Computer board detection method, device and system
Through a combination of software testing and micro-nano-level three-dimensional scanning, functional and physical data are integrated to identify micro-defects of computer boards, and through adaptive adjustment, the problem that existing detection methods cannot fully capture micro-defects and adapt to changes is solved. , achieving more efficient defect identification and detection system flexibility.
Patent Information
- Application Number
- CN202311513544.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-13
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-11-13
AI Technical Summary
Existing computer board detection methods cannot fully capture microscopic defects, and lack flexibility to adapt to changes in production processes or newly designed board designs.
Software test cases are used to simulate multiple operating environments, combine micro-nano-level three-dimensional scanner to scan surfaces and internal structures, integrate functional behavior data and physical structure data to identify potential defects, and adjust test cases and scanning parameters through adaptive adjustments , enhance the adaptability of the detection system.
It improves the accuracy of defect identification and the flexibility of the detection system, can quickly adapt to production line changes and new designs, reduce the risk of missed inspection and missed inspection, and improve production efficiency and economic benefits.
Smart Images

Figure CN118053218B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of computer board detection, and in particular to a method, device and system for detecting a computer board. Background Art
[0002] In the modern electronic manufacturing world, ensuring the quality and reliability of computer boards is critical. As technology continues to advance, the complexity of computer boards is also increasing, which has led to a continuous need for innovative inspection methods. Traditional inspection methods, such as visual inspection or testing based on specific functions, often fail to fully capture all potential quality issues, especially those microscopic defects that are not easily discovered under normal operating conditions. In addition, a single test method may lack flexibility and cannot adapt to changes in the production process or respond to newly introduced board designs.
[0003] Therefore, the industry urgently needs a testing method that can comprehensively evaluate the performance of computer boards under various operating conditions. Summary of the invention
[0004] The present application provides a computer board detection method, device and system to improve the accuracy of computer board defect identification.
[0005] The present application provides a computer board detection method, comprising:
[0006] Run software test cases to simulate the working state of the computer board under various operating environments including temperature, current and data transmission to obtain functional behavior data of the computer board;
[0007] While running the software test, a micro-nano three-dimensional scanner is used to scan the surface and internal structure of the computer board to collect physical structure data of the computer board, wherein the physical structure data is used to identify physical defects of the computer board at a microscopic scale;
[0008] Using an integrated algorithm to perform fusion analysis on the functional behavior data and the physical structure data, wherein the fusion analysis is based on the fact that the final performance of the computer board is affected by the functional state and the physical state, thereby identifying defects that cannot be detected by a single method;
[0009] According to the processing result of the fusion analysis processing, determining whether the computer board has defects;
[0010] If a computer board has defects, it is determined to be a defective product, and the defective computer board is removed from the production line, and the software test cases and the scanning parameters of the micro-nanoscale 3D scanner are adaptively adjusted; wherein the adaptive adjustment is based on the results of a detailed analysis of the defective products, so as to diagnose and incorporate computer board defect types that have not been fully covered before in the computer board inspection process, and ensure that adjustments are made to changes in the production line conditions of the computer boards or newly introduced board designs.
[0011] Furthermore, the computer board detection method further includes:
[0012] The detection data during the detection process is stored, and the detection data includes functional behavior data, physical structure data, and processing results of fusion analysis and processing. The detection data is used to continuously optimize software test cases and 3D scanning parameters to achieve self-calibration and self-optimization of the detection system, and enhance the system's adaptability to quality differences of various types of computer boards and responsiveness to changes in the production process.
[0013] Furthermore, the use of an integrated algorithm to perform fusion analysis on the functional behavior data and the physical structure data includes:
[0014] The functional behavior data and physical structure data are converted into standardized feature vectors using multidimensional scaling analysis and principal component analysis;
[0015] Using the standardized feature vectors, a deep learning-based feature learning model is used to identify and extract the potential associations between functional and physical features by learning the intrinsic structure of the data, thereby generating a comprehensive feature representation for each computer board.
[0016] inputting the comprehensive feature representation into a fuzzy logic-based ensemble learning framework, wherein the ensemble learning framework intelligently matches the comprehensive feature representation with a set of predefined defect patterns to identify potential defects that cannot be detected by a single test;
[0017] Employs causal reasoning-based interpretation algorithms to analyze the causes of potential defects and transforms complex data and patterns into intuitive reports and charts, providing technicians with a clear visual and theoretical basis to understand the inspection results for further decision-making and action.
[0018] Furthermore, the use of multidimensional scaling analysis and principal component analysis to convert the functional behavior data and physical structure data into standardized feature vectors includes:
[0019] Performing data normalization on the functional behavior data and physical structure data respectively, calculating the deviation of each data point in each data set relative to the mean of its data set, and using the deviation to normalize the original data points, thereby forming two independent and normalized functional behavior data sets and physical structure data sets;
[0020] Multidimensional scaling analysis was applied to the normalized functional behavior dataset and physical structure dataset, respectively, by constructing a distance matrix between data points and performing nonlinear dimensionality reduction techniques to generate corresponding low-dimensional space coordinates for each dataset;
[0021] Integrate the low-dimensional space coordinates corresponding to the functional behavior dataset and the physical structure dataset to generate a joint data matrix, and calculate the covariance matrix on the joint data matrix;
[0022] Performing principal component analysis using the covariance matrix to determine the main variation directions of the joint data set and extracting the eigenvectors corresponding to the maximum eigenvalues to form a comprehensive set of principal component eigenvectors;
[0023] The comprehensive principal component feature vector set is resized so that the features of each dimension have the same magnitude, so as to obtain a standardized feature vector.
[0024] Furthermore, the feature learning model includes:
[0025] Input layer, used to receive the normalized feature vector;
[0026] A dual-channel feature processing layer, comprising two parallel sub-networks, wherein the first sub-network is configured with a deep network structure optimized for time series data features, and the second sub-network is configured with a deep network structure optimized for spatial data features, each sub-network extracts the time series features and the spatial features in the fused feature vector respectively, and outputs two independent feature representations;
[0027] An association learning layer, configured with a custom attention mechanism, for integrating the output of the dual-channel feature processing layer to generate a feature representation by learning the potential association between the time series feature representation and the spatial feature representation;
[0028] A deep feature fusion layer, used for receiving the output of the association learning layer, further integrating and refining the comprehensive feature representation through a series of fully connected layers, and forming a comprehensive feature representation for characterizing the state of the computer board;
[0029] The output layer is configured with a fully connected layer and a Softmax activation function, which is used to convert the comprehensive feature representation of the deep feature fusion layer into a probability distribution or classification label describing the possible states of the computer board.
[0030] Furthermore, the loss function L of the feature learning model is defined as:
[0031]
[0032] Among them, C is the total number of categories of the output layer, y m is the one-hot encoded vector of the true label, is the corresponding model output, and m is the ordinal number of the sample;
[0033] Among them, L CE is the cross entropy loss, defined as:
[0034]
[0035] Among them, y mn Represents the true label vector y m The value of the nth element in ; Represents the model output vector The value of the nth element in ;
[0036] L WHD is a weighted loss for hard-to-detect defects, defined as:
[0037]
[0038] Among them, y mC is the element of the difficult-to-detect defect category in the true label vector, is the corresponding element in the prediction vector, and α and β are hyperparameters used to balance the two parts of the loss.
[0039] Furthermore, the customized attention mechanism includes:
[0040] The variational autoencoder is used to evaluate the correlation between time series features and spatial features, and determine the joint contribution of time series features and spatial features to defect detection;
[0041] Assign weights to each feature based on the evaluation results of the variational autoencoder;
[0042] Perform feature weighted fusion and merge time series features and spatial features according to the assigned weights to generate feature representation for defect detection.
[0043] Furthermore, the variational autoencoder further includes a regularization term, which penalizes the distribution deviation of the variational autoencoder latent space based on Kullback-Leibler divergence, ensuring that the generated feature representation not only accurately reflects the characteristics of the original data, but also has good generalization ability.
