Signal Processing and Speech Communication Laboratory
home › phd theses › Automated Anomaly Classification for the Post-Mortem System using Machine Learning

Automated Anomaly Classification for the Post-Mortem System using Machine Learning

Status
In work
Student
Philipp Hermüller
Mentors
Research Areas

As the size and complexity of future accelerators increases, the automated analysis and validation of machine protection functionalities will become more and more critical. The development of a fully automated analysis tool to classify machine-protection-relevant data in the LHC will serve as proof-of-concept for future high energy colliders. It will allow to identify important design requirements which are relevant for the early design phase of such a collider.

The project is structured around three main research axes.

  • The first is to classify the losses during LHC beam dumps using advanced machine-learning techniques. While the existing Beam Loss Analysis Module efficiently classifies losses as OK or NOT-OK based on individual thresholds for nearly 4000 Beam Loss Monitors (BLM), the type of anomaly must still be manually assessed by the expert. To address this gap, the successful candidate will investigate and implement a method to classify the type of beam loss anomaly. This will include:
    1. using a supervised neural network architecture to classify known anomalies,
    2. exploring the possibility of an unsupervised model that could detect unknown anomalies, and
    3. incorporating additional data into the training set, e.g. the full time series of the beam losses.

  • The second axis focuses on the automated classification of beam orbit movements before the beam dump. This involves the systematic acquisition, analysis and pre-processing of historical data from the Beam Position Monitors (BPM). Classification methods will then be developed and compared using both threshold-based and machine-learning approaches.

  • The goal of the third axis is to build an integral model that combines data from separate systems, such as BLMs, BPMs or the interlock systems, to classify the type of anomaly. This could be used to find similar beam dump events in the past based on similarity encodings and AI-based lookup tables.

The project is carried out in collaboration with the Signal Processing and Speech Communication Laboratory (SPSC Lab) at Graz University of Technology (TU Graz) and CERN.