1st. Korea-Japan KAGRA detector characterization meeting Date: 2013-06-25 (Tue) 11:00~ At SeeVogh Minutes taker: Yokozawa, T. Attendee; Edwin J. Son, John J. Oh, Kyungmin Kim, SangHoon Oh, Young-Min Kim, Shuhei Mano, Kazuhiro Hayama, Kenichi Ohara, Masato Kaneyama, Yousuke Itoh, Shuhei Mano, Takaaki Yokozawa, Hirotaka Yuzurihara, Kazuyuki Tanaka, Takahiro Yamamoto munites; Self introduction Send slide http://gwdoc.icrr.u-tokyo.ac.jp/cgi-bin/private/DocDB/ShowDocument?docid=1718 title Development of the multi-channel analysis for noise-source identification explain about slide by Hayama-san Short commissioning period, a half of a year, compared with LIGO (~2years) We should do commissioning very quickly Background ~10000 physics and environmental monitors in/around KAGRA Detcher group provides a systems for identifing noise sources and in order to do that, it forcuses on multivariate analysis The task of our project is written in http://gwdoc.icrr.u-tokyo.ac.jp/cgi-bin/private/DocDB/ShowDocument?docid=1718 to develop event trigger generation pipeline. to develop supervised/un-supervised multivariate classifier to identify glitches to develop a system to localize their noise sources. Writing a scientific paper using CLIO (or LIGO). Young-Min presentation Received by e-mail characterization work using CLIO data and do HHT analysis Since glitches are not gaussian, difficult to identify them. The goal include cleaning and monitoring data. Page 3 what do you mean "cleaning data" between glitch or data quality was high? Mean remove glitch from data? if removing glitch, data becomes good. high correlation between the GW channel and PE channels. Korean group was working Page 4 Page4 Could you explain lower left figure? There is n channel there, what is purple circle x1, x2… Time series? No, this is the trigger information. Is there m-th trigger information. there is m label. In each trigger, there is noise glitch And try to remove those, and motivate to make clean data with many trigger channels Page 5 shows 4 comparable components L1-L3 are combination of comparable components. you can see arXiv in detail. This is current job Using auxMCV method, try to improve efficiency Page8 blue line is ensemble network result right compare 4 components. Page 6 What do you do here? You have same data sets and trigger sets All networks are independent it means different computers Does it means several layers? No, several Networks. Take Ensemble of them left top shows efficiency is almost 100% bout others are about 50% efficiency We should study about feature selection Page 9 Where do you use this statistics Each related channels You tried to find responsible channels first? Yes. Page 10 ROC pictures. There is no improvements.. have to continue these works. opinion we need different input sources in this work So need Deferent(Different?) input information Page 12 there is several trigger generation methods. today not need to explain detail about HHT method move to page 18 Hibert method is Adaptive basis show plot of time series for each channel Page 19 IMF1 in CH:A IMF5 in CH:B why? just show one of explain we don't know how to choose it. discrepancy position will be well. CH:B is not so significant to see the data. So I asked Two simulation data of CBC signal see kind of peak at the same injection time Check CLIO data with help of Hayama-san and Yamamoto-san There seems to see many peaks Need to decide trigger or not -> need more studies. Presentation was finished. Page 10 The ROC curve was improved what causes it? performance of Ensemble network Ensemble network has more connection between ch That may be the most reason. What is your working plan? future plan? We want to improve the artificial neural networks ensemble HHT improvement Need more CLIO multi channel information Need some time, but we will provide data. We need to deice trigger condition, it would be great help for this We now be very preliminary stage So next time, we would like to show some presentation. Hayama-san has some glitch samples 10,000 glitch catalog by TAMA Glitch means time series date. Sanghoon made Webpage for this work but few people can access. Shall we make wiki page for this analysis? Korean group will improve. Next meeting would be 2 weeks later and start from 11:30~.