In the context of open source development or software evolution, developers often face test suites which have been developed with no apparent rationale and which may need to be augmented or refined to ensure sufficient dependability, or even reduced to meet tight deadlines. We refer to this process as the re-engineering of test suites. It is important to provide both methodological and tool support to help people understand the limitations of test suites and their possible redundancies, so as to be able to refine them in a cost effective manner. To address this problem in the case of black-box testing, we propose a methodology based on machine learning that has shown promising results on a case study.