[0044] The present application provides a computer board detection device, comprising:
[0045] A test unit, used to run software test cases to simulate the working state of the computer board under various operating environments including temperature, current and data transmission, so as to obtain functional behavior data of the computer board;
[0046] A scanning unit, used to scan the surface and internal structure of the computer board using a micro-nano three-dimensional scanner while running the software test, so as to collect physical structure data of the computer board, wherein the physical structure data is used to identify physical defects of the computer board at a microscopic scale;
[0047] An analysis unit, configured to perform fusion analysis processing on the functional behavior data and the physical structure data using an integrated algorithm, wherein the fusion analysis processing is based on the fact that the final performance of the computer board is jointly affected by the functional state and the physical state, thereby identifying defects that cannot be detected by a single method;
[0048] A judgment unit, used for judging whether the computer board has a defect according to the processing result of the fusion analysis processing;
[0049] An adjustment unit is used to determine that a computer board is defective, thereby removing the defective computer board from the production line and adaptively adjusting software test cases and scanning parameters of the micro-nanoscale three-dimensional scanner; wherein the adaptive adjustment is based on the results of a detailed analysis of the defective products, so as to diagnose and incorporate computer board defect types that have not been fully covered before in the computer board detection process, and ensure that adjustments are made to changes in production line conditions of computer boards or newly introduced board designs.
[0050] The present application provides a computer board detection system, comprising:
[0051] A data acquisition module, including a central processing unit and a storage unit, wherein the storage unit is pre-installed with operation instructions for simulating the working state of the computer board under various environmental conditions and generating corresponding functional behavior data;
[0052] A micro-nanoscale three-dimensional imaging device, operating synchronously with the data acquisition module, is used to scan the surface and internal structure of the computer board and collect physical structure data, which is used to identify microscopic scale physical defects of the computer board;
[0053] The analysis and processing unit is equipped with data processing hardware and software resources, receives data from the data acquisition module and the 3D imaging device, and uses integrated algorithms to perform fusion analysis to identify potential defects;
[0054] A defect detection module, connected to the analysis and processing unit, is used to evaluate the fusion analysis results and determine whether the computer board meets the quality standards;
[0055] User interface terminal, which provides an interactive interface for technical operators to display test results, integrate analysis feedback and real-time system status, and receive adjustment commands input by users;
[0056] The remote monitoring unit includes at least one network interface for remotely transmitting detection data, system status and alarm information to a central monitoring station or a mobile terminal.
[0057] The technical solution proposed in this application has the following beneficial technical effects:
[0058] (1) By integrating software testing and micro-nano 3D scanning technology, a comprehensive test of the functional behavior and physical structure of computer boards is achieved. This dual detection mechanism can greatly improve the accuracy and reliability of identifying various defects, especially those at the microscopic scale that are difficult to detect with traditional methods.
[0059] (2) Use integrated algorithms for fusion analysis and processing, which can comprehensively consider the joint influence of functional and physical data, thereby revealing potential defects that cannot be detected by a single test method. This all-round detection method makes defect identification more accurate, reduces the risk of missed detection and false detection, and thus improves the quality control level of the entire production line.
[0060] (3) Through detailed analysis of defective products, the method can adaptively adjust software test cases and scanner parameters to enhance the flexibility and adaptability of the inspection system. This dynamic adjustment mechanism enables the inspection system to quickly adapt to changes in production line conditions and the introduction of new board designs, ensuring continuous optimization and improvement of the inspection process.
[0061] (4) It improves production efficiency and economic benefits, reduces rework and waste caused by defects, and provides technicians with a clear basis for diagnosis and decision-making, which helps to improve maintenance efficiency and reduce maintenance costs. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 This is a flow chart of a computer board detection method provided in the first embodiment of the present application.
[0063] Figure 2 It is a schematic diagram of a computer board detection device provided in the second embodiment of the present application. DETAILED DESCRIPTION
[0064] Many specific details are described in the following description to facilitate a full understanding of the present application. However, the present application can be implemented in many other ways than those described herein, and those skilled in the art can make similar generalizations without violating the connotation of the present application, so the present application is not limited by the specific implementation disclosed below.
[0065] The first embodiment of the present application provides a method for detecting a computer board. Figure 1 , which is a schematic diagram of the first embodiment of the present application. Figure 1 The first embodiment of the present application provides a method for detecting a computer board for detailed description.
[0066] Step S101: running software test cases to simulate the working state of a computer board under various operating environments including temperature, current and data transmission, so as to obtain functional behavior data of the computer board.
[0067] When running software test cases to simulate the working status of computer boards in different operating environments, key test cases can be designed to cover multiple aspects to ensure comprehensive testing of the functionality and performance of the board. The following are some exemplary software test cases:
[0068] Temperature limit test: Use software to adjust the temperature control device connected to the computer board, set different temperature points, and simulate the working environment from low temperature to high temperature. For each temperature point, the software test case will record the response and performance of the board to detect whether it can maintain normal operation.
[0069] Voltage and current fluctuation test: Use software to change the power supply voltage and current of the computer board to test the computer board's ability to work under different power conditions and the robustness of the power circuit.
[0070] Computational stress testing: Runs intensive computing tasks, such as complex mathematical operations or graphics processing, to assess the performance limits of the processor and GPU.
[0071] Memory Access and Bandwidth Test: Tests the speed and efficiency of memory and its performance under heavy load by performing large amounts of data read and write operations.
[0072] Data transmission and interface testing: simulate large amounts of data being transmitted through various interfaces of the board (such as USB, SATA, PCIe) to evaluate its data throughput and transmission stability.
[0073] Long-term operation stability test: Test the reliability and error rate of computer boards under continuous long-term operation.
[0074] Power management and energy-saving feature testing: Evaluate the power management system of the computer board, its performance and conversion efficiency in different energy consumption modes.
[0075] Startup and Restart Test: Repeatedly start and restart the computer board to test the reliability of its startup logic and hardware initialization process.
[0076] Peripheral compatibility test: Connect various peripherals, such as storage devices, input devices, etc., to evaluate the compatibility of the board and the peripheral management capabilities.
[0077] Software and driver compatibility testing: Install and run different operating systems and drivers to test software compatibility and system integration capabilities.
[0078] Fault injection testing: artificially injecting faults into software or hardware, such as interrupting data flow or generating erroneous electrical signals, to test the error handling and recovery capabilities of the board.
[0079] Environmental factor testing: Taking into account the actual usage environment, the computer board is tested for environmental factors such as vibration, humidity, and static electricity.
[0080] These test cases can cover some important aspects that the computer board may encounter in its expected use, ensuring that it can work properly under various conditions. Each test case should produce detailed logs and result data that can be used to analyze the functionality and performance of the board. When designing test cases, it is also necessary to consider the specific application scenarios of the board and the specific needs of the user to ensure the practicality and relevance of the test results.
[0081] Functional behavior data refers to the data collected during the execution of software test cases, which reflects the performance and response of the computer board in a specific simulated operating environment. Functional behavior data usually includes but is not limited to the following:
[0082] Processor performance data: such as CPU usage, core temperature, number of executed instructions, processing speed, etc.
[0083] Memory usage: including memory occupancy, read and write speed, access latency, etc.
[0084] Power usage data: such as power consumption per unit time, current and voltage stability, performance changes under load, etc.
[0085] Data transmission and network performance indicators: including data throughput, transmission rate, error rate, connection stability, etc.
[0086] Temperature response data: the working status of the board under different temperature conditions, and the impact of temperature changes on performance.
[0087] Hardware interface response: the performance of various interfaces (such as USB, HDMI, PCIe, etc.) under high load or extreme conditions.
[0088] Software compatibility record: compatibility between the board and the running operating system, drivers and other software.
[0089] Error logs and system stability: Record any errors, system crashes, unexpected reboots, or other unusual behavior that occur during testing.
[0090] The purpose of functional behavior data is to evaluate the performance and stability of computer boards from the software level. By analyzing this data, it is possible to identify possible functional defects of computer boards, predict the problems they may encounter in actual use, and determine where improvements may be needed in their design and manufacturing processes.
