This paper introduces a new signal processing algorithm for vibration-based fault detection and diagnosis of roller bearings. The methodology proposed in this paper is based on the combination of two data-adaptive techniques which are further enhanced through the use of an automatic feature identification mechanism. The new technique, introduced as empirical mode envelope with minimum entropy, combines elements from the empirical mode decomposition (EMD) and minimum entropy deconvolution (MED) approaches with an energy moment technique to improve the feature selection stage of the EMD algorithm. This improvement allows the processing chain to identify early stage roller bearing faults in noisier signals. The energy moment technique is used to automatically identify the most appropriate intrinsic mode function from the EMD process prior to the MED algorithm being applied. This is in contrast to conventional approaches which tend to use the first mode or make selections based on traditional energy techniques. The combination of the adaptive techniques of EMD and MED allows the development of an improved technique for fault detection and diagnosis of signals. Combining these techniques with the energy moment approach allows further improved fault detection in complex non-stationary conditions. The processing chain has been tested using data obtained during laboratory testing. From the experimental results, it is shown that the new technique is capable of the detection of early stage (minor) roller and outer race defects found in tapered-roller-bearings rotating at a variety of speeds and noise scenarios.