►
From YouTube: NuFaucet [DEMO #6] (2015 Spring NuPIC Hackathon)
Description
Doing audio WAV analysis on faucet sounds to identify which faucet was turned on.
http://nupic2015spring.challengepost.com/submissions/37841-nufaucet
A
If
we
ever
wonder
we
conserve
water,
we
really
can't,
because
we
have
no
idea
how
much
water
we're
using
per
day
because
I,
don't
think
any
of
you
guys
know
if
I
wanted
to
know.
I
will
have
to
go
to
my
basement.
Take
my
meter
reading
every
day
or
have
the
way
from
my
utility
company
to
give
me
a
reading
every
three
months
and
for
new
york
city.
I
found
out
that
you
guys
paid
by
the
frontage.
A
So
so,
if
you
want
to
measure
how
much
water
you
use
you
can
put,
you
can
put
a
meter
like
you
can
put
a
flow
meter
in
there.
He
used
or
miss
the
utility
company
will
have
a
one
of
these
newer
electrical
ones,
digital
ones,
but
that
requires
them
doing
the
reading
and
the
data
is
not
available
to
you.
You
can
do
your
own
by
buying
one
of
these
from
alibaba,
but
if
you
want
to
do
that,
you
hire
him.
A
I
met
and
it's
expensive
do
that
so
and
I'm
sure
you
guys
are
aware
of
this.
This
picture
I
think
everybody
seeing
this
picture
enough
times
so
in
show
any
way
we
can
measure
those
four
faucets
without
putting
in
a
full
meter,
because
if
you
I
think
everybody's
know
what
it
sounds
like,
so
each
one
of
the
faucet
has
it
has
its
own
unique
sound.
A
So
the
question
is:
can
count
each
new
pic
to
to
detect
that
sound
and
tell
me
that
it's
that
which
false
has
been
used
eventually
want
to
know
what
the
duration
is
in,
which
user?
Because
they'll
have
their
own
unique
characteristic,
if
you
actually
so
that's
why
what's
doing
last
night
running
around
between
that
room
in
the
bathroom
taking
tons
of
recordings?
So
you
can
see
this
one.
This
is
the
left
sink.
A
The
hot
water
side
you
see
if
this
is
the
flow
it's
pretty
loud
and
you
see
a
consistent
number
and
this
is
actually
a
little
different.
This
is
like
this
is
the
F
of
T
of
fft,
so
this
called
septra
and
that's
actually
something
I
have
to
implement
eventually
to
to
reduce
the
total
amount
of
data
set.
A
The
system,
the
left
sink
of
water,
the
clicking
sound
it's
dumb
to
rattling
cap
because
it
was
a
little
loose,
but
those
are
the
stuff
that
you
will
want
to
use
to
be
able
to
identify
which
fall
so
do
this.
The
actual
sensor
is
pay,
it's
just
a
recorder
with
a
little
little
mic
and
it's
cool
down
to
the
right
sink
hot
water
line.
A
So
where
are
you
listening
through
the
feed
line
to
be
able
to
hear
which
mall
said
it
is
that
water
and
the
copper
axe
is
a
great
conduit,
and
this
one
is
the
right
sink
ko
Sai,
because
I
didn't
want
to
turn
out
the
line
with
my
kids
gets
kind
of
loud,
so
the
train.
Then
this
project
actually
is
based
off
Matt
Matt's
project,
their
new
pic
critic.
So
it
took
me
a
while
to
understand
how
to
get
it
to
work.
A
So,
basically,
what
we
found
out
was
the
initial
training
didn't
work
very
well
to
find
out.
They
needed
to
do
a
lot
of
data
to
feed
it
through,
so
they
ended
up
running
a
10
Paul
sets
and
then
had
to
look
at
ten
times
and
what
I
was
doing
these
data
generation
I
think
I
disturb
somebody
yesterday
what
I
was
sitting
there,
just
keep
hitting
the
faucet
and
the
person
was
waiting
inside.
They
want
to
come
out.
A
So
out
of
the
first
set
of
data
from
the
left
sink
hot
water,
faucet,
I,
rent
the
coal,
the
two
codes,
the
right
code,
because
that's
the
code,
the
code
one
they
didn't
want
to
hold
down
a
kept
bouncing
back.
You
hear
the
difference.
These
are
anomaly:
output
from
the
new
pic
and
there
it's
salt
that
and
it
gave
it
a
detection
saying:
that's
how
it
normally.
A
Are
you
seeing
a
non
anomaly
here,
because
new
Big
Data
saw
the
gap
because
when
the
water,
the
system
is
being
trained,
I
have
pulses
so
I
had
a
lot
of
flat
lines
between
so
detective
flatline
said
well.
That
was
a
detection
too.
So
for
the
actual
algorithm
to
work,
we'll
need
to
detect
for
that
and
then
take
that
out
of
the
equation.
Then
let's
stand
at
this
I'll.
You
know
what
needed
to
see
if
I
can
see
I
can
detect.
You
know
something
that
hasn't
seen
before,
so
it
throws
some
data
into
it.
A
A
Actually
it
came
out
clear
on
both
sides,
but
what
I
found
is
depends
on
the
actual
data
itself,
for
example,
these
guys
or
you
train
when
you
want
your
data
set
raining
on
the
Left
Cove,
the
amplitude
is
really
small,
so
judge
is
not
enough
data
for
its
system
to
learn,
and
then,
if
it
did,
is
really
short
like
this
one,
this
one
this
set
was
trained
on
these
guys.
Ten
of
those,
even
though
we
do
a
hundred,
it's
still
not
enough
data
for
new
pic
to
learn.
A
B
A
A
B
A
A
C
A
C
A
A
That's
actually
something
he
is
thinking
about
on
the
acoustic,
because
it's
has
to
do
with
this
fees,
guys
ability
to
navigate
through
here
the
higher
order.
So
if
you
see
these
guys
are
going
up
and
down
it's
kind
of
like
that,
but
this
feature
ability
to
follow
the
feature.
These
are
the
subtle
nuance
in
sort
of
acoustic
when
you
speak
in
boughs
and
then
seem
exactly
use
up
for
additional
features
directions.
I'm.