Intelligent data analysis system for semiconductor packaging detection
Through multimodal data analysis and convolutional neural network, the problem of single data source and fixed feature extraction in semiconductor packaging detection is solved, efficient and accurate packaging quality evaluation and production line optimization are achieved, and detection accuracy and production efficiency are improved.
Patent Information
- Application Number
- CN202510338274.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-21
- Publication Date
- 2025-07-08
AI Technical Summary
The existing semiconductor packaging detection technology relies on a single data source and fixed feature extraction method, resulting in low reliability of detection results and it is difficult to achieve ideal results in complex and changeable packaging environments.
A multimodal data analysis system is adopted, combining image data, temperature distribution data and ultrasonic reflected signal data, features are extracted through grayscale histograms, statistical analysis and spectrum analysis, and feature fusion and model training are used for attention mechanism and convolutional neural network to dynamically adjust the importance of features to achieve efficient detection.
It improves the accuracy and efficiency of inspection, can evaluate packaging quality in real time and automatically adjust production line parameters, significantly improving production efficiency and product quality.
Smart Images
Figure CN120277607A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of semiconductor package testing, and particularly to an intelligent data analysis system for semiconductor package testing. Background Art
[0002] The development of semiconductor packaging technology has evolved from simple pin - in - package to complex three - dimensional integrated packaging. With the continuous progress of integrated circuit technology, the integration degree of semiconductor devices is getting higher and higher, and the packaging technology is becoming increasingly complex. Traditional semiconductor package testing mainly relies on manual visual inspection and some basic physical testing methods. These methods have problems such as low detection efficiency, high mis - detection rate, and difficulty in quantification. In recent years, with the rapid development of machine learning and artificial intelligence technologies, intelligent data analysis systems have gradually been applied to the field of semiconductor package testing, significantly improving the accuracy and efficiency of detection.
[0003] Although the existing semiconductor package testing technologies have made certain progress, there are still some deficiencies. First, traditional testing methods often rely on a single type of data source, such as image data or temperature data, lacking the ability to comprehensively analyze multi - modal data. This single - data - source method cannot fully reflect various potential problems in the packaging process, resulting in relatively low reliability of the detection results. Second, the existing detection systems are relatively simple in feature extraction and fusion. Usually, fixed feature extraction methods are adopted, lacking in - depth exploration of the interaction between different features. This makes it difficult for the system to achieve an ideal detection effect when facing complex and variable packaging environments. Therefore, it is particularly necessary to develop an intelligent data analysis system that can integrate multiple data sources, dynamically adjust the importance of features, and have high - efficiency detection capabilities. Summary of the Invention
[0004] In view of the above existing problems, the present invention is proposed.
[0005] Therefore, the present invention provides an intelligent data analysis system for semiconductor package testing to solve the problems of insufficient feature fusion and low reliability of detection results caused by a single data source and fixed feature extraction methods in the prior art.
[0006] To solve the above technical problems, the present invention provides the following technical solutions:
[0007] The present invention provides an intelligent data analysis system for semiconductor package detection, which includes: a data acquisition and preprocessing module: collecting package data and preprocessing the package data; a feature extraction and fusion module: extracting package features from the preprocessed package data and fusing the extracted package features; a model construction module: constructing a package detection model based on historical package data; a detection and analysis module: inputting the fused package features into the package detection model to obtain a package quality result; a control and adjustment module: adjusting production line parameters based on the package quality result.
[0008] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: the package data includes image data, temperature distribution data, and ultrasonic reflection signal data during the packaging process;
[0009] The preprocessing of the package data includes data cleaning, data standardization, and data normalization.
[0010] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: the specific steps of extracting package features from the preprocessed package data are as follows,
[0011] Using a grayscale histogram, convert the image into a grayscale image, obtain the number of pixels at each grayscale level to form a grayscale histogram, and normalize the histogram to extract picture features;
[0012] Using statistical methods, perform statistical analysis on the preprocessed temperature distribution data to extract key temperature features;
[0013] Using spectral analysis, perform spectral analysis on the preprocessed ultrasonic reflection signal data to extract frequency domain features.
