Machine Learning for Fault Diagnosis in Centrifugal Pump Systems

Centrifugal pumps underpin a vast array of industrial and municipal fluid‐handling processes. Faults such as impeller imbalance, bearing wear, seal leakage and cavitation can lead to efficiency loss, unplanned downtime and increased maintenance costs. Traditional diagnostic techniques rely on manual inspection of vibration or electrical signals, often demanding expert intervention and extensive pre-processing. Machine learning offers a data-driven alternative, combining automated feature extraction with robust classification to detect and diagnose faults at an early stage. Vibration or current measurements are transformed into time-frequency representations—through wavelet, Stockwell or Fourier methods—and then encoded as images or statistical feature vectors. Classical algorithms such as support vector machines and k-nearest neighbours have been complemented by deep learning architectures, notably convolutional neural networks, which learn discriminant patterns from two-dimensional scalograms or spectrograms without handcrafted features. Recent advances include transfer learning with pre-trained visual networks, supervised contrastive learning to enhance class separability, and hybrid pipelines that fuse global and local representations. Such approaches are increasingly integrated with smart sensors and Internet of Things platforms, enabling real-time monitoring, predictive maintenance and adaptive control. Despite considerable progress, challenges remain in coping with limited annotated data, imbalanced fault classes and noisy industrial environments. Ongoing research seeks to consolidate multi-modal data, refine self-supervised pre-training and validate algorithms on full-scale pump installations, with the aim of delivering reliable, scalable diagnostic systems that reduce downtime and energy consumption across global pumping infrastructure.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Machine Learning for Fault Diagnosis in Centrifugal Pump Systems publication trend

The graph below shows the total number of articles in machine learning for fault diagnosis in centrifugal pump systems across all publications each year (not limited to Nature Index journals).

Technical terms

Vibration signal: Time-series data capturing mechanical oscillations of pump components, used to infer fault conditions.

Scalogram: A two-dimensional representation of signal energy over time and frequency, typically produced by continuous wavelet or Stockwell transforms.

Spectrogram: A visual display of a signal’s frequency spectrum as it varies with time, generated using Short-Time Fourier Transform.

Convolutional Neural Network: A deep learning architecture designed for grid-structured data, employing convolutional filters to learn spatial or temporal patterns.

Transfer Learning: A technique that adapts a model pre-trained on a large dataset to a new, related task, reducing training time and data requirements.

Wavelet Coherence Analysis: A method for assessing the correlation between two signals as a function of both time and frequency, revealing synchronous behaviour under fault conditions.

References

1. Centrifugal Pump Fault Diagnosis Based on a Novel SobelEdge Scalogram and CNN. Sensors (2023).
2. An Intelligent Framework for Fault Diagnosis of Centrifugal Pump Leveraging Wavelet Coherence Analysis and Deep Learning. Sensors (2023).
3. Deep Learning for Enhanced Fault Diagnosis of Monoblock Centrifugal Pumps: Spectrogram-Based Analysis. Machines (2023).
4. A Centrifugal Pump Fault Diagnosis Framework Based on Supervised Contrastive Learning. Sensors (2022).
5. Monitoring and Predictive Maintenance of Centrifugal Pumps Based on Smart Sensors. Sensors (2022).