Anomaly Detection in Cyber-Physical Systems: A Case Study on Pump Health Monitoring

In industry, pumps play a pivotal role, and their reliability is paramount for sustaining operational efficiency and safety. Despite their significance, pumps are susceptible to various faults, with cavitation emerging as a prevalent issue caused by factors such as misalignment, imbalance, and wear. Traditional anomaly detection methods relying on displacement measurements of the pump shaft are limited in their effectiveness due to external influences like varying loads and motor speeds. This study introduces a more robust approach to anomaly detection in pumps by utilizing vibration acceleration sensors. Vibration acceleration, measuring the rate of change of vibration displacement, proves to be a more sensitive indicator of pump health, less susceptible to external factors. Additionally, the importance of speed normalization in conjunction with vibration acceleration is highlighted for cross-comparison across different pumps and operational conditions. In the exploration of related works, this paper delves into secondary sources contributing to vibration and noise in pump systems. Beyond cavitation, factors such as radial misalignments, impeller pressure variations, and hydraulic sources of noise are considered. The subsequent methodology section outlines the utilization of unsupervised learning models such as Autoencoders, Variational Autoencoders, and Long Short-Term Memory networks for the early prediction of anomalous pump behavior.