Multistage Centrifugal Pump Fault Diagnosis Using Informative Ratio Principal Component Analysis

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Vibration signatures (VS) of multistage centrifugal pumps (MCP) often contain interference and background noise, complicating fault diagnosis.
• Traditional feature extraction methods may yield high-dimensional feature vectors with suboptimal discriminative power for MCP condition monitoring.

Purpose of the Study:

• To propose a novel fault diagnosis (FD) method for MCP that effectively handles noisy vibration signatures.
• To develop a feature extraction technique that identifies and utilizes discriminant features for accurate condition classification.

Main Methods:

• The proposed method involves three steps: 1) selecting a fault-specific frequency band (FSFB) from VS, 2) extracting statistical features in time, frequency, and wavelet domains to form a multi-domain feature pool (MDFP), and 3) introducing informative ratio principal component analysis (Ir-PCA).
• Ir-PCA assesses feature informativeness using the ratio of within-class scatteredness to between-class distance and applies PCA to high-informative features for dimensionality reduction.
• The reduced, discriminant features are then fed into a K-nearest neighbor (K-NN) classifier for MCP condition classification.

Main Results:

• The Ir-PCA method successfully reduces feature dimensions while retaining critical information for fault diagnosis.
• The proposed fault diagnosis method demonstrated superior performance compared to existing state-of-the-art techniques.
• Accurate classification of multistage centrifugal pump conditions was achieved using the novel Ir-PCA feature set.

Conclusions:

• The developed fault diagnosis method, incorporating Ir-PCA, offers a robust and accurate approach for monitoring the condition of multistage centrifugal pumps.
• This technique effectively overcomes noise interference in vibration signatures, leading to improved fault classification accuracy.
• The study highlights the potential of Ir-PCA as a powerful tool for feature extraction in rotating machinery diagnostics.