Integrated circuit packaging quality detection method and system based on test data

Abstract

The invention relates to the technical field of integrated circuits, and discloses an integrated circuit packaging quality detection method and system based on test data, and the method comprises the steps: obtaining various test data of integrated circuit packaging, carrying out the preprocessing of the collected test data, and obtaining the preprocessed test data; feature information is extracted from the preprocessed test data, and the extracted feature information is utilized to construct an integrated circuit packaging quality evaluation model; performing preprocessing and feature extraction on test data of the integrated circuit package to be detected, and inputting the test data into the quality evaluation model to obtain a quality evaluation result of the integrated circuit package; according to the invention, rapid, accurate and comprehensive detection of the packaging quality of the integrated circuit is realized, and compared with traditional manual visual detection and simple electrical test methods, the detection efficiency is improved, the labor cost and subjective errors are reduced, and the performance, reliability and service life of the integrated circuit are effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated circuits, and in particular to a method and system for detecting the quality of integrated circuit packaging based on test data. Background Art

[0002] With the rapid development of integrated circuit technology, integrated circuit packaging, as a key link in integrated circuit manufacturing, has a direct impact on the performance and reliability of integrated circuits. Existing integrated circuit packaging quality inspections mostly rely on manual visual inspection or a single inspection method, such as judging packaging quality solely through appearance inspection or simple electrical parameter testing. Manual visual inspection is inefficient, highly subjective, and difficult to detect subtle internal defects. A single inspection method cannot fully reflect packaging quality and is prone to missed detections and false detections, resulting in poor accuracy and reliability of test results, which cannot meet the growing demand for integrated circuit packaging quality inspection. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and to design a method and system for detecting the quality of integrated circuit packaging based on test data.

[0004] A first aspect of the present invention provides a method for detecting integrated circuit packaging quality based on test data, the method comprising the following steps:

[0005] Acquiring a variety of test data of the integrated circuit package, wherein the test data includes appearance image data, electrical parameter data, and thermal performance data;

[0006] Preprocessing the collected test data, including noise reduction, enhancement, and edge detection of appearance image data, as well as normalization of electrical parameter data and thermal performance data, to obtain preprocessed test data;

[0007] Extract feature information from the preprocessed test data and build an integrated circuit packaging quality assessment model using the extracted feature information;

[0008] After preprocessing and feature extraction, the test data of the integrated circuit package to be inspected is input into the quality assessment model to obtain the quality assessment result of the integrated circuit package.

[0009] Optionally, in a first implementation of the first aspect of the present invention, preprocessing the collected test data includes:

[0010] Define a sliding window of size 3×3 and place it on the first pixel of the appearance image in the test data;

[0011] Extract the grayscale values of all pixels in the sliding window and sort them in ascending order. Take the grayscale value in the middle position after sorting and assign it to the pixel in the center of the window to complete the noise reduction of this pixel.

[0012] With a step size of 1, the sliding window is moved rightward and downward on the image in sequence until all pixels on the image are processed to obtain the denoised image;

[0013] Obtain the frequency of each gray level in the denoised image, and calculate the probability and cumulative distribution function of each gray level based on the frequency;

[0014] Map the cumulative distribution function to a new grayscale range, and replace the grayscale value of each pixel in the image according to the new grayscale mapping relationship to obtain an enhanced image;

[0015] The Canny edge detection algorithm is used to perform edge detection processing on the enhanced image to obtain an image after edge detection;

[0016] The electrical parameter data and thermal performance data in the test data are normalized using the maximum-minimum normalization method to obtain normalized electrical parameter data and thermal performance data;

[0017] The preprocessed test data is obtained by integrating the edge detected image, normalized electrical parameter data and thermal performance data.

[0018] Optionally, in a second implementation of the first aspect of the present invention, performing edge detection processing on the enhanced image using a Canny edge detection algorithm to obtain an image after edge detection includes:

[0019] Define multiple Gaussian kernels of different scales, perform convolution operations on each Gaussian kernel and the enhanced image to obtain multiple groups of smoothed images, perform weighted fusion on the multiple groups of smoothed images, and obtain the final smoothed image through weighted summation;

[0020] For the smoothed image, the gradient is calculated using convolution kernels with adaptive window sizes in the horizontal and vertical directions respectively;

[0021] Calculate the mean and standard deviation of the gradient amplitude in the neighborhood of each pixel based on the local contrast enhancement factor;

[0022] Traverse each pixel in the image, obtain the gradient direction of the current pixel, and determine the two adjacent pixels in the gradient direction based on the gradient direction. If the gradient magnitude of the current pixel is not a local maximum and the weighted difference between it and the gradient magnitude of the adjacent pixel is less than the threshold, then set it to 0; otherwise, retain the gradient magnitude of the edge pixel;

[0023] Count the distribution of gradient amplitudes in the image, calculate the mean and standard deviation, dynamically set high and low thresholds, identify pixels with gradient amplitudes greater than the high threshold as edge pixels, and exclude pixels with gradient amplitudes less than the low threshold;

[0024] For pixel points whose gradient amplitude is between the high threshold and the low threshold, the edge connection is performed using the region growing method to obtain the image after edge detection.

