System Modelling
Artificial Intelligence & Machine Learning
Paper Title Page
TU1BCO01 A Workflow for Training and Deploying Machine Learning Models to EPICS 244
 
  • M.F. Leputa, K.R.L. Baker, M. Romanovschi
    STFC/RAL/ISIS, Chilton, Didcot, Oxon, United Kingdom
 
  The tran­si­tion to EPICS as the con­trol sys­tem for the ISIS Neu­tron and Muon Source ac­cel­er­a­tors is an op­por­tu­nity to more eas­ily in­te­grate ma­chine learn­ing into op­er­a­tions. But de­vel­op­ing high qual­ity ma­chine learn­ing (ML) mod­els is in­suf­fi­cient. In­te­gra­tion into crit­i­cal op­er­a­tions re­quires good de­vel­op­ment prac­tices to en­sure sta­bil­ity and re­li­a­bil­ity dur­ing de­ploy­ment and to allow ro­bust and easy main­te­nance. For these rea­sons we im­ple­mented a work­flow for train­ing and de­ploy­ing mod­els that uti­lize off-the-shelf, in­dus­try-stan­dard tools such as MLflow. Our ex­pe­ri­ence of how adop­tion of these tools can make de­vel­oper’s lives eas­ier dur­ing the train­ing phase of a pro­ject is dis­cussed. We de­scribe how these tools may be used in an au­to­mated de­ploy­ment pipeline to allow the ML model to in­ter­act with our EPICS ecosys­tem through Python-based IOCs within a con­tainer­ized en­vi­ron­ment. This re­duces the de­vel­oper ef­fort re­quired to pro­duce GUIs to in­ter­act with the mod­els within the ISIS Main Con­trol Room as tools fa­mil­iar to op­er­a­tors, such as Phoe­bus, may be used.  
slides icon Slides TU1BCO01 [3.370 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TU1BCO01  
About • Received ※ 05 October 2023 — Accepted ※ 12 October 2023 — Issued ※ 19 October 2023  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TU1BCO02 Integrating System Knowledge in Unsupervised Anomaly Detection Algorithms for Simulation-Based Failure Prediction of Electronic Circuits 249
 
  • F. Waldhauser, H. Boukabache, D. Perrin, S. Roesler
    CERN, Meyrin, Switzerland
  • M. Dazer
    Universität Stuttgart, Stuttgart, Germany
 
  Funding: This work has been sponsored by the Wolfgang Gentner Programme of the German Federal Ministry of Education and Research (grant no. 13E18CHA).
Ma­chine learn­ing al­go­rithms en­able fail­ure pre­dic­tion of large-scale, dis­trib­uted sys­tems using his­tor­i­cal time-se­ries datasets. Al­though un­su­per­vised learn­ing al­go­rithms rep­re­sent a pos­si­bil­ity to de­tect an evolv­ing va­ri­ety of anom­alies, they do not pro­vide links be­tween de­tected data events and sys­tem fail­ures. Ad­di­tional sys­tem knowl­edge is re­quired for ma­chine learn­ing al­go­rithms to de­ter­mine the na­ture of de­tected anom­alies, which may rep­re­sent ei­ther healthy sys­tem be­hav­ior or fail­ure pre­cur­sors. How­ever, knowl­edge on fail­ure be­hav­ior is ex­pen­sive to ob­tain and might only be avail­able upon pre-se­lec­tion of anom­alous sys­tem states using un­su­per­vised al­go­rithms. More­over, sys­tem knowl­edge ob­tained from eval­u­a­tion of sys­tem states needs to be ap­pro­pri­ately pro­vided to the al­go­rithms to en­able per­for­mance im­prove­ments. In this paper, we will pre­sent an ap­proach to ef­fi­ciently con­fig­ure the in­te­gra­tion of sys­tem knowl­edge into un­su­per­vised anom­aly de­tec­tion al­go­rithms for fail­ure pre­dic­tion. The method­ol­ogy is based on sim­u­la­tions of fail­ure modes of elec­tronic cir­cuits. Trig­ger­ing sys­tem fail­ures based on syn­thet­i­cally gen­er­ated fail­ure be­hav­iors en­ables analy­sis of the de­tectabil­ity of fail­ures and gen­er­a­tion of dif­fer­ent types of datasets con­tain­ing sys­tem knowl­edge. In this way, the re­quire­ments for type and ex­tend of sys­tem knowl­edge from dif­fer­ent sources can be de­ter­mined, and suit­able al­go­rithms al­low­ing the in­te­gra­tion of ad­di­tional data can be iden­ti­fied.
 
