Feature Extraction

1
Feature Extraction
Dmitry Chirkin, LBNL
IceCube Collaboration meeting in Berkeley, March
2005
2
What is Feature Extraction

• Given an ATWD or FADC waveform, determine arrival
  times of all photons which contributed
• hit series
• FEInfo combination or leading edge, width,
  charge (or amplitude)
• Also applicable to AMANDA TWR

3
Feature Extraction last fall (DFL data)
Fitted function p0A0 exp(-(t-t0)/s0)(1-exp(-(t
-t0)2/s02))
4
New features discovered since
5
Multi-peak fit
6
Multi-peak fit (cont.)

• Fit the sum of two SPE functions to the waveform

• repeat for all SPE terms with amplitude above
  the threshold until the quality of the fit stops
  improving

7
In-Ice Fits (low PEs)
8
In-Ice Fits (high PEs)
9
IceTop Fits
10
Other feature extraction features

• other fitting functions were tried log-normal
  (by Tom McCauley) ? provides a different
  description of the rising leading edge
• the undershooting is now fitted, so higher ATWD
  channels should be used not only for saturated
  values, but also for values close to 0
• zero-suppression road grader algorithm needs be
  modified to suppress the most-repeated value
  instead of 0
• the higher ATWD channels are narrower, creating
  extra mismatch peak at the trailing edge.
• ? higher-channel peaks need to be widened before
  they are combined with channel 0

11
Other feature extraction features

• a slewing correction (shift of the leading
  edge proportional to width) may need to be made
  to the leading edge to describe electronics
  delays
• Laser DFL or flasher in-ice calibration?
• another correction proportional to high-voltage
  needs must made to describe high-voltage-dependent
  delay of the developing signal in the PMT
• ? Laser DFL calibration should be sufficient?

12
IceTray FE implementation

• FeatureExtractor is a project on glacier, a part
  of
• OFFLINE-SOFTWARE
• FATDATA
• ? example script is in the fat-reader/resources/
  directory
• ? you can control
• MaxNumHits maximum number of separate SPE
  functions to be fit, if necessary (default 20)
• through the DataOptions of the fat-reader
  select hits that only pass a certain fraction of
  SPE threshold (--thrs)
• At this time hidden in the source code
• maximum SPE waveform width reduce it to split
  up large pulses into smaller ones (default 6
  bins)
• fixed parameters for the description of
  undershooting

13
FeatureExtractor usage and dataclasses

• ATWDChannelMerger must be plugged in to produce
  the CombinedATWD waveform used by the
  FeatureExtractor
• I3DOMCalibration class was modified to
  accommodate calibration and combining of the ATWD
  channels of different size
• ? now Set methods set by ATWD bin name, 0-127
  in reversed time order, as before
• ? now Get methods get by the time-ordered ATWD
  bin number, 0-127 in correct time order ? this
  changed
• ? need not worry about this if only combined
  ATWD traces or Feature-Extracted hits are used

14
Conclusions

• possibility to fit multi-peak waveforms was a
  highly-anticipated feature, which should be
  considered a major improvement
• precision of the multi-peak fits for complicated
  waveforms is proportional to the time one is
  willing to spend on extracting features from
  waveforms from a few milliseconds for 2-3 peaks
  to a few seconds for 10 to a few dozen seconds
  for 20.
• ATWDChannelMerger and I3DOMCalibration class
  were modified to accommodate for hits with
  different ATWD-channel sizes (e.g., currently for
  in-ice 128, 32, 32)
• FeatureExtractor is a part of both
  OFFLINE-SOTWARE, and FATDATA. For the
  FeatureExtractor development the FATDATA provides
  a more versatile environment, allowing for a fast
  selection of the high- or low-PE events.

15
Road-grader zero-suppression

• Common SPE-like waveform
• pedestal is shifted down compared to the value
  expected from calibration. This is a well-known
  (by now) effect and is corrected by the
  fat-reader
• ok to use road-grader as is

16
Road-grader zero-suppression

• A large-amplitude, saturated waveform
• undershoot is not recorded by the current
  road-grader implementation, but is a part of the
  waveform features

17
Road-grader zero-suppression

• Highly-saturated muti-PE waveform
• the undershooting and small pulses on top of the
  undershot tail are all suppressed by the
  road-grader. Both amount of the undershooting and
  small pulses are features of the waveform and are
  used/reconstructed by the FeatureExtractor

18
Road-grader proposed modifications

• find the most-repeated value, and compress all
  values above and below it (no more than a
  threshold-setting away)
• this requires one pass over the incoming
  waveform and a small (256 byte) memory buffer
• the zero-suppressed value itself should be
  encoded into the compressed data
• to make word length more uniform (11 bits all
  the time), prepend the 10-bit number of the next
  zero-suppressed words with 1, and all other
  (10-bit) values with 0. This is more uniform
  (and possibly efficient) than the current
  road-grader Huffman-encoding algorithm

19
Modified road-grader compression ratio