[0025] Optionally, in a third implementation of the first aspect of the present invention, extracting feature information from the preprocessed test data includes:

[0026] The pre-processed test data is divided into multimodal data blocks and spatiotemporal associations are constructed. A structural attention mechanism based on graph convolution is introduced to calculate the structural attention weight of each data block.

[0027] In the query calculation of the ProbSparse self-attention mechanism, the modal category information and spatiotemporal label information of the data block are introduced to perform weighted adjustment on the query, and the structural attention weight and the self-attention weight are weighted fused. The fusion weight is dynamically adjusted to obtain the hybrid attention weight.

[0028] A multi-layer feature extraction network is constructed. Each layer receives the features and mixed attention weights output by the previous layer. The features extracted by each layer are residually connected, and the features extracted by the current layer are added to the features of the previous layer as the input of the next layer. In the last layer of the network, the features extracted from different modalities are spliced and fused to obtain the final feature information.

[0029] Optionally, in a fourth implementation of the first aspect of the present invention, the performing multimodal data segmentation and spatiotemporal association construction on the preprocessed test data includes:

[0030] The pre-processed test data is divided into three categories according to the modality: appearance image data, electrical parameter data and thermal performance data;

[0031] The appearance image data is spatially divided into blocks according to different areas of the integrated circuit package, and each block forms an independent data unit;

[0032] For electrical parameter data and thermal performance data, the data are divided into equal intervals according to the time series, and overlapping windows are set between adjacent blocks;

[0033] Add spatiotemporal labels to each data block and construct the spatiotemporal correlation matrix between data blocks of different modalities.

[0034] Optionally, in a fifth implementation of the first aspect of the present invention, constructing an integrated circuit packaging quality assessment model using the extracted feature information includes:

[0035] Based on the extracted feature information, a feature semantic association graph is constructed, and the input feature information of the underlying model layer is constructed. A heterogeneous model fusion architecture is adopted in the middle layer, including a DNN network to capture the nonlinear relationship of features, a GNN network to mine the structural information in the feature semantic association graph, and a lightweight decision tree integration model for fast feature classification.

[0036] The weights of different network outputs in the middle layer are calculated through the attention mechanism to perform feature-level fusion. A fusion decision layer is set at the top layer to receive the fused features of the middle layer, and the cosine annealing algorithm is used for dynamic adjustment to output the final evaluation results.

[0037] Optionally, in a sixth implementation of the first aspect of the present invention, the dynamic adjustment using a cosine annealing algorithm includes:

[0038] Calculate the variance and entropy of the features after fusion in the intermediate layer to obtain statistics, determine the initial temperature parameters based on the feature statistics, and establish the association mapping between the temperature parameters and feature categories;

[0039] Define the base learning rate and calculate it based on the current training round and the total training rounds. Calculate the differences between different category features in the current training batch and adjust the base learning rate based on the differences.

[0040] Calculate the importance score of each network output feature in the middle layer, calculate the weight adjustment amount of each network output feature based on the importance score and the current temperature parameter, update the weight of each network output feature, and perform normalization.

[0041] A decision threshold is set. When the temperature parameter is lower than the decision threshold, the system enters the stable decision stage. In the stable decision stage, the two strategies of weighted voting and probability output are combined to determine the final evaluation result.

[0042] A second aspect of the present invention provides an integrated circuit packaging quality inspection system based on test data, the system comprising:

[0043] an acquisition module, configured to acquire a variety of test data of the integrated circuit package, wherein the test data includes appearance image data, electrical parameter data, and thermal performance data;

[0044] The preprocessing module is used to preprocess the collected test data, including noise reduction, enhancement, and edge detection processing of the appearance image data, as well as normalization processing of the electrical parameter data and thermal performance data, to obtain the preprocessed test data;

[0045] A construction module is used to extract feature information from the preprocessed test data and construct an integrated circuit packaging quality assessment model using the extracted feature information;

[0046] The input module is used to input the test data of the integrated circuit package to be tested into the quality assessment model after preprocessing and feature extraction, so as to obtain the quality assessment result of the integrated circuit package.

[0047] The third aspect of the present invention provides an integrated circuit packaging quality inspection device based on test data, the integrated circuit packaging quality inspection device based on test data includes a memory and at least one processor, the memory stores instructions; the at least one processor calls the instructions in the memory to enable the integrated circuit packaging quality inspection device based on test data to perform each step of the integrated circuit packaging quality inspection method based on test data as described in any one of the above items.

[0048] A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed by a processor, implement the various steps of the integrated circuit packaging quality inspection method based on test data as described in any of the above items.

