Vanilla LSTM Predictive Maintenance Model for Scientific Research Facilities

Area of Science:

• Scientific instrumentation
• Machine learning applications
• High-energy physics infrastructure

Background:

• Critical scientific equipment failures cause significant financial losses and disrupt research.
• Maintaining operational efficiency in nuclear physics and particle accelerator facilities is paramount.

Purpose of the Study:

• To develop and evaluate a predictive maintenance (PdM) framework for critical scientific equipment.
• To improve the reliability and operational efficiency of experimental facilities through early failure detection.

Main Methods:

• A two-layer Vanilla Long Short-Term Memory (LSTM) network was trained on multivariate sensor data (supply voltage, vibration velocity, differential pressure, rotational speed).
• Data was collected from January 2021 to December 2023 at NRF-iThemba LABS using industrial-grade transducers and a multi-channel data-acquisition system.
• A normalized failure score was developed to provide an interpretable health indicator for maintenance scheduling.

Main Results:

• The Vanilla LSTM model achieved a test-set F1-score of 75% and an AUC of 0.856.
• The model outperformed five competing machine learning architectures in predictive accuracy.
• The framework delivered a mean failure lead time of (42.3±7.2) hours, exceeding the 36-hour requirement for proactive maintenance.

Conclusions:

• The developed PdM framework effectively predicts equipment failures in critical research environments.
• LSTM-based predictive maintenance offers a reliable and efficient solution for operational continuity in scientific facilities.
• This approach supports timely interventions, reducing downtime and associated costs.