Centrifugal Pump Fault Diagnosis Using Discriminative Factor-Based Features Selection and K-Nearest Neighbors

Abstract

This paper proposes a new fault diagnosis framework for Centrifugal Pump (CP) fault diagnosis. To utilize the fault-related transients, the proposed fault diagnosis framework first preprocesses the vibration signal (VS) using wavelet packet transform (WPT). Instead of extracting features from a specific wavelet packet transform base (node), the proposed method utilizes all the bases of wavelet packet transform and extract features from all the bases. As the time domain features are suitable for representing weak faults, the proposed method also extracts features from vibration signals in the time domain (TD). All these features are merged into a combined hybrid feature pool (HFP). The combined hybrid feature pool results in a high dimensional space, moreover, some of the features might not be helpful for the classification of centrifugal pump working conditions. To select discriminant features, the proposed method uses a discriminative-factor-based feature selection method. The discriminative factor for a feature indicates within the class feature scatteredness and between classes feature distance. After selecting discriminant features, the selected features are then classified by the K-nearest neighbor (KNN) algorithm. The classification results obtained from the K-nearest neighbor (KNN) algorithm for our proposed method outperform already existing state-of-the-art methods.