slides icon Slides TU1BCO02 [2.541 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TU1BCO02  
About • Received ※ 02 October 2023 — Accepted ※ 12 October 2023 — Issued ※ 25 October 2023  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TU1BCO03 Systems Modelling, AI/ML Algorithms Applied to Control Systems 257
 
  • S.A. Mnisi
    SARAO, Cape Town, South Africa
 
  Funding: National Research Foundation (South Africa)
The 64 re­cep­tor (with 20 more being built) radio tele­scope in the Karoo, South Africa, com­prises a large num­ber of de­vices and com­po­nents con­nected to the Con­trol-and-Mon­i­tor­ing (CAM) sys­tem via the Karoo Array Tele­scope Com­mu­ni­ca­tion Pro­to­col (KATCP). KATCP is used ex­ten­sively for in­ter­nal com­mu­ni­ca­tions be­tween CAM com­po­nents and other sub­sys­tems. A KATCP in­ter­face ex­poses re­quests and sen­sors; sam­pling strate­gies are set on sen­sors, rang­ing from sev­eral up­dates per sec­ond to in­fre­quent on-change up­dates. The sen­sor sam­ples are of dif­fer­ent types, from small in­te­gers to text fields. The sam­ples and as­so­ci­ated time­stamps are per­ma­nently stored and made avail­able for sci­en­tists, en­gi­neers and op­er­a­tors to query and an­a­lyze. This is a pre­sen­ta­tion on how to apply Ma­chine Learn­ing tools which uti­lize data-dri­ven al­go­rithms and sta­tis­ti­cal mod­els to an­a­lyze sen­sor data sets and then draw in­fer­ences from iden­ti­fied pat­terns or make pre­dic­tions based on them. The al­go­rithms learn from the sen­sor data as they run against it, un­like tra­di­tional rules-based an­a­lyt­ics sys­tems that fol­low ex­plicit in­struc­tions. Since this in­volves data pre­pro­cess­ing, we will go through how the MeerKAT tele­scope data stor­age in­fra­struc­ture (called Kat­store) man­ages the vo­lu­mi­nous va­ri­ety, ve­loc­ity and vol­ume of this data.
 
slides icon Slides TU1BCO03 [1.647 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TU1BCO03  
About • Received ※ 06 October 2023 — Revised ※ 09 November 2023 — Accepted ※ 14 December 2023 — Issued ※ 21 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TU1BCO04 Laser Focal Position Correction Using FPGA-Based ML Models 262
 
  • J.A. Einstein-Curtis, S.J. Coleman, N.M. Cook, J.P. Edelen
    RadiaSoft LLC, Boulder, Colorado, USA
  • S.K. Barber, C.E. Berger, J. van Tilborg
    LBNL, Berkeley, California, USA
 
  Funding: This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of High Energy Physics under Award Number DE-SC 00259037.
High rep­e­ti­tion-rate, ul­tra­fast laser sys­tems play a crit­i­cal role in a host of mod­ern sci­en­tific and in­dus­trial ap­pli­ca­tions. We pre­sent a di­ag­nos­tic and cor­rec­tion scheme for con­trol­ling and de­ter­min­ing laser focal po­si­tion by uti­liz­ing fast wave­front sen­sor mea­sure­ments from mul­ti­ple po­si­tions to train a focal po­si­tion pre­dic­tor. This pre­dic­tor and ad­di­tional con­trol al­go­rithms have been in­te­grated into a uni­fied con­trol in­ter­face and FPGA-based con­troller on beam­lines at the Bella fa­cil­ity at LBNL. An op­tics sec­tion is ad­justed on­line to pro­vide the de­sired cor­rec­tion to the focal po­si­tion on mil­lisec­ond timescales by de­ter­min­ing cor­rec­tions for an ac­tu­a­tor in a tele­scope sec­tion along the beam­line. Our ini­tial proof-of-prin­ci­ple demon­stra­tions lever­aged pre-com­piled data and pre-trained net­works op­er­at­ing ex-situ from the laser sys­tem. A frame­work for gen­er­at­ing a low-level hard­ware de­scrip­tion of ML-based cor­rec­tion al­go­rithms on FPGA hard­ware was cou­pled di­rectly to the beam­line using the AMD Xil­inx Vitis AI tool­chain in con­junc­tion with de­ploy­ment scripts. Lastly, we con­sider the use of re­mote com­put­ing re­sources, such as the Sirepo sci­en­tific frame­work*, to ac­tively up­date these cor­rec­tion schemes and de­ploy mod­els to a pro­duc­tion en­vi­ron­ment.
* M.S. Rakitin et al., "Sirepo: an open-source cloud-based software interface for X-ray source and optics simulations" Journal of Synchrotron Radiation25, 1877-1892 (Nov 2018).
 
