| Thesis | |
| Report number | CERN-THESIS-2023-321 |
| Title | Interpretable Fault Prediction for CERN Energy Frontier Colliders |
| Author(s) | Obermair, Christoph (Graz, Tech. U.) |
| Publication | 151. |
| Thesis note | PhD : Graz, Tech. U. : 2023 |
| Thesis supervisor(s) | Wollmann, Daniel ; Pernkopf, Franz |
| Note | Presented 19 Jan 2024 |
| Subject category | Engineering ; Accelerators and Storage Rings |
| Accelerator/Facility, Experiment | CERN LHC |
| Abstract | The Large Hadron Collider (LHC) is the world’s highest energy particle collider, which has already delivered data for numerous physical discoveries. To continue this quest for discovering new physics, the Compact Linear Collider (CLIC) and the Future Circular Collider (FCC) aim to push the boundaries of fundamental physics at high collision energies. However, as their power, size, and complexity increases, so does the risk of failures and their associated downtime. Fault prediction is a way to minimize downtime by fixing faults in scheduled maintenance intervals before they occur. In the LHC, such fault prediction methods have been supporting system experts to decrease downtime since its start in 2008/9. There are many different scenarios of faults. Each of them occurs rarely, which is why the predictions cannot be validated by statistical tests alone. Nonetheless, the methods work reliably as their predictions are based on known fault indicators which are validated by experts. To use Machine Learning (ML) methods for fault prediction, the same approach is required: Predictions must be interpreted and validated by experts. Demonstrating the predictive capabilities of ML, this thesis presents three approaches for interpretable fault prediction. Firstly, a novel autoencoder-based method for explaining fault predictions to system experts is proposed. A survey of 73 potential users confirms its effectiveness when compared to two other popular methods. This explanation method is then used to interpret ML-based breakdown predictions in radio frequency cavities. The interpretation reveals that a pattern in the emitted electrons following an initial breakdown is closely related to the probability of another breakdown occurring shortly thereafter. This explanation is consistent with the findings of recent research. Secondly, non-negative matrix factorization, a ML method that is designed to be interpretable, is used to detect normal and abnormal behavior in the LHC main dipole magnets. Five dipole magnets with abnormal behavior are identified, of which one was confirmed to be damaged. Thirdly, a hybrid method is proposed, that allows experts to rely on their existing tools and still benefit from non-interpretable ML. The method is tested to predict faults in a protection system of the LHC main dipole magnets. The method captured 113 out of 116 faults, while only 99 out of 116 faults were captured with the existing tool. The presented approaches demonstrate the strength of interpretable ML for contributing to reliable operation of next-generation particle accelerators. Its applicability extends to numerous other collider components, including radio frequency cavities and dipole magnets in the FCC. |
Corresponding record in: INSPIRE
Email contact: [email protected]
Notice créée le 2024-01-23, modifiée le 2024-11-13
