[TriEmbed] Name these algorithms

[Justis Peters]

Thinking on how to make this robust with noisy instruments, I was
considering pointing you Kalman filters. Searching on topics about that and
anomaly detection yielded many promising results. This one caught my eye
(Liu et al 2004, "On-line outlier detection and data cleaning"):
http://www.ualberta.ca/dept/chemeng/personalweb/slshah/files/on_line_outlier_det.pdf

You might investigate the Kalman filter by itself, too. It's usually used
to smooth a noisy signal, but there's valuable information in the distance
between its prediction and the raw sensor value.

Kind regards,
Justis

On Sun, Oct 25, 2015 at 2:38 PM, Justis Peters <justis.peters at gmail.com>
wrote:
>
> Pete, the examples you give are all univariate timeseries. I think the
solution will be really easy, but I want to better understand your goals.
Is it enough to use a Z-score against a rolling window? You seem to be
optimizing for something operational here, since you're thinking in terms
of interrupts. What are you trying to optimize?
>
> Kind regards,
> Justis
>
> On Sat, Oct 24, 2015 at 5:49 PM, Pete Soper via TriEmbed <
triembed at triembed.org> wrote:
>>
>> Say I have a sensor for which a measured baseline signal can be used as
a steady state but some predefined deviation greater than a given amount
can can be defined "interesting" by virtue of an interrupt. With the
interrupt can come a sequence of samples (either polled or by reusing the
same or different deviation for an interrupt) before the sensor reading
changes are declared ended, returning to the old or a new normal baseline
level, and the process is set up to be repeated. Meanwhile, between
interrupts and sample capturing tasks, a set of samples is compared with
previously defined "signatures" to confirm that a set of samples is
"complete" and a <something> has or has not been detected. I realize that
in many cases there has to be a "moving window" to do with matching up
samples and signatures: I'm just
>>
>> I'm looking for the math taxonomy to do with this for making filters and
pattern matchers, but also practical examples that express specific
calculations for use cases.
>>
>> Example: A photodetector w a programmable comparator that triggers an
mcu interrupt when a light level is above or below threshold, and
arrangements for detecting these "signatures":
>>  . The moon obscured momentarily by a cloud
>>  . A passing car's headlights
>>  . Dawn
>>  . A cigarette lighter lighting a cigarette
>>  . A cigarette lighter used in "music concert style"
>>  . Lightning
>>
>> I guess the ideal would be a few Wikipedia references to help me begin
my education, but some practical examples that allow correlating the math
with code that implements the math would be much appreciated.
>>
>> Thanks,
>> Pete
>>
>>
>>
>>
>> _______________________________________________
>> Triangle, NC Embedded Computing mailing list
>> TriEmbed at triembed.org
>> http://mail.triembed.org/mailman/listinfo/triembed_triembed.org
>> TriEmbed web site: http://TriEmbed.org
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.triembed.org/pipermail/triembed_triembed.org/attachments/20151025/d52b30e1/attachment.htm>