slides icon Slides TU1BCO04 [1.876 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TU1BCO04  
About • Received ※ 06 October 2023 — Accepted ※ 14 November 2023 — Issued ※ 18 December 2023  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TU1BCO05 Model Driven Reconfiguration of LANSCE Tuning Methods 267
 
  • C.E. Taylor, P.M. Anisimov, S.A. Baily, E.-C. Huang, H.L. Leffler, L. Rybarcyk, A. Scheinker, H.A. Watkins, E.E. Westbrook, D.D. Zimmermann
    LANL, Los Alamos, New Mexico, USA
 
  Funding: National Nuclear Security Administration (NNSA)
This work pre­sents a re­view of the shift in tun­ing meth­ods em­ployed at the Los Alamos Neu­tron Sci­ence Cen­ter (LAN­SCE). We ex­plore the tun­ing cat­e­gories and meth­ods em­ployed in four key sec­tions of the ac­cel­er­a­tor, namely the Low-En­ergy Beam Trans­port (LEBT), the Drift Tube Linac (DTL), the side-Cou­pled Cav­ity Linac (CCL), and the High-En­ergy Beam Trans­port (HEBT). The study ad­di­tion­ally pre­sents the find­ings of em­ploy­ing novel soft­ware tools and al­go­rithms to en­hance each do­main’s beam qual­ity and per­for­mance. This study show­cases the ef­fi­cacy of in­te­grat­ing model-dri­ven and model-in­de­pen­dent tun­ing tech­niques, along with ac­cep­tance and adap­tive tun­ing strate­gies, to en­hance the op­ti­miza­tion of beam de­liv­ery to ex­per­i­men­tal fa­cil­i­ties. The re­search ad­di­tion­ally ad­dresses the prospec­tive strate­gies for aug­ment­ing the con­trol sys­tem and di­ag­nos­tics of LAN­SCE.
*R.W. Garnett, J. Phys.: Conf. Ser. 1021 012001
**A. Scheinker, Rev. ST Accel. Beams 16 102803 2013
***R. Keller, Proc of Part Accel Conf
****M. Oothoudt, Proc of Part Accel Conf, 2003, v4
 
slides icon Slides TU1BCO05 [2.886 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TU1BCO05  
About • Received ※ 06 October 2023 — Revised ※ 08 October 2023 — Accepted ※ 12 December 2023 — Issued ※ 13 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TU1BCO06 Disentangling Beam Losses in The Fermilab Main Injector Enclosure Using Real-Time Edge AI 273
 
  • K.J. Hazelwood, J.M.S. Arnold, M.R. Austin, J.R. Berlioz, P.M. Hanlet, M.A. Ibrahim, A.T. Livaudais-Lewis, J. Mitrevski, V.P. Nagaslaev, A. Narayanan, D.J. Nicklaus, G. Pradhan, A.L. Saewert, B.A. Schupbach, K. Seiya, R.M. Thurman-Keup, N.V. Tran
    Fermilab, Batavia, Illinois, USA
  • J.YC. Hu, J. Jiang, H. Liu, S. Memik, R. Shi, A.M. Shuping, M. Thieme, C. Xu
    Northwestern University, EVANSTON, USA
  • A. Narayanan
    Northern Illinois University, DeKalb, Illinois, USA
 
