Raw hit finding for dual phase Christoph Alt July 12 th , 2017 / protoDUNEs sim & reco meeting Christoph Alt Raw hit finding for dual phase Slide 1 / 20
Overview Raw hit finding for dual phase is ready to use: 1. Explaining the algorithm and the fcl parameters 2. Plans for further improvements 3. Where to find the code and .fcl’s 4. Performance check Christoph Alt Raw hit finding for dual phase Slide 2 / 20
Explaining the algorithm Goal: develop a hit finding algorithm for raw waveforms Data preparation: remove pedestal no deconvolution no noise filtering (for now) → In this presentation: white noise with RMS = 2.4 ADC total gain of 20 (= 10 per channel) Algorithm can be divided into 5 steps: 1. Peak finding 2. Peak grouping and merging and noise peak removal 3. Fitting 4. Re-fitting 5. Create a hit Christoph Alt Raw hit finding for dual phase Slide 3 / 20
1. Peak finding 1. Find bin with highest ADC count above MinSig (8 ADC) 2. Define t ini and t fin : walk along time axis until: a bin with ADC � 0 is found or a bin followed by TicksToStopPeakFinder consecutive bins with same/higher ADC count is found (default for TicksToStopPeakFinder is 4 , which is the rise time of elec. response function) 3. Repeat step 1 and 2 for remaining region Christoph Alt Raw hit finding for dual phase Slide 4 / 20
2.1 Peak grouping Group nearby peaks (...to fit the group instead of single peaks) if no bin with ADC � 0 is found between the peaks and if distance between the peaks � GroupMaxDistance ( 5 ) (this is mainly to catch low frequency noise which can ”connect” two seperated hits) and if number of peaks in this group � MaxMultiHit ( 6 ) (fitting long multi peaks is slow, ∽ 1s for 10 peaks, growing exponentially) example below: grouped the two peaks on the left, but not the one on the right Christoph Alt Raw hit finding for dual phase Slide 5 / 20
2.2 Peak merging example below: one peak misidentified as two peaks due to noise Solution: merge two peaks of a group into one peak before fitting if DoMergePeaks = true and if ADC sum peak 2 < MergeADCSumThreshold · ADC sum peak 1 and if height peak 2 < MergeMaxADCThreshold · height peak 1 → This reduces fake noise peaks in the tails of the waveform default values for both parameters: 0.2 Christoph Alt Raw hit finding for dual phase Slide 6 / 20
2.3 Removing noise peaks Want to get rid of noise peaks before fitting (see example below) Do not fit and remove peak if the width of a single peak/group of peaks is < MinWidth ( 10 ) or if ADC sum over width of this single peak/group of peaks is < MinADCSumOverWidth ( 4.8 = 2 · noise RMS) → MinWidth shouldn’t be larger than the width of elec. response ( ≈ 20 ticks) and MinADCSumOverWidth should scale with noise RMS Christoph Alt Raw hit finding for dual phase Slide 7 / 20
3.1 Fitting: single peak Fit single peak from t ini to t fin (inclusive): t − t 0 e τ 1 fit function f ( t ) = A · t − t 0 1 + e τ 2 t 0 : seed = t max - 2 ticks, limit: seed ± 5 ticks A : seed = 1.65 · maxADC, limit: 0.1 · seed to 2 · seed Christoph Alt Raw hit finding for dual phase Slide 8 / 20
3.2 fitting: grouped peaks Perform single fit for a group of peaks (black) Each summand of the fit function corresponds to a single peak (blue) τ 1 and τ 2 are identical for all peaks, amplitude and mean vary seeds same as for single peaks, ranges for t 0 , i do not overlap t − t 0 , i e τ 1 fit function f ( t ) = i A i · � t − t 0 , i 1 + e τ 2 Christoph Alt Raw hit finding for dual phase Slide 9 / 20
4. Re-fitting: check requirements Why re-fitting? Two overlapping peaks misidentified as one: ⇓ re-fit ⇓ Do re-fit if TryNplus1Fits = true and χ 2 /NDF > Chi2NDFRetry ( 25 ) and number of peaks < 2 · MaxMultiHit (2 · 6 ) → challenge: need good seeds for re-fit! Try to add a new peak or split an existing one. Christoph Alt Raw hit finding for dual phase Slide 10 / 20
4.1 Re-fitting: Add peak 1. Sort all peaks in a group by their χ 2 /NDF, start with highest 2. Determine the bin with the highest positive derivation between signal and fit for this peak. This bin is t max , 2 of the new peak. 3. Split peak halfway between t max , 1 and t max , 2 . 4. Group peaks by default and re-fit 5. If χ 2 /NDF improved with re-fit, keep added peak and check requirements for re-fitting again (previous slide) 6. Otherwise, start over with step 2 for peak with second highest χ 2 /NDF etc. ⇓ add (and split) peak ⇓ Christoph Alt Raw hit finding for dual phase Slide 11 / 20
