基于随机森林的X值输入灵敏度预测方法

Predicting X-Sensitivity of Circuit-Inputs Based on Random Forests

随着基于模块化的电路设计变得越来越复杂,未初始化的时序单元、设计中的黑盒、时钟域交叉以及模数转换器的错误行为等原因会导致电路中出现未知的逻辑值(X),降低电路测试集的测试覆盖率.为了快速确定电路中X值输入对测试覆盖率的影响,提出了一种基于机器学习的方法来预测X值输入的灵敏度.首先通过拓扑算法计算电路的各项基础结构参数;然后对电路进行区域划分,提取特定的电路特征参数作为原始数据集;最后利用随机森林模型对所有电路中得到的数据集进行训练和预测.实验选择ISCAS’89和ITC’99中的部分电路作为数据集来源,与现有的预测方法相比,该方法总体预测准确率达到90.27%,提高了14.69%,大型电路预测准确率达到93.32%,提高了19.49%.实验结果表明,该方法具有更高的准确率和更好的泛化能力.

As modular circuit design becomes more and more complex,uninitialized timing cells,black boxes in the design,clock-domain interface and erroneous behavior of analog-to-digital converters may produce unknown logic value(X)at various circuit nodes.The existence of X-value decreases the test coverage of the test set.For the X-sources that X-value exists on the certain primary or secondary inputs of a logic circuit,this paper presented a method based on machine learning to predict the X-sensitivity which measures the effect of X-source on test coverage.Firstly,the basic structure parameters of the circuit are calculated by topology algorithm.Then the circuit is divided into three parts,and the specific circuit characteristic parameters are extracted as the original data set.Finally,the random forest model is used to train and predict the datasets which obtained in all circuits.Some of the circuits in ISCAS’89 and ITC’99 are selected as the data set source.Compared with the existing prediction methods,the test set accuracy of proposed method is 90.27%with 14.69%higher,and the large circuits accuracy of proposed method is 93.32%with 19.49%higher.Experimental results show the higher accuracy and better generalization ability of proposed method.