基于随机森林的集成电路适应性测试方法研究

An Adaptive Test Method of IC Based on Random Forest

在集成电路测试过程中,随着测试时间的延长,会导致测试成本偏高。针对这种情况,提出一种基于随机森林的适应性测试方法。对于训练模型的芯片,通过计算基尼指数得出芯片参数测试过程中每个测试组对模型分类的重要程度,按照特征重要性对测试组进行重要度排序,筛选出特征重要性最高的测试组,并统计每个测试组测出的缺陷芯片数。对测试集中部分芯片进行测试,并通过删除部分测试组来减少测试时间,采用随机森林算法预测芯片质量,在保证预测准确率的基础上,尽量减少预测时间。实验结果表明:与KNN和逻辑回归算法相比,随机森林算法在预测准确率、测试逃逸水平和运行时间方面始终保持最优。与传统测试方法相比,随机森林算法在保证较低测试逃逸的情况下可以减少约28%的测试时间。与其他两种具有代表性的适应性测试方法相比,所提出的方法在测试时间方面表现更优。

An adaptive test method based on random forest was proposed to solve the problem of high test cost caused by increasing test time.For the chip of the training model,by calculating the Gini index,the importance of each test group to the model classification in the process of the chip parameter test was obtained,and the feature importance was used to quantify it.Then the importance of the test group was ranked to screen out the most important test groups for predicting the quality of the chip.At the same time,the number of defective chips that could be detected in each test group was counted.Part of the test was performed on the chips in the test set,the test time was reduced by deleting some test groups,and the random forest algorithm was used to predict the quality of the chips,so as to achieve a compromise between prediction accuracy and less time.Experimental results showed that compared with KNN and logistic regression algorithms,random forest could always maintains the best prediction accuracy,test escape level and running time.Compared with traditional test methods,the test time could be reduced by about 28% while maintaining a low test escape level.Compared with the other two representative adaptive testing methods,the proposed method performed better in reducing the test time.