[0091] Step S102: while running the software test, a micro-nano three-dimensional scanner is used to scan the surface and internal structure of the computer board to collect physical structure data of the computer board, and the physical structure data is used to identify physical defects of the computer board at a microscopic scale.
[0092] Step S102 includes using a high-resolution micro-nano three-dimensional scanner to perform a comprehensive scan of the surface and internal structure of the computer board. This step includes:
[0093] Select appropriate micro-nano 3D scanning technologies, such as laser scanning, optical interferometry or X-ray computed tomography, to obtain high-resolution images of the microstructure of computer boards. These technologies can reveal surface and internal details at nanoscale resolution, capturing tiny defects that may be missed by conventional inspections.
[0094] The 3D scanning needs to be run synchronously with the software test cases in step S101 to ensure that the physical state of the computer board can be captured in real time while it is running various operations. This synchronization is to capture physical defects that may only appear under certain working conditions.
[0095] Through high-precision scanning, physical structure data on multiple levels of the board are collected. The physical structure data of computer boards refers to the data on the detailed size, shape, material and relative position relationship of each component of the board obtained by micro-nano-level 3D scanners. This type of data reflects the physical characteristics of the board at a microscopic scale, including but not limited to the following aspects:
[0096] Appearance data: Detailed images and measurements of the board’s dimensions, edge flatness, surface roughness, and appearance defects such as scratches, dents, or bumps.
[0097] Interlayer structure data: the precise distance between the layers of a multilayer PCB (Printed Circuit Board), the integrity of the interlayer connections, and any interlayer short circuits or open circuits.
[0098] Solder joint and connector data: shape, size, distribution and possible soldering defects of solder joints (such as cold solder joints, false solder joints, over solder joints or solder balls); connector alignment and integrity of contact surfaces.
[0099] Material defect data: material uniformity, defects (such as microcracks, pores, impurities) and composition distribution.
[0100] Component installation data: the position accuracy, component fit and welding quality of various electronic components (such as chips, resistors, capacitors) installed on the board.
[0101] Wire and trace data: Wire width, trace spacing, trace integrity, and any etch defects or trace discontinuities.
[0102] Mechanical stress data: The mechanical stress or deformation that the board may be subjected to, especially near connectors or fixing points.
[0103] Thermal Stress Data: Structural changes due to thermal expansion or areas of thermal stress concentration that may lead to failure.
[0104] The physical structure data is obtained by a 3D scanner with micro-nano precision. The scanner may use a variety of different imaging techniques, such as laser confocal scanning, electron microscope imaging, X-ray tomography, etc., to provide high-resolution 3D images. These images can be subsequently converted into digital models and can be used for detailed analysis and calculations, providing a basis for detecting and evaluating the quality of computer boards. The physical structure data provides key information for discovering physical defects at the microscopic scale that may occur during board design and manufacturing.
[0105] Using 3D scanning data, many types of physical defects can be identified, such as cracks, delamination, poor solder joints, wire breaks, poor lamination, chip packaging defects, tiny foreign objects or bubbles, etc. This identification is based on a detailed analysis of the collected 3D images.
[0106] The raw data collected by 3D scanners usually needs to be post-processed by specialized software to generate images or models that are easy to analyze. Post-processing steps may include data filtering, denoising, image reconstruction, edge detection, and defect annotation.
[0107] The entire scanning process can be automated, reducing manual intervention and operator errors. At the same time, operators should have an intuitive user interface to monitor the scanning process, view real-time scan images, or receive preliminary analysis of the scanning results.
[0108] Identifying physical defects is critical to understanding the risk of performance degradation or failure of computer boards in real-world use, especially when the boards are subjected to extreme operating conditions. In some cases, functional issues not discovered in software testing may stem from tiny defects at the physical level, and 3D scanning can provide an explanation for such issues.
[0109] All scanning processes and results should be recorded and consistent with industry standards to ensure data comparability and reliability.
[0110] Scans may need to be performed under specific environmental conditions, such as temperature, humidity, and vibration control, to ensure data quality is not affected by external factors.
[0111] In the entire step S102, the synchronization and precision of scanning, as well as the synergy between scanning data and software test data are emphasized. This step is crucial to determine the overall quality and reliability of the board, and provides basic data for the next step of fusion analysis.
[0112] Step S103: Use an integrated algorithm to perform fusion analysis on the functional behavior data and the physical structure data. The fusion analysis is based on the fact that the final performance of the computer board is affected by the functional state and the physical state, so as to identify defects that cannot be detected by a single method.
[0113] Furthermore, the use of an integrated algorithm to perform fusion analysis on the functional behavior data and the physical structure data includes:
[0114] The functional behavior data and physical structure data are converted into standardized feature vectors using multidimensional scaling analysis and principal component analysis;
[0115] Using the standardized feature vectors, a deep learning-based feature learning model is used to identify and extract the potential associations between functional and physical features by learning the intrinsic structure of the data, thereby generating a comprehensive feature representation for each computer board.
[0116] inputting the comprehensive feature representation into a fuzzy logic-based ensemble learning framework, wherein the ensemble learning framework intelligently matches the comprehensive feature representation with a set of predefined defect patterns to identify potential defects that cannot be detected by a single test;
[0117] Employs causal reasoning-based interpretation algorithms to analyze the causes of potential defects and transforms complex data and patterns into intuitive reports and charts, providing technicians with a clear visual and theoretical basis to understand the inspection results for further decision-making and action.
[0118] The method of converting the functional behavior data and physical structure data into a standardized feature vector using multidimensional scaling analysis and principal component analysis includes:
[0119] Performing data normalization on the functional behavior data and physical structure data respectively, calculating the deviation of each data point in each data set relative to the mean of its data set, and using the deviation to normalize the original data points, thereby forming two independent and normalized functional behavior data sets and physical structure data sets;
[0120] Multidimensional scaling analysis was applied to the normalized functional behavior dataset and physical structure dataset, respectively, by constructing a distance matrix between data points and performing nonlinear dimensionality reduction techniques to generate corresponding low-dimensional space coordinates for each dataset;
[0121] Integrate the low-dimensional space coordinates corresponding to the functional behavior dataset and the physical structure dataset to generate a joint data matrix, and calculate the covariance matrix on the joint data matrix;
[0122] Performing principal component analysis using the covariance matrix to determine the main variation directions of the joint data set and extracting the eigenvectors corresponding to the maximum eigenvalues to form a comprehensive set of principal component eigenvectors;
[0123] The comprehensive principal component feature vector set is resized so that the features of each dimension have the same magnitude, so as to obtain a standardized feature vector.
[0124] First, the collected functional behavior data and physical structure data are normalized. Functional behavior data and physical structure data constitute their own data sets. In this step, for each data set, the deviation of the data point relative to its data set mean is calculated. Such deviation calculation aims to reduce the scale differences between the measurements of each feature in the data set, so that each feature can be treated fairly in subsequent analysis. In this way, two independent and normalized data sets are obtained, laying the foundation for further data analysis.
[0125] Next, multidimensional scaling (MDS) is performed on each of these normalized data sets. Multidimensional scaling is a technique for reducing the dimensionality of data by constructing a distance matrix between data points and using nonlinear dimensionality reduction techniques to reveal the structure of the data in a low-dimensional space. The purpose of this step is to provide a low-dimensional spatial coordinate representation for each data set that retains the intrinsic structure of the data while simplifying the complexity of the data.
[0126] Subsequently, the low-dimensional spatial coordinates corresponding to the functional behavior dataset and the low-dimensional spatial coordinates corresponding to the physical structure dataset are integrated to form a joint data matrix. The purpose of integrating these two datasets is to obtain a comprehensive data representation that combines the functional and physical properties of the board. On this joint data matrix, the covariance matrix is further calculated. The covariance matrix reflects the correlation between the various features in the dataset and is a key step in performing principal component analysis.
[0127] Using the resulting covariance matrix, a principal component analysis (PCA) is performed. PCA is a powerful statistical tool used to reveal the main directions of variation in a data set. Through eigenvalue decomposition, the eigenvectors corresponding to the largest eigenvalues are extracted. These eigenvectors form a comprehensive set of principal component eigenvectors, which reflect the main variability of the joint data set. This set is key to identifying the board status because it combines the most important information from different tests.
