Prognostics CoE Publications

Many diagnostic datasets suffer from the adverse
effects of spikes that are embedded in data and noise. For
example, this is true for electrical power system data where
the switches, relays, and inverters are major contributors to
these effects. Spikes are mostly harmful to the analysis of
data in that they throw off real-time detection of abnormal
conditions, and classification of faults. Since noise and
spikes are mixed together and embedded within the data,
removal of the unwanted signals from the data is not always
easy and may result in losing the integrity of the
information carried by the data. Additionally, in some
applications noise and spikes need to be filtered
independently. The proposed algorithm is a multi-resolution
filtering approach based on Haar wavelets that is capable of
removing spikes while incurring insignificant damage to
other data. In particular, noise in the data, which is a useful
indicator that a sensor is healthy and not stuck, can be
preserved using our approach. Presented here is the
theoretical background with some examples from a realistic
testbed.