[0049] In the technical solution provided by the present invention, a variety of test data of integrated circuit packaging are obtained, wherein the test data include appearance image data, electrical parameter data and thermal performance data; the collected test data are preprocessed, including noise reduction, enhancement, and edge detection processing of the appearance image data, and normalization processing of the electrical parameter data and thermal performance data, to obtain preprocessed test data; feature information is extracted from the preprocessed test data, and an integrated circuit packaging quality assessment model is constructed using the extracted feature information; the test data of the integrated circuit package to be inspected is input into the quality assessment model after preprocessing and feature extraction to obtain a quality assessment result of the integrated circuit package; the present invention improves the quality and efficiency of data processing, improves the accuracy and credibility of quality assessment, and shows stronger practicality and innovation in complex and changeable integrated circuit packaging quality inspection scenarios, effectively guarantees the performance, reliability and service life of the integrated circuit, so as to solve the problems of low efficiency, poor accuracy and incomplete inspection of integrated circuit packaging quality in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0051] Figure 1 A flow chart of a method for detecting integrated circuit packaging quality based on test data provided by an embodiment of the present invention;

[0052] Figure 2A structural diagram of an integrated circuit packaging quality inspection system based on test data provided by an embodiment of the present invention;

[0053] Figure 3 A schematic diagram of the structure of an integrated circuit packaging quality inspection device based on test data provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The terms "first," "second," "third," "fourth," and the like (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar objects and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, apparatus, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to these processes, methods, products, or apparatus.

[0055] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 A flowchart of a method for testing integrated circuit packaging quality based on test data provided by an embodiment of the present invention, the method specifically comprising the following steps:

[0056] Step 101: Acquire various test data of an integrated circuit package, wherein the test data includes appearance image data, electrical parameter data, and thermal performance data;

[0057] Step 102: preprocessing the collected test data, including noise reduction, enhancement, and edge detection processing of the appearance image data, and normalization processing of the electrical parameter data and thermal performance data, to obtain preprocessed test data;

[0058] In this embodiment, a sliding window of size 3×3 is defined and placed on the first pixel of the appearance image in the test data; the grayscale values of all pixels in the sliding window are extracted and sorted in ascending order; the grayscale value of the middle position after sorting is taken and assigned to the pixel at the center of the window to complete the noise reduction of the pixel; with a step size of 1, the sliding window is moved rightward and downward on the image in sequence until all pixels on the image are processed to obtain the noise-reduced image; the frequency of each grayscale level in the noise-reduced image is obtained, and the probability and probability of each grayscale level are calculated based on the frequency. Cumulative distribution function; mapping the cumulative distribution function to a new grayscale range, replacing the grayscale value of each pixel in the image according to the new grayscale mapping relationship to obtain an enhanced image; performing edge detection processing on the enhanced image using the Canny edge detection algorithm to obtain an edge-detected image; normalizing the electrical parameter data and thermal performance data in the test data using the maximum-minimum normalization method to obtain normalized electrical parameter data and thermal performance data; integrating the edge-detected image, the normalized electrical parameter data and the thermal performance data to obtain the preprocessed test data.

[0059] In this embodiment, multiple Gaussian kernels of different scales are defined. Each Gaussian kernel is convolved with the enhanced image to obtain multiple smoothed images. The multiple smoothed images are weightedly fused and a final smoothed image is obtained by weighted summation. For the smoothed image, gradients are calculated using convolution kernels with adaptive window sizes in the horizontal and vertical directions. The mean and standard deviation of the gradient amplitude within the neighborhood of each pixel are calculated based on the local contrast enhancement factor. Each pixel in the image is traversed to obtain the gradient direction of the current pixel. Based on the gradient direction, two adjacent pixels in the gradient direction are determined. If the gradient amplitude of the current pixel is not a local maximum and the weighted difference between the gradient amplitude of the current pixel and the gradient amplitude of the adjacent pixel is less than a threshold, the gradient amplitude is set to 0; otherwise, the gradient amplitude of the edge pixel is retained. The distribution of the gradient amplitude in the image is statistically analyzed, the mean and standard deviation are calculated, and a high threshold and a low threshold are dynamically set. Pixels with gradient amplitudes greater than the high threshold are determined as edge pixels, and pixels with gradient amplitudes less than the low threshold are excluded. For pixels with gradient amplitudes between the high threshold and the low threshold, edge connection is performed using region growing to obtain an image after edge detection.

[0060] In this embodiment, in the data preprocessing stage, differentiated and efficient processing algorithms are designed for appearance image data and electrical and thermal performance data. Appearance image data is processed through median filtering for noise reduction, histogram equalization for enhancement, and improved Canny edge detection to accurately remove noise, enhance image details, and extract edge information, effectively avoiding detection errors caused by image quality issues; electrical parameter data and thermal performance data are normalized through the minimum-maximum normalization algorithm to unify the data format and dimension, laying the foundation for subsequent data analysis. At the same time, the ProbSparse self-attention mechanism of Informer is improved in the feature extraction link. Through multimodal data segmentation, hybrid attention mechanism construction, feature hierarchical extraction and enhancement, and adaptive feature screening, the spatiotemporal correlation and potential features in the data are fully explored. Compared with traditional methods, more comprehensive and accurate feature information can be extracted, greatly improving the quality and efficiency of data processing.

