Author: Pernkopf, F.
Paper Title Page
MOPAB344 Machine Learning Models for Breakdown Prediction in RF Cavities for Accelerators 1068
 
  • C. Obermair, A. Apollonio, T. Cartier-Michaud, N. Catalán Lasheras, L. Felsberger, W.L. Millar, W. Wuensch
    CERN, Geneva, Switzerland
  • C. Obermair, F. Pernkopf
    TUG, Graz, Austria
  
  Radio Fre­quency (RF) break­downs are one of the most preva­lent lim­its in RF cav­i­ties for par­ti­cle ac­cel­er­a­tors. Dur­ing a break­down, field en­hance­ment as­so­ci­ated with small de­for­ma­tions on the cav­ity sur­face re­sults in elec­tri­cal arcs. Such arcs de­grade a pass­ing beam and if they occur fre­quently, they can cause ir­repara­ble dam­age to the RF cav­ity sur­face. In this paper, we pro­pose a ma­chine learn­ing ap­proach to pre­dict the oc­cur­rence of break­downs in CERN’s Com­pact LIn­ear Col­lider (CLIC) ac­cel­er­at­ing struc­tures. We dis­cuss state-of-the-art al­go­rithms for data ex­plo­ration with un­su­per­vised ma­chine learn­ing, break­down pre­dic­tion with su­per­vised ma­chine learn­ing, and re­sult val­i­da­tion with Ex­plain­able-Ar­ti­fi­cial In­tel­li­gence (Ex­plain­able AI). By in­ter­pret­ing the model pa­ra­me­ters of var­i­ous ap­proaches, we go fur­ther in ad­dress­ing op­por­tu­ni­ties to elu­ci­date the physics of a break­down and im­prove ac­cel­er­a­tor re­li­a­bil­ity and op­er­a­tion.  
DOI • reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2021-MOPAB344  
About • paper received ※ 20 May 2021       paper accepted ※ 16 July 2021       issue date ※ 11 August 2021  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
MOPAB345 Machine Learning with a Hybrid Model for Monitoring of the Protection Systems of the LHC 1072
 
  • C. Obermair, A. Apollonio, Z. Charifoulline, M. Maciejewski, A.P. Verweij
    CERN, Geneva, Switzerland
  • C. Obermair, F. Pernkopf
    TUG, Graz, Austria
  
  The LHC is the world’s largest par­ti­cle ac­cel­er­a­tor and uses a com­plex set of so­phis­ti­cated and highly re­li­able ma­chine pro­tec­tion sys­tems to en­sure a safe op­er­a­tion with high avail­abil­ity for par­ti­cle physics pro­duc­tion. The data gath­ered dur­ing sev­eral years of suc­cess­ful op­er­a­tion allow the use of data-dri­ven meth­ods to as­sist ex­perts in find­ing anom­alies in the be­hav­ior of those pro­tec­tion sys­tems. In this paper, we de­rive a model that can ex­tend the ex­ist­ing sig­nal mon­i­tor­ing ap­pli­ca­tions for the LHC pro­tec­tion sys­tems with ma­chine learn­ing. Our hy­brid model com­bines an ex­ist­ing thresh­old-based sys­tem with a SVM by using sig­nals, man­u­ally val­i­dated by ex­perts. Even with a lim­ited amount of data, the SVM learns to in­te­grate the ex­pert knowl­edge and con­tributes to a bet­ter clas­si­fi­ca­tion of safety-crit­i­cal sig­nals. Using this ap­proach, we an­a­lyze his­tor­i­cal sig­nals of quench heaters, which are an im­por­tant part of the quench pro­tec­tion sys­tem for su­per­con­duct­ing mag­nets. Par­tic­u­larly, it is pos­si­ble to in­cor­po­rate ex­pert de­ci­sions into the clas­si­fi­ca­tion process and to im­prove the fail­ure de­tec­tion rate of the ex­ist­ing quench heater dis­charge analy­sis tool.  
DOI • reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2021-MOPAB345  
About • paper received ※ 20 May 2021       paper accepted ※ 19 July 2021       issue date ※ 01 September 2021  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)