Pump Fault Detection Using Autoencoding Neural Network

Abstract

This paper deals with the fault detection of centrifugal pumps, based on measured radial vibrations. The detection method compares the vibration signature of the equipment during normal behavior with the current recorded vibration signal. It raises an alarm if a distance function of the resulted residuum exceeds a predefined threshold. The normal signature and the threshold are learned through a machine learning procedure, based on autoencoding neural networks (NN). Two versions of NNs are trained and evaluated. The detection method proved to be reliable in an industrial application, even when using a single low-cost accelerometer for vibration sensing.