[0014] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: the specific process of fusing the extracted package features is as follows,
[0015] Using feature interpolation technology to adjust different types of package features of image data, ultrasonic reflection signal data, and temperature distribution data to the same dimension;
[0016] Stitch the adjusted package features to obtain preliminarily fused package features;
[0017] Using an MLP to add cross-connections to the package features of the same dimension in the feature fusion network to enhance the interaction between package features;
[0018] Using an attention mechanism to calculate the importance scores of different package features, and the expression is:
[0019]
[0020] Among them, L i represents the importance score of the i-th encapsulation feature, and w i represents the weight vector of the i-th encapsulation feature in the attention mechanism, and w j represents the weight vector of the j-th encapsulation feature in the attention mechanism, T represents the transpose of the weight vector, and d i represents the bias term of the i-th encapsulation feature in the attention mechanism, and d j represents the bias term of the j-th encapsulation feature in the attention mechanism, R represents the preliminary fusion of the encapsulation features, n represents the number of encapsulation features, and j and i are index variables of the encapsulation features;
[0021] By using the importance score L of each encapsulation feature i as the weight to weight different features, the importance of different encapsulation features is dynamically adjusted;
[0022] Based on the importance of different encapsulation features, the encapsulation features of different modalities are fused using the splicing method to obtain the fused encapsulation features.
[0023] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: the steps of constructing the package detection model based on historical package data are as follows,
[0024] Use a convolutional neural network as the basic model;
[0025] The input layer is set to match the dimension of the Input and the fused historical package features;
[0026] Use multiple convolutional layers to extract the local information of the package features;
[0027] Reduce the size of the feature map through the pooling layer and retain the important feature information;
[0028] The fully connected layer maps the features extracted by the convolutional layer and the pooling layer to a high-dimensional feature space;
[0029] Use the Softmax activation function as the output layer to output the package quality score;
[0030] Use the cross-entropy loss function and the adaptive learning mechanism to train the package detection model.
[0031] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: the steps of training the package detection model using the cross-entropy loss function and the adaptive learning mechanism are as follows,
[0032] Divide the historical package data into a training set and a validation set by the random sampling method;
[0033] Install the deep learning framework using pip;
[0034] Define the cross-entropy loss function by using the built-in functions provided by the deep learning framework;
[0035] The deep learning framework provides the Adam optimizer. Initialize the relevant parameters according to the requirements of the Adam optimizer, create an initialized parameter instance using the constructor of the Adam optimizer, initialize the adaptive learning mechanism based on the initialized parameter instance through the constructor of the Adam optimizer, and dynamically adjust the learning rate of the initialized adaptive learning mechanism;
[0036] Start training the encapsulated detection model by inputting the training set through the defined cross-entropy loss function and the adjusted adaptive learning mechanism. Within each Epoch, the encapsulated detection model will traverse the entire training set, calculate the loss between the predicted value and the true value, and at the end of each Epoch, use the validation set to test the performance of the encapsulated detection model.
[0037] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: Dynamically adjust the learning rate of the initialized adaptive learning mechanism, and the specific steps are as follows,
[0038] Set the initial learning rate and hyperparameters, calculate the gradient of the current parameters in each training step, use the Adam optimizer to accumulate the historical gradient information, and dynamically adjust the learning rate based on the historical gradient information and the current gradient using the Adam optimizer.
[0039] As a preferred solution of the intelligent data analysis system for semiconductor package detection described in the present invention, wherein: The steps of inputting the fused package features into the package detection model to obtain the package quality result are as follows,
[0040] Input the fused package features into the package detection model, and calculate the package quality score. The expression is:
[0041]
[0042] Wherein, f(F) is the package quality score, f is the fused package feature, ∥f∥2 is the L2 Euclidean norm of the fused package feature f, W1 is the weight matrix of the linearly combined and normalized package features, S(f) is the attention mechanism function, W2 is the weight matrix of the package features processed by the linearly combined attention mechanism, and b is the bias vector;
[0043] Set the package quality level, and based on the package quality score f(F), judge the package quality by setting a threshold to obtain the package quality result.
