectiveness of the detection method.

[0115] In summary, the electronic component defect detection method in this embodiment effectively enhances defect and deformation features through multi-scale feature enhancement preprocessing; simultaneously extracts the defect body and surrounding microstructural deformation features through a dual-feature recognition model that integrates spatiotemporal correlation; quantitatively represents the relationship between defects and deformation through spatiotemporal correlation feature extraction and defect-deformation correlation feature matrix construction; accurately quantifies defect and deformation regions through dual-region quantitative analysis; adaptively solves the dynamic influence coefficient through gradient descent iterative calculation; comprehensively determines the fusion of defects and deformation through multi-level fusion judgment logic; and achieves dual optimization of the model and parameters by combining a two-dimensional model update step. The entire method forms a complete, closed-loop electronic component defect detection system, effectively capturing the correlation information between the defect itself and the surrounding microstructural deformation. It solves the limitation of traditional detection technologies that only focus on the defect itself, and can accurately judge the actual impact of defects and deformation on component performance. This avoids the problem of some components with potential performance risks being misjudged as qualified, improving the accuracy and comprehensiveness of electronic component defect detection. At the same time, each step of the method is based on mature algorithms and technologies, which has strong operability and practicality. It can be adapted to the internal defect detection of various electronic components and meet the testing needs of the electronic manufacturing industry chain for product reliability and safety of use.

[0116] Example 2

[0117] Please see Figure 5 The present invention also provides an electronic component defect detection system, comprising:

[0118] The image acquisition module is used to acquire visual inspection images of the internal structure of the electronic components to be inspected;

[0119] The multi-scale preprocessing module, connected to the image acquisition module, is used to perform multi-scale feature enhancement preprocessing on the visualized detection image of the electronic components to be detected. It extracts the basic features of the multi-scale image through Gaussian pyramid downsampling, reconstructs the upsampled image using Laplacian pyramid, and outputs the preprocessed image by combining neighborhood pixel gray-level difference enhancement processing.

[0120] The dual-feature recognition module is used to input the preprocessed image into a dual-feature recognition model that integrates spatiotemporal correlation, and simultaneously extract the category and location information of the defect body, as well as the texture features and degree level information of the surrounding microstructural deformation area;

[0121] The associated feature extraction module is used to extract spatiotemporal associated features based on the formation process rules and spatial geometric relationship between the defect body and the microstructure deformation region, and to quantize and encode the spatiotemporal associated features to construct a defect-deformation associated feature matrix.

[0122] The dual-region quantization module is used to perform dual-region quantization analysis based on the characteristics of the defect body and the microstructure deformation region respectively, to obtain the quantization results of the defect body and the deformation region.

[0123] The dynamic coefficient generation module is used to generate the dynamic influence coefficient between the defect body and the microstructure deformation by gradient descent iterative calculation based on the defect-deformation correlation feature matrix.

[0124] The fusion judgment module is used to apply multi-level fusion judgment logic, combine the defect body quantification result, the deformation region quantification result and the dynamic influence coefficient, and determine the performance risk level and qualification conclusion of the electronic component to be tested.

[0125] The dual-dimensional model update module is used to receive data samples generated during system operation and simultaneously optimize and adjust the model parameters of the dual-feature recognition module and the judgment parameters of the fusion judgment module.

[0126] Specifically, the image acquisition module is an integrated module of industrial CT imager and high-magnification microscopic imager. It can automatically switch imaging modes according to the type and size of the electronic components to be inspected. After completing the visualization imaging of the internal structure of the components, the detection image data is transmitted to the multi-scale preprocessing module in real time.

[0127] The multi-scale preprocessing module has built-in algorithms for Gaussian pyramid downsampling, Laplacian pyramid reconstruction, and neighborhood pixel gray-level difference enhancement. After receiving the detection image transmitted by the image acquisition module, it automatically performs multi-scale feature enhancement preprocessing, generates a preprocessed image, and transmits it to the dual-feature recognition module.

