Application of Support Vector Machine (SVM) and Proximal Support Vector Machine (PSVM) for fault classification of monoblock centrifugal pump. (4th December 2009)

• Record Type: Journal Article
• Authors: Sakthivel, N.R.; Sugumaran, V.; Nair, Binoy B.
• Is Part Of: International journal of data analysis techniques and strategies. Volume 2:Number 1(2010)
• Page Start: 38
• Page End: 61
• Publication Date: 2009-12-04
• Subjects: monoblock centrifugal pumps -- bearing faults -- seal faults -- impeller faults -- cavitation -- CAV -- support vector machines -- proximal SVM -- PSVM -- decision trees -- fault diagnosis -- fault classification -- vibration signals; Electronic data processing -- Periodicals; Database searching -- Periodicals; 005
• Languages: English
• ISSNs: 1755-8050

Abstract

Monoblock centrifugal pumps are widely used in a variety of applications. Defects and malfunctions (faults) of these pumps result in significant economic loss. Therefore, the pumps must be under constant monitoring. When a possible fault is detected, diagnosis is carried out to pinpoint it. In many applications, the role of monoblock centrifugal pumps is critical and condition monitoring is essential. Vibration-based condition monitoring and analysis using the machine-learning approach is gaining momentum. In particular, Artificial Neural Networks (ANNs), fuzzy logic and roughsets have been employed for condition monitoring and fault diagnosis. While it is difficult to train the neural network-based fault classifier, the classification accuracy in case of fuzzy logic- and roughest-based fault classifiers is not very high. This paper presents the use of Support Vector Machines (SVMs) and Proximal Support Vector Machines (PSVMs) for classifying faults using statistical features extracted from vibration signals under good and faulty conditions of a monoblock centrifugal pump. The Decision Tree (DT) algorithm is used to select prime features. These features are fed as inputs for training and testing SVMs and PSVMs and their fault classification accuracy is compared. The results are found to be better than neural network-, fuzzy- and roughest-based methods.