[0128] Finally, the integrated principal component feature vector set is scaled. The scale adjustment process can be standardization (Z-score normalization) or normalization (Min-Max scaling). This step ensures that the contribution of different dimensional features in the model is balanced, so that the features of each dimension have the same magnitude. This standardized feature vector set provides a reliable input for subsequent machine learning models or other statistical analysis methods, which helps to accurately classify the status of the board, such as detecting and identifying potential defects.
[0129] Using the standardized feature vectors, a deep learning-based feature learning model is used to identify and extract the potential associations between functional and physical features by learning the intrinsic structure of the data, thereby generating a comprehensive feature representation for each computer board.
[0130] The structure of the feature learning model provided in this embodiment includes:
[0131] (1) Input layer:
[0132] Input: The fused normalized feature vector containing preprocessed functional behavior and physical structure data.
[0133] Implementation: The input layer serves as the starting point of the model and directly passes the data to the next layer without any complex processing.
[0134] (2) Dual-channel feature processing layer:
[0135] Input: The fused feature vector from the input layer.
[0136] Implementation: Design two parallel sub-networks or channels, each of which is a deep network structure but optimized for different types of features. One sub-network can focus on processing features related to time series data, while the other sub-network targets spatial data features. This design allows the network to process and learn the intrinsic features of both types of data in parallel.
[0137] Output: Feature representations of the two sub-networks, each capturing the temporal and spatial information in the input data.
[0138] (3) Association learning layer:
[0139] Input: Output of the two-channel feature processing layer.
[0140] Implementation: The purpose of this layer is to integrate the outputs of the two sub-networks and learn the potential correlation between them. Custom attention mechanisms or joint embedding strategies can be used to emphasize the interaction and correlation between the two features.
[0141] Output: A feature representation that integrates the association of functional and physical features.
[0142] (4) Deep feature fusion layer:
[0143] Input: Output of the association learning layer.
[0144] Implementation: Use fully connected layers and deep neural network structures, such as multi-layer perceptron (MLP), to further integrate and refine feature representations so that they can better characterize the state of the computer board.
[0145] Output: Comprehensive feature representation for defect identification.
[0146] (5) Output layer:
[0147] Input: Output of the deep feature fusion layer.
[0148] Implementation: Classify based on the board status (e.g., normal, abnormal, specific defect type), possibly using a fully connected layer with a Softmax activation function.
[0149] Output: probability distribution of possible states of the board, or direct classification label.
[0150] Custom loss function and optimizer:
[0151] Choose or design a loss function that places extra emphasis on the recognition accuracy of key features and penalizes defects that may be missed by a single test.
[0152] Let C be the total number of categories, y m is the one-hot encoded vector of the true label, is the corresponding model output, m is the ordinal number of the sample, and the loss function L is defined as:
[0153]
[0154] Among them, L CE is the cross entropy loss, defined as:
[0155]
[0156] Among them, y mn Represents the true label vector y m The value of the nth element in ; Represents the model output vector The value of the nth element in ;
[0157] L WHD is a weighted loss for hard-to-detect defects, defined as:
[0158]
[0159] Here, y mC is the element of the difficult-to-detect defect category in the true label vector, is the corresponding element in the prediction vector, and α and β are hyperparameters used to balance the two parts of the loss. The prediction accuracy of hard-to-detect defects is emphasized, and the role of β is to amplify the impact of prediction errors in this category. With this definition, the loss function not only penalizes classification errors, but also places special emphasis on the identification of hard-to-detect defects, so that the model pays more attention to the accurate prediction of these key categories during training.
[0160] Training this feature learning model typically involves the following steps:
[0161] Data preparation:
[0162] Prepare a set of labeled training data, which contains the fused standardized feature vectors of multiple instances and their corresponding labels (such as the status of the board).
[0163] Model initialization:
[0164] Initialize the layers and parameters of the model, which may include random initialization of weights and setting of biases.
[0165] Choice of loss function and optimizer:
[0166] Define the loss function, which will be used to evaluate the difference between the model output and the actual label. The loss function should reflect the model's recognition accuracy of key features and give greater penalties to defects that are difficult to detect.
[0167] Choose an optimizer, such as Adam, which will be used to update the model parameters to minimize the loss function.
[0168] Forward propagation:
[0169] For each training instance, the model performs a forward pass, computing the activity of each layer from the input layer to the output layer.
[0170] Calculate the loss:
[0171] The loss function is used to calculate the error between the predicted result and the true label.
[0172] Back propagation and parameter update:
[0173] Through the back-propagation algorithm, the gradient of the loss function is propagated back to each layer of the model, and then the optimizer is used to update the model parameters.
[0174] Iterative training:
[0175] The process of forward propagation, loss calculation, backpropagation, and parameter update is repeated until the model reaches a certain accuracy on the training data, or until the number of iterations reaches a preset threshold.
[0176] Verify and adjust:
[0177] Use the validation dataset to test the performance of the model and adjust the parameters or structure of the model as needed.
[0178] Evaluate and optimize:
[0179] Finally, the performance of the model is evaluated on an independent test set, and necessary optimization is performed based on the evaluation results.
[0180] During the training process, strategies such as batch processing, regularization techniques (such as Dropout), and early stopping may be used to improve training efficiency, prevent overfitting, and ensure the generalization ability of the model. Through these steps, the feature learning model can effectively learn how to predict the status of the computer board from the training data.
[0181] Below is the reference code for the feature learning model:
[0182]
[0183]
[0184]
[0185] Furthermore, the customized attention mechanism includes:
[0186] The variational autoencoder is used to evaluate the correlation between time series features and spatial features, and determine the joint contribution of time series features and spatial features to defect detection;
[0187] Based on the evaluation results of the variational autoencoder, weights are assigned to each feature, emphasizing the features that are more important for defect detection;
[0188] Perform feature weighted fusion and merge time series features and spatial features according to the assigned weights to generate feature representation for defect detection.
[0189] Applications of Variational Autoencoder (VAE):
[0190] Purpose: VAE is used to analyze and evaluate the correlation between two types of data features extracted from computer boards: time series features and spatial features. These features represent the functional parameters of the board that change over time (such as current or temperature) and the physical properties of the board (such as component layout).
[0191] Process: VAE learns deep representations of these features through encoding and decoding processes, thereby revealing complex structures and association patterns hidden in the original data. During the encoding process, VAE compresses the high-dimensional features of the input into low-dimensional potential representations. During the decoding process, it attempts to reconstruct the input data from these potential representations, and in the process learns the intrinsic correlation of the data.
[0192] Based on the analysis results of VAE, weights are dynamically assigned to each feature to highlight those that are crucial for detecting computer board defects.
[0193] Weighted feature fusion:
[0194] Purpose: To generate a feature representation optimized for detecting defects in computer boards by combining time series and spatial features and performing weighted fusion according to their weights.
[0195] Process: Weighted fusion involves multiplying each feature by its corresponding weight, which reflects the relative importance of the features. In this weighted manner, the combined feature representation can more accurately describe the health of the board and more effectively point out potential defects.
[0196] The following is the Python reference implementation code of this custom attention mechanism:
[0197]
[0198]
[0199]
[0200] #In the above code, defect_detection_features is the feature representation optimized for defect detection and can be used for further detection model training.
[0201] Furthermore, the variational autoencoder further includes a regularization term, which penalizes the distribution deviation of the variational autoencoder latent space based on Kullback-Leibler divergence, ensuring that the generated feature representation not only accurately reflects the characteristics of the original data, but also has good generalization ability.
[0202] In the design of variational autoencoders, a key component is the regularization term, which guides the latent space representation of the variational autoencoder to conform to the preset probability distribution (for example, multi-dimensional Gaussian distribution) as much as possible during the optimization process. The regularization term uses the Kullback-Leibler divergence (KL divergence) to measure the deviation between the latent space distribution in the variational autoencoder model and the preset distribution, and adjusts the model parameters by penalizing this deviation.