[0128] The dual-feature recognition module is equipped with a pre-trained dual-feature recognition model that integrates spatiotemporal correlation. It incorporates a ResNet50 basic feature extraction network, a Faster R-CNN defect recognition branch, and a deformation feature extraction branch that integrates the CBAM attention mechanism. After receiving the pre-processed image, it automatically extracts features and outputs the category and location information of the defect ontology, as well as the texture features and degree level information of the microstructure deformation region. At the same time, it transmits the feature information to the correlation feature extraction module and the dual-region quantization module respectively.

[0129] The associated feature extraction module has a built-in electronic component manufacturing process library and a spatiotemporal associated feature extraction algorithm. After receiving feature information, it automatically retrieves process rules, calculates spatial geometric features, extracts spatiotemporal associated features and completes quantization encoding, constructs a defect-deformation associated feature matrix, and transmits the matrix data to the dynamic coefficient generation module.

[0130] The dual-region quantization module incorporates multiple geometric shape fitting algorithms and region contour closure fitting techniques. After receiving feature information, it performs dual-region quantization analysis on the defect body and the microstructure deformation region respectively, generating the quantization results of the defect body and the deformation region, and then transmits them to the dynamic coefficient generation module and the fusion judgment module.

[0131] The dynamic coefficient generation module has a built-in gradient descent iterative algorithm. After receiving the defect-deformation correlation feature matrix, the defect body quantization result, and the deformation region quantization result, it automatically calculates and generates dynamic influence coefficients and transmits the coefficients to the fusion judgment module.

[0132] The fusion judgment module has built-in multi-level fusion judgment logic and preset thresholds and risk matrices. After receiving the defect ontology quantification results, deformation area quantification results and dynamic influence coefficients, it automatically performs basic judgment, correlation quantification and performance risk judgment, and outputs the performance risk level and pass / fail conclusion of the electronic component to be tested. At the same time, it transmits the judgment results and various data in the testing process to the dual-dimensional model update module.

[0133] The dual-dimensional model update module has a built-in data standardization and organization program, incremental learning algorithm, and matching degree calculation program. After receiving various data samples generated during system operation, it automatically completes standardization and organization, performs incremental learning optimization on the model parameters of the dual-feature recognition module, and calculates the matching degree between the detection judgment result and the actual result by combining the actual performance test data of the components. Based on the matching degree deviation, it dynamically adjusts the judgment threshold and weight parameters of the fusion judgment module to achieve dual optimization of the model and parameters.

[0134] Example 3

[0135] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principle of this invention is further explained below in conjunction with a specific application scenario.

[0136] This embodiment selects multilayer ceramic chip capacitors (MLCCs), which are widely used in the field of electronic manufacturing, as the electronic component to be tested and conducts practical application verification. Multilayer ceramic chip capacitors have a fine internal structure and high packaging density. During the manufacturing process, internal defects such as cracks and pinholes are prone to occur. Moreover, such defects are often accompanied by microstructure warping and deformation of the surrounding dielectric layer. They are typical electronic components for which the correlation between defects and microstructure deformation needs to be considered. This verification uses the detection method of Embodiment 1 and the detection system of Embodiment 2 to detect internal defects in a batch of multilayer ceramic chip capacitors with the serial number MLCC-202501. There are 50 samples to be tested in this batch. The detection scenario is the finished product inspection stage of an electronic component manufacturing enterprise. The detection requirements are to comply with the industry standards for civilian electronic components. Components that are deemed qualified must meet the requirements of having no obvious internal defects, no microstructure deformation affecting performance, and a low overall performance risk.

[0137] Before testing, the parameters of the testing system are preset. Based on the industry standard for civilian multilayer ceramic chip capacitors, a single defect threshold vector for the quantitative results of the defect body is set. The qualified threshold of fusion quantization value Risk matrix threshold , That is, the fusion quantization value < For low risk, < Medium risk. To mitigate the high risk, all modules of the detection system were debugged to normal operating conditions, and the manufacturing process library for multilayer ceramic chip capacitors was imported into the associated feature extraction module.

