Testing method and system for stacked package memory
By processing and analyzing multi-dimensional data of stacked packaged memory, the problem of inaccurate testing and comprehensive evaluation in existing technologies has been solved, enabling accurate testing and comprehensive evaluation of memory chips and improving data processing accuracy and detection efficiency.
Patent Information
- Application Number
- CN202410935108.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-12
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-12
AI Technical Summary
Existing technologies are insufficient for accurate testing and comprehensive evaluation of stacked packaged memories, resulting in problems such as insufficient data processing accuracy, low testing efficiency, and inability to fully evaluate chip performance.
Advanced technologies such as multi-dimensional data processing, feature extraction, cluster analysis, anomaly detection, and heatmap analysis are employed to collect electrical performance data of stacked packaged memory, perform comprehensive data processing and analysis, and generate test evaluation reports.
It enables precise testing and comprehensive evaluation of memory chips, improves data processing accuracy and testing efficiency, and ensures the overall quality and performance stability of memory chips.
Smart Images

Figure CN118471320B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of semiconductors, and more particularly to a testing method and system for stacked package memory. Background Art
[0002] With the rapid development of electronic products, memory technology is increasingly used in electronic devices, especially stacked package memory, which plays an important role in improving storage density and performance. Stacked package memory greatly increases the storage capacity per unit volume and improves data transmission speed by stacking multiple memory chips vertically together. It is widely used in high-performance computing, smart phones and other fields.
[0003] In the relevant technical means, memory chips are tested independently, their electrical performance data is collected, and then the performance of memory chips is judged through traditional data processing methods, such as mean processing and simple anomaly detection. These methods can evaluate the basic performance of memory chips and screen out unqualified chips to a certain extent, thereby ensuring the overall quality of the memory.
[0004] With respect to the above technical solution, the basic performance of memory chips can be evaluated and screened through existing testing methods. However, when dealing with complex stacked packaged memories, it is difficult to effectively process and analyze multi-dimensional electrical performance data, and it is impossible to accurately test and comprehensively evaluate stacked packaged memories. There are problems such as insufficient data processing accuracy, low detection efficiency, and inability to comprehensively evaluate chip performance. Summary of the invention
[0005] In order to improve the inability to accurately test and comprehensively evaluate stacked packaged memories, there are problems such as insufficient data processing accuracy, low detection efficiency, and inability to comprehensively evaluate chip performance. The present application provides a testing method and system for stacked packaged memories.
[0006] The present invention provides a test method for a stacked package memory, comprising: collecting electrical performance data of each memory chip in the stacked package memory, inputting all the electrical performance data into a preset test model, and outputting electrical characteristic parameters; performing multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and performing feature extraction on the standardized data set to obtain a feature vector; performing cluster analysis on the feature vector to obtain a clustering result, and screening the clustering result based on a preset performance standard to obtain a performance data set; performing anomaly detection on the performance data set to generate an anomaly detection report, and performing pattern recognition processing on the anomaly detection report to obtain a pattern recognition result; performing heat map analysis on the pattern recognition result to generate heat map data, and performing heat analysis on the performance data set based on the heat map data to generate a heat distribution map; and performing comprehensive evaluation on the clustering result, the performance data set, the pattern recognition result, and the heat distribution map to obtain a test evaluation report.
[0007] As a preferred scheme, the steps of collecting electrical performance data of each memory chip in the stacked package memory, inputting all the electrical performance data into a preset test model, and outputting electrical characteristic parameters include: connecting each memory chip in the stacked package memory to an electrical testing device one by one, applying a preset test voltage and test current to each memory chip according to the electrical testing device, and collecting the electrical performance data of each memory chip in real time through the electrical testing device; preprocessing all the electrical performance data; wherein the preprocessing includes denoising, filtering and normalization; inputting all the preprocessed electrical performance data into a preset test model, and outputting corresponding electrical characteristic parameters.
[0008] As a preferred embodiment, the steps of performing multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and extracting features from the standardized data set to obtain a feature vector include: performing data cleaning on the electrical characteristic parameters; wherein the data cleaning includes removing outliers and filling missing values; normalizing the cleaned electrical characteristic parameters to convert data of different dimensions of the electrical characteristic parameters into the same dimension; performing dimensionality reduction processing on the normalized electrical characteristic parameters using principal component analysis to obtain a standardized data set; extracting representative features from the standardized data set based on a convolutional neural network, and integrating the extracted representative features to obtain a feature vector; wherein the representative features refer to characteristic values reflecting the key performance of the memory chip.
[0009] As a preferred scheme, the step of performing cluster analysis on the feature vectors to obtain clustering results, and screening the clustering results based on preset performance standards to obtain a performance data set includes: inputting the feature vectors into a preset cluster analysis model, and performing cluster analysis on the feature vectors by utilizing the K-means clustering algorithm in the cluster analysis model to obtain clustering results; screening the clustering results based on power consumption thresholds, voltage stability, and response time limits of preset performance standards to obtain clusters that meet the performance standards and clusters that do not meet the performance standards, removing the clusters that do not meet the performance standards, and recombining the cluster data that meet the preset performance standards to obtain a performance data set.