[0203] Specifically, the encoder part of the variational autoencoder maps the input features to a latent space whose distribution parameters (such as mean and variance) are learned by a neural network. The decoder part attempts to reconstruct the input features from this latent space. In this process, the KL divergence regularization term calculates the difference between the actual distribution of the latent space and the target distribution (usually assumed to be a standard normal distribution). This difference is added to the loss function of the VAE, so that while the model learns the data representation, it must also ensure that the distribution of these representations matches the expected distribution.
[0204] This approach has two main benefits: first, it avoids any part of the latent space from being ignored (i.e., it ensures that every direction in the space is meaningful), and second, it prevents overfitting because it forces the representation learned by the model to conform to a certain distribution, thus having better generalization ability. When implementing it, the weight of the regularization term can be adjusted to balance the reconstruction accuracy and the degree of matching of the latent space distribution.
[0205] In this way, the generated feature representation can more comprehensively capture the key characteristics of time series data and spatial data, which is of great significance for detecting complex defects on computer boards. For example, it can help distinguish those tiny or potential defects that are not easy to be found under a single test, making defect detection more accurate and reliable.
[0206] In summary, the variational autoencoder model provided in this embodiment introduces a regularization term based on KL divergence, which ensures the generalization and reliability of these representations while learning the deep feature representations of the data, thereby greatly improving the performance of computer board defect detection.
[0207] Overall, this custom attention mechanism uses advanced machine learning techniques to deeply analyze and optimize the feature analysis process to achieve more accurate defect detection. The variational autoencoder provides a deep understanding of feature associations, the adaptive learning algorithm ensures that feature weights reflect their true importance to defect detection, and feature weighted fusion combines this information into a powerful tool to support defect identification in the decision-making process.
[0208] Furthermore, the comprehensive feature representation output by the deep feature fusion layer of the feature learning model is input into an integrated learning framework based on fuzzy logic, and the integrated learning framework intelligently matches the comprehensive feature representation with a set of predefined defect patterns to identify potential defects that cannot be detected by a single test.
[0209] This embodiment provides a method for defect detection of computer boards through an integrated learning framework based on fuzzy logic. The core of the framework is to intelligently match the comprehensive feature representation generated by the deep feature learning model with a set of predefined defect patterns. This matching process can reveal complex or hidden defects that may not be detected by a single test.
[0210] First, an ensemble learning framework is initialized, which contains multiple fuzzy logic systems, each of which specializes in identifying a specific type of defect pattern. Fuzzy logic systems use the concepts of fuzzy sets and fuzzy rules to account for uncertainty and ambiguity in the comprehensive feature representation.
[0211] In this framework, a set of defect modes are defined, each corresponding to a specific type of defect that may occur on a board. These modes are derived based on previous data collection and expert knowledge, and cover a wide range of failure phenomena from microscopic cracks to circuit shorts.
[0212] A set of fuzzy logic rules is established for each defect mode. These rules relate various aspects of the feature representation to the membership of the corresponding defect mode. For example, if a feature value is within a predetermined range, it may indicate the presence of a specific defect.
[0213] The input comprehensive feature representation is fuzzified, that is, it is converted into a fuzzy set according to the importance of each feature and its contribution to the defect. This process includes assigning a membership function to each feature to determine its membership to each defect mode.
[0214] The fuzzy logic system is used to intelligently match the fuzzy feature representation. The matching process is based on the previously established fuzzy logic rules, calculates the membership of each defect mode, and generates a fuzzy reasoning output based on this, which reflects the possible defect type of the board.
[0215] Through fuzzy logic reasoning, the membership of each defect mode obtained from intelligent matching is converted into a clear defect diagnosis result. This process may include a defuzzification step, that is, selecting the defect mode with the highest membership, or combining multiple defect modes with high membership to determine the final defect status of the board.
[0216] The following is a reference implementation using Python and the scikit-fuzzy library, which includes a detailed fuzzy logic system initialization, defect mode definition, fuzzy rule establishment, and fuzzy reasoning and result interpretation process.
[0217]
[0218]
[0219] In the above implementation, fuzzy logic variables for feature inputs and defect level variables for outputs are defined. Membership functions are defined for each variable and how rules are applied to evaluate possible defect levels based on the fuzzified values of the input features.
[0220] Employs causal reasoning-based interpretation algorithms to analyze the causes of potential defects and transforms complex data and patterns into intuitive reports and charts, providing technicians with a clear visual and theoretical basis to understand the inspection results for further decision-making and action.
[0221] Causal inference generally involves determining cause-effect relationships between variables. In the context of defect detection, this might mean finding out which features indicate a particular defect type. Here are the steps to perform this type of analysis:
[0222] (1) Data collection and preprocessing
[0223] Collect a large amount of historical data, including the characteristic values of computer boards and the corresponding defect records.
[0224] Perform preprocessing on the data, such as cleaning, normalization, and denoising.
[0225] (2) Mapping of defect characteristics and possible causes
[0226] Use statistical analysis methods or machine learning algorithms to analyze the relationship between features and defects. For example, decision tree algorithms can reveal feature decision paths, which can be used as preliminary cause-effect mapping.
[0227] (3) Establishment of causal reasoning model
[0228] Use methods such as structural equation modeling (SEM) or Bayesian networks to build a model that can represent the potential causal relationships between variables.
[0229] Define the structure and conditional probabilities in the model based on historical data and expert knowledge.
[0230] 3.1 Choose the right model type
[0231] Structural equation modeling (SEM): SEM is a type of statistical model that allows multiple systems of equations to be estimated simultaneously. SEM is often used for causal inference because it can reveal direct and indirect effects between variables.
[0232] Bayesian network: This is a probabilistic graphical model that uses probabilistic reasoning to predict the relationship between different variables. Bayesian networks are particularly suitable for dealing with uncertainty and high complexity of causal relationships.
[0233] 3.2 Define the model structure
[0234] Variable selection: Determine which variables will be included in the model. These variables should include all relevant feature data and defect records.
[0235] Structural definition: Based on theoretical knowledge and expert opinion, define the underlying causal structure between variables. For example, experts may know that there is a causal relationship between certain features and a specific type of defect.
[0236] 3.3 Establishing model parameters
[0237] Conditional Probability Table: In a Bayesian network, the probability of each variable depends on its parent variables. A conditional probability table (CPT) is created for each variable to express this relationship.
[0238] Regression coefficients: In SEM, the parameters in an estimated equation, such as regression coefficients, quantify the effect of one variable on another variable.
[0239] 3.4 Model Coding and Implementation
[0240] Use a programming language (such as Python) and related libraries (e.g., pgmpy for Bayesian networks and semopy for structural equation models) to encode model structure and parameters.
[0241] Write code to estimate model parameters, such as using maximum likelihood estimation or Bayesian inference.
[0242] 3.5 Initial model estimation and adjustment
[0243] Use historical data to estimate the initial parameters of the model.
[0244] Adjust the model structure and parameters based on the results of parameter estimation, which may involve adding or removing variables, or changing the relationship between variables.
[0245] 3.6 Model Optimization and Iteration
[0246] The iterative process may need to be performed multiple times, each time optimizing the model based on statistical fit indices and expert opinion.
[0247] Use metrics such as the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC) to assess the goodness of fit of the model.
[0248] (4) Verification of causal inference model
[0249] Use datasets to train and validate the model to ensure it reflects true causal relationships.
[0250] Perform cross-validation or use an independent test set to evaluate the accuracy of the model.
[0251] (5) Explanation and visualization of causal relationships
[0252] Provide technicians with a clear visual and rationale for understanding test results, for example, using graphical representations such as cause-and-effect diagrams to show cause-and-effect relationships between variables.
[0253] Use visualization tools such as heat maps, scatter plots, and bar charts to visually display the association between feature values and defect types.
[0254] (6) Report Generation
[0255] Develop an algorithm to automatically generate interpretive reports based on model outputs and visualizations.
[0256] The report should describe in detail which features are associated with specific defects and the strength of their association.
[0257] (7) Decision support
[0258] Integrate interpretive reports into decision support systems to help technicians make data-driven repair or quality control decisions.