[0138] During the inspection process, the image acquisition module first performs industrial CT imaging on 50 samples to obtain visualized inspection images of the internal structure with a resolution of 2048×2048 pixels. Then, the inspection and judgment are completed sequentially through a multi-scale preprocessing module, a dual-feature recognition module, a correlation feature extraction module, a dual-region quantization module, a dynamic coefficient generation module, and a fusion judgment module. Finally, the dual-dimensional model update module completes the data sample processing and preliminary model parameter fine-tuning. The entire inspection process is completed automatically by the system, and the inspection of a single sample takes about 10 seconds. Compared with traditional manual inspection and single defect inspection technology, the inspection efficiency is effectively improved.

[0139] The test results of 50 samples to be tested were compiled in this verification. Typical test results are shown in Table 1. The table presents core information such as sample number, defect type, degree of microstructure deformation, key spatiotemporal correlation characteristics (after normalization), fusion quantification value, preliminary criteria, performance risk level and final qualification conclusion.

[0140] Table 1

[0141]

[0142] The test results show that the testing method and system of this application can accurately identify the type of internal defects and the degree of microstructural deformation in multilayer ceramic chip capacitors. Simultaneously, it can effectively capture the spatiotemporal correlation characteristics between the two. Through fusion quantization and multi-level judgment, accurate performance risk levels and compliance conclusions are obtained. For example, sample MLCC-202501-08 was found to have pinhole defects accompanied by slight dielectric layer deformation. Its fusion quantization value was 0.08, lower than the compliance threshold of 0.2, indicating a low performance risk, and was ultimately judged as qualified. The defects and deformation of this sample had a relatively small impact on performance, meeting the requirements for civilian use. Sample MLCC-202501-15 was also found to have pinhole defects, but accompanied by moderate dielectric layer deformation. Its fusion quantization value was 0.15. Although initially judged as qualified, the fusion quantization value was close to the compliance threshold, indicating a medium performance risk. The final result was deemed unqualified. This result fully demonstrates the advantage of this application in considering the influence of microstructure deformation, and avoids the misjudgment of such components with potential performance risks as qualified. On the other hand, samples MLCC-202501-22 and MLCC-202501-37 were found to have crack defects and were initially judged as unqualified. Moreover, the fusion quantification value of both exceeded the qualified threshold, indicating a high risk of performance. Therefore, they were ultimately judged as unqualified. Crack defects themselves have a significant impact on the performance of chip multilayer ceramic capacitors, and even with slight deformation, they cannot meet the usage requirements.

[0143] In this practical application verification, the automated operation of the detection system enabled rapid detection of defects in multilayer ceramic chip capacitors. The multi-level fusion judgment logic of the detection method ensured the accuracy of the judgment results, effectively identifying components with potential performance risks and solving the misjudgment problem caused by traditional detection techniques that only focus on the defect itself. Simultaneously, the standardized data samples of 50 samples generated during the detection process were added to the training set of the dual-feature recognition model, completing an incremental learning process. This further improved the model's accuracy in recognizing defects and deformation features of multilayer ceramic chip capacitors, providing a better model foundation for the subsequent detection of similar components.

[0144] This application verification shows that the electronic component defect detection method and system of this application can be effectively applied to actual electronic component defect detection scenarios, especially for electronic components with fine internal structures and where defects are correlated with the deformation of surrounding microstructures. It has the characteristics of high detection efficiency, accurate judgment results, and strong adaptability, which can meet the actual needs of electronic manufacturing for component defect detection, effectively improve the reliability and safety of electronic component products, and has broad practical application value.

