PDA

View Full Version : RE: time-series averaging



Ton Van Den Bogert
02-05-1990, 02:49 AM
John Scholz writes:

>I would like to know if anyonehas an algorithm or source code for a
>program that they would be willing to share which averages two or more
>time series collected under the same experimental conditions: e.g.,
>several gait cycles or reaches to a target under constrained conditions.
>The program would have to run on an IBM-386 compatible computer.

I don't have a ready-to-run program, but I could send you some Fortran
code. My approach is as follows: first, a program called CUT is used
to cut a series of strides (or trials) into pieces. Each stride is
then represented by a fixed number of equidistant samples. This is
done by the subroutine EQUISEL (see listing below). If you
have a stride with length of 92 samples, and you want to create a
record of 100 samples, you say:

call EQUISEL(data_raw, 92, data_norm, 100)

Where data_raw is an array with the raw samples. The result (time-
normalized data) is stored in the array data_norm.

Output of CUT is a file with all strides stored as a fixed-length
record. These records can then be plotted for visual detection of
abnormal strides. After that, a program AVFIL is used to create an
average of all `normal' strides from this file. This program also
calculates the variation between strides. This variation is
quantified in three different ways:

1. A time-average of the standard deviation.
2. The CV (coefficient of variation), see paper by D.A. Winter, Human
Movement Science 3:51-76(1984).
3. The VR (variance ratio), see paper by M.P. Kadaba, J. Orthopaed. Res.
3:350-359(1985).

I tend to prefer the VR as a measure of variability, as it is not
affected by a DC shift of the signal. Fortran source for AVFIL
is available.

John Wann made a good point about suitable
criteria for detecting begin and end of trials. In gait studies
however, the instant of foot impact is sufficiently well-defined to
avoid such difficulties. In my work on equine gait, I routinely
use an accelerometer on the hoof, which provides an excellent timing
trigger.

Good luck,

Ton van den Bogert
University of Utrecht, Netherlands.

--------------------------------------------------------
subroutine equisel(x,n,xsel,nsel)
c
c select nsel equidistant points from n points of array x
c by linear interpolation
c mapping: 1 --> 1
c n+1 --> nsel+1
c
c Input: x(n)........Raw data
c n...........Number of samples in raw data
c nsel........Number of samples you want in time-normalized data
c Output: xsel(n).....Time-normalized data
c
real xsel(nsel),x(n)
double precision step,frac
c
step = 1d0*n/nsel
iptr = 1
frac = 0d0
do j=1,nsel
xsel(j) = (1d0-frac)*x(iptr) + frac*x(iptr+1)
frac = frac + step
do while (frac .gt. 1d0)
frac = frac - 1d0
iptr = iptr + 1
end do
end do
return
end