Option One Title Here
ANLY-601
Advanced Pattern Recognition
Spring 2018
L5b — Example sequential test
Detection of Marine Sensor Biofouling
Antonio Baptista, Cynthia Archer, Haiming Zheng
2
Conductivity Sensor
Biofouling
CT1448, 9/28/01
CT1459, 9/30/02
CT1449, 8/28/01
3
Conductivity Sensor
Biofouling
CT1448, 9/28/01
CT1459, 9/30/02
CT1449, 8/28/01
4
Biofouling
Detection
• Detect onset of biofouling within several diurnal cycles
• Challenges
– Variability of fouling signature
– Very few examples of fouling onset (many days of
lost conductivity data, but few episodes) – can’t use
clean/fouled discriminators.
– Distinguish natural variation from sensor
degradation
5
Detection Algorithm
,
,
( | )
( ) ln
( | )
fouled
Now
n n
n Now n n
clean
p x T fouled
h Now
p x T clean
Sequential likelihood ratio test
Now is current time
Now- is start of fouling event (unknown)
xn is salinity at time n
Tn is vector of local water temperatures at time n
is detection threshold (set by specifying false alarm rate)
p(x|clean) is Gaussian with mean E[s|T] dependent on observed tidal variation of
local temperature (i.e. predict salinity from temperature) and variance var(s|T)
p(x|fouled) is Gaussian with mean decreasing linearly (slope m).
6
Example of On-Line Detector Signals
Onset labeled by human7
Impact and Further
Development
• Initial detectors placed on-line in spring of 2001.
• Detectors eventually placed at all observing sites.
• Data Preservation
– Prior to the deployment (four years of data from 1997
through the summer of 2001), CORIE salinity sensors
suffered a 68% data loss due to bio-fouling.
– Post-deployment (spring/summer 2001 through February
2003), data loss due to bio-fouling dropped to 35%. This
includes delays in responding to the event detection.
– If all sensors were attended to immediately following a
detected event, the data loss would have dropped to
17%.
– DETECTORS CUT CORIE BIOFOULING-INDUCED
DATA LOSS IN HALF
8
• One false alarm due to precipitation; one
unexplained false alarm.
• Seasonal changes in river/ocean
temperature profile require different
predictive models at different times of
year.
– We developed mixture models for prediction
that automatically adjust to current
temperature profile conditions in the river-
estuary-ocean system.
Impact and Further
Development
9
References
Leen, T.K.; Archer, C; Baptista, A. Parameterized Novelty Detector for
Environmental Sensor Monitoring. Advances in Neural Information
Processing Systems 16, Thrun, Saul, and Scholkopf (eds.), The Mit
Press, 2004.
Archer, C; Baptista, A; Leen, T.K. Fault detection for salinity sensors in
the Columbia River Estuary. Water Resources Research, 39, 1060,
2003.
10