[0145] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0146] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting defects in electronic components, characterized in that, Includes the following steps: Acquire visual inspection images of the internal structure of the electronic components to be inspected; Multi-scale feature enhancement preprocessing is performed on the visualized inspection image of the electronic components to be inspected. The basic features of the multi-scale image are extracted by Gaussian pyramid downsampling, and the upsampled image is reconstructed by Laplacian pyramid. The preprocessed image is output by combining the gray-level difference enhancement processing of neighboring pixels. The preprocessed image is input into a dual-feature recognition model that integrates spatiotemporal correlation, and the category and location information of the defect body, as well as the texture features and degree level information of the surrounding microstructural deformation region are extracted simultaneously. Based on the formation process rules and spatial geometric relationship between the defect body and the microstructure deformation region, spatiotemporal correlation features are extracted, and the spatiotemporal correlation features are quantized and encoded to construct a defect-deformation correlation feature matrix; Dual-region quantization analysis based on the characteristics of the defect body and the microstructure deformation region is performed on the defect body and the microstructure deformation region respectively to obtain the quantization results of the defect body and the deformation region. Based on the defect-deformation correlation feature matrix, the dynamic influence coefficient between the defect body and the microstructure deformation is generated by gradient descent iteration. By applying multi-level fusion judgment logic and combining the defect body quantification result, the deformation region quantification result, and the dynamic influence coefficient, the performance risk level and qualification conclusion of the electronic component under test are determined.

2. The method for detecting defects in electronic components according to claim 1, characterized in that, The specific steps of the multi-scale feature enhancement preprocessing are as follows: The original visualized detection image is continuously downsampled through a Gaussian pyramid to obtain image pyramid levels at different resolutions to characterize the basic structural features. In the process of Laplacian pyramid reconstruction, interpolation is performed on images of adjacent levels to extract high-frequency detail information, and interpolation algorithms are used to complete the reconstruction of the upsampled image. Within a local window of the reconstructed image, the grayscale difference between the center pixel and its neighboring pixels is calculated. Based on the grayscale difference, the gradient features of the defect body edge and the subtle texture features of the surrounding microstructure deformation are linearly compensated to suppress random noise signals in the image background.

3. The method for detecting defects in electronic components according to claim 1, characterized in that, The dual-feature recognition model that integrates spatiotemporal correlation includes a basic feature extraction network, a defect recognition branch, and a deformation feature extraction branch; The basic feature extraction network is used to extract the low-level shared features of the preprocessed image; The defect identification branch locates and classifies the defect entity based on the underlying shared features; The deformation feature extraction branch runs in parallel with the defect recognition branch. It captures the surrounding area affected by defect formation through an attention mechanism, identifies the location of the microstructure deformation area, extracts texture feature vectors from the microstructure deformation area, and classifies the deformation performance of the microstructure deformation area into different degree levels.

4. The method for detecting defects in electronic components according to claim 1, characterized in that, The specific process for extracting spatiotemporal correlation features includes: In the time dimension, the process rules corresponding to the defect body are retrieved, and the sequential characteristics of defect generation and deformation occurrence in the manufacturing process are matched. In the spatial dimension, the Euclidean distance characteristics from the edge of the defect body to the geometric center of the microstructure deformation region are calculated, the positional overlap characteristics of the defect body and the microstructure deformation region on the two-dimensional projection plane are analyzed, and the spatial envelopment ratio characteristics of the microstructure deformation region on the defect body are calculated.

5. The method for detecting defects in electronic components according to claim 1, characterized in that, The method for constructing the defect-deformation correlation feature matrix is ​​as follows: The extracted temporal features, spatial distance features, spatial overlap features, and spatial wrapping features are mapped to a unified quantization space. The spatiotemporal correlation features are converted into feature vectors using a numerical encoding method, and the feature vectors are arranged according to a preset topological structure to form a dedicated correlation feature matrix that represents the correlation between the defect body and the microstructure deformation region.

6. The method for detecting defects in electronic components according to claim 1, characterized in that, The specific implementation method of the dual-region quantitative analysis is as follows: For the defect entity, a corresponding geometric shape fitting algorithm is selected according to its category to calculate the area, perimeter, and aspect ratio features of the defect entity, and generate the quantification result of the defect entity. For the microstructure deformation region, the influence range involved in the deformation is calculated using the region contour closure fitting technique, and the texture pixel distribution density within the deformation region is statistically analyzed. Combined with the identified degree level, the quantification result of the deformation region is generated by weighted summarization.