[0010] As a preferred scheme, the step of performing cluster analysis on the feature vectors by using the K-means clustering algorithm in the cluster analysis model to obtain clustering results includes: using a distance calculation method to measure the distance of the feature vectors to obtain a distance matrix between the feature vectors; clustering and updating the distance matrix to obtain cluster centers, and calculating the cluster to which the feature vector of each memory chip in the memory belongs based on the cluster centers; performing performance screening based on the clusters and preset standards, eliminating clustering results that do not meet the requirements, and obtaining a screened performance data set; comparing and analyzing the screened performance data set and the eliminated non-performance data set to extract features with obvious performance differences and obtain clustering results.
[0011] As a preferred solution, the steps of performing anomaly detection on the performance data set, generating an anomaly detection report, performing pattern recognition processing on the anomaly detection report, and obtaining a pattern recognition result include: inputting the performance data set into an anomaly detection model, analyzing the performance data set by using an isolation forest algorithm, identifying abnormal feature data, and generating an anomaly detection report based on the abnormal feature data; based on the parameters and model structure for pattern recognition set in a deep learning algorithm, performing multi-level analysis and learning on the abnormal data of the anomaly detection report, extracting feature patterns of the abnormal data, and generating a pattern recognition result based on the feature patterns.
[0012] As a preferred scheme, the steps of performing multi-level analysis and learning on the abnormal data of the abnormal detection report, extracting the characteristic pattern of the abnormal data, and generating a pattern recognition result based on the characteristic pattern include: performing feature decomposition on the abnormal detection report by a data decomposition method, extracting the main characteristic components of the abnormal data, and reconstructing the main characteristic components by a feature reconstruction algorithm to obtain abnormal features; performing deep feature learning on the abnormal data according to the abnormal features to obtain a deep feature representation corresponding to the abnormal data; classifying the abnormal data based on the deep feature representation to confirm the category information of the abnormal data; performing health status assessment on a memory chip by using the deep feature representation and the category information to obtain a health status score; and generating a pattern recognition result by combining the health status score and the category information with the working parameters and usage environment of the memory.
[0013] As a preferred scheme, the steps of performing a heat map analysis on the pattern recognition results to generate heat map data, performing a thermal analysis on the performance data set based on the heat map data to generate a heat distribution map include: inputting the pattern recognition results into a heat map analysis model, analyzing the characteristic patterns and abnormality types in the pattern recognition results using a heat calculation algorithm through the heat map analysis model to generate heat map data; combining the heat map data with the electrical characteristic parameters in the performance data set to obtain a fused data set, performing a multi-dimensional thermal analysis on the fused data set, calculating the heat values of each characteristic point of the fused data set, and generating a corresponding heat distribution map.
[0014] As a preferred scheme, the step of comprehensively evaluating the clustering results, the performance data set, the pattern recognition results and the heat distribution map to obtain a test evaluation report includes: uniformly formatting the clustering results, the performance data set, the pattern recognition results and the heat distribution map based on multi-source data fusion technology; performing correlation analysis on the clustering results, the performance data set, the pattern recognition results and the heat distribution map after the uniform formatting to obtain correlation analysis results; inputting the correlation analysis results into a preset comprehensive evaluation model to evaluate the overall performance of the stacked package memory to obtain a test evaluation report.
[0015] The present application also provides a test system for a stacked package memory, comprising: an acquisition unit, used to acquire electrical performance data of each memory chip in the stacked package memory, input all the electrical performance data into a preset test model, and output electrical characteristic parameters; a processing unit, used to perform multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and perform feature extraction on the standardized data set to obtain a feature vector; an analysis unit, used to perform cluster analysis on the feature vector to obtain a clustering result, and screen the clustering result based on a preset performance standard to obtain a performance data set; a detection unit, used to perform anomaly detection on the performance data set, generate an anomaly detection report, and perform pattern recognition processing on the anomaly detection report to obtain a pattern recognition result; a generation unit, used to perform heat map analysis on the pattern recognition result to generate heat map data, and perform heat analysis on the performance data set based on the heat map data to generate a heat distribution map; an evaluation unit, used to perform comprehensive evaluation on the clustering result, the performance data set, the pattern recognition result, and the heat distribution map to obtain a test evaluation report.
[0016] Compared with the prior art, the present application has the following beneficial effects: high test accuracy and comprehensive evaluation. Through comprehensive collection, processing and analysis of the electrical performance data of stacked package memory, accurate testing and comprehensive evaluation of memory chips are achieved; advanced technical means such as multi-dimensional data processing, feature extraction, cluster analysis, anomaly detection and heat map analysis are adopted to improve data processing accuracy and detection efficiency, ensure the overall quality and performance stability of memory chips, and conduct in-depth analysis and visual display of chip performance through pattern recognition and heat map analysis, improve data processing accuracy and detection efficiency, ensure the overall quality and performance stability of stacked package memory, and improve the inability to achieve accurate testing and comprehensive evaluation of stacked package memory, and the problems of insufficient test accuracy and inability to comprehensively evaluate chip performance. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0018] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.
