PaperlayerView exact revision on arXiv

Source-linked AI summary

Novel technique for monitoring the performance of the LAT instrument on board the GLAST satellite

D. Paneque, A. Borgland, A. Bovier, E. Bloom, Y. Edmonds, S. Funk, G. Godfrey, R. Rando, L. Wai, P. Wang

arXiv:0706.1544v1astro-ph

TL;DR

Monitoring LAT performance requires a way to detect differences between reference and newly acquired data and identify their likely parameters. The paper applies Random Forest classification using classification error and Z-scores to pre-launch event classes, clearly separating faulty data and locating the affected tracker-hit variable. The method is presented as a fast, potentially useful on-orbit monitoring tool, but deployment depends on tuning the reference data, variables, and filters.

  • Problem

    LAT performance must be monitored accurately so potential problems can be corrected promptly despite the instrument’s many detector channels and parameters.

    Source: METHODOLOGY: USAGE OF CLASSIFICATION… · p. 1

  • Method

    Random Forest classification compares reference and newly acquired data, using classification error to measure differences and Z-scores to identify discriminating parameters.

    Source: METHODOLOGY: USAGE OF CLASSIFICATION… · p. 1

  • Results

    The method clearly separates faulty A-B2 data from reference data, while identifying Tkr1Hits with a Z-score of 41; 100 trees provide good separation in this example.

    Source: DATA SETS · p. 2

  • Takeaways & Limitations

    Random Forest is presented as a fast and efficient potential tool for monitoring LAT performance during on-orbit operations.

    Source: CONCLUSIONS · p. 2

    On-orbit success depends on correctly selecting the reference data set, variables, and filters, which require tuning before and during early space operations.

    Source: CONCLUSIONS · p. 2

Abstract

from arXiv · show

The Gamma-ray Large Area Space Telescope (GLAST) is an observatory designed to perform gamma-ray astronomy in the energy range 20 MeV to 300 GeV, with supporting measurements for gamma-ray bursts from 10 keV to 25 MeV. GLAST will be launched at the end of 2007, opening a new and important window on a wide variety of high energy astrophysical phenomena . The main instrument of GLAST is the Large Area Telescope (LAT), which provides break-through high-energy measurements using techniques typically used in particle detectors for collider experiments. The LAT consists of 16 identical towers in a four-by-four grid, each one containing a pair conversion tracker and a hodoscopic crystal calorimeter, all covered by a segmented plastic scintillator anti-coincidence shield. The scientific return of the instrument depends very much on how accurately we know its performance, and how well we can monitor it and correct potential problems promptly. We report on a novel technique that we are developing to help in the characterization and monitoring of LAT by using the power of classification trees to pinpoint in a short time potential problems in the recorded data. The same technique could also be used to evaluate the effect on the overall LAT performance produced by potential instrumental problems.

METHODOLOGY: USAGE OF CLASSIFICATION TREES TO COMPARE 2 DATA

The method compares reference and newly acquired LAT data in a selected parameter space using Random Forest classification. Classification error measures the overall difference, while Z-scores identify parameters responsible for separation.

METHODOLOGY: USAGE OF CLASSIFICATION TREES TO COMPARE 2 DATA

  • Random Forest compares reference and newly acquired data sets represented in an N-dimensional space of selected parameters.The approach complements monitoring based on parameter values, correlations, and time evolution.Source: METHODOLOGY: USAGE OF CLASSIFICATION… · p. 1
  • Classification error quantifies the magnitude of data-set differences, with approximately 50% indicating no separability and lower values indicating separability.The error is estimated from Out-Of-Bag events during forest growth.Source: METHODOLOGY: USAGE OF CLASSIFICATION… · p. 1
  • Z-scores pinpoint parameters driving differences by estimating each variable’s statistical significance in distinguishing the two event classes.A high Z-score, such as > 5, implies a large statistically significant difference in that variable.Source: METHODOLOGY: USAGE OF CLASSIFICATION… · p. 1
  • Figure 1 compares classification errors for A-B1 and A-B2 and shows Tkr1Hits distributions for event classes A, B1, and B2.The left panel concerns classification error, while the right panel concerns the Tkr1Hits distribution.Source: Figure 1 · p. 2

DATA SETS

Pre-launch cosmic-ray data define a correctly operating reference class and two evaluation classes, one correctly operating and one containing missing tracker information. Random Forest tests distinguish the faulty class and identify the affected parameter.

DATA SETS

  • A-B2 data contain missing information from half layer 0 of tracker tower 10, whereas A and B1 represent supposedly correct LAT operation.The test uses cosmic-ray data, mostly muons, processed with standard LAT event reconstruction software.Source: DATA SETS · p. 2
  • A-B1 classification shows no effective separation from A at approximately 50% error, while A-B2 is clearly separable, indicating differences between those classes.The comparison uses separate forests for A-B1 and A-B2.Source: DATA SETS · p. 2
  • 100 trees are sufficient for good separation in this example, allowing the forest-growing step to run 10 times faster.Each forest otherwise used 10,000 events, 1,000 trees, 80 variables, and 4 variables per node.Source: DATA SETS · p. 2
  • The highest A-B2 Z-score is 41 for Tkr1Hits, whose distribution shows more odd hit-plane counts in B2 than in A or B1.The excess is attributed to missing information from plane 0 in some events.Source: DATA SETS · p. 2

CONCLUSIONS

The authors conclude that Random Forest is a fast and efficient potential tool for monitoring LAT performance during on-orbit operations. Practical deployment requires tuning the reference data, variables, and filters.

CONCLUSIONS

  • Random Forest can be a useful tool for monitoring LAT performance during on-orbit operations, based on a fast and efficient pre-launch test.Applying the method to low-level data could improve hardware-problem discovery but would require more computing power.Source: CONCLUSIONS · p. 2
  • On-orbit application is not straightforward because success depends on selecting an appropriate reference data set and suitable variables and filters.These choices were expected to be tuned before launch and during the first months of space operation.Source: CONCLUSIONS · p. 2
Loading 0706.1544v1…