  The Fer­mi­lab Main In­jec­tor en­clo­sure houses two ac­cel­er­a­tors, the Main In­jec­tor and Re­cy­cler Ring. Dur­ing nor­mal op­er­a­tion, high in­ten­sity pro­ton beams exist si­mul­ta­ne­ously in both. The two ac­cel­er­a­tors share the same beam loss mon­i­tors (BLM) and mon­i­tor­ing sys­tem. De­ci­pher­ing the ori­gin of any of the 260 BLM read­ings is often dif­fi­cult. The (Ac­cel­er­a­tor) Real-time Edge AI for Dis­trib­uted Sys­tems pro­ject, or READS, has de­vel­oped an AI/ML model, and im­ple­mented it on fast FPGA hard­ware, that dis­en­tan­gles mixed beam losses and at­trib­utes prob­a­bil­i­ties to each BLM as to which ma­chine(s) the loss orig­i­nated from in real-time. The model in­fer­ences are then streamed to the Fer­mi­lab ac­cel­er­a­tor con­trols net­work (ACNET) where they are avail­able for op­er­a­tors and ex­perts alike to aid in tun­ing the ma­chines.  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TU1BCO06  
About • Received ※ 06 October 2023 — Revised ※ 11 October 2023 — Accepted ※ 15 November 2023 — Issued ※ 06 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUMBCMO13 Applications of Artificial Intelligence in Laser Accelerator Control System 372
 
  • F.N. Li, K.C. Chen, Z. Guo, Q.Y. He, C. Lin, Q. Wang, Y. Xia, M.X. Zang
    PKU, Beijing, People’s Republic of China
 
  Funding: the National Natural Science Foundation of China (Grants No. 11975037, NO. 61631001 and No. 11921006), and the National Grand Instrument Project (No. 2019YFF01014400 and No. 2019YFF01014404).
Ul­tra-in­tense laser-plasma in­ter­ac­tions can pro­duce TV/m ac­cel­er­a­tion gra­di­ents, mak­ing them promis­ing for com­pact ac­cel­er­a­tors. Peking Uni­ver­sity is con­struct­ing a pro­ton ra­dio­ther­apy sys­tem pro­to­type based on PW laser ac­cel­er­a­tors, but tran­sient processes chal­lenge sta­bil­ity con­trol, crit­i­cal for med­ical ap­pli­ca­tions. This work demon­strates ar­ti­fi­cial in­tel­li­gence’s (AI) ap­pli­ca­tion in laser ac­cel­er­a­tor con­trol sys­tems. To achieve mi­cro-pre­ci­sion align­ment be­tween the ul­tra-in­tense laser and tar­get, we pro­pose an au­to­mated po­si­tion­ing pro­gram using the YOLO al­go­rithm. This real-time method em­ploys a con­vo­lu­tional neural net­work, di­rectly pre­dict­ing ob­ject lo­ca­tions and class prob­a­bil­i­ties from input im­ages. It en­ables pre­cise, au­to­matic solid tar­get align­ment in about a hun­dred mil­lisec­onds, re­duc­ing ex­per­i­men­tal prepa­ra­tion time. The YOLO al­go­rithm is also in­te­grated into the safety in­ter­lock­ing sys­tem for anti-tail­ing, al­low­ing quick emer­gency re­sponse. The in­tel­li­gent con­trol sys­tem also en­ables con­ve­nient, ac­cu­rate beam tun­ing. We de­vel­oped high-per­for­mance vir­tual ac­cel­er­a­tor soft­ware using "OpenXAL" and GPU-ac­cel­er­ated multi-par­ti­cle beam trans­port sim­u­la­tions. The soft­ware al­lows real-time or cus­tom pa­ra­me­ter sim­u­la­tions and fea­tures con­trol in­ter­faces com­pat­i­ble with op­ti­miza­tion al­go­rithms. By de­sign­ing tai­lored ob­jec­tive func­tions, de­sired beam size and dis­tri­b­u­tion can be achieved in a few it­er­a­tions.
 
slides icon Slides TUMBCMO13 [1.162 MB]  
poster icon Poster TUMBCMO13 [1.011 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUMBCMO13  
About • Received ※ 04 October 2023 — Revised ※ 12 October 2023 — Accepted ※ 23 November 2023 — Issued ※ 23 November 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUMBCMO14 Initial Test of a Machine Learning Based SRF Cavity Active Resonance Control 379
 
