This paper primarily focuses on detecting electrical faults in turbine generator sets by monitoring torsional vibrations with the help of the non-contact measurement technique and analysing the data acquired from torsional vibration meter. Torsional vibrations in shaft trains can be excited by periodic excitation due to a variety of electromagnetic disturbances or unsteady flow in large steam turbine generator sets and may cause failure in shaft trains. A method is presented to use wavelet packet analysis for fault detection. It is achieved by decomposing the torsional vibration signals in the wavelet packet space at different levels to give finer details. Shannon wavelets have compact supports in frequency domain, it allows the analysis to be carried out in frequency bands of interest. Its effectiveness is verified by experimental results. Moreover, the application of the proposed method can be extended to analysis for transient conditions.