[0061] Step 103: extracting feature information from the pre-processed test data, and constructing an integrated circuit packaging quality assessment model using the extracted feature information;

[0062] In this embodiment, the preprocessed test data is divided into multimodal data blocks and spatiotemporal associations are constructed, a structural attention mechanism based on graph convolution is introduced, and the structural attention weight of each data block is calculated; in the Query calculation of the ProbSparse self-attention mechanism, the modal category information and spatiotemporal label information of the data block are introduced, the Query is weighted and adjusted, the structural attention weight and the self-attention weight are weightedly fused, and the fusion weight is dynamically adjusted to obtain the hybrid attention weight; a multi-layer feature extraction network is constructed, each layer receives the features and hybrid attention weights output by the previous layer, the features extracted by each layer are residually connected, the features extracted by the current layer are added to the features of the previous layer as the input of the next layer, and in the last layer of the network, the features extracted by different modalities are spliced and fused to obtain the final feature information.

[0063] In this embodiment, the preprocessed test data is divided into three categories according to the modality: appearance image data, electrical parameter data, and thermal performance data. The appearance image data is spatially blocked according to different areas of the integrated circuit package, and each block forms an independent data unit. The electrical parameter data and thermal performance data are equally spaced and blocked according to the time series, and overlapping windows are set between adjacent blocks. A spatiotemporal label is added to each data block, and a spatiotemporal correlation matrix between data blocks of different modalities is constructed.

[0064] In this embodiment, a feature semantic association graph is constructed based on the extracted feature information, and the input feature information of the bottom model layer is constructed. A heterogeneous model fusion architecture is adopted in the middle layer, including a DNN network to capture the nonlinear relationship of features, a GNN network to mine the structural information in the feature semantic association graph, and a lightweight decision tree integration model for fast feature classification. The weights of the outputs of different networks in the middle layer are calculated through the attention mechanism to perform feature-level fusion. A fusion decision layer is set at the top layer to receive the fused features of the middle layer, and the cosine annealing algorithm is used for dynamic adjustment to output the final evaluation results.

[0065] In this embodiment, the variance and entropy of the features after fusion in the intermediate layer are calculated to obtain statistics, the initial temperature parameters are determined based on the feature statistics, and an association mapping between the temperature parameters and feature categories is established; a basic learning rate is defined, and the basic learning rate is calculated based on the current training round and the total training rounds, the difference between different categories of features in the current training batch is calculated, and the basic learning rate is adjusted based on the difference; the importance score of each network output feature in the intermediate layer is calculated, and the weight adjustment amount of each network output feature is calculated based on the importance score and the current temperature parameter, the weight of each network output feature is updated, and normalization is performed. A decision threshold is set, and when the temperature parameter is lower than the decision threshold, the stable decision stage is entered. In the stable decision stage, the two strategies of weighted voting and probability output are combined to determine the final evaluation result.

[0066] In this embodiment, the innovative construction of the integrated circuit packaging quality assessment model forms a complete and advanced system, from feature data enhancement and heterogeneous model fusion to dynamic weight adjustment training and result calibration and interpretation. Feature data enhancement and semantic association construction increase data diversity and semantic connections between features, making model training more comprehensive. The heterogeneous model fusion architecture combines the advantages of deep neural networks, graph neural networks, and decision tree ensemble models, combining a layered fusion structure and attention mechanism to effectively integrate different types of feature information. The dynamic weight adjustment training mechanism optimizes model parameters in real time based on training conditions, enhancing the model's adaptability and generalization capabilities. The result calibration and interpretation module ensures that the assessment results are accurate, reliable, and interpretable, significantly improving the accuracy and credibility of quality assessments compared to traditional single models or simple assessment methods.

[0067] In this embodiment, improvements to the Canny edge detection algorithm and cosine annealing algorithm specifically address key issues in integrated circuit packaging inspection. The improved Canny edge detection algorithm, through multi-scale Gaussian weighted smoothing, adaptive gradient calculation, improved non-maximum suppression, and dynamic threshold decision-making, can better adapt to the complex characteristics of package images and accurately detect edges. The improved cosine annealing algorithm dynamically adjusts temperature parameters, multi-dimensional learning rate adjustment, and adaptive weight updates based on data features, enabling the model to more rationally integrate intermediate layer features when making top-level decisions. This dynamically optimizes the decision-making process based on data changes, effectively improving the accuracy and stability of evaluation results, and demonstrating greater practicality and innovation in complex and ever-changing integrated circuit packaging quality inspection scenarios.

