Multi-sensor observer-based residual learning with Auto-Permutation Feature Importance for fault diagnosis of...

Area of Science:

• Mechanical Engineering
• Machine Learning
• Signal Processing

Background:

• Deep learning models for fault diagnosis require extensive labeled data, which is difficult to obtain for industrial machinery like centrifugal pumps.
• Existing methods struggle with data scarcity, imbalance, and high-dimensional features, hindering interpretability and efficiency.
• Collecting fault data is challenging due to the risk of damaging equipment.

Purpose of the Study:

• To develop a sensor-fused, data-efficient framework for centrifugal pump fault diagnosis under varying operating pressures.
• To address limitations of deep learning models in terms of data requirements and feature extraction.
• To provide a reliable, interpretable, and efficient fault diagnosis solution.

Main Methods:

• Utilized an autoregressive (AR) observer for normal-class signal modeling and fault residual extraction across multiple sensors.
• Computed statistical and spectral descriptors (e.g., RMS, band power) from residuals.
• Employed Auto-Permutation Feature Importance (Auto-PFI) for dimensionality reduction and feature selection.
• Applied Gaussian Mixture Model (GMM) for class-wise density estimation and fault classification.

Main Results:

• Achieved classification accuracies exceeding 99% across multiple pressure levels (3, 3.5, and 4 bar).
• Demonstrated superior performance compared to single-sensor setups and existing state-of-the-art methods.
• Validated robustness and generalization using t-SNE, ROC curves, and confusion matrices.

Conclusions:

• The proposed framework integrating AR-based residual modeling, Auto-PFI, and GMM offers a reliable and interpretable fault diagnosis approach.
• The method is effective even with limited or imbalanced data, making it suitable for real-world industrial applications.
• This sensor-fused approach enhances diagnostic accuracy and efficiency for centrifugal pumps.