A Centrifugal Pump Fault Diagnosis Framework Based on Supervised Contrastive Learning.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Centrifugal pumps (CPs) are critical in industrial operations.
• Non-stationary vibration signals from CPs complicate traditional fault diagnosis.
• Accurate fault diagnosis is essential for preventing equipment failure and operational downtime.

Purpose of the Study:

• To propose a novel intelligent fault diagnosis method for centrifugal pumps.
• To address the limitations of traditional time and frequency domain analyses for non-stationary vibration data.
• To develop a robust and highly accurate fault identification system for CPs.

Main Methods:

• Utilized kurtogram images to visualize fault-related impulses in vibration data.
• Employed a deep learning convolutional encoder (CE) with supervised contrastive loss for feature extraction.
• Pretrained the CE on kurtograms to infer contrasting features, followed by training a linear classifier on frozen representations.

Main Results:

• Achieved a high classification accuracy of 99.1% on real industrial testbed data.
• Demonstrated a low error rate of less than 1% in fault identification.
• Validated the model's robustness on CP data with varying inlet pressures (3.0 and 3.5 bar).

Conclusions:

• The proposed intelligent method effectively diagnoses centrifugal pump faults using vibration data and deep learning.
• The approach overcomes challenges posed by non-stationary signals, offering superior performance.
• This method provides a reliable and accurate solution for centrifugal pump condition monitoring.