[0068] Step 104 : After preprocessing and feature extraction, the test data of the integrated circuit package to be inspected is input into a quality assessment model to obtain a quality assessment result of the integrated circuit package.

[0069] In this embodiment, the data source is determined, covering the data collection points of each key link in the integrated circuit packaging production line; with the help of various sensors, such as temperature sensors, pressure sensors, current sensors, voltage sensors, etc., the original data of the integrated circuit packaging process, such as packaging temperature, packaging pressure, pin welding current, pin welding voltage, etc., are collected; high-precision testing equipment, such as electrical performance testers, microscopes, etc., are used to carry out electrical performance tests and physical performance tests on the integrated circuits after packaging, and obtain electrical performance test data (such as on-resistance, insulation resistance, capacitance, inductance, etc.) and physical performance test data (such as pin dimensional accuracy, package appearance defects, etc.); appearance image data preprocessing and noise reduction are performed. Processing (using median filtering algorithm) defines a sliding window of size n×n (such as 3×3, 5×5), and places it on the first pixel of the appearance image; extracts the grayscale values of all pixels in the sliding window, and then arranges these grayscale values in ascending order; selects the grayscale value in the middle position after sorting, and assigns it to the pixel in the center of the window, thereby completing the noise reduction of the pixel; with a step size of 1, moves the sliding window to the right and downward on the image in turn, repeating the above sorting and assignment steps until all pixels on the image are processed and the noise-reduced image is obtained; enhancement processing (using histogram equalization algorithm) counts the frequency of each grayscale in the original noise-reduced image, that is, calculates the frequency of each grayscale value in the original noise-reduced image. The number of times it appears in the image; the probability of each gray level is calculated based on the frequency; the cumulative distribution function of the gray level is calculated, and the cumulative distribution function is mapped to a new gray level range. According to the new gray level mapping relationship, the gray value of each pixel in the image is replaced to obtain the enhanced image; edge detection processing (using the improved Canny edge detection algorithm) multi-scale Gaussian weighted smoothing processing defines multiple Gaussian kernels of different scales, and each Gaussian kernel is convolved with the enhanced image to obtain multiple groups of smoothed images; multiple groups of smoothed images are weighted fused, and the weights are dynamically adjusted according to the local features of the image; the variance of the local area of the image is calculated. The larger the variance, the richer the details of the area. The weight calculation formula is set. The final smoothed image is obtained by weighted summation. For the smoothed image, the gradient is calculated in the horizontal and vertical directions using a convolution kernel with an adaptive window size. The image is divided into multiple sub-regions, and the convolution kernel size is dynamically adjusted according to the grayscale variance of the pixels in the sub-region. If the grayscale variance of the sub-region is greater than the threshold, the convolution kernel size is increased to better capture large-scale edges; otherwise, the convolution kernel size is reduced. The gradient of the image in the horizontal and vertical directions is calculated using the adjusted convolution kernel (such as the adaptive Sobel operator), and the gradient magnitude and gradient direction are calculated. A local contrast enhancement factor is introduced and calculated based on the gradient magnitude distribution of the local region. For each pixel point, the mean and standard deviation of the gradient magnitude in its neighborhood are calculated.Improved non-maximum suppression: traverse each pixel in the image, obtain its gradient direction for the current pixel, determine the two adjacent pixels in the gradient direction according to the gradient direction, introduce a distance weighting coefficient when comparing the gradient amplitude of the current pixel with the gradient amplitude of the two adjacent pixels; the pixel points closer to the current pixel point are given a higher weight, if the gradient amplitude of the current pixel point is not a local maximum, and the weighted difference between the gradient amplitude of the current pixel point and the gradient amplitude of the adjacent pixel points is less than the threshold, then it is set to 0; otherwise, the gradient amplitude of the edge pixel point is retained, dynamic threshold decision and edge connection statistics are used to calculate the distribution of the gradient amplitude in the image, calculate the mean and standard deviation, dynamically set the high threshold and low threshold, and set the gradient amplitude greater than the high threshold. Pixels with a gradient amplitude less than a certain threshold are determined as edge pixels; pixels with a gradient amplitude less than a low threshold are excluded; for pixels with a gradient amplitude between a high threshold and a low threshold, edge connection is performed using a region growing method, and the determined edge pixels are used as seed points. Pixels that meet the connection conditions (such as the gradient amplitude difference is less than a certain threshold, and the gradient direction angle is less than a certain angle) are searched in their neighborhoods and included in the edge pixel set until no more pixels that meet the conditions can be connected, and finally an image after edge detection is obtained; the electrical parameter data and thermal performance data are preprocessed using a minimum-maximum normalization algorithm, which traverses the electrical parameter data or thermal performance data set to find the minimum and maximum values in the data. For each number in the data set, Normalization calculations are performed on the data points, mapping the data to the [0,1] interval to obtain normalized data. The above calculations are completed for all data