Wavelet coherence-aware multi-branch deep ensemble for fault identification in centrifugal pumps

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Centrifugal pump fault diagnosis is hindered by complex vibration signals.
• Early fault signatures are often weak and overlap with other defects.
• Nonstationary signal characteristics pose significant challenges.

Purpose of the Study:

• To develop a robust framework for reliable centrifugal pump fault diagnosis.
• To integrate time-frequency analysis with deep learning for enhanced feature extraction.
• To address challenges of weak signals and overlapping fault characteristics.

Main Methods:

• A wavelet coherence-aware multi-branch deep ensemble framework was proposed.
• Multi-channel vibration signals were converted into 2D wavelet coherence maps.
• Three diverse convolutional neural networks were trained in parallel and fused using soft-voting.

Main Results:

• The framework achieved consistent and reliable fault discrimination across different operating pressures.
• Strong class separability was confirmed via Receiver Operating Characteristic analysis.
• The proposed method demonstrated effectiveness in identifying mechanical defects.

Conclusions:

• The wavelet coherence-aware deep ensemble framework is effective for centrifugal pump fault diagnosis.
• The integration of time-frequency coupling and deep features enhances diagnostic performance.
• The study validates the framework's reliability under investigated experimental conditions.