  • F.Y. Wang, J. Cruz
    SLAC, Menlo Park, California, USA
 
  We’ll in­tro­duce a high pre­ci­sion ac­tive mo­tion con­troller based on ma­chine learn­ing (ML) tech­nol­ogy and elec­tric piezo ac­tu­a­tor. The con­troller will be used for SRF cav­ity ac­tive res­o­nance con­trol, where a data-dri­ven model for sys­tem mo­tion dy­nam­ics will be de­vel­oped first, and a model pre­dic­tive con­troller (MPC) will be built ac­cord­ingly. Sim­u­la­tion re­sults as well as ini­tial test re­sults with real SRF cav­i­ties will be pre­sented in the paper.  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUMBCMO14  
About • Received ※ 03 October 2023 — Revised ※ 14 November 2023 — Accepted ※ 27 November 2023 — Issued ※ 09 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUMBCMO15 Enhancing Electronic Logbooks Using Machine Learning 382
 
  • J. Maldonado, S.L. Clark, W. Fu, S. Nemesure
    BNL, Upton, New York, USA
 
  Funding: Work supported by Brookhaven Science Associates, LLC under Contract No. DE-SC0012704
The elec­tronic log­book (elog) sys­tem used at Brookhaven Na­tional Lab­o­ra­tory’s Col­lider-Ac­cel­er­a­tor De­part­ment (C-AD) al­lows users to cus­tomize log­book set­tings, in­clud­ing spec­i­fi­ca­tion of fa­vorite log­books. Using ma­chine learn­ing tech­niques, cus­tomiza­tions can be fur­ther per­son­al­ized to pro­vide users with a view of en­tries that match their spe­cific in­ter­ests. We will uti­lize nat­ural lan­guage pro­cess­ing (NLP), op­ti­cal char­ac­ter recog­ni­tion (OCR), and topic mod­els to aug­ment the elog sys­tem. NLP tech­niques will be used to process and clas­sify text en­tries. To an­a­lyze en­tries in­clud­ing im­ages with text, such as screen­shots of con­trols sys­tem ap­pli­ca­tions, we will apply OCR. Topic mod­els will gen­er­ate entry rec­om­men­da­tions that will be com­pared to pre­vi­ously tested lan­guage pro­cess­ing mod­els. We will de­velop a com­mand line in­ter­face tool to ease au­toma­tion of NLP tasks in the con­trols sys­tem and cre­ate a web in­ter­face to test entry rec­om­men­da­tions. This tech­nique will cre­ate rec­om­men­da­tions for each user, pro­vid­ing cus­tom sets of en­tries and pos­si­bly elim­i­nate the need for man­ual search­ing.
 
slides icon Slides TUMBCMO15 [0.905 MB]  
poster icon Poster TUMBCMO15 [4.697 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUMBCMO15  
About • Received ※ 04 October 2023 — Revised ※ 12 October 2023 — Accepted ※ 24 November 2023 — Issued ※ 10 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP020 Summary Report on Machine Learning-Based Applications at the Synchrotron Light Source Delta 537
 
  • D. Schirmer, S. Khan, A. Radha Krishnan
    DELTA, Dortmund, Germany
 
  In re­cent years, sev­eral con­trol sys­tem ap­pli­ca­tions using ma­chine learn­ing (ML) tech­niques have been de­vel­oped and tested to au­to­mate the con­trol and op­ti­miza­tion of the 1.5 GeV syn­chro­tron ra­di­a­tion source DELTA. These ap­pli­ca­tions cover a wide range of tasks, in­clud­ing elec­tron beam po­si­tion cor­rec­tion, work­ing point con­trol, chro­matic­ity ad­just­ment, in­jec­tion process op­ti­miza­tion, as well as CHG-spec­tra (co­her­ent har­monic gen­er­a­tion) analy­sis. Var­i­ous ma­chine learn­ing tech­niques have been used to im­ple­ment these pro­jects. This re­port pro­vides an overview of these pro­jects, sum­ma­rizes the cur­rent re­sults, and in­di­cates ideas for fu­ture im­prove­ments.  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP020  
About • Received ※ 04 October 2023 — Accepted ※ 06 December 2023 — Issued ※ 13 December 2023  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP102 Leveraging Local Intelligence to Industrial Control Systems through Edge Technologies 793
 
  • A. Patil, F. Ghawash, B. Schofield, F. Varela
    CERN, Meyrin, Switzerland
  • D. Daniel, K. Kaufmann, A.S. Sündermann
    SAGÖ, Vienna, Austria
  • C. Kern
    Siemens AG, Corporate Technology, München, Germany
 