[0044] As a preferred solution of the intelligent data analysis system for semiconductor package detection according to the present invention, wherein: the package quality is judged by the package quality score f(F), and the specific steps are as follows.
[0045] Set three package quality levels of high quality, medium quality and low quality, and set thresholds t1 and t2 based on historical package detection data;
[0046] Judge the package quality according to the set thresholds and the package quality score. The expression is:
[0047]
[0048] Wherein, represents the package quality result, A represents high quality, B represents medium quality, C represents low quality, t1 is the threshold for low quality, and t2 is the threshold for high quality.
[0049] As a preferred solution of the intelligent data analysis system for semiconductor package detection according to the present invention, wherein: based on the package quality result, the control system adjusts the production line parameters, and the specific steps are as follows.
[0050] When the control system receives a low-quality result, it sends a stop instruction through the PLC and calls the automated inspection process to inspect the low-quality packages;
[0051] According to the inspection results, the control system adjusts the production line parameters through a composite function. The expression is:
[0052]
[0053] Wherein, P new represents the adjusted production parameter value, P old represents the current production parameter value, α represents the adjustment coefficient of the update step size, m is the number of production factors, w e is the weight of the e-th production factor, g e (Q e ) represents the influence function of the e-th production factor, Q e represents the actual value of the e-th production factor, and e is the index variable of the production factor.
[0054] The beneficial effects of the present invention are as follows: through the attention mechanism in the feature extraction and fusion module, the dynamic adjustment of the importance of different package features is realized, improving the accuracy of feature fusion and the accuracy of detection; through the composite function in the control adjustment module to adjust the production line parameters, the present invention realizes the rapid response and precise adjustment of low-quality packages, improving the production efficiency and product quality. Description of the Drawings
[0055] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.
[0056] Figure 1 It is the system diagram of the intelligent data analysis system for semiconductor package detection in Embodiment 1.
[0057] Figure 2 It is the flowchart for constructing the package detection model in Embodiment 1. Specific Embodiments
[0058] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will provide a detailed description of the specific embodiments of the present invention in conjunction with the drawings of the specification.
[0059] In the following description, many specific details are set forth to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Those skilled in the art can make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below.
[0060] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that can be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment that is mutually exclusive with other embodiments.
[0061] Embodiment 1, referring to Figure 1 and Figure 2 , is the first embodiment of the present invention. This embodiment provides an intelligent data analysis system for semiconductor package detection, including the following steps:
[0062] The package data includes image data, temperature distribution data, and ultrasonic reflection signal data during the packaging process.
[0063] It should be noted that a high-definition camera is used to collect the image data during the packaging process, an infrared sensor is used to collect the temperature distribution data, and an ultrasonic probe is used to collect the ultrasonic reflection signal data.
[0064] The preprocessing of the package data includes data cleaning, data standardization, and data normalization.
[0065] It should be noted that for data cleaning, samples including a large number of missing values are deleted, outliers in the data are identified and corrected, and duplicate records are deleted to ensure the uniqueness of the data;
[0066] Data standardization, calculate the mean and standard deviation of each feature respectively, and standardize each feature using the mean and standard deviation;
[0067] Data normalization, calculate the maximum and minimum values of each feature respectively, and normalize each feature using the maximum and minimum values.
[0068] Extract encapsulated features from the preprocessed encapsulated data.
[0069] Use the grayscale histogram to convert the image into a grayscale image, obtain the number of pixels at each grayscale level, form a grayscale histogram, normalize the histogram, and extract image features.
[0070] It should be noted that the image features obtained using the grayscale histogram are expressed as:
[0071]
[0072] where Fu represents the image features, and I gray (u, v) is the grayscale value of the grayscale image at (u, v), M is the total number of grayscale levels, l represents the grayscale level, and (u, v) is the pixel position coordinates in the grayscale image.
