Re: Computerised Golw Curve Deconvolution

[Orville & Dianne Cypret <ocypret@apex.net>]

TOSH USHINO wrote:
> 
> In response to Doug Minnema, Dr. Charles and John Simpson, I purchased Peakfit
> 3.18 (DOS version) from Jandel Scientific few years before they got bought out
> by SPSS Inc. I intended to use the program for Panasonic TLD glowcurve
> deconvolution. However, unlike other programs (2-D and 3-D curve fitting
> programs) we purchased from them, the program proved to be not intuitive and
> required more manipulation than I liked. For us to develop a TLD program to
> automatically deconvolute the glowcurves with a link to Peakfit 3.18 appeared
> to be too great a challenge for our limit resource (time & $$, and time=$$).
> (The program is interesting, and would like to spend more time with it but
> can't). Perhaps the new version of the program is much easier to implement. I
> actually started writing my own code for glowcurve deconvolution at home but
> gave up because whether we used Peakfit or my own code, it would have required
> our in-house professional programmers to code it and to document it... a very
> expensive process which would have driven up the cost-per-badge and would have
> made our in-house TLD processing not competitive.
> 
> I have a question for others.... I have been interested in using neural
> network software to spot abnormal glowcurves (e.g., hydrogen sulfide exposure
> on lithium borate phosphors) and TLD element response patterns. I know there
> have been some work done using neural networks. What kind of success have you
> had? Did you write your own neural network code? [I know, I know...if I try to
> implement it, it will drive up the cost-per-badge].
> 
> Tosh Ushino
> Health Physics Engineer
> San Onofre Nuclear Generating Station
> ushinot@songs.sce.com

This isn't exactly a direct reply to the deconvolution question, but it
does speak to Tosh's ending question.

Around 1991 I trained a neural network to recognize glow curve
anomalities for a Panasonic system.  At the time I was working at
Arkansas Nuclear One.  We tested the system pretty extensively offline,
i.e., we downloaded glow curves to a second computer on which the net
was installed and ran the curves through it.  As I recall, it did a good
job recognizing problem curves.  Very, very fast on a 386.  The intent
was to make the net small enough to include it in the processing
software w/o adding any overhead.  This was intended to reduce the cost
per badge in that it would do a pre-screening for anomalous curves w/o
in lieu of a tecnician doing the review; the tech would only evaluate
the curves which failed the pre-screening.  

Sorry, I don't know what the status of the application is today.  I left
shortly after the testing was complete and the head of the Dosimetry
section left a few months after I did, so I don't know if they ever did
anything with the program.  I think I used a neural net package called
Brainmaker.  Some of the parameters I used were chip ratios, peak height
and location, and ?? (been too long, but I know there were more than
these).  You might contact Tom Rolniak at the plant and see if he knows
anything about it.

I did some other things there with AI including help with skin dose
calcs and rad posting decisions.  They were using them when I left, but,
again, I don't know what they're doing today.

If you want to contact me about it, please send private email to
cypretow@ornl.gov.

Orvile Cypret