15.08.2013 Views

Journal of Accident Investigation

Journal of Accident Investigation

Journal of Accident Investigation

SHOW MORE
SHOW LESS

Create successful ePaper yourself

Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.

JOSEPH GREGOR<br />

USING CURVE-FITTING TECHNIQUES TO<br />

OBTAIN A FINAL TRANSFORM<br />

Figure 8. Overlay <strong>of</strong> FDR(n) [designated by x’s] with CVR(n) [designated by o’s] with<br />

the x-axis expanded to show the goodness <strong>of</strong> fit between the two data sets.<br />

The exact form <strong>of</strong> the transform required to convert from<br />

elapsed FDR time to elapsed CVR time may be found most<br />

generally by performing a simple curve fit on the time-aligned<br />

FDR and CVR data. If we subtract the value for C in Eq. (1) –<br />

obtained using the cross-correlation function – the times found<br />

for each corresponding on-key/<strong>of</strong>f-key event should be identical<br />

to within experimental measurement error. The magnitude<br />

<strong>of</strong> this error is primarily driven by the 1 Sa/s sample rate 7 <strong>of</strong><br />

the FDR on-key/<strong>of</strong>f-key data. The goodness <strong>of</strong> the resulting<br />

transform may be seen most easily by producing a scatter plot <strong>of</strong><br />

the data. Figure 9 shows such a plot for data corresponding to<br />

the simplest case, where the time-base operates at an identical,<br />

constant rate in both units.<br />

Each data point in this figure corresponds to a unique<br />

on-key/<strong>of</strong>f-key event, with elapsed FDR time plotted on the<br />

x-axis and elapsed CVR time plotted on the y-axis. The solid<br />

line corresponds to a minimum least squares fit to the data using<br />

a linear transform corresponding to Eq. (2). The results <strong>of</strong> this<br />

curve fit yield b = 1.0001 and C = 0.6 7, indicating that the<br />

timebases in both the CVR and the FDR were operating at the<br />

same rate to within a small fraction <strong>of</strong> a percent. Since the<br />

7 Several other sources <strong>of</strong> error may apply in setting the error bars for this<br />

calculation, but the sampling error is considered by far the largest source.<br />

data are normalized prior to performing the curve fit, we find<br />

C ≈ 0 ±1 s – well within the known error bar <strong>of</strong> ±1 Sa. Also<br />

plotted in this figure is a variable, called Delta, representing the<br />

difference between the elapsed CVR time as calculated using<br />

the resulting linear curve fit, and the corresponding elapsed time<br />

actually measured and plotted on the graph. This comparison<br />

is performed for each on-key/<strong>of</strong>f-key event to quantify the error<br />

in calculating elapsed CVR time from the elapsed FDR time<br />

using the calculated transform. The resulting error should fall<br />

within the ±1 s error bar established by the 1 Sa/s sample rate;<br />

the exact distribution is dependent on the relative phase <strong>of</strong> the<br />

actual on-key/<strong>of</strong>f-key event, the FDR sample acquisition time,<br />

and the CVR sample acquisition time [since the CVR data<br />

were also sampled at a 1 Hz rate to obtain a vector CVR(n)<br />

appropriate for correlating with FDR(n)].<br />

The situation highlighted in figure 10 is typical for a properly<br />

functioning solid-state CVR and FDR. Each unit in this case<br />

is operating from a precisely controlled digital time-base. In<br />

this case, the transform required to move from elapsed FDR<br />

time to elapsed CVR time should take the form <strong>of</strong> Eq. (1), with<br />

C given by the cross-correlation function. The curve fit in<br />

figure 9 serves as a validation that the data in FDR(n) and<br />

CVR(n) were acquired accurately. Any problem data would<br />

show up in the figure as an outlier, signalling the need for closer<br />

scrutiny and possibly a re-run <strong>of</strong> the algorithm once the problem<br />

with the data has been resolved.<br />

0 NTSB JOURNAL OF ACCIDENT INVESTIGATION, SPRING 2006; VOLUME 2, ISSUE 1

Hooray! Your file is uploaded and ready to be published.

Saved successfully!

Ooh no, something went wrong!