Fault Localization Method Based On Model Combination

Fault localization is one of the most expensive activity and time-consuming process of debugging. To address this expense, researchers have presented lots of automated fault localization techniques to provide automated assistance in finding the faults that cause executions to produce incorrect outputs.These methods can be utilized to improve the efficiency of Fault localization, however, it is still far from mature. Firstly, single Fault localization methods usually rely on specific state information of a program, hence only corresponding types of errors can be handled. And it is a difficult task to decide which approach to use in practice. Secondly, most of the fault localization available is based on single type of coverage information, such as statements and branches. Thus, the precision of Fault localization is limited due to the incomplete utilization of available information.In this paper, we detail the framework of the fault localization that combines multiple fault localization techniques based of types of coverage, inspired by the idea of ensemble learning. The study focuses on the combinations of different types of spectrum-based fault localization. The proposed framework includes two parts, the collection of program coverage and the algorithm of combining multiple fault localization techniques.Based on LLVMcompiler, a new automatic instrument tool is developed to collect the coverage information. We analyze the LLVM compiler framework and its Pass system, and implement the methods of collect the coverage of statements and predicate based on the character of the program. Furthermore, we implement the instrument tool in order to collect the coverage information of statements and predicate.In addition, we propose a method of combining two fault localization model based on the statement and predicate coverage information. Because of the difference of the models, we select Tarantula method and SOBER method as the candidate. We implement three methods which are Simple-mix Rank, Max-mix Rank and Avg-mix Rank. The experimental results of Siemens assembly show that, the method of combining the models of fault localization based on different types of coverage, Simple-mix Rank and Max-mix Rank, compared to the method based on single type of coverage, have a higher efficiency and a stronger generalization ability.