EBSCN: An Error Backtracking Method for Soft Errors Based on Clustering and a Neural Network

With the development of integrated circuit design technology, soft errors have become an important threat to system reliability, and software-based fault-tolerant techniques are gradually attracting people's attention. In many cases, researchers use fault injection techniques that are less observable and less controllable to verify system reliability, and at this point, analysing where soft errors occur requires considerable work. In this paper, we present EBSCN, an error backtracking method. The EBSCN method sorts functions by suspiciousness by analysing the erroneous output results, which will help researchers reduce the amount of work required for analysis. The EBSCN method includes a feature extraction method based on clustering and a feature analysis method based on a deep neural network. This paper introduces the principle of the two methods as well as methods to improve and extend them with program-related information. We discuss the effect of the scale of the output result and the severity of the error on the EBSCN method through experiments and verify the effect of the EBSCN method. The results showed that the proportion of the function in which the soft error actually occurs in the ranking of the top 25% of the suspiciousness sequence is no less than 82%, and the proportion ranked in the top 50% is no less than 97%.