[0073] Use statistical methods to perform statistical analysis on the preprocessed temperature distribution data and extract key temperature features.
[0074] It should be noted that the key temperature features extracted using statistical methods are expressed as:
[0075]
[0076] where TF is the temperature feature, N is the number of temperature data samples, T a is the ath temperature data point in the preprocessed temperature distribution data vector, is the mean value of the temperature data, s(T) is the standard deviation of the temperature data, and a is the index variable of the temperature data point.
[0077] Use spectral analysis to perform spectral analysis on the preprocessed ultrasonic reflection signal data and extract frequency domain features.
[0078] It should be noted that the frequency domain features extracted using spectral analysis are obtained by performing a Fourier transform on the ultrasonic reflection signal to convert it from the time domain to the frequency domain and calculating the amplitude at each frequency point, and the expression is:
[0079]
[0080] where |X(y k )| represents the frequency domain information at frequency yk The amplitude at, D is the number of sampling points of the time-domain signal, x(r) is the r-th sampling point of the time-domain signal, y k is the k-th frequency point, r is the sampling point index, and k is the frequency point index;
[0081] Calculate the amplitudes of each frequency point and finally form the frequency-domain features.
[0082] Fuse the extracted package features.
[0083] Use feature interpolation technology to adjust the package features of different types such as image data, ultrasonic reflection signal data, and temperature distribution data to the same dimension.
[0084] Specifically, perform feature extraction on each type of data to obtain their respective feature vectors, and use the linear interpolation method to adjust these feature vectors to the same dimension.
[0085] Concatenate the adjusted package features to obtain the preliminarily fused package features;
[0086] Use MLP to add cross-connections in the feature fusion network for the package features of the same dimension to enhance the interaction between package features.
[0087] Specifically, concatenate all the package features of the same dimension into a high-dimensional feature vector, perform non-linear transformation through a multi-layer perceptron (MLP), and add cross-connections between the hidden layers of the MLP to enhance the interaction between different features.
[0088] Use the attention mechanism to calculate the importance scores of different package features, and the expression is:
[0089]
[0090] Among them, L i represents the importance score of the i-th package feature, w i represents the weight vector of the i-th package feature in the attention mechanism, w j represents the weight vector of the j-th package feature in the attention mechanism, T represents the transpose of the weight vector, d i represents the bias term of the i-th package feature in the attention mechanism, d j represents the bias term of the j-th package feature in the attention mechanism, R represents the preliminarily fused package features, n represents the number of package features, and j and i are the index variables of the package features;
[0091] By taking the importance score L of each package feature i as the weight, weight different features, and dynamically adjust the importance of different package features;
[0092] Based on the importance of different packaging features, the splicing method is used to fuse packaging features of different modalities to obtain fused packaging features.
[0093] Build a packaging detection model based on historical packaging data
[0094] Use a convolutional neural network as the basic model;
[0095] The input layer is set to match the dimension of the Input with the fused historical packaging features;
[0096] Use multiple convolutional layers to extract local information of packaging features;
[0097] Reduce the size of the feature map through the pooling layer and retain important feature information;
[0098] The fully connected layer maps the features extracted by the convolutional layer and the pooling layer to a high-dimensional feature space;
[0099] Use the Softmax activation function as the output layer to output the packaging quality score;
[0100] Use the cross-entropy loss function and the adaptive learning mechanism to train the packaging detection model.
[0101] It should be noted that building a packaging detection model with a convolutional neural network (CNN) as the basic model lies in that CNN can automatically extract local features and spatial hierarchical structures in images, is suitable for identifying patterns and details in images, and is very effective for defect recognition, edge detection, and texture analysis in packaging detection tasks, improving the accuracy of detection.
[0102] Use the cross-entropy loss function and the adaptive learning mechanism to train the packaging detection model.
