Nowadays time series are generated relatively more easily and in larger quantity than ever, by the advances of IoT and sensor applications. Training a prediction model effectively using such big data streams poses certain challenges in machine learning. Data sampling has been an important technique in handling over-sized data in pre-processing which converts the huge data streams into a manageable and representative subset before loading them into a model induction process. In this paper a novel data conversion method, namely Kennard-Stone Balance (KSB) Algorithm is proposed. In the past decades, KS has been used by researchers for partitioning a bounded dataset into appropriate portions of training and testing data in cross-validation. In this new proposal, we extend KS into balancing the sub-sampled data in consideration of the class distribution by round-robin. It is also the first time KS is applied on time-series for the purpose of extracting a meaningful representation of big data streams, for improving the performance of a machine learning model. Preliminary simulation results show the advantages of KBS. Analysis, discussion and future works are reported in this short paper. It is anticipated that KBS brings a new alternative of data sampling to data stream mining with lots of potentials.