points in the data set to obtain preprocessed electrical parameter data and thermal performance data. Feature extraction, multimodal data segmentation and spatiotemporal association construction divide the preprocessed test data into three categories according to modality: appearance image data, electrical parameter data, and thermal performance data. For appearance image data, spatial segmentation is performed according to different areas of the integrated circuit package (such as the pin area and the package surface area), and each segment forms an independent data unit. For electrical parameter data and thermal performance data, equal intervals are segmented according to time series, and a certain overlapping window is set between adjacent segments to preserve the temporal continuity of the data. Add spatiotemporal labels to each data block, build a spatiotemporal association matrix between data blocks of different modalities, and clarify the position and association relationship of each data block in the overall data; the hybrid attention mechanism construction introduces a structural attention mechanism based on graph convolution: based on the spatiotemporal association matrix, a graph structure of data blocks is constructed, each data block is a node in the graph, and the edges between nodes are weighted according to the spatiotemporal association strength; through graph convolution operations, the structural attention weight of each data block based on the overall structure is calculated; Improved ProbSparse self-attention mechanism: In the query calculation of the original ProbSparse self-attention mechanism, the modal category information and spatiotemporal label information of the data block are introduced to perform weighted adjustment on the query;When calculating the attention score, the attention score is corrected by combining the local statistical features of the data block (such as the mean and variance of the electrical parameter data block, and the grayscale entropy of the image data block); the structural attention weight is weightedly fused with the improved ProbSparse self-attention weight, and the fusion weight is dynamically adjusted according to the data modality and the characteristics of the data block to obtain the hybrid attention weight; feature hierarchical extraction and enhancement construct a multi-layer feature extraction network, and each layer receives the features and hybrid attention weights output by the previous layer; in each layer, for the image data block, the improved convolutional neural network (CNN) is used in combination with the hybrid attention weight to extract image features, and during the convolution operation, the parameters of the convolution kernel are dynamically adjusted according to the attention weight ; For electrical parameter data and thermal performance data blocks, a fully connected neural network based on the attention mechanism is used for feature extraction, and the input data is weighted by the mixed attention weight; the features extracted in each layer are residually connected, and the features extracted in the current layer are added to the features in the previous layer as the input of the next layer to enhance the expressive power of the features; in the last layer of the network, the features extracted from different modalities are spliced and fused to obtain the final feature vector; adaptive feature screening and optimization are used to calculate the importance score of each feature in the final feature vector, and a comprehensive evaluation index is constructed by combining the weight of the feature in the hybrid attention mechanism, the variance contribution of the feature in different modal data, and the correlation between the features; dynamic screening is set according to the comprehensive evaluation index The threshold is adaptively adjusted according to the overall distribution of the data and the number of features; features with importance scores higher than the threshold are retained, redundant and unimportant features are removed, and simplified feature information is obtained; Step 4: Quality assessment The simplified feature information obtained after feature extraction is input into the trained integrated circuit package quality assessment model; the model adopts a heterogeneous model fusion architecture, and the bottom layer is composed of a deep neural network (DNN), a graph neural network (GNN) and a decision tree integration model (such as a random forest). The middle layer calculates the weights of different network outputs through the attention mechanism for feature-level fusion. The top layer sets a fusion decision layer and uses the improved cosine annealing algorithm for dynamic adjustment; at the bottom layer of the model, each basic model performs input The feature information is preliminarily processed; deep neural networks capture the nonlinear relationship of features, graph neural networks mine the structural information in the feature semantic association map, and decision tree integration models perform rapid feature classification; the middle layer receives the output features of each underlying model, calculates the weights of different network outputs through the attention mechanism, and splices and fuses the features; the top fusion decision layer receives the fused features of the middle layer and uses the improved cosine annealing algorithm for dynamic adjustment; the algorithm first determines the initial temperature parameter based on the distribution characteristics of the fused features of the middle layer, and establishes an association mapping between the temperature parameter and the feature category; then a multi-dimensional learning rate adjustment strategy is adopted, combining the feature difference adjustment factor and the performance of the model on the validation set to adjust the learning rate;Then, through an adaptive weight update mechanism, the weights of each network output feature are updated based on the feature importance score and the current temperature parameter. Finally, in the stable decision phase, a combination of weighted voting and probabilistic output strategies is used to determine the final evaluation result. The model outputs the quality assessment result of the integrated circuit package, clearly determining whether it is qualified. If it is unqualified, further data analysis is performed to locate the specific quality issues.