[0259] Provides action suggestions, such as which parts require further inspection or replacement.
[0260] To facilitate understanding of this step, an example is given below for illustration.
[0261] Consider an example where a tiny crack in a computer board may cause intermittent circuit failure. This defect may not always be active, so a single functional test (such as a current test) or a single physical inspection (such as a visual inspection) may not reliably identify it. Here is an operational flow of how an integrated algorithm might identify such a defect:
[0262] Data Collection: Both functional testing and physical structural testing are performed.
[0263] Functional testing may include current and voltage measurements, as well as signal integrity testing.
[0264] Physical structure detection may be achieved by capturing fine structural images of the board using a high-resolution micro-nanoscale 3D scanner.
[0265] Data preprocessing: Standardize functional test data and physical structure data, eliminate scale differences and prepare for in-depth analysis.
[0266] Feature extraction: Use deep learning methods to extract features from physical structure images; at the same time, extract change patterns and abnormal indicators from functional test data.
[0267] Data fusion: Combine the extracted functional behavior features and physical structure features and use an integration algorithm to create a comprehensive feature representation that captures the functional and physical conditions of the board.
[0268] Defect Identification: The comprehensive feature vector is input into a fuzzy logic-based ensemble learning framework that can handle the uncertainty and complexity in feature representation and intelligently match the features to predefined defect patterns, including intermittent circuit problems caused by tiny cracks that may be missed by a single approach.
[0269] Results analysis and interpretation: If the model detects a potential defect, use causal reasoning tools to trace the possible causes of the defect and help technicians understand and interpret the results by generating easy-to-understand reports and visualizations.
[0270] Self-optimization: Based on these findings, test cases and scan parameters are adjusted to improve detection of similar defects, and a closed-loop feedback mechanism is implemented to continuously optimize the detection process.
[0271] In this example, the integrated algorithm exploits the complementary properties of functional and physical structural data. While a single test may not reveal a problem, combining data from both can identify defects by revealing correlations between intermittent failures and tiny physical defects. For example, current fluctuations in a functional test may be associated with tiny cracks observed in a physical scan. This integrated approach provides a more comprehensive inspection method, increasing the chances of identifying subtle defects under complex real-world conditions.
[0272] Step S104: judging whether the computer board has defects according to the processing result of the fusion analysis processing.
[0273] This step involves a detailed analysis of the fusion analysis processing results produced by the integrated algorithm.
[0274] By analyzing the fused data, technicians or automated systems can detect and identify inconsistencies and abnormal patterns between functional behavior and physical structure data, which may indicate potential defects. For example, in step S104, if the integrated learning framework identifies a potential defect that cannot be detected by a single test, then it is determined that the computer board is defective.
[0275] Step S105: If a defect is identified in the computer board, it is determined to be a defective product, and the defective computer board is removed from the production line, and the software test cases and three-dimensional scanning parameters are adaptively adjusted; wherein the adaptive adjustment is based on the results of a detailed analysis of the defective products, so as to diagnose and incorporate computer board defect types that have not been fully covered before in the computer board inspection process, and ensure that adjustments are made to changes in the production line conditions of the computer boards or newly introduced board designs.
[0276] The Adaptation Process is a comprehensive quality control approach designed to improve the sensitivity and adaptability of detection methods to new or previously under-recognized defect types in the production process. The following are detailed steps of the Adaptation Process to ensure that technicians can implement and comply with possible patent requirements.
[0277] 1. Analyze defective products
[0278] Detailed defect recording: Use high-precision scanning and test results to record the exact location, size and shape of each defect on the board.
[0279] Cause Analysis: Analyze the potential causes of the defect, which may include material defects, environmental factors during the production process, or operational errors.
[0280] Severity Assessment: Assess the severity of the defect based on its impact on the functionality of the board.
[0281] Historical comparison: Compare with historical defect data to find patterns and correlations.
[0282] 2. Data Collection
[0283] Comprehensive Data Archive: Collects detailed test logs, scan images and production process parameters to create a comprehensive data archive for in-depth analysis.
[0284] Environmental parameter records: Record the production environment parameters of the defective boards in detail, such as temperature and humidity, machine settings, etc.
[0285] Operation log review: Review the operation logs during the production process to look for non-standard operations that may be related to the generation of defects.
[0286] 3. Improve test cases
[0287] Test case design: Based on the defect analysis results, design new test cases or adjust existing test scripts to capture similar defects more accurately.
[0288] Automated testing updates: Integrate new test cases into the automated testing framework to ensure they are automatically executed in production in the future.
[0289] 4. Adjust scanning parameters
[0290] Scanning accuracy adjustment: Increase scanning resolution to capture tiny defects more clearly.
[0291] Parameter Optimization: Optimize scanning speed and focal length to obtain the best image quality and defect detection rate.
[0292] Automated Adjustment Implementation: Write automated scripts or adjust the control system so that scanning parameters are automatically adjusted based on board type and expected defects.
[0293] Furthermore, the computer board detection method further includes:
[0294] The detection data during the detection process is stored, and the detection data includes functional behavior data, physical structure data, and processing results of fusion analysis and processing. The detection data is used to continuously optimize software test cases and 3D scanning parameters to achieve self-calibration and self-optimization of the detection system, and enhance the system's adaptability to quality differences of various types of computer boards and responsiveness to changes in the production process.
[0295] This is a self-calibrating and self-optimizing inspection method that uses the data collected during the inspection process to optimize the test cases and 3D scanning parameters of the computer board. This process ensures that the inspection system can adapt to the differences in board quality and changes in the production process. It includes the following steps:
[0296] (1) Data storage
[0297] Create a database or data warehouse to store all the inspection data collected during each inspection cycle.
[0298] The detection data should include functional behavior data, physical structure data and results from fusion analysis processing.
[0299] Ensure the security, integrity and retrievability of stored data.
[0300] (2) Data classification and labeling
[0301] Classify the collected data and label it as good, defective or waste.
[0302] Record and mark the test results corresponding to each board, including the test cases and scan parameters used during the test process.
[0303] (3) Data analysis
[0304] Analyze stored inspection data to identify patterns of inspection failures and common defect types.
[0305] Use statistical analysis or machine learning algorithms to identify factors that may lead to detection vulnerabilities.
[0306] (4) Test case optimization
[0307] Based on the results of data analysis, adjust and optimize software test cases to improve the detection capabilities of specific problems.
[0308] Add new test cases or modify existing ones to cover the vulnerabilities revealed by data analysis.
[0309] (5) Scan parameter adjustment
[0310] Adjust the scanning parameters of the 3D scanner, such as resolution, scanning speed, and scanning path, to better capture defects.
[0311] For new or complex defect types detected, refine the scanning parameters to increase detection accuracy.
[0312] (6) Self-calibration process
[0313] Establish a closed-loop control system that automatically adjusts inspection parameters based on past inspection results and optimization measures.
[0314] Implementing a continuous learning mechanism, the system continuously adjusts its own behavior by analyzing new detection data.
[0315] (7) Performance evaluation and iteration
[0316] Regularly evaluate the performance of the detection system to ensure the effectiveness of optimization measures.
[0317] Based on the results of the performance evaluation, the optimization process is iterated to continuously improve the test cases and scanning parameters.
[0318] In the above-mentioned embodiment, a method for detecting a computer board is provided. Correspondingly, the present application also provides a device for detecting a computer board. Figure 2 , which is a schematic diagram of an embodiment of a computer board detection device of the present application. Since this embodiment, i.e., the second embodiment, is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The system embodiment described below is only illustrative.