[0019] Figure 1 It is a flowchart of a test method of a stacked package memory provided by an embodiment of the present invention;
[0020] Figure 2 It is a schematic block diagram of the structure of a test system for a stacked package memory provided by an embodiment of the present invention.
[0021] Description of reference numerals:
[0022] 10. A test system for stacked package memory; 11. An acquisition unit; 12. A processing unit; 13. An analysis unit; 14. A detection unit; 15. A generation unit; 16. An evaluation unit. DETAILED DESCRIPTION
[0023] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0024] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.
[0025] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.
[0026] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.
[0027] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.
[0028] Embodiment 1:
[0029] like Figure 1 As shown, the test method of the stacked package memory provided by the embodiment of the present application includes steps S100 to S600.
[0030] Step S100: Collect electrical performance data of each memory chip in the stacked package memory, input all electrical performance data into a preset test model, and output electrical characteristic parameters.
[0031] In this step, electrical test instruments are used to conduct comprehensive electrical performance tests on each memory chip, collecting multiple parameters including current, voltage, frequency, signal integrity, etc. Specifically, the different working states of each chip are tested in turn through automated test equipment (ATE), and its electrical performance data under different temperatures and load conditions is recorded to ensure the comprehensiveness and accuracy of the data.
[0032] For example, high-precision oscilloscopes and logic analyzers from Tektronix or Agilent are used to test the voltage fluctuations and signal delays of each chip in various states such as reading, writing, and standby. These data are summarized and input into a dedicated test model to output electrical characteristic parameters such as operating voltage range, maximum operating frequency, etc.
[0033] Step S200: perform multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and perform feature extraction on the standardized data set to obtain a feature vector.
[0034] In this step, the collected electrical characteristic parameters are normalized to eliminate data deviations under different test conditions; specifically, the principal component analysis (PCA) method is used to reduce the dimension of the normalized data to extract the feature vector that best represents the chip performance.
[0035] For example, the scikit-learn library in the Python programming language is used to standardize the electrical characteristic parameter data set, and the PCA algorithm is applied to reduce the original multi-dimensional electrical performance data into several main eigenvectors. These eigenvectors can reflect the main electrical performance characteristics of the memory chip, such as signal integrity and power supply noise.
[0036] Step S300: cluster analysis is performed on the feature vectors to obtain clustering results, and the clustering results are screened based on preset performance standards to obtain a performance data set.
[0037] In this step, the extracted feature vectors are clustered using K-means or hierarchical clustering algorithms to identify groups of chips with similar performance. Specifically, an appropriate number of clusters is set, and the clustering algorithm is executed to divide chips with similar electrical characteristics into several categories, and chips that meet the requirements are screened out based on preset performance standards.
[0038] For example, the K-means clustering algorithm is used, the number of clusters is set to 5, and the feature vectors are clustered to obtain five chip groups. Then, the central value of each group is compared with the preset performance standards (such as the maximum allowable delay and the minimum allowable voltage) to screen out chips with stable performance and meet the standards, forming a performance data set.
[0039] Step S400: perform anomaly detection on the performance data set, generate an anomaly detection report, perform pattern recognition processing on the anomaly detection report, and obtain a pattern recognition result.
[0040] In this step, statistical methods and machine learning algorithms are used to detect anomalies in the performance data set and identify abnormal data points. Specifically, the Isolation Forest algorithm or a statistically based anomaly detection method is used to mark samples in the data set that are significantly deviated from the normal range, and a detailed anomaly detection report is generated.
[0041] For example, the isolation-forest library in Python is used to perform anomaly detection on performance data sets, identify chips with abnormal electrical characteristic parameters, and generate an anomaly detection report that lists in detail the specific parameters and degree of deviation of each abnormal sample. The support vector machine (SVM) algorithm is then used to perform pattern recognition on the anomaly detection report to identify abnormal patterns and causes, such as specific production batches or process problems.
[0042] Step S500: Perform heat map analysis on the pattern recognition results to generate heat map data, and perform heat analysis on the performance data set based on the heat map data to generate a heat distribution map.
[0043] In this step, the pattern recognition results are visualized using heat map analysis technology to more intuitively analyze data distribution and anomalies. Specifically, the pattern recognition results are processed using a heat map generation tool to generate heat map data that reflects the distribution of different electrical characteristic parameters, and further thermal analysis is performed based on these data to generate a heat distribution map.
[0044] For example, the heatmap function in the Matplotlib library is used to perform heat map analysis on the pattern recognition results to generate a heat map showing the distribution of various electrical characteristic parameters among different chips. By analyzing the hot spots in the heat map, it is determined which electrical performance parameters have the greatest impact on chip performance. Finally, a heat distribution map is generated to intuitively display the distribution of important parameters in the performance data set.
[0045] Step S600: Comprehensively evaluate the clustering results, performance data set, pattern recognition results, and heat distribution map to obtain a test evaluation report.