  In­dus­trial processes often use ad­vanced con­trol al­go­rithms such as Model Pre­dic­tive Con­trol (MPC) and Ma­chine Learn­ing (ML) to im­prove per­for­mance and ef­fi­ciency. How­ever, de­ploy­ing these al­go­rithms can be chal­leng­ing, par­tic­u­larly when they re­quire sig­nif­i­cant com­pu­ta­tional re­sources and in­volve com­plex com­mu­ni­ca­tion pro­to­cols be­tween dif­fer­ent con­trol sys­tem com­po­nents. To ad­dress these chal­lenges, we show­case an ap­proach lever­ag­ing in­dus­trial edge tech­nolo­gies to de­ploy such al­go­rithms. An edge de­vice is a com­pact and pow­er­ful com­put­ing de­vice placed at the net­work’s edge, close to the process con­trol. It ex­e­cutes the al­go­rithms with­out ex­ten­sive com­mu­ni­ca­tion with other con­trol sys­tem com­po­nents, thus re­duc­ing la­tency and load on the cen­tral con­trol sys­tem. We also em­ploy an an­a­lyt­ics func­tion plat­form to man­age the life cycle of the al­go­rithms, in­clud­ing mod­i­fi­ca­tions and re­place­ments, with­out dis­rupt­ing the in­dus­trial process. Fur­ther­more, we demon­strate a use case where an MPC al­go­rithm is run on an edge de­vice to con­trol a Heat­ing, Ven­ti­la­tion, and Air Con­di­tion­ing (HVAC) sys­tem. An edge de­vice run­ning the al­go­rithm can an­a­lyze data from tem­per­a­ture sen­sors, per­form com­plex cal­cu­la­tions, and ad­just the op­er­a­tion of the HVAC sys­tem ac­cord­ingly. In sum­mary, our ap­proach of uti­liz­ing edge tech­nolo­gies en­ables us to over­come the lim­i­ta­tions of tra­di­tional ap­proaches to de­ploy­ing ad­vanced con­trol al­go­rithms in in­dus­trial set­tings, pro­vid­ing more in­tel­li­gent and ef­fi­cient con­trol of in­dus­trial processes.  
poster icon Poster TUPDP102 [3.321 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP102  
About • Received ※ 06 October 2023 — Revised ※ 21 October 2023 — Accepted ※ 05 December 2023 — Issued ※ 12 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP114 Machine Learning Based Noise Reduction of Neutron Camera Images at ORNL 841
 
  • I.V. Pogorelov, J.P. Edelen, M.J. Henderson, M.C. Kilpatrick
    RadiaSoft LLC, Boulder, Colorado, USA
  • S. Calder, B. Vacaliuc
    ORNL RAD, Oak Ridge, Tennessee, USA
  • R.D. Gregory, G.S. Guyotte, C.M. Hoffmann, B.K. Krishna
    ORNL, Oak Ridge, Tennessee, USA
 
  Funding: This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Science under Award Number DE-SC0021555.
Neu­tron cam­eras are uti­lized at the HB2A pow­der dif­frac­tome­ter to image the sam­ple for align­ment in the beam. Typ­i­cally, neu­tron cam­eras are quite noisy as they are con­stantly being ir­ra­di­ated. Re­moval of this noise is chal­leng­ing due to the ir­reg­u­lar na­ture of the pixel in­ten­sity fluc­tu­a­tions and the ten­dency for it to change over time. Ra­di­a­Soft has de­vel­oped a novel noise re­duc­tion method for neu­tron cam­eras that in­scribes a lower en­ve­lope of the image sig­nal. This process is then sped up using ma­chine learn­ing. Here we re­port on the re­sults of our noise re­duc­tion method and de­scribe our ma­chine learn­ing ap­proach for speed­ing up the al­go­rithm for use dur­ing op­er­a­tions.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP114  
About • Received ※ 07 October 2023 — Revised ※ 22 October 2023 — Accepted ※ 11 December 2023 — Issued ※ 16 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP115 Machine Learning for Compact Industrial Accelerators 846
 
  • J.P. Edelen, J.A. Einstein-Curtis, M.J. Henderson, M.C. Kilpatrick
    RadiaSoft LLC, Boulder, Colorado, USA
  • J.A. Diaz Cruz, A.L. Edelen
    SLAC, Menlo Park, California, USA
 
