基于特征融合的语句级别软件故障定位

Statement-level Software Fault Localization Based on Feature Fusion

摘要:

针对软件故障定位任务中提取的特征不全面、将不同特征对故障的贡献无差化的问题,提出了一种基于特征融合的语句级别软件故障定位方法；对每条语句进行语义上的扩展,采用Doc2Vec技术提取扩展后语句的语义信息；选用六种基于频谱的故障定位公式来获取频谱信息,选用两种基于变异的故障定位公式来获取变异信息；采用注意力机制对三种来自不同信息源的特征进行融合处理,自动学习对故障最有效的特征；在Defects4J数据集的五个真实项目进行了实验,采用基于注意力机制的多特征融合在Top-K(K=1,3,5)上能够多定位11~16个故障,在MRR上提高了5.17%,实验结果表明,所提方法能够有效提高模型的定位性能。

Abstract:

Aiming at the problem that the features extracted in the software fault localization task are not comprehensive and the contribution of different features to the fault is indifferent, a statement-level software fault localization method based on feature fusion is proposed. Each sentence was semantically extended, and the Doc2Vec technology was used to extract the semantic information of the extended sentence. Six spectrum-based fault location formulas were used to obtain spectrum information, and two mutation-based fault location formulas were used to obtain mutation information. The attention mechanism was used to fuse the features from three different information sources, and the most effective features for faults were automatically learned. Experiments were carried out on five real projects of Defects4J dataset. The multi-feature fusion based on attention mechanism can locate 11 to 16 faults on Top-K(K=1,3,5), and improve the MRR by 5.17%. The experimental results show that the proposed method can effectively improve the localization performance of the model.