[0046] In this step, all previous analysis results are comprehensively considered, a comprehensive performance evaluation is conducted, and a final test evaluation report is generated. Specifically, the overall performance of the stacked package memory is evaluated in detail based on the cluster analysis results, the overall performance of the performance data set, the abnormal patterns in the pattern recognition results, and the hot spots in the heat distribution map, and a test evaluation report containing all the analysis content is generated.
[0047] For example, use Excel or dedicated data analysis software to summarize and cross-analyze clustering results, performance data sets, pattern recognition results, and heat distribution maps to form a detailed test evaluation report. The report includes detailed performance indicators of each chip, abnormal detection results, abnormal cause analysis, and overall performance distribution, providing data support for subsequent quality control and process improvements.
[0048] In the above embodiment, the electrical performance data of each memory chip in the stacked package memory is collected, and all the electrical performance data are input into a preset test model to output electrical characteristic parameters; then, multi-dimensional data processing is performed on the electrical characteristic parameters to obtain a standardized data set, and feature vectors are extracted therefrom; then, cluster analysis is performed on the feature vectors to obtain clustering results, and the clustering results are screened based on preset performance standards to obtain a performance data set; further, anomaly detection is performed on the performance data set to generate an anomaly detection report, and pattern recognition processing is performed on it to obtain a pattern recognition result; next, a heat map analysis is performed on the pattern recognition result to generate heat map data, and a heat analysis is performed on the performance data set based on the heat map data to generate a heat distribution map; finally, the clustering results, the performance data set, the pattern recognition results and the heat distribution map are comprehensively evaluated to generate a test evaluation report.
[0049] By comprehensively collecting and processing the electrical performance data of stacked package memory, using preset test models, multi-dimensional data processing, feature extraction and cluster analysis, accurate testing and comprehensive evaluation of memory chips are achieved. Compared with traditional methods, this method can more accurately identify and screen out chips with superior performance, while effectively detecting and reporting abnormal data. Through pattern recognition and heat map analysis, in-depth analysis and visual display of chip performance are carried out, which improves data processing accuracy and detection efficiency, ensures the overall quality and performance stability of stacked package memory, and improves the problem of inability to achieve accurate testing and comprehensive evaluation of stacked package memory, insufficient test accuracy and inability to comprehensively evaluate chip performance.
[0050] Embodiment 2:
[0051] In step S100, each memory chip in the stacked package memory is connected to an electrical testing device one by one, a preset test voltage and test current are applied to each memory chip according to the electrical testing device, and the electrical performance data of each memory chip is collected in real time through the electrical testing device; all electrical performance data are preprocessed; wherein the preprocessing includes denoising, filtering and normalization processing; all preprocessed electrical performance data are input into a preset test model, and the corresponding electrical characteristic parameters are output.
[0052] Each memory chip is tested one by one using electrical testing equipment to ensure the consistency of test conditions and data accuracy of each chip. Specifically, an automated test system (ATE) is used to connect each memory chip, and the electrical performance of each chip under different working conditions is gradually tested according to the preset test voltage and current. The electrical performance data is collected and recorded in real time, and then data preprocessing is performed, including removing noise signals, applying bandpass filters to remove high-frequency and low-frequency interference, and normalizing data of different dimensions to ensure data consistency and comparability.
[0053] For example, the Agilent 3458A multimeter is used for high-precision voltage and current measurements. In combination with the NIPXI automatic test system, the test process is controlled by writing LabVIEW scripts. The electrical performance data such as current and voltage of each chip in the reading, writing and idle states are measured one by one, and the collected data is denoised, filtered and normalized. The processed data is input into the test model built based on the TensorFlow framework to obtain the electrical characteristic parameters of each chip.
[0054] In step S200, data cleaning is performed on the electrical characteristic parameters; wherein the data cleaning includes removing outliers and filling missing values; the cleaned electrical characteristic parameters are normalized, and data of different dimensions of the electrical characteristic parameters are converted into the same dimension; the normalized electrical characteristic parameters are reduced in dimension using principal component analysis to obtain a standardized data set; representative features in the standardized data set are extracted based on a convolutional neural network, and the extracted representative features are integrated to obtain a feature vector; wherein the representative features refer to characteristic values that reflect the key performance of the memory chip.
[0055] Data quality and processing efficiency are improved by cleaning, normalizing and reducing data dimensions. Specifically, the IQR (Interquartile Range) method is first used to identify and remove outliers, and linear interpolation is applied to fill missing values. The cleaned data is then normalized to convert all electrical characteristic parameters into the same dimension. The PCA method is then used to reduce the dimension of the normalized data to reduce the data dimension and retain important features. Finally, a convolutional neural network (CNN) is used to extract and integrate features from the standardized data set after dimensionality reduction to obtain a feature vector that can reflect the key performance of the chip.
