ABSTRACT
This thesis approaches the development of distributed detection algorithms in
monosensor systems. Traditionally, the distributed detection has been applied to
the decision fusion taken by space dispersed sensors. Nevertheless, it can be useful
for the implementation of detectors that try to take advantage of the different types
of information on the existing differences among the two hypotheses for taking
decision. The objective is to substitute a unique complex detector, fairly
unapproachable, for a combination of simpler detectors, to be implemented from
the optimal decision theory.
On the contrary of a general approach, the thesis has been focused in two
applications especially interesting for the research group, the detection of forest
fires by means of infrared signals and the detection of echoes in granular noise, in
non destructive evaluation by ultrasounds.
For the detection of forest fires, a detection scheme that fuses two types of
detectors that we have denominated persistence detector and growth detector has
been proposed. The underlying idea is to try to obtain the same information used by
the watchman supposed to be replaced. The persistence detector is based on the
decision fusion of small consecutive observing intervals. In each interval an adapted
filter detector in subspace is implemented which tries to exploit the time
persistence behaviour of the fire in short term; opposed to other impulsive effects
which also can produce sporadic increments of temperature in the surveillance cell
(cars, people, sun,…). On the other hand the growth detector tries to take
advantage, by an adapted filter to lineal growth, the expected growth in
temperature, in a middle and long term, that should produce an uncontrolled fire,
compared to controlled fires that could trigger the system wrongly. The outlined
scheme allows adjusting the probability of false alarm, obtaining good achievement
of the detection probability, as we have demonstrated with simulated and real
signals.
Regarding with the detection of ultrasonic echoes in granular noise, we have
developed a scheme that combines the underlying philosophy of the called “splitspectrum”
with the distributed detection. The “split-spectrum” techniques exploit
implicitly the frequency diversity, comparing the answers of a signal echo at
different frequencies with the answers of the granular noise. The first ones should
be much more stable and the seconds much more irregular. The proposed scheme
consists on placing a detector “tuned” to each frequency band and then to fuse all
the detections. The tuning in each detector is carried out by a filter adapted to a
“band-pass” subspace. We have developed the whole analysis required to adjust
the wanted false alarm probability, guaranteeing the detection probability
maximization. Results from simulated signals and experiments with laboratory
specimens of aluminium pieces have been obtained. In addition, we have applied
the techniques proposed here, in the determination of the interface profiles
between layers corresponding to the “Cúpula de la Basílica de la Virgen de los
Desamparados de Valencia”, as a support in their restoration.