[0319] The second embodiment of the present application provides a computer board detection device, characterized by comprising:
[0320] The test unit 201 is used to run software test cases to simulate the working state of the computer board under various operating environments including temperature, current and data transmission to obtain functional behavior data of the computer board;
[0321] The scanning unit 202 is used to scan the surface and internal structure of the computer board using a micro-nano three-dimensional scanner while running the software test, so as to collect physical structure data of the computer board, wherein the physical structure data is used to identify physical defects of the computer board at a microscopic scale;
[0322] The analysis unit 203 is used to perform fusion analysis processing on the functional behavior data and the physical structure data using an integrated algorithm, wherein the fusion analysis processing is based on the fact that the final performance of the computer board is affected by the functional state and the physical state, thereby identifying defects that cannot be detected by a single method;
[0323] A judgment unit 204 is used to judge whether the computer board has a defect according to the processing result of the fusion analysis processing;
[0324] The adjustment unit 205 is used to determine that the computer board is defective if it is defective, and then remove the defective computer board from the production line, and adaptively adjust the software test cases and the scanning parameters of the micro-nano-level three-dimensional scanner; wherein the adaptive adjustment is based on the results of the detailed analysis of the defective products, so as to diagnose and incorporate the computer board defect types that have not been fully covered in the computer board detection process before, and ensure that adjustments are made to the changes in the production line conditions of the computer board or the newly introduced board design.
[0325] A third embodiment of the present application provides a computer board detection system, comprising:
[0326] A data acquisition module, including a central processing unit and a storage unit, wherein the storage unit is pre-installed with operation instructions for simulating the working state of the computer board under various environmental conditions and generating corresponding functional behavior data;
[0327] A micro-nanoscale three-dimensional imaging device, operating synchronously with the data acquisition module, is used to scan the surface and internal structure of the computer board and collect physical structure data, which is used to identify microscopic scale physical defects of the computer board;
[0328] The analysis and processing unit is equipped with data processing hardware and software resources, receives data from the data acquisition module and the 3D imaging device, and uses integrated algorithms to perform fusion analysis to identify potential defects;
[0329] A defect detection module, connected to the analysis and processing unit, is used to evaluate the fusion analysis results and determine whether the computer board meets the quality standards;
[0330] User interface terminal, which provides an interactive interface for technical operators to display test results, integrate analysis feedback and real-time system status, and receive adjustment commands input by users;
[0331] The remote monitoring unit includes at least one network interface for remotely transmitting detection data, system status and alarm information to a central monitoring station or a mobile terminal.
[0332] A fourth embodiment of the present application provides an electronic device, the electronic device comprising:
[0333] processor;
[0334] The memory is used to store a program, and when the program is read and executed by the processor, the computer board detection method provided in the first embodiment of the present application is executed.
[0335] The fifth embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the computer board detection method provided in the first embodiment of the present application is executed.
[0336] Although the present application is disclosed as above in the form of a preferred embodiment, it is not intended to limit the present application. Any technical personnel in this field may make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the scope of protection of the present application shall be based on the scope defined by the claims of the present application.
Claims
1. A method for detecting a computer board, characterized in that: include: Run software test cases to simulate the working state of the computer board under various operating environments including temperature, current and data transmission to obtain functional behavior data of the computer board; While running the software test, a micro-nano three-dimensional scanner is used to scan the surface and internal structure of the computer board to collect physical structure data of the computer board, wherein the physical structure data is used to identify physical defects of the computer board at a microscopic scale; Using an integrated algorithm to perform fusion analysis on the functional behavior data and the physical structure data, wherein the fusion analysis is based on the fact that the final performance of the computer board is affected by the functional state and the physical state, thereby identifying defects that cannot be detected by a single method; According to the processing result of the fusion analysis processing, determining whether the computer board has defects; If the computer board has defects, it is determined to be a defective product, and the defective computer board is removed from the production line, and the software test cases and the scanning parameters of the micro-nano-level three-dimensional scanner are adaptively adjusted; wherein the adaptive adjustment is based on the results of the detailed analysis of the defective products, so as to diagnose and incorporate computer board defect types that have not been fully covered in the computer board detection process before, and ensure that adjustments are made to changes in the production line conditions of the computer board or newly introduced board designs; The method of using an integrated algorithm to perform fusion analysis on the functional behavior data and the physical structure data includes: The functional behavior data and physical structure data are converted into standardized feature vectors using multidimensional scaling analysis and principal component analysis; Using the standardized feature vectors, a deep learning-based feature learning model is used to identify and extract the potential associations between functional and physical features by learning the intrinsic structure of the data, thereby generating a comprehensive feature representation for each computer board. inputting the comprehensive feature representation into a fuzzy logic-based ensemble learning framework, wherein the ensemble learning framework intelligently matches the comprehensive feature representation with a set of predefined defect patterns to identify potential defects that cannot be detected by a single test; Employs causal reasoning-based interpretation algorithms to analyze the causes of potential defects and convert complex data and patterns into intuitive reports and charts, providing technicians with a clear visual and theoretical basis to understand the inspection results for further decision-making and action; The method of converting the functional behavior data and physical structure data into standardized feature vectors by using multidimensional scaling analysis and principal component analysis includes: Performing data normalization on the functional behavior data and physical structure data respectively, calculating the deviation of each data point in each data set relative to the mean of its data set, and using the deviation to normalize the original data points, thereby forming two independent and normalized functional behavior data sets and physical structure data sets; Multidimensional scaling analysis was applied to the normalized functional behavior dataset and physical structure dataset, respectively, by constructing a distance matrix between data points and performing nonlinear dimensionality reduction techniques to generate corresponding low-dimensional space coordinates for each dataset; Integrate the low-dimensional space coordinates corresponding to the functional behavior dataset and the physical structure dataset to generate a joint data matrix, and calculate the covariance matrix on the joint data matrix; Performing principal component analysis using the covariance matrix to determine the main variation directions of the joint data set and extracting the eigenvectors corresponding to the maximum eigenvalues to form a comprehensive set of principal component eigenvectors; Performing a scale adjustment process on the comprehensive principal component feature vector set so that the features of each dimension have the same magnitude, so as to obtain a standardized feature vector; Wherein, the feature learning model includes: Input layer, used to receive the normalized feature vector; A dual-channel feature processing layer, including two parallel sub-networks, the first sub-network is configured with a deep network structure optimized for time series data features, and the second sub-network is configured with a deep network structure optimized for spatial data features. Each sub-network extracts the time series features and spatial features in the fused feature vector and outputs two independent feature representations; An association learning layer, configured with a custom attention mechanism, for integrating the output of the dual-channel feature processing layer to generate a feature representation by learning the potential association between the time series feature representation and the spatial feature representation; A deep feature fusion layer, used for receiving the output of the association learning layer, further integrating and refining the comprehensive feature representation through a series of fully connected layers, and forming a comprehensive feature representation for characterizing the state of the computer board; The output layer is configured with a fully connected layer and a Softmax activation function, which is used to convert the comprehensive feature representation of the deep feature fusion layer into a probability distribution or classification label describing the possible states of the computer board.
2. The computer board detection method according to claim 1, characterized in that: Also includes: The detection data during the detection process is stored, and the detection data includes functional behavior data, physical structure data, and processing results of fusion analysis and processing. The detection data is used to continuously optimize software test cases and 3D scanning parameters to achieve self-calibration and self-optimization of the detection system, and enhance the system's adaptability to quality differences of various types of computer boards and responsiveness to changes in the production process.
3. The computer board detection method according to claim 1, characterized in that: The loss function of the feature learning model Defined as: Among them, C is the total number of categories of the output of the output layer, is the one-hot encoded vector of the true label, is the corresponding model output, and m is the ordinal number of the sample; in, is the cross entropy loss, defined as: in, Represents the true label vector The value of the nth element in ; Represents the model output vector The value of the nth element in ; is a weighted loss for hard-to-detect defects, defined as: in, is the element of the difficult-to-detect defect category in the true label vector, is the corresponding element in the prediction vector, , is a hyperparameter used to balance the two parts of loss.
4. The computer board detection method according to claim 1, characterized in that: The custom attention mechanism includes: The variational autoencoder is used to evaluate the correlation between time series features and spatial features, and determine the joint contribution of time series features and spatial features to defect detection; Assign weights to each feature based on the evaluation results of the variational autoencoder; Perform feature weighted fusion and merge time series features and spatial features according to the assigned weights to generate feature representation for defect detection.
5. The computer board detection method according to claim 4, characterized in that: The variational autoencoder further includes a regularization term, which penalizes the distribution deviation of the variational autoencoder latent space based on Kullback-Leibler divergence, ensuring that the generated feature representation not only accurately reflects the characteristics of the original data, but also has good generalization ability.