[0056] For example, the pandas library in Python is used for data cleaning. First, the IQR method is used to remove outliers that exceed 1.5 times the IQR range. The interpolate module in the SciPy library is used to perform linear interpolation on missing data. Then, the MinMaxScaler in the scikit-learn library is used for normalization, and the data is scaled to the range of 0 to 1. The PCA algorithm is then used to reduce the dimension of the normalized data, retaining 95% of the variance information. Finally, a convolutional neural network is constructed using the Keras library to extract and integrate features of the reduced-dimensional data to obtain feature vectors.
[0057] In step S300, the feature vector is input into a preset cluster analysis model, and the feature vector is clustered by using the K-means clustering algorithm in the cluster analysis model to obtain a clustering result; the clustering result is screened based on the power consumption threshold, voltage stability, and response time limit of the preset performance standard to obtain clusters that meet the performance standard and clusters that do not meet the performance standard, remove the clusters that do not meet the performance standard, and recombine the cluster data that meet the preset performance standard to obtain a performance data set.
[0058] By performing cluster analysis on the feature vectors, memory chips that meet the preset performance standards are screened. Specifically, the K-means clustering algorithm is used to divide the feature vectors into several categories. A distance matrix is formed based on the calculated Euclidean distance between the feature vectors. The cluster centers are updated iteratively to finally obtain the clustering results. The clustering results are then screened based on preset standards such as power consumption threshold, voltage stability, and response time limit. Clusters that do not meet the standards are removed, and cluster data that meet the standards are recombined to form the final performance data set.
[0059] For example, the KMeans algorithm in Python is used for cluster analysis. The number of clusters is set to 5. The Euclidean distance between the eigenvectors is calculated to obtain the distance matrix. The cluster centers are updated through 100 iterations to eventually form 5 clustering results. Then, based on the preset performance standards of power consumption not exceeding 10mW, voltage fluctuation less than 5%, and response time not exceeding 10ns, clusters that meet the standards are screened out, cluster data that do not meet the standards are removed, and data that meet the standards are recombined to obtain the final performance data set.
[0060] Among them, the step of obtaining clustering results by clustering feature vectors using the K-means clustering algorithm in the clustering analysis model includes using a distance calculation method to measure the distance of feature vectors to obtain a distance matrix between feature vectors; clustering and updating the distance matrix to obtain a cluster center, and calculating the cluster to which the feature vector of each memory chip in the memory belongs according to the cluster center; performing performance screening according to the cluster and preset standards, eliminating clustering results that do not meet the requirements, and obtaining a screened performance data set; comparing and analyzing the screened performance data set and the eliminated non-performance data set, extracting features with obvious performance differences, and obtaining clustering results.
[0061] By using the distance calculation method and clustering update method, the feature vectors are clustered and screened in detail. Specifically, the Euclidean distance between the feature vectors is calculated to obtain the distance matrix, and the cluster centers are continuously adjusted through the iterative update algorithm until the clustering results are stable. Then, the cluster to which each chip feature vector belongs is calculated according to each cluster center. Screening is performed based on preset standards such as power consumption, voltage stability, and response time. Clustering results that do not meet the requirements are eliminated to obtain a screened performance data set. The performance data set and the eliminated non-performance data set are compared and analyzed in detail to extract features with obvious performance differences.
[0062] For example, the Euclidean distance between feature vectors is calculated using the NumPy library to obtain a distance matrix, which is then iteratively updated using the KMeans algorithm. By comparing the distance between the cluster center and the sample point, the position of the cluster center is continuously adjusted until the clustering result is stable. Next, the distance between each chip feature vector and the cluster center is used to determine the cluster to which the chip belongs. The cluster results that do not meet the requirements are screened based on preset standards such as power consumption, voltage stability, and response time, and the filtered performance data set is obtained. The performance data set and the non-performance data set are compared and analyzed using the Pandas library to extract features with obvious differences in power consumption and response time.
[0063] In step S400, the performance data set is input into the anomaly detection model, and the performance data set is analyzed by using the isolation forest algorithm to identify abnormal feature data, and an anomaly detection report is generated based on the abnormal feature data; based on the parameters and model structure of pattern recognition set in the deep learning algorithm, the abnormal data of the anomaly detection report is subjected to multi-level analysis and learning, the characteristic patterns of the abnormal data are extracted, and the pattern recognition results are generated based on the characteristic patterns.
[0064] Anomaly detection is performed using the isolation forest algorithm, and pattern recognition is performed using the deep learning algorithm. Specifically, the isolation forest algorithm is used to analyze the performance data set, identify and mark abnormal feature data, and generate an anomaly detection report. Then, based on the deep learning algorithm, multi-level analysis and learning are performed on the abnormal data in the anomaly detection report to extract the main feature patterns of the abnormal data. The pattern recognition results are generated through classification and clustering analysis of the feature patterns.
[0065] For example, the IsolationForest algorithm in the scikit-learn library is used to detect anomalies in performance data sets, identify outliers in feature data such as power consumption, voltage fluctuations, and response time, and generate a detailed anomaly detection report that lists the specific characteristics and causes of each abnormal data point. Then, based on the deep learning model in TensorFlow, a multi-level analysis of the abnormal data is performed to extract abnormal patterns. The categories and patterns of the abnormal data are identified through classification and clustering algorithms, and finally a pattern recognition result is generated.