[0103] Divide the historical packaging data into a training set and a validation set by the random sampling method;
[0104] Use pip to install the deep learning framework;
[0105] Define the cross-entropy loss function by using the built-in functions provided by the deep learning framework;
[0106] The deep learning framework provides the Adam optimizer. Initialize the relevant parameters according to the requirements of the Adam optimizer, and use the constructor of the Adam optimizer to create an initialized parameter instance. Based on the initialized parameter instance, initialize the adaptive learning mechanism through the constructor of the Adam optimizer, and dynamically adjust the learning rate of the initialized adaptive learning mechanism.
[0107] It should be noted that the requirements of the Adam optimizer include the learning rate, the decay rate of the first-moment estimate, the decay rate of the second-moment estimate, and the smoothing term.
[0108] The training set is input into the encapsulation detection model for training by means of the defined cross-entropy loss function and the adjusted adaptive learning mechanism. Within each Epoch, the encapsulation detection model traverses the entire training set, calculates the loss between the predicted value and the true value. At the end of each Epoch, the validation set is used to test the performance of the encapsulation detection model.
[0109] It should be noted that the cross-entropy loss function is used to calculate the difference between the predicted value and the true label. Each sample in the training set is traversed. For each sample, each category is traversed, the logarithmic loss between the true label and the predicted probability is calculated, and the logarithmic losses of all samples and categories are summed up. Finally, the average value is taken to obtain the average loss between the final predicted value and the true value.
[0110] It should also be noted that each sample in the validation set is input into the encapsulation detection model to generate a prediction result, and then the loss value between the predicted value and the true label is calculated using the same cross-entropy loss function.
[0111] Dynamically adjust the learning rate of the adaptive learning mechanism after initialization.
[0112] Set the initial learning rate and hyperparameters, calculate the gradient of the current parameters in each training step, use the Adam optimizer to accumulate historical gradient information, and dynamically adjust the learning rate based on the historical gradient information and the current gradient using the Adam optimizer.
[0113] It should be noted that the learning rate is dynamically adjusted using the Adam optimizer, and the expression is:
[0114]
[0115] where, θ t is the parameter at the t-th iteration, is the initial learning rate, is the first moment estimate after bias correction, is the second moment estimate after bias correction, and ∈ is a small constant.
[0116] Input the fused encapsulation features into the encapsulation detection model to obtain the encapsulation quality result.
[0117] Input the fused encapsulation features into the encapsulation detection model to calculate the encapsulation quality score, and the expression is:
[0118]
[0119] Among them, f(F) is the encapsulation quality score, F is the fused encapsulation feature, ∥F∥2 is the L2 Euclidean norm of the fused encapsulation feature F, W1 is the weight matrix of the linearly combined and normalized encapsulation feature, S(F) is the attention mechanism function, W2 is the weight matrix of the encapsulation feature processed by the linearly combined attention mechanism, and b is the bias vector;
[0120] Set the encapsulation quality level. Based on the encapsulation quality score f(F), judge the encapsulation quality by setting thresholds, and obtain the encapsulation quality result.
[0121] It should be noted that the weight matrix W1 of the linearly combined and normalized encapsulation feature has the following expression:
[0122]
[0123] Among them, h is the output feature dimension and q is the input feature dimension;
[0124]
[0125] Among them, z is the dimension of the feature processed by the attention mechanism.
[0126] Judge the encapsulation quality through the encapsulation quality score f(F).
[0127] Set three encapsulation quality levels: high quality, medium quality, and low quality. Set thresholds t1 and t2 based on historical encapsulation detection data;
[0128] Judge the encapsulation quality according to the set thresholds and the encapsulation quality score. The expression is:
[0129]
[0130] Among them, represents the encapsulation quality result, A represents high quality, B represents medium quality, C represents low quality, t1 is the threshold for low quality, and t2 is the threshold for high quality.
[0131] It should be noted that the thresholds t1 and t2 are set based on historical encapsulation detection data. Statistical analysis is performed on the historical encapsulation detection data to find out the distribution of different quality levels. Based on the statistical analysis results, appropriate thresholds t1 and t2 are set.
