In this work, we develop a fault detection methodology based on the use of selected narrow band time series signals. It is based on the Kullback Leibler Divergence (KLD) and selected components obtained from the Empirical Mode Decomposition (EMD) applied to non stationary time series vibration signals. The EMD decomposes the signal into narrow frequency bands components called Intrinsic Mode Functions (IMFs). A first selection of the most energised and consequently the most sensitive IMFs to fault occurrence is proposed, thanks to the computation of the Signal to Noise Ratio of the IMFs. Thus, the retained components are analysed using the Kullback Leibler Divergence to proceed the fault detection. A quantitative sensitivity criteria is derived to evaluate the fault detection performances and confirmed by a probabilistic analysis. The proposed methodology is validated using an experimental dataset from the Case Western Reserve University with three different fault severities and operating load conditions. With this proposed methodology a 100% probability of detection is obtained with each of the first six selected IMFs, the best results being achieved with IMF