[0066] Among them, the steps of performing multi-level analysis and learning on the abnormal data of the abnormal detection report, extracting the characteristic pattern of the abnormal data, and generating pattern recognition results based on the characteristic pattern include performing feature decomposition on the abnormal detection report by a data decomposition method, extracting the main characteristic components of the abnormal data, reconstructing the main characteristic components by a feature reconstruction algorithm, and obtaining abnormal features; performing deep feature learning on the abnormal data according to the abnormal features, and obtaining a deep feature representation corresponding to the abnormal data; classifying and processing the abnormal data based on the deep feature representation, and confirming the category information of the abnormal data; using the deep feature representation and category information to evaluate the health status of the memory chip and obtain a health status score; and generating pattern recognition results through the health status score and category information, combined with the working parameters and usage environment of the memory.
[0067] Through deep feature learning, we can better understand the intrinsic characteristics and patterns of abnormal data. Specifically, deep learning models (such as deep neural networks) are used to train abnormal data to learn the correlation and regularity between abnormal data. For example, using model structures such as convolutional neural networks (CNN) or recurrent neural networks (RNN), features of abnormal data are extracted and abstractly represented, thereby obtaining a deep feature representation of abnormal data. In this process, by training the network layer by layer, the high-order features of abnormal data are abstracted and learned, making the deep feature representation of abnormal data more discriminative and expressible.
[0068] Furthermore, based on deep feature representation, abnormal data can be classified and processed to confirm the category information of abnormal data. For example, using a trained deep learning model, abnormal data can be classified and predicted into different abnormal types or situations to more accurately understand the nature and source of abnormal data. Through this step, abnormal data can be analyzed more carefully and accurately, providing a more reliable basis for subsequent health status assessment.
[0069] The health status of memory chips can be evaluated by using deep feature representation and category information. For example, a comprehensive health status evaluation model can be constructed by combining the deep feature representation and classification information of abnormal data, as well as the working parameters and usage environment of the memory. By comprehensively considering the abnormal conditions, working conditions and environmental factors of the memory, the health status of the memory chip is quantitatively evaluated to obtain the corresponding health status score. This score can help users fully understand the working status and performance of the memory chip, and provide guidance and reference for subsequent maintenance and optimization.
[0070] Pattern recognition results can be generated by combining the health status score and category information with the working parameters and usage environment of the memory. For example, the corresponding pattern recognition results can be generated based on the health status score of the memory and the category information of the abnormal data, combined with the working parameters (such as temperature, voltage, frequency, etc.) and usage environment (such as workload, usage time, etc.) of the memory. This result can help users understand the working conditions and performance characteristics of the memory more comprehensively, and provide an important reference for the management and optimization of the memory.
[0071] In step S500, the pattern recognition results are input into the thermal map analysis model, and the characteristic patterns and abnormal types in the pattern recognition results are analyzed by the thermal map analysis model using a thermal calculation algorithm to generate thermal map data; the thermal map data are combined with the electrical characteristic parameters in the performance data set to obtain a fused data set, and the fused data set is subjected to a multi-dimensional thermal analysis to calculate the thermal values of each characteristic point of the fused data set to generate a corresponding thermal distribution map.
[0072] The pattern recognition results are further analyzed through the heat map analysis model, and multi-dimensional thermal analysis is performed in combination with the electrical characteristic parameters in the performance data set to evaluate the working status and performance of the memory. Specifically, the characteristic patterns and abnormal types in the pattern recognition results are analyzed using the heat calculation algorithm to generate heat map data, which is then fused with the performance data set. The heat value of each feature point is calculated to generate the corresponding heat distribution map, which intuitively displays the heat distribution of each feature of the memory.
[0073] For example, the Matplotlib library is used to perform heat map analysis on the characteristic patterns and anomaly types in the pattern recognition results, and heat map data is generated according to the density and distribution of the feature points. The heat map data is then combined with the electrical characteristic parameters in the performance data set. The Seaborn library is used to perform multi-dimensional thermal analysis on the fused data set, calculate the heat value of each feature point, and draw the corresponding heat distribution map to intuitively display the working status and performance of the memory.
[0074] In step S600, the clustering results, performance data sets, pattern recognition results and heat distribution maps are uniformly formatted based on multi-source data fusion technology; correlation analysis is performed on the uniformly formatted clustering results, performance data sets, pattern recognition results and heat distribution maps to obtain correlation analysis results; the correlation analysis results are input into a preset comprehensive evaluation model to evaluate the overall performance of the stacked package memory to obtain a test evaluation report.
[0075] By utilizing multi-source data fusion technology, the data generated in each step are integrated and uniformly formatted for subsequent comprehensive analysis. Specifically, the clustering results, performance data sets, pattern recognition results, and heat distribution maps are uniformly formatted to eliminate the differences between the data. Then, association analysis is performed to explore the association relationship between different data. The association analysis results are input into a comprehensive evaluation model to evaluate the overall performance of the memory and generate a test evaluation report to provide guidance and basis for further optimization of the memory.