[0132] Based on the encapsulation quality result, the control system adjusts the production line parameters.
[0133] When the control system receives a low-quality result, it sends a stop instruction through the PLC and calls the automated inspection process to inspect the low-quality encapsulation;
[0134] According to the inspection result, the control system adjusts the production line parameters through a composite function. The expression is:
[0135]
[0136] Among them, P new represents the adjusted production parameter value, P old represents the current production parameter value, α represents the adjustment coefficient of the update step size, m is the number of production factors, and w e is the weight of the e-th production factor, and g e (Q e ) represents the influence function of the e-th production factor, and Q e represents the actual value of the e-th production factor, and e is the index variable of the production factor.
[0137] It should be noted that the influence function g e (Q e ) of the e-th production factor has the following expression:
[0138]
[0139] Among them, Q min,e is the minimum value of the e-th production factor, and Q max,e is the maximum value of the e-th production factor.
[0140] In summary, the present invention realizes efficient feature extraction and fusion of multi-modal packaging data by introducing an attention mechanism and a convolutional neural network, significantly improving the accuracy and reliability of packaging detection. The attention mechanism dynamically adjusts the importance of different features, enhancing the robustness and generalization ability of the model. The convolutional neural network automatically learns and extracts complex features through multi-layer convolution and pooling operations, improving the accuracy and efficiency of feature extraction. Dynamically adjusting the learning rate optimizes the model training process and speeds up the convergence speed. In practical applications, the present invention can evaluate the packaging quality in real time, timely detect and correct quality problems, automatically adjust the production line parameters through the control system, optimize the production process, significantly improve the production efficiency and product quality, reduce the generation of defective products, and lower the production cost.
[0141] Example 2. Referring to Table 1, this is the second example of the present invention. To further verify the technical solution of the present invention, experimental simulation data of an intelligent data analysis system for semiconductor packaging detection is given.
[0142] In the experimental preparation stage, three different types of semiconductor package samples were first selected: standard package (SampleA), package with minor defects (Sample B), and package with severe defects (Sample C). To ensure the accuracy and consistency of the data, high-precision devices such as high-resolution industrial cameras, thermal imagers, and ultrasonic flaw detectors were used to collect image data, temperature distribution data, and ultrasonic reflection signal data during the packaging process respectively. Subsequently, these raw data were cleaned, standardized, and normalized through specially developed data preprocessing software to eliminate noise and outliers and ensure the quality of the data. In addition, algorithm tools for feature extraction and fusion, as well as a package detection model trained based on historical packaging data, were also prepared. The entire preparation stage was carried out strictly in accordance with the requirements of the experimental design, laying a solid foundation for subsequent detection and analysis.
[0143] First, the experiment collected image data, temperature distribution data, and ultrasonic reflection signal data of three different types of semiconductor package samples through high-precision devices and preprocessed these data to ensure the accuracy and consistency of the data. Second, key package features were extracted from the preprocessed data, and these features were fused through feature interpolation technology, multi-layer perceptron (MLP), and attention mechanism to form the final fused package features. Then, the fused package features were input into the package detection model trained based on historical data to obtain the package quality results of each sample, and the production line parameters were adjusted according to these results. Finally, by comparing the detection accuracy, speed, and error rate of the intelligent data analysis system of the present invention with traditional detection methods, the significant advantages of the present invention in improving detection efficiency and quality were verified.
[0144] The traditional method specifically refers to traditional manual or semi-automatic detection methods, which usually include manual visual inspection, manual measurement of temperature distribution, and the use of traditional ultrasonic detection equipment.
