Cavitation Detection of Centrifugal Pumps Using SVM and Statistical Features

Abstract

AbstractCentrifugal pump is one type of pump that is widely used in the industry. Its mechanism which creates pressure change may cause cavitation. Generally, the cavitation increases noise and vibration level, if not properly maintained leads to catastrophic failure and total stop of the whole process. Therefore, it is needed to develop a method that can detect cavitation as early as possible. Recently, pattern recognition-based fault detection is gaining attention from many researchers due to its superior detection and classification performance. Support vector machine (SVM) is one of the pattern recognition techniques which requires statistical features as input for building the model. However, the selection of statistical features is arbitrary. In this study, ten statistical features are extracted from the time-domain vibration signal and selected using Relief Feature Selection. The selected features are used as input for two types of SVM, binary, and multi-class, to classify new vibration data. Each multi-class classification result is optimized by the Bayesian Optimization algorithm and Grid Search Method. The result shows that root mean square, standard deviation, variance, entropy, and standard error are several features that indicate the best plot. The feature selection process reveals that variance, root mean square, and standard error are the best feature to use for SVM classification. The binary SVM method shows the best plot on early cavitation with an accuracy of 99%. The Bayesian optimization algorithm with multi-class SVM is the best combination to classify all pump conditions with an overall accuracy of 100%.KeywordsCentrifugal pumpSupport vector machineRelief feature selectionBayesian optimizationGrid search method