  Funding: This material is based upon work supported by the DOE Accelerator R&D and Production under Award Number DE-SC0023641.
The in­dus­trial and med­ical ac­cel­er­a­tor in­dus­try is an ever-grow­ing field with ad­vance­ments in ac­cel­er­a­tor tech­nol­ogy en­abling its adop­tion for new ap­pli­ca­tions. As the com­plex­ity of in­dus­trial ac­cel­er­a­tors grows so does the need for more so­phis­ti­cated con­trol sys­tems to reg­u­late their op­er­a­tion. More­over, the en­vi­ron­ment for in­dus­trial and med­ical ac­cel­er­a­tors is often harsh and noisy as op­posed to the more con­trolled en­vi­ron­ment of a lab­o­ra­tory-based ma­chine. This en­vi­ron­ment makes con­trol more chal­leng­ing. Ad­di­tion­ally, in­stru­men­ta­tion for in­dus­trial ac­cel­er­a­tors is lim­ited mak­ing it dif­fi­cult at times to iden­tify and di­ag­nose prob­lems when they occur. Ra­di­a­Soft has part­nered with SLAC to de­velop new ma­chine learn­ing meth­ods for con­trol and anom­aly de­tec­tion for in­dus­trial ac­cel­er­a­tors. Our ap­proach is to de­velop our meth­ods using sim­u­la­tion mod­els fol­lowed by test­ing on ex­per­i­men­tal sys­tems. Here we pre­sent ini­tial re­sults using sim­u­la­tions of a room tem­per­a­ture s-band sys­tem.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP115  
About • Received ※ 06 October 2023 — Accepted ※ 05 December 2023 — Issued ※ 18 December 2023  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP116 Machine Learning Based Sample Alignment at TOPAZ 851
 
  • M.J. Henderson, J.P. Edelen, M.C. Kilpatrick, I.V. Pogorelov
    RadiaSoft LLC, Boulder, Colorado, USA
  • S. Calder, B. Vacaliuc
    ORNL RAD, Oak Ridge, Tennessee, USA
  • R.D. Gregory, G.S. Guyotte, C.M. Hoffmann, B.K. Krishna
    ORNL, Oak Ridge, Tennessee, USA
 
  Funding: This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Science under Award Number DE-SC0021555.
Neu­tron scat­ter­ing ex­per­i­ments are a crit­i­cal tool for the ex­plo­ration of mol­e­c­u­lar struc­ture in com­pounds. The TOPAZ sin­gle crys­tal dif­frac­tome­ter at the Spal­la­tion Neu­tron Source stud­ies these sam­ples by il­lu­mi­nat­ing sam­ples with dif­fer­ent en­ergy neu­tron beams and record­ing the scat­tered neu­trons. Dur­ing the ex­per­i­ments the user will change tem­per­a­ture and sam­ple po­si­tion in order to il­lu­mi­nate dif­fer­ent crys­tal faces and to study the sam­ple in dif­fer­ent en­vi­ron­ments. Main­tain­ing align­ment of the sam­ple dur­ing this process is key to en­sur­ing high qual­ity data are col­lected. At pre­sent this process is per­formed man­u­ally by beam­line sci­en­tists. Ra­di­a­Soft in col­lab­o­ra­tion with the beam­line sci­en­tists and en­gi­neers at ORNL has de­vel­oped a new ma­chine learn­ing based align­ment soft­ware au­tomat­ing this process. We uti­lize a fully-con­nected con­vo­lu­tional neural net­work con­fig­ured in a U-net ar­chi­tec­ture to iden­tify the sam­ple cen­ter of mass. We then move the sam­ple using a cus­tom python-based EPICS IOC in­ter­faced with the mo­tors. In this talk we pro­vide an overview of our ma­chine learn­ing tools and show our ini­tial re­sults align­ing sam­ples at ORNL.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP116  
About • Received ※ 06 October 2023 — Accepted ※ 05 December 2023 — Issued ※ 11 December 2023  
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP117 Classification and Prediction of Superconducting Magnet Quenches 856
 
  • J.A. Einstein-Curtis, J.P. Edelen, M.C. Kilpatrick, R. O’Rourke
    RadiaSoft LLC, Boulder, Colorado, USA
  • K.A. Drees, J.S. Laster, M. Valette
    BNL, Upton, New York, USA
 