[0145] As shown in Table 1 below:
[0146] Table 1 Comparison Table of Semiconductor Package Detection Performance
[0147]
[0148]
[0149] Through the data analysis of the above table, it can be clearly seen that the intelligent data analysis system of the present invention has significant advantages over traditional methods in many aspects. The present invention realizes higher detection accuracy, faster detection speed and lower error rate through advanced data preprocessing, feature extraction and fusion technologies, and a deep learning-based package detection model. For example, for the standard package (Sample A), the detection accuracy of the present invention reaches 99.4%, while the traditional method is only 94.5%; in terms of detection speed, the present invention can reach 238 samples per minute, while the traditional method is only 155 samples / minute; in terms of error rate, the present invention is only 0.6%, while the traditional method is as high as 4.8%. These data fully demonstrate the innovation and superiority of the present invention in improving the efficiency and quality of semiconductor package detection.
[0150] The intelligent data analysis system of the present invention has significantly improved the detection accuracy (such as the standard package detection accuracy has increased from 94.5% to 99.4%), accelerated the detection speed (such as the standard package detection speed has increased from 155 samples / minute to 238 samples / minute), reduced the error rate (such as the standard package error rate has decreased from 4.8% to 0.6%) in semiconductor package detection, and through multi-modal data fusion and automatic adjustment of production line parameters, a more efficient and reliable production process has been achieved.
[0151] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.
Claims
1. An intelligent data analysis system for semiconductor package detection, characterized in that: including a data acquisition and preprocessing module, which acquires encapsulated data and preprocesses the encapsulated data; a feature extraction and fusion module, which extracts encapsulated features from the preprocessed encapsulated data and fuses the extracted encapsulated features; a model construction module, which constructs an encapsulation detection model based on historical encapsulated data; a detection and analysis module, which inputs the fused encapsulated features into the encapsulation detection model to obtain the encapsulation quality result; a control and adjustment module, which adjusts the production line parameters based on the encapsulation quality result.
2. The intelligent data analysis system for semiconductor package detection according to claim 1, characterized in that: The encapsulated data includes image data, temperature distribution data, and ultrasonic reflection signal data during the encapsulation process; The preprocessing of the encapsulated data includes data cleaning, data standardization, and data normalization.
3. The intelligent data analysis system for semiconductor package detection according to claim 2, wherein: The steps of extracting encapsulated features from the preprocessed encapsulated data are as follows: Using a grayscale histogram, convert the image to a grayscale image, obtain the number of pixels for each grayscale level to form a grayscale histogram, normalize the histogram, and extract image features; Using statistical methods, perform statistical analysis on the preprocessed temperature distribution data to extract key temperature features; Using spectral analysis, perform spectral analysis on the preprocessed ultrasonic reflection signal data to extract frequency domain features.
4. The intelligent data analysis system for semiconductor package detection according to claim 3, wherein: The process of fusing the extracted encapsulated features is as follows: Use feature interpolation technology to adjust encapsulated features of different types, such as image data, ultrasonic reflection signal data, and temperature distribution data, to the same dimension; Concatenate the adjusted encapsulated features to obtain preliminarily fused encapsulated features; Use an MLP to add cross connections to the encapsulated features of the same dimension in the feature fusion network to enhance the interaction between encapsulated features; Use an attention mechanism to calculate the importance scores of different encapsulated features. The expression is: Among them, L i represents the importance score of the i-th encapsulation feature, w i represents the weight vector of the i-th encapsulation feature in the attention mechanism, w j represents the weight vector of the j-th encapsulation feature in the attention mechanism, T represents the transpose of the weight vector, d i represents the bias term of the i-th encapsulation feature in the attention mechanism, d j represents the bias term of the j-th encapsulation feature in the attention mechanism, R represents the preliminary fusion of encapsulation features, n represents the number of encapsulation features, and j and i are index variables of the encapsulation features; By taking the importance score L of each encapsulation feature i as the weight to weight different features, dynamically adjust the importance of different encapsulation features; Based on the importance of different encapsulated features, use the concatenation method to fuse encapsulated features of different modalities to obtain fused encapsulated features.
