Ekstraksi Parameter Statistik Domain Waktu dan Domain Frekuensi untuk Mendeteksi Kavitasi pada Pompa Sentrifugal Berbasis Principal Component Analysis (PCA)

Abstract

Cavitation phenomenon frequently occurs on the centrifugal pump which may decrease its performance. It may cause a catastrophic failure which leads to a total breakdown of the piping system if the presence of cavitation is not immediately detected and solved. Recently, the popular method used to detect cavitation is based on pattern recognition. The use of pattern recognition technique requires statistical features which are used as input for building the classifier. The extraction of statistical features is usually taken from the vibration signal which consists of time domain and frequency domain. Previous research tends to use the statistical features extracted from the time domain or the frequency domain solely. There is a research gap that can be explored by combining statistical features extracted from both time domain and frequency domain. In this study, Principal Component Analysis (PCA) is used as a feature’s selection and fault classification. PCA linearly transforms statistical features from the original coordinate system into a new coordinate system called principal components (PCs). The first few PCs are a set of selected features which can be used as a classifier. The classifier evaluates and classifies the new set of vibration data then decides whether it falls into normal condition or cavitation category. The vibration signal is taken from the cavitation test-rig under normal condition by opening the valve, level 1 cavitation by opening 75% of valve, level 2 by 50%, and level 3 by 25%. The data is extracted into 7 statistical features from the time domain and 5 from the frequency domain. Five hundred sets of vibration data are recorded using an accelerometer which was then divided into 400 set for training and 100 set for testing. The study shows that the classifier using statistical features taken from the time domain and frequency domain gives promising results where the clustering effect between normal and cavitation condition is clearly observed.