  Funding: This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of High Energy Physics under Award Number DE-SC0021699.
Ro­bust and re­li­able quench de­tec­tion for su­per­con­duct­ing mag­nets is in­creas­ingly im­por­tant as fa­cil­i­ties push the bound­aries of in­ten­sity and op­er­a­tional run­time. Ra­di­a­Soft has been work­ing with Brookhaven Na­tional Lab on quench de­tec­tion and pre­dic­tion for su­per­con­duct­ing mag­nets in­stalled in the RHIC stor­age rings. This pro­ject has an­a­lyzed sev­eral years of power sup­ply and beam po­si­tion mon­i­tor data to train au­to­mated clas­si­fi­ca­tion tools and au­to­mated quench pre­cur­sor de­ter­mi­na­tion based on input se­quences. Clas­si­fi­ca­tion was per­formed using su­per­vised mul­ti­layer per­cep­tron and boosted de­ci­sion tree ar­chi­tec­tures, while mod­els of the ex­pected op­er­a­tion of the ring were de­vel­oped using a va­ri­ety of au­toen­coder ar­chi­tec­tures. We have con­tin­ued ef­forts to max­i­mize area under the re­ceiver op­er­at­ing char­ac­ter­is­tic curve for the mul­ti­ple clas­si­fi­ca­tion prob­lem of real quench, fake quench, and no-quench events. We have also begun work on long short-term mem­ory (LSTM) and other re­cur­rent ar­chi­tec­tures for quench pre­dic­tion. Ex­am­i­na­tions of fu­ture work uti­liz­ing more ro­bust ar­chi­tec­tures, such as vari­a­tional au­toen­coders and Siamese mod­els, as well as meth­ods nec­es­sary for un­cer­tainty quan­tifi­ca­tion will be dis­cussed.
 
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP117  
About • Received ※ 08 October 2023 — Revised ※ 22 October 2023 — Accepted ※ 05 December 2023 — Issued ※ 07 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)  
 
TUPDP138 Exploratory Data Analysis on the RHIC Cryogenics System Compressor Dataset 907
 
  • Y. Gao, K.A. Brown, R.J. Michnoff, L.K. Nguyen, A.Z. Zarcone, B. van Kuik
    BNL, Upton, New York, USA
  • A.D. Tran
    FRIB, East Lansing, Michigan, USA
 
  Funding: Work supported by Brookhaven Science Associates, LLC under Contract No. DE-SC0012704 with the U.S. Department of Energy.
The Rel­a­tivis­tic Heavy Ion Col­lider (RHIC) Cryo­genic Re­frig­er­a­tor Sys­tem is the cryo­genic heart that al­lows RHIC su­per­con­duct­ing mag­nets to op­er­ate. Parts of the re­frig­er­a­tor are two stages of com­pres­sion com­posed of ten first and five sec­ond-stage com­pres­sors. Com­pres­sors are crit­i­cal for op­er­a­tions. When a com­pres­sor faults, it can im­pact RHIC beam op­er­a­tions if a spare com­pres­sor is not brought on­line as soon as pos­si­ble. The po­ten­tial of ap­ply­ing ma­chine learn­ing to de­tect com­pres­sor prob­lems be­fore a fault oc­curs would greatly en­hance Cryo op­er­a­tions, al­low­ing an op­er­a­tor to switch to a spare com­pres­sor be­fore a run­ning com­pres­sor fails, min­i­miz­ing im­pacts on RHIC op­er­a­tions. In this work, var­i­ous data analy­sis re­sults on his­tor­i­cal com­pres­sor data are pre­sented. It demon­strates an au­toen­coder-based method, which can catch early signs of com­pres­sor trips so that ad­vance no­tices can be sent for the op­er­a­tors to take ac­tion.
 
poster icon Poster TUPDP138 [2.897 MB]  
DOI • reference for this paper ※ doi:10.18429/JACoW-ICALEPCS2023-TUPDP138  
About • Received ※ 05 October 2023 — Revised ※ 22 October 2023 — Accepted ※ 30 November 2023 — Issued ※ 11 December 2023
Cite • reference for this paper using ※ BibTeX, ※ LaTeX, ※ Text/Word, ※ RIS, ※ EndNote (xml)