Adaptive test method for integrated circuit parameter test

Abstract

The invention discloses an adaptability test method for an integrated circuit parameter test, and the method is used for the integrated circuit parameter test, and comprises the steps: analyzing the contribution degree of a test item to a test result and a maximum information coefficient between the test items, reasoning an optimal test subset, and obtaining a test result. And on the basis of the quality prediction result of the quality prediction model on the qualified products, the qualified products are divided into a plurality of quality grades more deeply, and potential fault products are further screened out. According to the adaptive test method, the test cost of parameter test can be reduced to the greatest extent when the adaptive test method is applied to integrated circuit products of different types and different production scales, and an extremely low test escape rate is always kept.

Description

Technical Field

[0001] The invention relates to an adaptability testing method, in particular to an adaptability testing method for integrated circuit parameter testing. Background Art

[0002] With the development of integrated circuit technology, the process of integrated circuit products is becoming more and more complex, the testing process is becoming more and more complicated, and the testing cost is also increasing. As an important branch of integrated circuit testing, parameter testing has long accounted for a large proportion of the manufacturing cost of integrated circuit products, which has seriously limited the market competitiveness and profit margins of integrated circuit products.

[0003] In order to meet this challenge, the industry is actively exploring ways to reduce the testing cost of parameter testing by introducing adaptive testing methods to replace traditional testing methods. For example, in test strategy optimization, the fixed test process is dynamically adjusted according to the test results to detect potential defects of the product at an earlier stage and reduce unnecessary test steps; in test process optimization, the test items are deleted based on the correlation between the test items and the test results, and the test results are statistically analyzed to predict product quality, which significantly reduces the test cost by sacrificing a certain degree of test accuracy.

[0004] Existing test strategy optimization methods and test process optimization methods start from evaluation indicators related to individual test items, such as fault detection rate, correlation with test results, and statistical analysis results. According to the contribution of test items to test results, test items with low contribution are deleted, and test items with high fault detection rate are prioritized, which effectively reduces the test cost. However, existing methods often only consider the contribution of test items to test results, and do not consider the correlation between test items. This may result in test items with high contribution but redundancy being retained, failing to reduce test costs to the greatest extent; test items with low contribution but not redundancy are deleted, sacrificing fault coverage to reduce test costs, and there is a risk of test escape.

[0005] Existing test process optimization methods rely on modeling of specific test objects for quality prediction, which requires samples to be large enough to have relatively accurate prediction results. For chip-level testing, existing methods can significantly reduce test costs and maintain a high quality prediction accuracy; however, for small-batch, customized integrated circuit products, overly complex modeling makes the existing methods very limited in reducing test costs. At the same time, the small product scale makes the existing methods less accurate in quality prediction, and there is a high test escape rate in actual applications.

[0006] Based on the various characteristics of the above-mentioned actual engineering and the application defects of existing related methods, the present invention proposes an adaptive testing method for integrated circuit parameter testing, which aims to minimize the testing cost and maintain an extremely low test escape rate when dealing with parameter testing of integrated circuit products of different types and production scales. Summary of the invention

[0007] The present invention solves the problem that the test cost of integrated circuit parameter testing is difficult to effectively reduce, and makes up for the shortcomings of the above-mentioned existing adaptive testing technology, and proposes an adaptive testing method for integrated circuit parameter testing. This method can always maintain an extremely low test escape rate when facing integrated circuit products of different types and different production scales, and reasonably reduce and reorder the test items, and infer the optimal test subset, thereby effectively reducing the test cost of integrated circuit parameter testing and significantly improving the test efficiency.

[0008] The adaptive testing method is oriented to integrated circuit parameter testing. Specifically, the method firstly infers the optimal test subset, that is, by analyzing the correlation information contained in the evaluation indicators of different dimensions such as the maximum information coefficient of the test items and the process capability index, the correlation information between the test items and the test results, and the test items is gradually mined, and then the test items are reasonably reduced and reordered to improve the test efficiency and reduce the test cost; then, based on the quality prediction results of qualified products by the quality prediction model, the qualified products are further divided into multiple quality levels, and potential faulty products are further screened out, so as to control the test escape rate at an extremely low level.

[0009] The characteristics of the present invention are:

[0010] (1) The adaptive testing method is based on the correlation between test items and test results, and the mutual correlation between test items. It performs comprehensive analysis, reasonably reduces and reorders the test items, and infers the optimal test subset, which can optimize the test process to the greatest extent, improve the test efficiency, and reduce the test cost of integrated circuit parameter testing;

[0011] (2) The adaptive testing method uses a machine learning method to predict the quality of qualified products and classifies the quality prediction results by a clustering algorithm, which can effectively identify and screen qualified products with potential faults, thereby maintaining an extremely low test escape rate and ensuring the accuracy and reliability of parameter testing;

[0012] (3) The adaptive testing method described herein only relies on a small amount of test data obtained through sampling, does not require historical test data or prior knowledge, and can autonomously infer the optimal test subset based on the actual test data of the product to be tested, and train the optimal quality assessment model. It has wide applicability and good versatility, and can cope with integrated circuit products of different types and different production scales. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 Adaptive test method flow for integrated circuit parameter testing

[0014] Figure 2 Reasoning process for optimal test subset based on contribution redundancy analysis

[0015] Figure 3 Screening process for potentially faulty products based on quality grade DETAILED DESCRIPTION

[0016] An adaptive testing method for integrated circuit parameter testing provided by the present invention is described in detail below with reference to the accompanying drawings.

