The machine learning approach to fault diagnosis consists of a chain of activities such as data acquisition, feature extraction, feature selection and classification. Each one is equally important in fault diagnosis. As machine learning is a soft science, there is a lot of scope for finding mathematical reasoning which otherwise researchers do it arbitrarily or heuristically. Minimum number of samples required to separate faulty conditions, with statistical stability is one such important factor. This paper provides a method for determination of minimum sample size using power analysis. A typical bearing fault diagnosis problem is taken as a case for illustration and the results are compared with that of entropy-based algorithm (J48) for determining minimum sample size. The results will serve as a guideline for researchers working in fault diagnosis area to choose appropriate sample size. © 2010 Elsevier Ltd. All rights reserved.