[0076] For example, the Pandas library is used to uniformly format clustering results, performance data sets, pattern recognition results, and heat distribution maps into DataFrames to ensure consistency in data fields and data types. Then, correlation coefficient analysis methods are used to perform association analysis on each data to explore the correlation and influencing factors between different data. Finally, the association analysis results are input into the preset comprehensive evaluation model to evaluate the overall performance of the storage device and generate a comprehensive test evaluation report to provide a reference for performance improvement and problem optimization of the storage device.
[0077] In this embodiment, by implementing the above steps, accurate analysis and comprehensive evaluation of multi-dimensional electrical performance data of the stacked package memory are ensured, effectively improving the accuracy and comprehensiveness of memory chip testing.
[0078] Embodiment 3:
[0079] like Figure 2 As shown, the present application also provides a test system 10 for stacked package memory, and the test system 10 for stacked package memory includes a collection unit 11, a processing unit 12, an analysis unit 13, a detection unit 14, a generation unit 15 and an evaluation unit 16.
[0080] The acquisition unit 11 is mainly used to acquire the electrical performance data of each memory chip in the stacked package memory, input all the electrical performance data into a preset test model, and output electrical characteristic parameters.
[0081] The processing unit 12 is mainly used to perform multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and perform feature extraction on the standardized data set to obtain a feature vector.
[0082] The analysis unit 13 is mainly used to perform cluster analysis on the feature vectors to obtain cluster results, and to screen the cluster results based on preset performance standards to obtain a performance data set.
[0083] The detection unit 14 is mainly used to perform anomaly detection on the performance data set, generate anomaly detection report, perform pattern recognition processing on the anomaly detection report, and obtain pattern recognition results.
[0084] The generating unit 15 is mainly used to perform heat map analysis on the pattern recognition result to generate heat map data, perform heat analysis on the performance data set based on the heat map data, and generate a heat distribution map.
[0085] The evaluation unit 16 is mainly used to perform a comprehensive evaluation on the clustering results, performance data sets, pattern recognition results and heat distribution diagrams to obtain a test evaluation report.
[0086] In this embodiment, a highly automated and integrated stacked package memory test system is constructed to achieve comprehensive electrical performance evaluation and testing of memory chips. Specifically, the test system uses advanced electronic test equipment and data processing technology to start with collecting electrical performance data, go through multiple processing stages, and finally generate a detailed test evaluation report.
[0087] It should be noted that technicians in the relevant technical field can clearly understand that for the convenience and simplicity of description, the specific working process of the system and each unit described above can refer to the corresponding process in the aforementioned stacked package memory test method embodiment, and will not be repeated here.
[0088] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for testing a stacked package memory, characterized in that: include: Collecting electrical performance data of each memory chip in the stacked package memory, inputting all the electrical performance data into a preset test model, and outputting electrical characteristic parameters; Performing multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and performing feature extraction on the standardized data set to obtain a feature vector; The feature vector is input into a preset cluster analysis model, and the feature vector is measured by distance calculation method to obtain a distance matrix between the feature vectors; the distance matrix is clustered and updated to obtain a cluster center, and the cluster to which the feature vector of each memory chip in the memory belongs is calculated according to the cluster center; performance screening is performed according to the cluster and preset standards, and clustering results that do not meet the requirements are eliminated to obtain a screened performance data set; Comparing and analyzing the screened performance data set and the eliminated non-performance data set, extracting features with obvious performance differences, and obtaining clustering results; The clustering results are screened based on the power consumption threshold, voltage stability, and response time limit of the preset performance standard to obtain clusters that meet the performance standard and clusters that do not meet the performance standard, the clusters that do not meet the performance standard are removed, and the cluster data that meet the preset performance standard are recombined to obtain a performance data set; The performance data set is input into an anomaly detection model, the performance data set is analyzed by using an isolation forest algorithm to identify abnormal feature data, and an anomaly detection report is generated based on the abnormal feature data; based on the parameters and model structure of pattern recognition set in the deep learning algorithm, the feature decomposition of the anomaly detection report is performed by a data decomposition method, the main feature components of the abnormal data are extracted, and the main feature components are reconstructed by a feature reconstruction algorithm to obtain abnormal features; deep feature learning is performed on the abnormal data according to the abnormal features to obtain a deep feature representation corresponding to the abnormal data; the abnormal data is classified and processed based on the deep feature representation to confirm the category information of the abnormal data; the health status of the memory chip is evaluated by using the deep feature representation and the category information to obtain a health status score; the pattern recognition result is generated by combining the health status score and the category information with the working parameters and usage environment of the memory; Performing a heat map analysis on the pattern recognition result to generate heat map data, and performing a heat analysis on the performance data set based on the heat map data to generate a heat distribution map; A comprehensive evaluation is performed on the clustering results, the performance data set, the pattern recognition results and the heat distribution map to obtain a test evaluation report.
