This paper investigates the performance of the 12 entropy-based features for the monitoring and detection of bearing faults. These entropy measures were proposed both in time, frequency and time–frequency domain. Probability mass function (PMF) was extracted from the time waveforms using four different methods: (i) via power spectral density, (ii) via ordinal pattern distribution, (iii) via wavelet packet tree and iv) ensemble empirical mode decomposition. Three different entropy measures were used in the article: (i) Shannon entropy, (ii) Rényi entropy and (iii) Jensen–Rényi divergence. A new bearing produces a vibration time series characterised by random noise without prominent periodic content. As soon as a fault develops, impulses are produced, what excites structural resonances generating a train of impulse responses. As defect grows, it becomes a distributed fault, and then no sharp impulses are generated but rather an amplitude modulated random noise signal. The proposed methodology has been applied to detect bearing faults by the analysis of two real bearing datasets, from run-to-failure experiments. Three bearings that presented different defects in the test (inner race fault, rolling elements fault and outer race fault) were analysed to validate the performance of the entropy-based features. The modified Z-score has been implemented and used as an index to detect changes of the entropy features. The results clearly demonstrate that the proposed approach represents a valuable non-parametric tool for early detection of anomalies in bearings vibration signals.