Multi Fault Diagnosis of Centrifugal Pumps with Time, Frequency and Wavelet-Based Features Using Support Vector Machines

Abstract

The centrifugal pumps (CPs) are most commonly chosen fluid machines for industrial and domestic applications. They constitute vital components to sustain the process flow of the plants, and hence their failure may lead to a significant monetary loss to the plant. Failures in CPs may be due to mechanical faults and/or fluid flow anomalies. In the current work, it is attempted to build an adaptable support vector machine (SVM) based algorithm to identify critical faults, like flow restrictions (with changing severities), impeller cracks, dry run and cover plate faults. Furthermore, co-existence of mechanical and fluid flow faults is studied. Experimentally generated CP vibration data and motor current data is utilized for the fault diagnosis. The classifier parameters are chosen optimally by means of cross-validation method along with grid-search technique, and effective fault feature selection is done using the wrapper model. Furthermore, a comparative analysis on the performance of the methodology for features extracted from three domains, namely: time, frequency and wavelets (wavelet packet transform, time-frequency domain) of the raw signals is presented at a range of CP operating speeds. The analysis results presented that the developed methodology could identify fifteen CP fault conditions successfully based on features from all three domains at all CP speeds.