2. The method for testing a stacked package memory according to claim 1, wherein: The step of collecting electrical performance data of each memory chip in the stacked package memory, inputting all the electrical performance data into a preset test model, and outputting electrical characteristic parameters includes: Connecting each memory chip in the stacked package memory to an electrical testing device one by one, applying a preset test voltage and test current to each memory chip according to the electrical testing device, and collecting electrical performance data of each memory chip in real time through the electrical testing device; Preprocessing all the electrical performance data; wherein the preprocessing includes denoising, filtering and normalization processing; All pre-processed electrical performance data are input into the preset test model and the corresponding electrical characteristic parameters are output.
3. The method for testing a stacked package memory according to claim 1, wherein: The step of performing multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and performing feature extraction on the standardized data set to obtain a feature vector includes: Performing data cleaning on the electrical characteristic parameters; wherein the data cleaning includes removing outliers and filling missing values; Normalizing the electrical characteristic parameters after cleaning, and converting data of different dimensions of the electrical characteristic parameters into the same dimension; Performing dimensionality reduction processing on the normalized electrical characteristic parameters using principal component analysis to obtain a standardized data set; Based on a convolutional neural network, representative features in the standardized data set are extracted, and the extracted representative features are integrated to obtain a feature vector; wherein the representative features refer to characteristic values that reflect the key performance of the memory chip.
4. The method for testing a stacked package memory according to claim 1, wherein: The step of performing heat map analysis on the pattern recognition result to generate heat map data, and performing heat analysis on the performance data set based on the heat map data to generate a heat distribution map comprises: Inputting the pattern recognition results into a heat map analysis model, and using the heat map analysis model to analyze the characteristic patterns and abnormal types in the pattern recognition results using a heat calculation algorithm to generate heat map data; The thermal map data is combined with the electrical characteristic parameters in the performance data set to obtain a fused data set, a multi-dimensional thermal analysis is performed on the fused data set, the thermal value of each feature point of the fused data set is calculated, and a corresponding thermal distribution map is generated.
5. The method for testing a stacked package memory according to claim 1, wherein: The step of comprehensively evaluating the clustering result, the performance data set, the pattern recognition result, and the heat distribution map to obtain a test evaluation report includes: Performing unified formatting processing on the clustering results, the performance data set, the pattern recognition results and the heat distribution map based on multi-source data fusion technology; Performing correlation analysis on the clustering results, the performance data set, the pattern recognition results, and the heat distribution map after unified formatting to obtain a correlation analysis result; The correlation analysis result is input into a preset comprehensive evaluation model to evaluate the overall performance of the stacked package memory to obtain a test evaluation report.
6. A test system for stacked package memory, characterized in that: include: A collection unit, used to collect electrical performance data of each memory chip in the stacked package memory, input all the electrical performance data into a preset test model, and output electrical characteristic parameters; A processing unit, configured to perform multi-dimensional data processing on the electrical characteristic parameters to obtain a standardized data set, and perform feature extraction on the standardized data set to obtain a feature vector; The analysis unit is used to input the feature vector into a preset cluster analysis model, perform distance measurement on the feature vector using a distance calculation method, and obtain a distance matrix between the feature vectors; cluster update the distance matrix to obtain a cluster center, and calculate the cluster to which the feature vector of each memory chip in the memory belongs according to the cluster center; perform performance screening according to the cluster and preset standards, and eliminate cluster results that do not meet the requirements to obtain a screened performance data set; Comparing and analyzing the screened performance data set and the eliminated non-performance data set, extracting features with obvious performance differences, and obtaining clustering results; The clustering results are screened based on the power consumption threshold, voltage stability, and response time limit of the preset performance standard to obtain clusters that meet the performance standard and clusters that do not meet the performance standard, the clusters that do not meet the performance standard are removed, and the cluster data that meet the preset performance standard are recombined to obtain a performance data set; A detection unit is used to input the performance data set into an anomaly detection model, analyze the performance data set by using an isolation forest algorithm, identify abnormal feature data, and generate an anomaly detection report based on the abnormal feature data; based on the parameters and model structure of pattern recognition set in the deep learning algorithm, feature decomposition of the abnormal detection report is performed by a data decomposition method, the main feature components of the abnormal data are extracted, and the main feature components are reconstructed by a feature reconstruction algorithm to obtain abnormal features; deep feature learning is performed on the abnormal data according to the abnormal features to obtain a deep feature representation corresponding to the abnormal data; classification processing is performed on the abnormal data based on the deep feature representation to confirm the category information of the abnormal data; the health status of the memory chip is evaluated by using the deep feature representation and the category information to obtain a health status score; the pattern recognition result is generated by combining the health status score and the category information with the working parameters and usage environment of the memory; A generating unit, configured to perform a heat map analysis on the pattern recognition result to generate heat map data, and perform a heat analysis on the performance data set based on the heat map data to generate a heat distribution map; An evaluation unit is used to comprehensively evaluate the clustering results, the performance data set, the pattern recognition results and the heat distribution map to obtain a test evaluation report.
