Efficient sampling techniques for ensemble learning and diagnosing bearing defects under class imbalanced condition

Document Type

Conference Proceeding

Publication Date


Publication Title

2016 IEEE Symposium Series on Computational Intelligence, SSCI 2016


This paper focuses on sampling techniques to rebalance class distribution in a way that major and minor classes reach to almost equal number of the samples. A novel iterative over-sampling technique has been proposed which initially induces the missing values on the set of samples of the minor class and, then, imputes the missing scores to generate new synthetic samples of the minor class, in order to re-balance the class distribution. Two variations of the proposed oversampling framework have been developed which make use of the Expectation Maximization and k-Nearest Neighbors imputation strategies. Moreover, the proposed over-sampling technique, which generates new samples for the minor class, has been integrated with a random under-sampling technique, which aims to simultaneously reduce the number of samples for the major class to speed up the process. The proposed sampling procedures have been used along with the ensemble of classifiers forming a diagnostic system. The constructed diagnostic scheme can efficiently diagnose multiple bearing defects in induction motors under class imbalanced condition.