[0070] In this embodiment, through the innovative design and synergy of each link, rapid, accurate and comprehensive detection of integrated circuit packaging quality is achieved. Compared with traditional manual visual inspection and simple electrical testing methods, it not only greatly improves the detection efficiency and reduces labor costs and subjective errors, but also can detect internal defects, material quality and other problems that are difficult to detect with traditional methods, effectively ensuring the performance, reliability and service life of the integrated circuit, and has broad application prospects and significant economic and social benefits in the field of integrated circuit manufacturing.

[0071] See also Figure 2 , a schematic diagram of the structure of an integrated circuit packaging quality inspection system based on test data provided by an embodiment of the present invention, the system includes:

[0072] An acquisition module 201 is configured to acquire various test data of an integrated circuit package, wherein the test data includes appearance image data, electrical parameter data, and thermal performance data;

[0073] The preprocessing module 202 is used to preprocess the collected test data, including noise reduction, enhancement, and edge detection processing of the appearance image data, and normalization processing of the electrical parameter data and thermal performance data, to obtain preprocessed test data;

[0074] A construction module 203 is configured to extract feature information from the pre-processed test data and construct an integrated circuit packaging quality assessment model using the extracted feature information;

[0075] The input module 204 is used to input the test data of the integrated circuit package to be inspected into the quality assessment model after preprocessing and feature extraction, so as to obtain the quality assessment result of the integrated circuit package.

[0076] above Figure 2 The integrated circuit packaging quality inspection device based on test data in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The integrated circuit packaging quality inspection device based on test data in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0077] Figure 3FIG. 6 is a schematic diagram of the structure of a test data-based integrated circuit packaging quality inspection device provided by an embodiment of the present invention. The test data-based integrated circuit packaging quality inspection device 600 may vary significantly due to different configurations or performance. The device may include one or more processors (central processing units, CPUs) 610 (e.g., one or more processors), a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing application programs 633 or data 632. The memory 620 and storage medium 630 may be either transient or persistent storage. The program stored in the storage medium 630 may include one or more modules (not shown), each of which may include a series of instruction operations in the test data-based integrated circuit packaging quality inspection device 600. Furthermore, the processor 610 may be configured to communicate with the storage medium 630, and execute the series of instruction operations in the storage medium 630 on the test data-based integrated circuit packaging quality inspection device 600 to implement the method provided in the above embodiment.

[0078] The integrated circuit packaging quality inspection device 600 based on test data may further include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input and output interfaces 660, and / or one or more operating devices 631, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the integrated circuit packaging quality inspection equipment based on test data shown does not constitute a limitation on the computer equipment provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0079] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the various steps of the integrated circuit packaging quality inspection method based on test data provided in the above embodiments.

[0080] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0081] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc., various media that can store program code.

[0082] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the quality of integrated circuit packaging based on test data, characterized in that: The method comprises the following steps: Acquiring a variety of test data of the integrated circuit package, wherein the test data includes appearance image data, electrical parameter data, and thermal performance data; Preprocessing the collected test data, including noise reduction, enhancement, and edge detection of appearance image data, as well as normalization of electrical parameter data and thermal performance data, to obtain preprocessed test data; Extract feature information from the preprocessed test data and build an integrated circuit packaging quality assessment model using the extracted feature information; After preprocessing and feature extraction, the test data of the integrated circuit package to be inspected is input into the quality assessment model to obtain the quality assessment result of the integrated circuit package.

2. The method for testing integrated circuit packaging quality based on test data according to claim 1, wherein: The preprocessing of the collected test data includes: Define a sliding window of size 3×3 and place it on the first pixel of the appearance image in the test data; Extract the grayscale values of all pixels in the sliding window and sort them in ascending order. Take the grayscale value in the middle position after sorting and assign it to the pixel in the center of the window to complete the noise reduction of this pixel. With a step size of 1, the sliding window is moved rightward and downward on the image in sequence until all pixels on the image are processed to obtain the denoised image; Obtain the frequency of each gray level in the denoised image, and calculate the probability and cumulative distribution function of each gray level based on the frequency; Map the cumulative distribution function to a new grayscale range, and replace the grayscale value of each pixel in the image according to the new grayscale mapping relationship to obtain an enhanced image; The Canny edge detection algorithm is used to perform edge detection processing on the enhanced image to obtain an image after edge detection; The electrical parameter data and thermal performance data in the test data are normalized using the maximum-minimum normalization method to obtain normalized electrical parameter data and thermal performance data; The preprocessed test data is obtained by integrating the edge detected image, normalized electrical parameter data and thermal performance data.

3. The method for testing integrated circuit packaging quality based on test data according to claim 2, wherein: The method of performing edge detection on the enhanced image using the Canny edge detection algorithm to obtain an image after edge detection includes: Define multiple Gaussian kernels of different scales, perform convolution operations on each Gaussian kernel and the enhanced image to obtain multiple groups of smoothed images, perform weighted fusion on the multiple groups of smoothed images, and obtain the final smoothed image through weighted summation; For the smoothed image, the gradient is calculated using convolution kernels with adaptive window sizes in the horizontal and vertical directions respectively; Calculate the mean and standard deviation of the gradient amplitude in the neighborhood of each pixel based on the local contrast enhancement factor; Traverse each pixel in the image, obtain the gradient direction of the current pixel, and determine the two adjacent pixels in the gradient direction based on the gradient direction. If the gradient magnitude of the current pixel is not a local maximum and the weighted difference between it and the gradient magnitude of the adjacent pixel is less than the threshold, then set it to 0; otherwise, retain the gradient magnitude of the edge pixel; Count the distribution of gradient amplitudes in the image, calculate the mean and standard deviation, dynamically set high and low thresholds, identify pixels with gradient amplitudes greater than the high threshold as edge pixels, and exclude pixels with gradient amplitudes less than the low threshold; For pixel points whose gradient amplitude is between the high threshold and the low threshold, the edge connection is performed using the region growing method to obtain the image after edge detection.