5. The intelligent data analysis system for semiconductor package detection according to claim 4, wherein: The steps of constructing an encapsulation detection model based on historical encapsulated data are as follows: Use a convolutional neural network as the basic model; The input layer is set to match the dimension of the Input with the fused historical encapsulated features; Use multiple convolutional layers to extract local information of the encapsulated features; Reduce the size of the feature map through a pooling layer and retain important feature information; The fully connected layer maps the features extracted by the convolutional layer and the pooling layer to a high-dimensional feature space; Use the Softmax activation function as the output layer to output the encapsulation quality score; Use the cross-entropy loss function and an adaptive learning mechanism to train the encapsulation detection model.
6. The intelligent data analysis system for semiconductor package detection according to claim 5, wherein: The steps of using the cross-entropy loss function and an adaptive learning mechanism to train the encapsulation detection model are as follows: Divide the historical encapsulated data into a training set and a validation set by the random sampling method; Use pip to install a deep learning framework; Define the cross-entropy loss function by using the built-in functions provided by the deep learning framework; The deep learning framework provides an Adam optimizer. Initialize the relevant parameters according to the requirements of the Adam optimizer, and use the constructor of the Adam optimizer to create an initialized parameter instance. Based on the initialized parameter instance, initialize the adaptive learning mechanism through the constructor of the Adam optimizer, and dynamically adjust the learning rate of the initialized adaptive learning mechanism. The training set input is encapsulated into the detection model for training using the defined cross-entropy loss function and the adjusted adaptive learning mechanism. Within each Epoch, the encapsulated detection model traverses the entire training set, calculates the loss between the predicted value and the true value, and at the end of each Epoch, the validation set is used to test the performance of the encapsulated detection model.
7. The intelligent data analysis system for semiconductor package detection according to claim 6, wherein: The learning rate of the adaptive learning mechanism after initialization is dynamically adjusted as follows. Set the initial learning rate and hyperparameters, calculate the gradient of the current parameters in each training step, use the Adam optimizer to accumulate historical gradient information, and dynamically adjust the learning rate using the Adam optimizer based on the historical gradient information and the current gradient.
8. The intelligent data analysis system for semiconductor package detection according to claim 7, wherein: The above-mentioned step of inputting the fused encapsulation features into the encapsulated detection model to obtain the encapsulation quality result is as follows. Input the fused encapsulation features into the encapsulated detection model, and calculate the encapsulation quality score. The expression is: where f(F) is the encapsulation quality score, F is the fused encapsulation feature, ∥F∥2 is the L2 Euclidean norm of the fused encapsulation feature F, W1 is the weight matrix of the linearly combined and normalized encapsulation features, S(F) is the attention mechanism function, W2 is the weight matrix of the linearly combined encapsulation features processed by the attention mechanism, and b is the bias vector; Set the encapsulation quality level. Based on the encapsulation quality score f(F), judge the encapsulation quality by setting a threshold to obtain the encapsulation quality result.
9. The intelligent data analysis system for semiconductor package detection according to claim 8, wherein: Judge the encapsulation quality through the encapsulation quality score f(F), and the specific steps are as follows. Set three encapsulation quality levels: high quality, medium quality, and low quality, and set thresholds t1 and t2 based on historical encapsulation detection data; Judge the encapsulation quality according to the set threshold and the encapsulation quality score. The expression is: Among them, represents the encapsulation quality result, A represents high quality, B represents medium quality, C represents low quality, t1 is the threshold for low quality, and t2 is the threshold for high quality.
10. The intelligent data analysis system for semiconductor package detection according to claim 9, characterized in that: The above-mentioned step of the control system adjusting the production line parameters based on the encapsulation quality result is as follows. When the control system receives a low-quality result, it sends a stop instruction through the PLC and calls the automated inspection process to inspect the low-quality encapsulation; According to the inspection result, the control system adjusts the production line parameters through a composite function. The expression is: Among them, P new represents the adjusted production parameter value, P old represents the current production parameter value, a represents the adjustment coefficient of the update step size, m is the number of production factors, w e is the weight of the e-th production factor, g e (Q e ) represents the influence function of the e-th production factor, Q e represents the actual value of the e-th production factor, and e is the index variable of the production factor.