7. The method for detecting defects in electronic components according to claim 1, characterized in that, The steps for generating the dynamic influence coefficient include: Obtain the initial weight coefficients corresponding to the defect ontology, and establish an objective function for calculating the influence coefficients based on the feature vector distribution in the defect-deformation correlation feature matrix; The gradient descent algorithm is used to perform multiple iterations for optimization. The coefficients are dynamically corrected based on the influence range and degree of deformation in the quantification results of the deformation region. When the objective function converges to the preset range, the dynamic influence coefficient representing the weight of the deformation on the defect performance is output.

8. The method for detecting defects in electronic components according to claim 1, characterized in that, The multi-level fusion judgment logic includes a basic judgment layer, a correlation quantization layer, and a performance risk layer: The basic judgment layer compares the quantification result of the defect ontology with a preset single defect threshold to obtain a preliminary qualification criterion. The correlation quantization layer performs multiplication and accumulation operations on the defect ontology quantization result, the deformation region quantization result, and the dynamic influence coefficient to calculate the defect-deformation fusion quantization value. The performance risk layer matches a preset risk matrix based on the size range of the defect-deformation fusion quantization value to determine the performance risk level of the component, and outputs a final judgment conclusion in conjunction with the preliminary qualification criteria.

9. The method for detecting defects in electronic components according to claim 1, characterized in that, The method also includes a two-dimensional model update step: The acquired original image, preprocessed image, dual-feature recognition result, associated feature matrix, quantitative analysis result, dynamic influence coefficient and judgment conclusion are integrated into a standardized data sample and added to the training set of the dual-feature recognition model for incremental learning. Simultaneously, the actual performance test data of the components after testing is obtained, the matching degree between the test judgment result and the performance test data is calculated, and the judgment threshold and weight parameters in the multi-level fusion judgment logic are dynamically adjusted according to the deviation of the matching degree.

10. An electronic component defect detection system, applied to the electronic component defect detection method according to any one of claims 1-9, characterized in that, include: The image acquisition module is used to acquire visual inspection images of the internal structure of the electronic components to be inspected; The multi-scale preprocessing module, connected to the image acquisition module, is used to perform multi-scale feature enhancement preprocessing on the visualized detection image of the electronic components to be detected. It extracts the basic features of the multi-scale image through Gaussian pyramid downsampling, reconstructs the upsampled image using Laplacian pyramid, and outputs the preprocessed image by combining neighborhood pixel gray-level difference enhancement processing. The dual-feature recognition module is used to input the preprocessed image into a dual-feature recognition model that integrates spatiotemporal correlation, and simultaneously extract the category and location information of the defect body, as well as the texture features and degree level information of the surrounding microstructural deformation area; The associated feature extraction module is used to extract spatiotemporal associated features based on the formation process rules and spatial geometric relationship between the defect body and the microstructure deformation region, and to quantize and encode the spatiotemporal associated features to construct a defect-deformation associated feature matrix. The dual-region quantization module is used to perform dual-region quantization analysis based on the characteristics of the defect body and the microstructure deformation region respectively, to obtain the quantization results of the defect body and the deformation region. The dynamic coefficient generation module is used to generate the dynamic influence coefficient between the defect body and the microstructure deformation by gradient descent iterative calculation based on the defect-deformation correlation feature matrix. The fusion judgment module is used to apply multi-level fusion judgment logic, combine the defect body quantification result, the deformation region quantification result and the dynamic influence coefficient, and determine the performance risk level and qualification conclusion of the electronic component to be tested. The dual-dimensional model update module is used to receive data samples generated during system operation and simultaneously optimize and adjust the model parameters of the dual-feature recognition module and the judgment parameters of the fusion judgment module.

Images