Introductory Use of Probabilistic Neural Networks for Spike Detection from an On-Line Vibration Diagnostic System

Abstract

We developed a two-part algorithm for detecting spikes in trended data from a vibra tion diagnostic monitoring system. Deviation from an exponential forecasting technique provides an initial screening of data for abnormal amplitudes. These abnormal events are classified as NOR MAL (i.e., no spikes), positive-going spikes (PSPIKE), or negative-going spikes (NSPIKE) by a PNN calculation trained on synthetic data using a Gaussian probability density function (PDF) esti mation. This screening reduces computational time and results in fewer false spike classifications compared with a probabilistic neural network (PNN) calculation alone. Results from sample field data sets showed that spike detection can be readily achieved, but that steps in the data lead to false spike classifications.

References

Branagan, L.A. 1990. "Experience with an On-Line Vibration Diagnostic Momtoring System Applied to a Turbine Generator Unit", Electric Power Research Institute 4th Incipient Failure Detection Conference, Philadelphia

Branagan, L.A. and P.D. Wasserman. 1992. "Introductory Use of Probabilistic Neural Networks for Spike Detection from an On-Line Vibration Diagnostic System ; Intelligent Engineering Systems through Artificial Neural Networks, C. H. Dagli, L. I. Burke and Y. C. Shin, eds., New York: ASME Press, 2:719-724.

1992. MATHCAD 3.1: UserS Guide, Windows Version Cambridge MA: MathSoft, Inc.

Montgomery, D.C. , L.A. Johnson and J.S. Gardmer. 1990. Forecasting & Time Series Analysis San Francisco: McGraw-Hill Inc.

Specht, D.F. 1990. "Probabilistic Neural Networks", Neural Networks, 3(1)(12):109-118.

Specht, D.F. 1992. "Enhancements to Probabilistic Neural Networks" , Proceedings of the IEEE International Joint Conference on Neural Networks, IEEE.

Wasserman, P.D. 1989. Neural Computing New York: Van Nostrand Reinhold.