带有故障性质预测的自动测试向量求解模型  

Automatic Test Pattern Solving with Fault Property Prediction

作　　者：贺丽媛 黄俊华 陶继平 HE Li-yuan;HUANG Jun-hua;TAO Ji-ping(Department of Automation,Xiamen University,Xiamen,Fujian 361102,China;Huawei Noah’s Ark Laboratory,Shenzhen,Guangdong 518000,China)

机构地区：[1]厦门大学自动化系,福建厦门361102 [2]华为诺亚方舟实验室,广东深圳518000

出　　处：《电子学报》2023年第12期3540-3548,共9页Acta Electronica Sinica

基　　金：福建省自然科学基金(No.2020J01053)。

摘　　要：基于布尔满足模型的自动测试向量生成是芯片故障检测的关键环节,相应布尔问题的求解已然成为整个故障检测过程的效率瓶颈.本文研究了主流自动测试向量求解框架中不同算子对求解效率的影响,在保证测试向量求解流程完备性的同时引入基于深度学习的故障分析机制,并将分析结果用于算子的自动选择和初始求解状态的确定,旨在优化整体求解进程.针对因真实电路故障数据不足导致模型学习效果欠佳的问题,本文利用生成对抗网络实现数据增广,结合多层图卷积神经网络促进高效表征学习,从而提高故障性质的预测精度.在若干真实电路上的实验结果表明,本文所提出的新框架与原有框架相比,平均求解效率提升近20%.Automatic test pattern generation(ATPG)based on the Boolean satisfaction model plays a key part in chip fault detection flow,in which solving the corresponding Boolean satisfiability problem(SAT)becomes the efficacy bottleneck of the whole process.In this paper,the influence of different operators on the solution efficiency in the mainstream automatic test pattern solution framework is studied.While ensuring the integrity of the test pattern solution process,a fault analysis mechanism based on deep learning is introduced,and the output vectors are used for the automatic selection of operators and the determination of initial solution states so as to accelerate the overall solution process.To alleviate poor performance mainly caused by insufficient real-world circuit fault data,a generative adversarial network(GAN),followed by a multi-layer graph convolutional neural network(GCN)which is designed to boost representation learning,is leveraged for data augmentation.Experimental results on several real circuits show that the proposed new framework,compared with the original version,has an average solution improvement of nearly 20%.

关 键 词：自动测试向量生成 图神经网络 生成对抗网络 数据增广 算子选择 

分 类 号：TN407[电子电信—微电子学与固体电子学] TP183[自动化与计算机技术—控制理论与控制工程]