Centrifugal Pump Fault Detection with Convolutional Neural Network Transfer Learning.

Area of Science:

• Engineering
• Machine Learning
• Data Science

Background:

• Centrifugal pumps are critical in industrial and domestic applications, but failures can cause significant financial losses and safety risks.
• Early fault diagnosis and predictive monitoring are essential for reliable pump operation.
• Existing machine learning fault detection methods often rely on synthetic or lab-based data, limiting real-world applicability.

Purpose of the Study:

• To develop and evaluate a machine learning-based system for detecting faults in real operational centrifugal pumps.
• To address the limitations of using synthetic or simulated data in previous fault detection research.
• To apply transfer learning from image detection to a practical engineering problem.

Main Methods:

• Utilized data collected from actual operational centrifugal pumps in diverse locations.
• Employed binary classification of visual features from DQ/Concordia patterns.
• Implemented a residual network (ResNet-34) with transfer learning from the image detection domain.

Main Results:

• Achieved a classification accuracy of up to 85.51% in detecting centrifugal pump faults.
• Demonstrated the effectiveness of using real operational data for fault diagnosis.
• Successfully applied transfer learning to solve a real-world engineering fault detection problem.

Conclusions:

• The proposed method effectively detects centrifugal pump faults using real operational data and residual networks.
• Transfer learning enhances the performance of fault detection systems in engineering applications.
• This research provides a viable approach for predictive maintenance of centrifugal pumps in industrial settings.