Acoustic Landmine Detection -- Progress as of March 1998

1. Original Design

In our original system, which grew out of Jonathan Wolff's senior project here at Harvard,
a piezo driver and a miniature accelerometer were located at the tips of a two-prod 
device which was pushed into the ground against the surface of a buried object.  An 
acoustical signature spectrum was obtained by driving the piezo "pill" to vibrate the 
surface of the object, and measuring the response on the accelerometer.  The output of the 
receiver was filtered, digitized, FIFO-buffered, and passed through an RS-232 serial
link to a laptop backend.  The computer parsed the time-domain data, performed thirty-two 
discrete windowed Fast Fourier Transforms, and averaged them to produce the spectrum 
of the test object.  This spectrum was then cross-correlated against templates of known 
landmines to determine the identity of the object.  In laboratory tests this system
discriminated mines from innocuous objects most of the time; it was not tested under
realistic field conditions however.

2. New Improved Plan

We talked to various deminers, and we experimented with the dual prod in our sandbox 
and in a miniature minefield that we created outside our laboratory window.  We decided 
that pushing two prods into the ground was going to create problems in too many real-
world scenarios (stony soil, minefields with tangled roots, hardened soil, etc.).  So we 
decided to redesign the "smart prod" to look and feel like the simple prods that deminers 
now rely upon, and feel comfortable with.

During the summer months we built a dozen prototypes of an improved prod.  The first 
models used a 0.25" diameter piezo driver and a Knowles accelerometer (as in the earlier 
system), attached to long coaxial tubes to keep the working diameter small.  There were 
lots of details to be worried about -- suppressing coupling between the sleeves while 
maintaining stiff coupling to the probe tips, preventing infiltration of dirt and water, 
providing means for independently spring-loading the driver and receiver against the 
object being probed, and so on.  The culmination of this effort was an elegant, but 
complicated, triaxial prod.  However, the mechanical complexity of this device made its 
use in a field situation impractical.

At this point (late summer) we decided to try something completely different:  a single 
prod system, with time-shared driver/receiver.  This meant abandoning the continuous 
noise ensonification scheme in favor of a pulsed system, so that the same piezo pill serves 
as both transmitter and receiver.  In addition, we decided to move the single piezo pill to 
the tip of the prod in order to minimize acoustic transmission effects of the prod itself.  

3. Pulsed Uniprod

The new system uses a single prod, with a small (1/8" diameter) piezo driver (shared as 
transmitter and receiver) right at the tip.  This makes for a simple mechanical system, 
easily repaired or cleaned in the field.  The entire prod assembly consists of a half-dozen 
relatively simple pieces, fed by a single coax line; the single piezo pill costs a dollar in 
quantity, replacing a $25 accelerometer in an assembly consisting of literally dozens of 
interlocking parts.

We did tests early on to see if the new system produces comparable information.  It seems 
to do at least as well, and often better -- the tip-located sensor arrangement avoids the 
relatively low frequency resonances of the long coaxial designs, and the pulsed scheme 
means that there is no driving signal during receive intervals.

As a further step in simplification we decided to take advantage of the capabilities of 
contemporary laptop computers, in particular the "sound card."  It was not an entirely 
trivial task to get it to do what we wanted, but we succeeded in getting Windows 95 to let 
us use one output channel to trigger the pulser, while the other output channel drives 
headphones for complete "hands-off" operation.  An input channel digitizes the received 
echo waveform, and the laptop software performs FFTs and post-spectral recognition.

Because the digitizing is done real-time on the laptop, there is no need for the hardware 
preprocessor we used earlier, with its A/D, FIFO buffer, and (slow) serial port.  The data 
bottleneck is gone, and so is most of the external hardware:  The only hardware other 
than the prod assembly itself is a small box containing a 500V pulse amplifier and a 
preamp whose gain is set by computer control; this hardware package is only a few cubic 
inches in size, and runs on a single 9V battery for about 400 hours.  The optional display 
on the laptop lets you see incoming time series, resultant power spectra (single or 
averaged), its guess as to the identity of the object under test, and lots of other useful 
diagnostic information.  You can even push a button on the probe handle to use the prod
as a simple microphone, listening to the "clunk" versus "tinkle" of various objects
(the human auditory recognition firmware has a big headstart on our software).

The acoustic signatures we see from buried plastic mines -- in particular the presence of
a relatively strong resonance in the 1-2 kHz range, combined with little response at
higher frequencies -- is apparently due to the elastic resonance in the prod-mine
system, rather than any intrinsic resonances of the mine itself.  Thus hard objects
(rock, metal, glass bottle) produce a much higher frequency resonance; soft objects (root,
soda can, thin plastic container) show a spectrum nearly devoid of resonant features.
(Although we were unaware of their work, a group based at the University of Alberta
exploited this principle in a prototype mine detection scheme.)

4. Current Status

We are satisfied, finally, with the hardware; we have a very clean and compact uniprod 
design, easy to make and repair.  Now our primary focus is collecting data and using it to 
refine the spectrum recognition algorithms.  To collect data, we bury our mine samples 
and innocuous test objects -- wood stumps, metal pieces, large rocks, plastic objects, 
bottles and cans, and so on -- during the current thaw (in a very mild New England 
winter), and take dozens of spectra.  We also take spectra of naturally buried stuff we find 
(tree roots, small stones, etc.).

Currently we are busy working on the algorithms, mostly heuristically -- we look at the 
spectra of different test objects, and try to see what distinguishes them best to the eye; 
then we code those discriminants, and try it on our growing library of test spectra.

The results are encouraging:  Plastic mines (we have only a PMN, a VS-MkII, and one other
plastic AP mine) are "easy," generally; the wooden box mine (PMD-6) is tough -- it behaves
like the hunk of wood that it is, and does not stand out from all the other wood in the
ground.

At the moment, we are using data collected from two sets of outdoor tests performed 
during the last two weeks; in total, this includes 36 spectra of mines and 95 spectra of 
harmless objects.  Using the current algorithms, we find for the PMN a probability of 
detection of around 97% with a false alarm rate of around 3%.  For the wood PMD-6 
mine, a similar probability of detection results in a much higher false alarm rate of around 
25%.  We are still revising the detection algorithms in hope of improving these statistics.

We are hoping to visit a government test minefield by late spring, to see if hardware and 
algorithms that work in the Physics Department's garden work also in a more realistic 
testing environment.