A Framework for Centrifugal Pump Diagnosis Using Health Sensitivity Ratio Based Feature Selection and KNN.

A new framework for the fault diagnosis of centrifugal pumps (CP) is presented in this paper. Time domain (TD) features obtained from the vibration signal (VS) of the CP are vulnerable to severe faults and can affect the fault classification accuracy of the classifier. To address this issue, the proposed method selects a healthy reference signal (HRS) and extracts raw statistical features from this signal and the vibration signals of the CP obtained under different operating conditions in the time and frequency domain (FD). The Pearson correlation coefficient is calculated by cross-correlating the time and frequency domain features of the healthy reference signal with the time and frequency domain features extracted from the vibration signal of the CP under different operating conditions. The Pearson correlation coefficient results in a new feature vector, however, some of the coefficients may not be the best to identify the ongoing conditions of the centrifugal pump. To overcome this problem, the proposed method uses a new health sensitivity ratio for the selection of CP health-sensitive features. The health sensitivity ratio (HSR) assesses per-class feature compactness and between-class distance of the features. The selected health-sensitive features are provided to KNN for the identification of centrifugal pump health conditions. The proposed method has achieved a classification accuracy of 97.13%, surpassing that of the conventional methods for CP fault diagnosis.