2. Measurement Uncertainties and Inconsistencies Dr. Richard Young Optronic Laboratories, Inc.

3. Introduction  The concept of accuracy is generally understood.  “…an accuracy of 1%.”  What does this mean? 99% inaccurate?99% inaccurate?

4. Introduction  The confusion between the concept and the numbers has lead national laboratories to abandon the term accuracy.  Except in qualitative terms e.g. high accuracy.  The term now used is uncertainty.  “…an uncertainty of 1%.”

5. Introduction  Sometimes…  Users do not know the uncertainty of their results.  They interpret any variations as inconsistencies.

6. Uncertainty vs. Inconsistency  Laboratories give different values, but the difference is within their combined uncertainties…  Pure chance.  Laboratories give different values, and the difference is outside their combined uncertainties…  Inconsistency.

7. What is uncertainty?  “…an uncertainty of 1%.”  But is 1% the maximum, average or typical variation users can expect?  Uncertainty is a statistical quantity based on the average and standard deviation of data.

8. Statistics “There are three types of lies: lies, damned lies and statistics. ” -attributed to Benjamin Disraeli “The difference between statistics and experience is time.” -Richard Young “The difference between statistics and experience is time.” -Richard Young Statistics uses past experience to predict likely future events.

9. Statistics  We toss a coin:  It is equally likely to be heads or tails.  We toss two coins at the same time:  There are 4 possible outcomes: Head + Head Head + Tail Tail + Head Tail + Tail These 2 are the same and hence twice as likely to happen as the others.

10. Statistics  Now let us throw 10 coins.  There are 1024 possibilities (2 10 ).  What if we threw them 1024 times, and counted each time a certain number of heads resulted…

11. Statistics  Although the outcome of each toss is random… ...not every result is equally likely.  If we divide the number of occurrences by the total number of throws…  We get probability.

12. Statistics  Here is the same plot, but shown as probability.  Probability is just a number that describes the likelihood between:  0 = never happens  1 = always happens

13. Statistics  Gauss described a formula that predicted the shape of any distribution of random events.  Shown in red  It uses just 2 values:  The average  The standard deviation

14. Statistics  Now throw 100 coins… We have an average = 50 And a standard deviation = 5 And the familiar bell-shaped distribution. The Gaussian curve fits exactly.

15. Confidence  Now throw 100 coins… Since the total probability must =1, the standard deviation marks off certain probabilities.

16. Confidence  Now throw 100 coins… Since the total probability must =1, the standard deviation marks off certain probabilities. About 67% of all results lie within  1 standard deviation. “I am 67% confident that a new throw will give between 45 and 55 heads.”

17. Confidence  Now throw 100 coins… Since the total probability must =1, the standard deviation marks off certain probabilities. About 95% of all results lie within  2 standard deviations. “I am 95% confident that a new throw will give between 40 and 60 heads.”

18. Real Data  Real data, such as the result of a measurement, is also characterized by an average and standard deviation.  To determine these values, we must make measurements.

19. Real Data  NVIS radiance measurements are unusual.  The signal levels at longer wavelengths can be very low – close to the dark level of the system.  The signal levels at longer wavelengths dominate the NVIS radiance result.  The uncertainty in results close to the dark level can be dominated by PMT noise.  Therefore: Variations in NVIS results can be dominated by PMT noise.

20. Real Data  The net signal from the PMT is used to calculate the spectral radiance.  Dark current, which is subtracted from each current reading during a scan, contains PMT noise.  Scans at low signals contain PMT noise.

21. Real Data  PMT noise present in each of these current readings does not have the same effect on results:  A high or low dark reading will raise or lower ALL points.  Current readings during scans contain highs and lows that cancel out to some degree.

22. Real Data Excel: “= average()”  2E-12 Excel: “= stdev()”  1E-13

23. Real Data

26. Calculations  We can describe the effects of noise on class A NVIS radiance mathematically:   s is the standard deviation of the noise  C( ) is the calibration factors  G A ( ) is the relative response of class A NVIS Dark subtraction Signal averaging

27. Calculations  A similar equation, but using NVIS class B response instead of class A, can give the standard deviation in NVISb radiance.  The standard deviations should be scaled to the luminance to give the expected variations in scaled NVIS radiance.

28. Calculations  Noise can be reduced by multiple measurements.  If we generalize the equation to include multiple dark readings (N D ) and scans (S): Brain overload

29. Spreadsheet Moving on to the benefits… Introducing The Spreadsheet