[0017] The present invention provides an adaptive testing method for integrated circuit parameter testing. The adaptive testing method is oriented to integrated circuit parameter testing, and aims to fully explore the correlation between test items and test results, and between test items, optimize the test process, and screen potential faulty products through quality prediction and quality grade classification, without being limited by the type and scale of integrated circuit products, to reduce the test cost of parameter testing to the greatest extent, and always maintain an extremely low test escape rate.

[0018] The adaptive testing method for integrated circuit parameter testing is mainly divided into two steps: optimal test subset reasoning based on contribution redundancy analysis and potential fault product screening based on quality level classification. Figure 1 As shown, the optimal test subset reasoning is based on the contribution redundancy analysis of the sampled test data. First, the test set is reasonably reduced and reordered. While deleting redundant test items, important test items are tested first to improve test efficiency and reduce test costs. Potential faulty product screening then predicts quality based on the test data obtained using the optimal test subset, and divides qualified products into multiple quality levels to further screen out potential faulty products, so that the test escape rate of parameter testing is always maintained at an extremely low level.

[0019] In the optimal test subset reasoning step, each test item is based on the sampled test data, and the correlation between the test items and the test results and the test items is mined according to the multi-dimensional information such as statistical correlation index, process capability coefficient, and cross-correlation index, and the test items are further reasonably reduced and re-ordered. Figure 2 The optimal test subset reasoning process is explained in detail.

[0020] First, the products to be tested are divided into a training set and a test set according to a preset sampling rate β (10% ≤ β ≤ 30%), and the products to be tested in the training set are fully tested, that is, whether the products to be tested are faulty products or not, they are fully tested without skipping any test items, and the sampled test data is obtained. Specifically, the sampled test data contains the test data of each product to be tested in the training set, including the test item data set U = {U1, U2, ..., U n ,…,U N} and the test result data set R = {r1, r2, …, r n ,…,r N}, where N represents the total number of products to be tested in the training set, and U n Represents the test item data of the nth product to be tested, r n Indicates the test result of the nth product to be tested. If r n =0, indicating that the test result is not a fault. n =1, indicating the test result is a fault.

[0021] After obtaining the sample test data, the correlation analysis between the test items and the test results and between the test items will be conducted. In the correlation analysis between the test items and the test results, the Kendall correlation coefficient analysis and process capability index analysis will be combined with the test items and the test results to further calculate the contribution of the test items to the test results. Specifically, for any test item T i , its contribution to the test results G(T i ) is defined as:

[0022]

[0023] In the formula, Kendall (T i ,R) represents the test item T i Kendall correlation coefficient with test results, C pk (T i ) represents the test item T i The process capability index. The contribution of a test item to the test result ranges from 0 to 1. The higher the contribution, the greater the impact of the test item on the test result. In the correlation analysis between test items, the test items will be analyzed for maximum information coefficients in pairs to obtain the maximum information coefficient matrix M of the test items. Specifically, M i,j represents the maximum information coefficient between the i-th test item and the j-th test item. The maximum information coefficient ranges from 0 to 1. i,j The larger the value is, the stronger the correlation between the i-th test item and the j-th test item is.

[0024] After obtaining the contribution G of all test items to the test results and the maximum information matrix M of the test items, the test set T = {T1, T2, …, T n ,…,T N}Reasonably reduce and reorder the test items and infer the optimal test subset T′={T′1,T′2,…,T′ n ,…,T′ L}, where N and L represent the total number of test items in the test set and the optimal test subset, respectively. Specifically, the optimal test subset inference rule contains the following two sub-rules:

[0025] Rule 1: For any test item T i , if there is another test item T j , test item T j Contribution to the test result is greater than test item T i Contribution to the test results, and test item T j With test item T i The maximum information coefficient is greater than the test item T i Contribution to the test results, then the test item T i These are redundant test items and should not be included in the optimal test subset.

[0026] Rule 2: For any two test items T in the optimal test subset T′ i ′ and test item T j ',satisfy:

[0027]

[0028] In the formula, G(T′ i ) and G(T′ j ) represent the test items T′ i With the test item T′ j Contribution to the test results.

[0029] In the potential fault product screening step, the quality prediction model is trained using sample test data, the quality of qualified products is predicted based on the test data of the optimal test subset, and the quality prediction results are clustered and analyzed using a clustering algorithm to achieve multi-level quality classification of qualified products. Figure 3 The process of screening products with potential faults is explained in detail.

