Author: Tennant, C.
Paper Title Page
FRXC01 Superconducting Radio-Frequency Cavity Fault Classification Using Machine Learning at Jefferson Laboratory 4535
 
  • C. Tennant, A. Carpenter, T. Powers, L.S. Vidyaratne
    JLab, Newport News, Virginia, USA
  • K.M. Iftekharuddin, M. Rahman
    ODU, Norfolk, Virginia, USA
  • A.D. Shabalina
    STFC/DL, Daresbury, Warrington, Cheshire, United Kingdom
 
  Funding: This work supported by the U.S. Department of Energy, Office of Science, Office of Nuclear Physics under Contract No. DE-AC05-06OR23177.
We re­port on the de­vel­op­ment of ma­chine learn­ing mod­els for clas­si­fy­ing C100 su­per­con­duct­ing ra­diofre­quency (SRF) cav­ity faults in the Con­tin­u­ous Elec­tron Beam Ac­cel­er­a­tor Fa­cil­ity (CEBAF) at Jef­fer­son Lab. Of the 418 SRF cav­i­ties in CEBAF, 96 are de­signed with a dig­i­tal low-level RF sys­tem con­fig­ured such that a cav­ity fault trig­gers record­ings of RF sig­nals for each of eight cav­i­ties in the cry­omod­ule. Sub­ject mat­ter ex­perts an­a­lyze the col­lected time-se­ries data and iden­tify which of the eight cav­i­ties faulted first and clas­sify the type of fault. This in­for­ma­tion is used to find trends and strate­gi­cally de­ploy mit­i­ga­tions to prob­lem­atic cry­omod­ules. How­ever, man­u­ally la­bel­ing the data is la­bo­ri­ous and time-con­sum­ing. By lever­ag­ing ma­chine learn­ing, near real-time - rather than post­mortem - iden­ti­fi­ca­tion of the of­fend­ing cav­ity and clas­si­fi­ca­tion of the fault type has been im­ple­mented. We dis­cuss the per­for­mance of the ma­chine learn­ing mod­els dur­ing a re­cent physics run. We also dis­cuss ef­forts for fur­ther in­sights into fault types through un­su­per­vised learn­ing tech­niques and pre­sent pre­lim­i­nary work on cav­ity and fault pre­dic­tion using data col­lected prior to a fail­ure event.
 
DOI • reference for this paper ※ https://doi.org/10.18429/JACoW-IPAC2021-FRXC01  
About • paper received ※ 16 May 2021       paper accepted ※ 01 July 2021       issue date ※ 13 August 2021  
Export • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)