Abstract

The paper reports and demonstrates a computationally efficient method for machine fault detection in industrial turbine systems. Empirical mode decomposition (EMD) and Savitzky–Golay smoothing filters are used for signal denoising, with a resulting noise index being developed. By comparing the noise index with a power index (also derived in the paper), obtained from the detection of transients using a spectral analysis of the rate-of-change of unit power, three operational conditions are identifiable viz. normal operation, transient operation and operation when subject to emerging machine faults. The accommodation of transient operational conditions of the unit, so as not to create excessive ‘false alerts’, provides a valuable alternative to more traditional techniques, based on PCA for instance, that can only provide reliable information during steady-state operation. The efficacy of the proposed approaches is demonstrated through the use of experimental trials on sub-15MW gas turbines.