4.2 Re-fitting: Split peak If adding a new peak didn’t work, do the following: 1. Sort all peaks in a group by their χ 2 /NDF, start with highest 2. Split this peak in the middle. t max , 1 and t max , 2 are in the center of their respective peaks. 3. Group peaks by default and re-fit 4. If χ 2 /NDF improved with re-fit, keep added peak and check requirements for re-fitting again (previous slide) 5. Otherwise, start over with step 2 for peak with second highest χ 2 /NDF etc. ⇓ split (and add) peak ⇓ Christoph Alt Raw hit finding for dual phase Slide 12 / 20
5. Creating a hit Save fit parameters for each peak: amplitude: A from fit mean: max of fit (see plot) → causes problem for BackTracker (working on it) width: calculate FWHM (full width at half maximum) from τ 1 � save FWHM/(2 · 2 · ln ( 2 ) ) as width → corresponds to sigma of a Gaussian fit: � σ = FWHM/(2 · 2 · ln ( 2 ) ) energy: integral of fit from −∞ to + ∞ Christoph Alt Raw hit finding for dual phase Slide 13 / 20
Plans for further improvements 1. Re-fitting sometimes deliver unreasonably small hits → remove hits with an integral < threshold → only accept re-fit when chi2/NDF is e.g. 10 % better than before 2. When re-fitting, hits sometimes get split into two hits with almost the same t 0 → merge hits with e.g. ∆ t 0 < 1 tick after fitting (clustering?) 3. very long hits are split into several groups because of MaxMultiHit threshold → substract fit of first group from waveform before fitting second group → try to implement this in the event display Christoph Alt Raw hit finding for dual phase Slide 14 / 20
Where to find the code Module: larreco/HitFinder/DPRawHitFinder module.cc table: "dpraw hitfinder" in larreco/HitFinder/hitfindermodules.fcl protodune dual phase reco .fcl: dunetpc/fcl/protodunedp/reco/ rawhitfinding reco protodunedp.fcl 3x1x1 reco .fcl (not tuned): dunetpc/fcl/3x1x1dp/reco/ rawhitfinding reco 3x1x1dp.fcl For easier tuning, the .fcl parameters for ptotodune dp and 3x1x1 are in the respective reco. fcl’s. Once tuning is fnished, will move them to: srcs/dunetpc/dune/HitFinderDUNE/hitfindermodules dune.fcl Christoph Alt Raw hit finding for dual phase Slide 15 / 20
Performance of raw hit finding: data set 100 single µ − with P µ − = 500 MeV in protoDUNE dp geo isotropic distribution within 45 ◦ in both planes Top view: Side view: Christoph Alt Raw hit finding for dual phase Slide 16 / 20
Performance of raw hit finding: before re-fitting raw hit finding gauss hit finding Average χ 2 / NDF : 34 (deconvoluted) vs. 13 (raw) raw Number of hits: deconvoluted = 102 % You get these histograms when you run the raw hit finding/gauss hit finding. Useful for tuning of Chi2NDFRetry . Christoph Alt Raw hit finding for dual phase Slide 17 / 20
Performance of raw hit finding: re-fitting zoom after re-fitting (raw) Reminder: Chi2NDFRetry is set to 25 Re-fitting fails only for a few hits → will look at these and see if algorithm can be further improved Christoph Alt Raw hit finding for dual phase Slide 18 / 20
Performance of raw hit finding: tracking efficiency Efficiency definition for track reconstruction: 1. Completeness � 50 % 2. Purity � 50 % 3. 75 % � L reco L truth � 125 % more details: https: //indico.fnal.gov/getFile.py/access?contribId=57& sessionId=16&resId=0&materialId=slides&confId=10641 Results: Gauss: 97 % Raw: 100 % (standard linecluster or trajcluster, pmtrack) → check with higher statistics → will probably not be that good for tracks in drift direction Christoph Alt Raw hit finding for dual phase Slide 19 / 20
Summary and outlook Raw hit finding and fitting for dual phase is ready to use! (pushed one hour ago) Track reconstruction works well, shower reconstruction not tested yet Further improvements will only have a small impact Will tune it for 3x1x1 data next As usual: many thanks to Robert and Dorota (got many ideas from their hit fitting algorithm for ICARUS) Christoph Alt Raw hit finding for dual phase Slide 20 / 20
Recommend
More recommend