4. The method for testing integrated circuit packaging quality based on test data according to claim 1, wherein: The extracting feature information from the preprocessed test data includes: The pre-processed test data is divided into multimodal data blocks and spatiotemporal associations are constructed. A structural attention mechanism based on graph convolution is introduced to calculate the structural attention weight of each data block. In the query calculation of the ProbSparse self-attention mechanism, the modal category information and spatiotemporal label information of the data block are introduced to perform weighted adjustment on the query, and the structural attention weight and the self-attention weight are weighted fused. The fusion weight is dynamically adjusted to obtain the hybrid attention weight. A multi-layer feature extraction network is constructed. Each layer receives the features and mixed attention weights output by the previous layer. The features extracted by each layer are residually connected, and the features extracted by the current layer are added to the features of the previous layer as the input of the next layer. In the last layer of the network, the features extracted from different modalities are spliced and fused to obtain the final feature information.

5. The method for testing integrated circuit packaging quality based on test data according to claim 4, wherein: The multimodal data segmentation and spatiotemporal association construction of the pre-processed test data includes: The pre-processed test data is divided into three categories according to the modality: appearance image data, electrical parameter data and thermal performance data; The appearance image data is spatially divided into blocks according to different areas of the integrated circuit package, and each block forms an independent data unit; For electrical parameter data and thermal performance data, the data are divided into equal intervals according to the time series, and overlapping windows are set between adjacent blocks; Add spatiotemporal labels to each data block and construct the spatiotemporal correlation matrix between data blocks of different modalities.

6. The method for testing integrated circuit packaging quality based on test data according to claim 1, wherein: The method of constructing an integrated circuit packaging quality assessment model using the extracted feature information includes: Based on the extracted feature information, a feature semantic association graph is constructed, and the input feature information of the underlying model layer is constructed. A heterogeneous model fusion architecture is adopted in the middle layer, including a DNN network to capture the nonlinear relationship of features, a GNN network to mine the structural information in the feature semantic association graph, and a lightweight decision tree integration model for fast feature classification. The weights of different network outputs in the middle layer are calculated through the attention mechanism to perform feature-level fusion. A fusion decision layer is set at the top layer to receive the fused features of the middle layer, and the cosine annealing algorithm is used for dynamic adjustment to output the final evaluation results.

7. The method for testing integrated circuit packaging quality based on test data according to claim 6, wherein: The dynamic adjustment using the cosine annealing algorithm includes: Calculate the variance and entropy of the features after fusion in the intermediate layer to obtain statistics, determine the initial temperature parameters based on the feature statistics, and establish the association mapping between the temperature parameters and feature categories; Define the base learning rate and calculate it based on the current training round and the total training rounds. Calculate the differences between different category features in the current training batch and adjust the base learning rate based on the differences. Calculate the importance score of each network output feature in the middle layer, calculate the weight adjustment of each network output feature based on the importance score and the current temperature parameter, update the weight of each network output feature, and perform normalization; A decision threshold is set. When the temperature parameter is lower than the decision threshold, the system enters the stable decision stage. In the stable decision stage, the two strategies of weighted voting and probability output are combined to determine the final evaluation result.

8. An integrated circuit packaging quality inspection system based on test data, characterized in that: The system includes: an acquisition module, configured to acquire a variety of test data of the integrated circuit package, wherein the test data includes appearance image data, electrical parameter data, and thermal performance data; The preprocessing module is used to preprocess the collected test data, including noise reduction, enhancement, and edge detection processing of the appearance image data, as well as normalization processing of the electrical parameter data and thermal performance data, to obtain the preprocessed test data; A construction module is used to extract feature information from the preprocessed test data and construct an integrated circuit packaging quality assessment model using the extracted feature information; The input module is used to input the test data of the integrated circuit package to be tested into the quality assessment model after preprocessing and feature extraction, so as to obtain the quality assessment result of the integrated circuit package.

9. An integrated circuit packaging quality inspection device based on test data, characterized in that: The integrated circuit packaging quality inspection device based on test data includes a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the integrated circuit packaging quality inspection device based on test data performs the various steps of the integrated circuit packaging quality inspection method based on test data as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the various steps of the integrated circuit packaging quality detection method based on test data as described in any one of claims 1 to 7 are implemented.

Images