6. A computer board detection device, characterized in that: include: A test unit, used to run software test cases to simulate the working state of the computer board under various operating environments including temperature, current and data transmission, so as to obtain functional behavior data of the computer board; A scanning unit, used to scan the surface and internal structure of the computer board using a micro-nano three-dimensional scanner while running the software test, so as to collect physical structure data of the computer board, wherein the physical structure data is used to identify physical defects of the computer board at a microscopic scale; An analysis unit, configured to perform fusion analysis processing on the functional behavior data and the physical structure data using an integrated algorithm, wherein the fusion analysis processing is based on the fact that the final performance of the computer board is affected by both the functional state and the physical state, thereby identifying defects that cannot be detected by a single method; A judgment unit, used for judging whether the computer board has a defect according to the processing result of the fusion analysis processing; An adjustment unit, used to determine if a computer board has defects, as a defective product, and then remove the defective computer board from the production line, and adaptively adjust the software test cases and the scanning parameters of the micro-nano-level three-dimensional scanner; wherein the adaptive adjustment is based on the results of the detailed analysis of the defective products, so as to diagnose and incorporate computer board defect types that have not been fully covered in the computer board detection process before, and ensure that adjustments are made to changes in production line conditions of computer boards or newly introduced board designs; Wherein, the analysis unit is specifically used for: The functional behavior data and physical structure data are converted into standardized feature vectors using multidimensional scaling analysis and principal component analysis; Using the standardized feature vectors, a deep learning-based feature learning model is used to identify and extract the potential associations between functional and physical features by learning the intrinsic structure of the data, thereby generating a comprehensive feature representation for each computer board. inputting the comprehensive feature representation into a fuzzy logic-based ensemble learning framework, wherein the ensemble learning framework intelligently matches the comprehensive feature representation with a set of predefined defect patterns to identify potential defects that cannot be detected by a single test; Employs causal reasoning-based interpretation algorithms to analyze the causes of potential defects and convert complex data and patterns into intuitive reports and charts, providing technicians with a clear visual and theoretical basis to understand the inspection results for further decision-making and action; The method of converting the functional behavior data and physical structure data into standardized feature vectors by using multidimensional scaling analysis and principal component analysis includes: Performing data normalization on the functional behavior data and physical structure data respectively, calculating the deviation of each data point in each data set relative to the mean of its data set, and using the deviation to normalize the original data points, thereby forming two independent and normalized functional behavior data sets and physical structure data sets; Multidimensional scaling analysis was applied to the normalized functional behavior dataset and physical structure dataset, respectively, by constructing a distance matrix between data points and performing nonlinear dimensionality reduction techniques to generate corresponding low-dimensional space coordinates for each dataset; Integrate the low-dimensional space coordinates corresponding to the functional behavior dataset and the physical structure dataset to generate a joint data matrix, and calculate the covariance matrix on the joint data matrix; Performing principal component analysis using the covariance matrix to determine the main variation directions of the joint data set and extracting the eigenvectors corresponding to the maximum eigenvalues to form a comprehensive set of principal component eigenvectors; Performing a scale adjustment process on the comprehensive principal component feature vector set so that the features of each dimension have the same magnitude, so as to obtain a standardized feature vector; Wherein, the feature learning model includes: Input layer, used to receive the normalized feature vector; A dual-channel feature processing layer, including two parallel sub-networks, the first sub-network is configured with a deep network structure optimized for time series data features, and the second sub-network is configured with a deep network structure optimized for spatial data features. Each sub-network extracts the time series features and spatial features in the fused feature vector and outputs two independent feature representations; An association learning layer, configured with a custom attention mechanism, for integrating the output of the dual-channel feature processing layer to generate a feature representation by learning the potential association between the time series feature representation and the spatial feature representation; A deep feature fusion layer, used for receiving the output of the association learning layer, further integrating and refining the comprehensive feature representation through a series of fully connected layers, and forming a comprehensive feature representation for characterizing the state of the computer board; The output layer is configured with a fully connected layer and a Softmax activation function, which is used to convert the comprehensive feature representation of the deep feature fusion layer into a probability distribution or classification label describing the possible states of the computer board.
7. A computer board detection system, characterized in that: include: A data acquisition module, including a central processing unit and a storage unit, wherein the storage unit is pre-installed with operation instructions for simulating the working state of the computer board under various environmental conditions and generating corresponding functional behavior data; A micro-nanoscale three-dimensional imaging device, operating synchronously with the data acquisition module, is used to scan the surface and internal structure of the computer board and collect physical structure data, which is used to identify microscopic scale physical defects of the computer board; The analysis and processing unit is equipped with data processing hardware and software resources, receives data from the data acquisition module and the 3D imaging device, and uses integrated algorithms to perform fusion analysis to identify potential defects; A defect detection module, connected to the analysis and processing unit, is used to evaluate the fusion analysis results and determine whether the computer board meets the quality standards; User interface terminal, which provides an interactive interface for technical operators to display test results, integrate analysis feedback and real-time system status, and receive adjustment commands input by users; A remote monitoring unit, including at least one network interface for remotely transmitting detection data, system status and alarm information to a central monitoring station or a mobile terminal; Wherein, the analysis and processing unit is specifically used for: The functional behavior data and physical structure data are converted into standardized feature vectors using multidimensional scaling analysis and principal component analysis; Using the standardized feature vectors, a deep learning-based feature learning model is used to identify and extract the potential associations between functional and physical features by learning the intrinsic structure of the data, thereby generating a comprehensive feature representation for each computer board. inputting the comprehensive feature representation into a fuzzy logic-based ensemble learning framework, wherein the ensemble learning framework intelligently matches the comprehensive feature representation with a set of predefined defect patterns to identify potential defects that cannot be detected by a single test; Employs causal reasoning-based interpretation algorithms to analyze the causes of potential defects and convert complex data and patterns into intuitive reports and charts, providing technicians with a clear visual and theoretical basis to understand the inspection results for further decision-making and action; The method of converting the functional behavior data and physical structure data into standardized feature vectors by using multidimensional scaling analysis and principal component analysis includes: Performing data normalization on the functional behavior data and physical structure data respectively, calculating the deviation of each data point in each data set relative to the mean of its data set, and using the deviation to normalize the original data points, thereby forming two independent and normalized functional behavior data sets and physical structure data sets; Multidimensional scaling analysis was applied to the normalized functional behavior dataset and physical structure dataset, respectively, by constructing a distance matrix between data points and performing nonlinear dimensionality reduction techniques to generate corresponding low-dimensional space coordinates for each dataset; Integrate the low-dimensional space coordinates corresponding to the functional behavior dataset and the physical structure dataset to generate a joint data matrix, and calculate the covariance matrix on the joint data matrix; Performing principal component analysis using the covariance matrix to determine the main variation directions of the joint data set and extracting the eigenvectors corresponding to the maximum eigenvalues to form a comprehensive set of principal component eigenvectors; Performing a scale adjustment process on the comprehensive principal component feature vector set so that the features of each dimension have the same magnitude, so as to obtain a standardized feature vector; Wherein, the feature learning model includes: Input layer, used to receive the normalized feature vector; A dual-channel feature processing layer, including two parallel sub-networks, the first sub-network is configured with a deep network structure optimized for time series data features, and the second sub-network is configured with a deep network structure optimized for spatial data features. Each sub-network extracts the time series features and spatial features in the fused feature vector and outputs two independent feature representations; An association learning layer, configured with a custom attention mechanism, for integrating the output of the dual-channel feature processing layer to generate a feature representation by learning the potential association between the time series feature representation and the spatial feature representation; A deep feature fusion layer, used for receiving the output of the association learning layer, further integrating and refining the comprehensive feature representation through a series of fully connected layers, and forming a comprehensive feature representation for characterizing the state of the computer board; The output layer is configured with a fully connected layer and a Softmax activation function, which is used to convert the comprehensive feature representation of the deep feature fusion layer into a probability distribution or classification label describing the possible states of the computer board.
