thanks + how to detect a persistent signal anomaly

[gst]

(i'm posting here because i could adopt both an hardware or software
solution (having a fixed point dsp available); _if you think i'm
off-topic_ and that the solution lies on the software side, i can
re-post the question on comp.dsp, virtually "closing" the thread
here...)


hello

i would like to thank you for the answers to this thread
http://groups.google.com/group/sci....51e91dfcff1/399783a22eb72bb4#399783a22eb72bb4

which have been *extremely* useful and effective both in understanding
my problem better and as a solution; after some exploration we chose a
specific gaussian filter which - for the kind of application i'm
working on - behaves "spectacularly well" ;-)



now to _my current problem_ (i'm looking for *any generic hint about
the subject*, then i will google about what you suggest):

i would like to recognize an incoming persistent anomaly of the signal
(for example "sudden offset increase": what should be "zero" suddenly
becomes +2 because of sudden voltage skew and stays like that for a
reasonably long amount of time (not a "spike"), or sudden "white noise"
because a wire gets loose, or....)

i would like to do that "as fast as possible" (where "fast" means "with
the shortest latency/delay possible with respect to the appearance of
the anomaly on the signal")


does exist a standard way of dealing with anomalies? a "toolset of
procedures" i can study to understand what i can do?


since i think that the kind of problems i can face could be related to
the sensor technologies: i'm using cheap micromachined capacitive
accelerometers. the system can tolerate the specific sensor noise but i
don't know how they can _malfunction_.


thank you again, guys!
gst

 

[Ian Stirling]

i would like to recognize an incoming persistent anomaly of the signal
(for example "sudden offset increase": what should be "zero" suddenly
becomes +2 because of sudden voltage skew and stays like that for a
reasonably long amount of time (not a "spike"), or sudden "white noise"
because a wire gets loose, or....)

i would like to do that "as fast as possible" (where "fast" means "with
the shortest latency/delay possible with respect to the appearance of
the anomaly on the signal")

What's your signal?
Google "maximal likelyhood detection".
You need to model your signal, perhaps using more than one sensor, and
model.

Now, generate a signal with differences between your model (where
'difference' is dependant on your measuring) and your signal.

For example, with a diesel engine, if your input is a vibration sensor
next to the crankshaft, and an additional sensor is a voltage divisor on
an injector, with a crank position sensor as a third, you might have
four seperate models running.
One which uses all three sensors, and one each with a sensor missing.
You flag an error on significant divergance.

 

[gst]

What's your signal?
Google "maximal likelyhood detection".
You need to model your signal, perhaps using more than one sensor, and
model.

right now i'm googling for MLE/MLD thanks

the signal:

- i have the 1 accelerometer for each axis (thus no redundancy right
now, but i could ask for more...)

- basically i have to quickly detect a mechanical shock, and
distinguish it from an anomaly.

- the shock detection must be very very quick, thus i think i cannot
see enough signal to find peculiar frequencies if they exist (and being
the shock a sort of "prolonged impulse" i don't think they are there, i
cannot spot them using fft analysis or "eyeball detection" ;-) ). right
now i distinguish it from normal behaviour using only time/amplitude
information.

- the shock is itself something very different from the
usual/statistical behaviour of the system (where everything is usually
"quiet")

- i think this could be a problem to distinguish it from a quick
voltage increase caused by an hardware problem.....