Protection of Superconducting Industrial Machinery Using RNN-Based Anomaly Detection for Implementation in Smart...

Area of Science:

• Superconducting technology
• Machine learning applications
• High-energy physics instrumentation

Background:

• Safe operation of superconducting installations, like the Large Hadron Collider (LHC), relies on detecting voltage anomalies (quenches).
• Current quench detection methods at the LHC use custom rules, which are not scalable and require manual tuning.

Purpose of the Study:

• To explore the use of embedded Recurrent Neural Networks (RNNs) in superconducting protection devices.
• To develop a scalable and automatically configurable solution for anomaly detection in superconducting magnets.

Main Methods:

• Implementation of a two-layer Gated Recurrent Unit (GRU) model with 64 and 32 cells.
• Utilizing a custom data compression scheme for the GRU model.
• Evaluating the model's performance for anomaly/non-anomaly classification and memory footprint.

Main Results:

• The GRU model achieved an accuracy of 0.93 for anomaly classification.
• Data compression reduced the memory footprint by four times with minimal performance loss.
• The proposed solution is suitable for hardware implementation in protection devices.

Conclusions:

• Embedded RNNs, specifically GRUs, offer a scalable and efficient alternative to rule-based systems for superconducting quench detection.
• The developed model demonstrates high accuracy and reduced memory requirements, making it practical for real-world superconducting systems.