[0030] First, the quality prediction model is trained using the sampled test data obtained in the optimal test subset inference step. Considering the serious imbalance in the number of samples of qualified products and faulty products in the sampled test data, the traditional machine learning classification model tends to predict the majority class (qualified products) and cannot effectively identify the minority class (faulty products). The actual quality prediction model uses the KNN regression model (K-Nearest Neighbors Regression) to calculate the distance between any two products through the feature weighted standardized Euclidean distance (Feature Weighted Standardized Euclidean Distance) to improve the accuracy of quality prediction. Specifically, for any product A and product B, their weighted standardized Euclidean distance Defined as:

[0031]

[0032] Where n represents the number of test items in the optimal test subset; ω i Represents the contribution of the i-th test item in the optimal test subset, which is used to fully consider the contribution of each test item to the test results to improve the prediction accuracy and model robustness; The standardized Euclidean distance of only a single test item is considered to eliminate the problem of inconsistent dimensions and scales between different test items. Specifically, for any product A, the quality score S(A) predicted by the quality prediction model is defined as:

[0033]

[0034] In the formula, k represents the k value selected by the KNN regression model, B i represents the product that is the ith nearest neighbor of product A. The higher the quality score predicted by the quality prediction model, the better the quality of the product.

[0035] After the quality prediction model is trained, the quality of qualified products in the test set will be predicted based on the test data obtained by testing only the optimal test subset, and the quality score of each qualified product will be obtained. The output quality score data set S = {s1, s2, …, s n ,…,s N}, where N represents the total number of qualified products in the test set. Considering that inappropriate k value selection may lead to overfitting or underfitting of the quality prediction model, a fixed k value setting is difficult to ensure that the quality prediction model trained for integrated circuit products of different types and sizes can make the most accurate quality prediction. The actual k value is adaptively selected based on the normalized entropy of the sampled test data. Specifically, for any k value, the normalized entropy of the sampled test data Hnorm (k) is defined as:

[0036]

[0037] In the formula, N1, N2, and N represent the number of qualified products in the training set, the number of faulty products in the training set, and the total number of products to be tested in the test set, respectively; n1 and n2 represent the number of qualified products and faulty products in the k nearest neighbor products of any product to be tested x in the test set, respectively; ω1 and ω2 represent the weights of qualified products and unqualified products in calculating information entropy, respectively, and their values ​​are the inverse of the total number of qualified products and unqualified products in the training set, respectively. Normalized information entropy is a non-negative number. The lower the value of normalized information entropy, the lower the uncertainty of the quality prediction result of the quality prediction model for qualified products in the test set under the current k value. Specifically, by exploring the normalized information entropy of the sampled test data in the predetermined k value candidate set, and adaptively selecting the k value that minimizes the normalized information entropy of the sampled test data, the trained quality prediction model can give the most accurate quality prediction result when facing integrated circuit products of different types and sizes.

[0038] After obtaining the quality score data set, the K-Means clustering algorithm is used to classify the quality grades of qualified products in the test set. In this process, the number of clusters of the K-Means clustering algorithm is set to 3, and the quality score values ​​of the 25th, 50th and 75th percentiles in the quality score data set are initialized as cluster centers C = {c1, c2, c3}, and the Euclidean distance between each qualified product in the quality score data set and the cluster center is calculated. Each qualified product is assigned to the cluster center with the nearest distance, and the cluster center is continuously updated and the qualified products are redistributed according to the clustering results until the cluster center no longer changes. Specifically, the K-Means clustering algorithm will divide the qualified products in the test set into three clusters according to the quality score distribution in the quality score data set, and output three cluster labels L = {l1, l2, l3} as the quality level of each cluster, among which the cluster identified by the l1 label is high-quality qualified products, the cluster identified by the l2 label is general quality qualified products, and the cluster identified by the l3 label is low-quality qualified products. Low-quality qualified products represent qualified products that may have potential faults. Qualified products classified in the low-quality qualified product cluster will be fully tested to further screen out potential faulty products.

Claims

1. An adaptive testing method for integrated circuit parameter testing, characterized by: The adaptive testing method for integrated circuit parameter testing is aimed at integrated circuit parameter testing and aims to fully explore the correlation between test items and test results, and between test items, optimize the test process, and screen potential faulty products through quality prediction and quality grade classification. It is not limited by the type and scale of integrated circuit products, minimizes the test cost of parameter testing, and always maintains an extremely low test escape rate.

2. According to claim 1, the adaptive testing method for integrated circuit parameter testing is characterized by: The adaptive testing method for integrated circuit parameter testing is mainly divided into two steps: optimal test subset reasoning based on contribution redundancy analysis and potential fault product screening based on quality level division.

3. According to claim 2, the adaptive testing method for integrated circuit parameter testing is characterized by: Regarding the optimal test subset reasoning step based on contribution redundancy analysis, this step uses sample test data analysis to reduce and reorder test items through the contribution of test items to test results and the maximum information coefficient matrix between test items, infer the optimal test subset, improve test efficiency, and reduce test costs.

4. According to claim 2, the adaptive testing method for integrated circuit parameter testing is characterized by: For the step of screening potential faulty products based on quality grade classification, this step uses sample test data to train the KNN regression model for quality prediction, adaptively adjusts the KNN regression model through normalized information entropy analysis, and uses the K-Means clustering algorithm to classify the quality grades of qualified products, screen potential faulty products, and control the test escape rate at an extremely low level.

Images