Following the presentation by David Paneque at the previous CAL calibration meeting,  and after discussion within CAL group, we recommend the list of modifications

  • slide 2 - please read again the recommendations of Eric Grove posted at the CAL monitoring page
  • slide 6
    • for LEX8/HEX8 - change left boundary of fitting function to -2.5 sigma
    • keep both methods (truncated average and fit) and set alarm on when the difference between two RMS valus is bigger than some limit (TBD)
  • slide 7 - these plots are useless, to be removed; keep only the difference between pedestal for current run and the standard pedestal from database (used for data processing)
  • slide 8 - add the list of outliers (RMS > 6.5). Trend the RMS for outliers and set an alarm on the unmebr of outliers
  • silde 9 - chisq  is difficult to use as it is too sensitive to statistics and model quality. It's better to use a difference between the fit sigma and trancated RMS
  • slide 10 - not useful, becasue not sensitive to real changes
  • slide 12 - useful, but would be better to express as a fractional difference (trunc - fit)/fit
  • slide 13 - strange thing (in HEX8 and HEX1): truncated average has smaller RMS than fit - couldn't be correct
    • verify and possibly modify truncation limits
  • slide 15 - keep zoomed (right) plot with Y axis in log scale
  • slide 16
    • we seem to understand the reason for  the difference  between fit and  truncated  RMS
      • it is becasue of shaped readout noise
      • to decrease the effect significantly, let's apply the cut GemDeltaEventTime*0.05 > 100 us (may be more - depends on the event rate)
      • also useful - to exclude periodic trigger events following the 4 range readout
    • it is proposed to  plot  pedestal histograms  for  different bins in McIlwainL parameter (geomagnetic latitude)
  •  slide 17 - increase the cut upto >500 LEX1 units for RPp and RMm histograms.
  • slide 29 - plot LAC histogram in ADC scale (with binwidth =1 adc unit), connverted to energy using fixed coefficient - to avoid unphysical fluctuations.


  • No labels


  1. Comments on the points address by Sasha:

    Slide 7: Those plots are not reviewed by the shifters. Those are plots that ONLY experts will inspect whenever they consider necessary. I think we do not need to remove them; they might be useful if we want to dig into some particular problem. They do not harm and do not occupy space (effectively) since it is one plot (for each quantity) for all channels together.

    Setting alarms to those values is a different story. Perhaps we can put only very conservative values (50 ADC or 2000 ADC...) that will be violated only when we have real troubles (with that particualr channel)

    Slide 9: Chi2 give us an evaluation of the fit procedure (in the used range). If the Chi2/NDF is high, the the fit parameters are not reliable. I think this is a useful info. At some point we have to define a criteria to accept or do not accept the fit results.

    Slide 13: It is not that strange. For a perfectly gaussian distribution the truncated average RMS will be ALWAYS smaller than the Sigma of the fit, since I am narrowing the distribution by "truncating" the tails of the data. The truncation is currntly only 2%, so the difference should be small; but surely will be there. Besides that, the bins are very coarse for HEX1, which means that the fit might be taking an "artificially" wider values. The truncation is done with the unbinned data. This might be another reason for having different reasons.

    In addition to those "naturally" and surely existing reasons, there might be other technical aspects related to the fit procedure. Luca could comment further on that.

    Slide 16: The cut GemDeltaEvtTime >100 microseconds is already implemented in the current monitoring code. The cut on evts following a 4 range readout event is not implemented. Such implementation is quite work demanding, since it does not fit in the current framework where things are done on the "current event". On the other hand, we have not got convincing arguments of the impact of implementing such cut.

    Pedestal histograms (or any other quantity) can be easily correlated with McIlwainL parameter (or ridigity cutoff) using the web applications. I actually already did that exercise and saw a very clear correlation. Pedestals DO CHANGE with the Rigidity cutoff (i.e. with the event rate). See presentation I gave on July 15th on the analysis meeting:

    Slide 17: Already implemented in the current data monitoring code.

    Slide 19: The distributions in ADC counts are already available in the current monitoring code. The distributions in MeV (with the "spikes") were kept. A simple rebinning would remove those spikes. In order to find the LAC thresholds, I would suggest doing it on ADC counts, and then (if necessary) convert them to MeV using the proper Mev/ADC conversion factors for the single channels.

  2. I want to comment David's statment about slide 9 (chisq distributions) - if you have million events in the histograms the chisq/NDF will always be bad becasue real distribution could be systematically slightly defferent from gaussian. At high statistics (very small statistical errors) such  a small difference become statistically significant. But parameters of the gaussian are still reliable, as before, at least at acceptable level. So, the bad chisq just means that distribution isn't a perfect gaussian, but parameters are still meaningful.