Branlard, J. Eichler, A. Timm, J.H.K. Walker, N. JACoW Publishing Geneva, Switzerland 2226-0366 978-3-95450-215-8 10.18429/JACoW-LINAC2022-THPOPA26 English 798-801 cavity FEL operation software hardware Contribution to a conference proceedings 2022 2022-09 https://doi.org/10.18429/JACoW-LINAC2022-THPOPA26 https://jacow.org/linac2022/papers/thpopa26.pdf A server-based quench detection system is used since the beginning of operation at the European XFEL (2017) to stop driving superconducting cavities if they experience a quench. While this approach effectively detects quenches, it also generates false positives, tripping the accelerating stations when failures other than quenches occur. Using the post-mortem data snapshots generated for every trip, an additional signal (referred to as residual) is systematically computed based on the standard cavity model. Following an initial training on a set of such residuals derived from quench as well as non-quench events, two independent machine learning engines analyze routinely the trip snapshots and their residuals to identify if a trip was indeed triggered by a quench or has another root cause. The outcome of the analysis is automatically appended to the data snapshots and distributed to a team of experts. This constitutes a fully deployed example of machine-learning-assisted failure classification to identify quenches, supporting experts in their daily routine of monitoring and documenting the accelerator uptime and availability.