Looking for a Correct Solution of Anomaly Detection in the LHC Machine Protection System

Demand for high availability and performance of the LHC accelerator imposes stringent constraints on safety-critical systems handled by Machine Protection and Electrical Integrity Group at CERN. Consequently, unique means and equipment were introduced to monitor and protect superconducting magnets of the LHC. This work intends to ameliorate available algorithms by using RNN (Recurrent Neural Network) architectures. The proposed solution is equipped with a series of architecture modifications and optimizations which enable efficient deployment of the RNN model on embedded systems. A set of experiments were conducted which shows that the proposed system outperforms Isolation Forest and OC-SVM (One-Class SVM) algorithms reaching the accuracy of 0.93 for the two-layer LSTM (Long Short-Term Memory) model. Furthermore, the proposed approach allows for significant reduction of input data stream to few bits which enable low latency response of the module.