Experimental fault diagnosis for known and unseen operating conditions of centrifugal pumps using MSVM and WPT based analyses

Abstract

Centrifugal pumps (CPs) are extensively used in various industries and their consistent operation ensures the keep-up of the plant’s work-flow. CPs most commonly fail due to mechanical component malfunction/breakdown and/or abnormalities in the fluid flow patterns. But these mechanical/hydraulic CP faults may not always be devoid of one another. Also, their coexistence may accelerate the CP failure. Hence, in the present study, both individually existing and coexisting hydraulic and mechanical faults with varying severities are simulated. On the whole, thirty three distinct faults of the CP system are considered. Motor current and CP vibration signatures are utilized to monitor the faults. In addition, the wavelet packet transform (WPT) analysis is used to pre-process the time-domain signal data in order to take into consideration the transients and spectral frequencies produced because of the complex fluid-structure interactions of mechanical and hydraulic faults. The faults classification is aimed at serving two purposes, (i) segregating different fault families and (ii) identifying the severity of a particular fault. Multiclass support vector machine algorithm (MSVM) is employed in classification of the fault data. Two methodologies based on best energy (BE) criterion and principal component analysis (PCA), that is, respectively, the WPT-BE-MSVM and WPT-PCA-MSVM are proposed and their performances are compared. The feature combinations giving the highest performance for each of these methodologies are identified. The algorithms are tested for wide operating conditions of the CP. An intermediate speed fault classification case is also studied to accommodate for the absence of data (unseen data case) at a particular operating speed of the CP. The classification accuracies obtained from both the methodologies are promising.