Machine Learning Framework for Anomaly Detection and Maintenance Optimization in Large-Scale Cryogenic Systems

Abstract This work presents a machine learning framework designed for anomaly detection and prescriptive maintenance tested on the helium compression system installed at CERN. The Large Hadron Collider (LHC) features an extensive cryogenic infrastructure, essential for its operation. Our framework leverages data acquired from 20 custom motor-compressor systems, with comprehensive tracking of operational parameters and vibration data since their commissioning. Exploiting deep autoencoders, the system detects anomalies by reconstructing operational signals and identifying deviations from nominal behaviour. Experimental results demonstrate the framework’s reliability in detecting and labeling anomalies. Moreover, we extended the capabilities of the system to provide an estimate of the Remaining Useful Life (RUL) of the compression system machinery. The result is a framework that effectively allows to improve the infrastructure’s availability and reduce maintenance costs. Designed for computational efficiency, the algorithm can be smoothly scaled down to be integrated into edge devices for on-site, real-time data processing and integration of federated learning. Such infrastructure can be developed leveraging CERN’s secure federated learning platform CAFEIN FL [1], allowing the developed framework to further enhance its robustness and accuracy through collaborative model improvement across multiple devices and facilities without compromising data privacy.