Signal processing for earthen dam measurements analysis WS 2011-11 Pyayt.pdfSignal processing for...
Transcript of Signal processing for earthen dam measurements analysis WS 2011-11 Pyayt.pdfSignal processing for...
UrbanFloodUrbanFlood
Signal processing for earthen dam
measurements analysis
2011 UrbanFlood Workshop
Amsterdam, The Netherlands, November 3, 2011
Alexander Pyayt [email protected]
Ilya Mokhov [email protected]
Natalia Melnikova [email protected]
Valeria Krzhizhanovskaya [email protected]
Alexey Kozionov [email protected]
Victoria Kusherbaeva [email protected]
Artem Ozhigin [email protected]
2A. PyaytUrbanFlood
The Netherlands and Flood
3200 km of primary dikes
14 000 km of secondary dikes
3A. PyaytUrbanFlood
The Netherlands and Flood
3200 km of primary dikes
14 000 km of secondary dikes
Overtopping
4A. PyaytUrbanFlood
Dike failures
Wilnis 2003 – drought (peat) & uplift
http://wwwen.uni.lu/research/fstc/research_unit_in_ engineering_science_rues/members/stefan_van_baars/research/dike_engineer ing
• 1735 dike failures in the
Netherlands between 1134 and 2006
• rare visual inspections
• old dikes
Ijkdijk piping
34%30%
20%
10%6%
0%
10%
20%
30%
40%
Ove
rtopp
ing
Foun
datio
n
def
ect
s
Pip
ing
and
seepa
ge
Con
duits
and
valv
es
Oth
er
* Percantage of flood events in the USA
5A. PyaytUrbanFlood
UrbanFlood decision support system workflow
6A. PyaytUrbanFlood
Abnormal Behaviour Detection Approach
Analytical redundancy
Physical redundancy (the same placement)
Physical redundancy (type of sensor)
Committee
Feature extraction
Cl1
Sen
sor
mea
sure
men
ts
Con
fiden
ce v
alue
s
-- NORMAL BEHAVIOUR-- ABNORMAL BEHAVIOUR
-- NORMAL BEHAVIOUR-- ABNORMAL BEHAVIOUR
X1
X2
-5 -4 -3 -2 -1 0 1 2 3 4 5
x 10-3
-4
-2
0
2
4
6x 10
-3
Cl2
Clk
Logical groups
Pre-processing
Feature Extraction
Cl1
Cl2
Clk
Decision support
7A. PyaytUrbanFlood
Knowledge about anomaliesReal measurements
+ Real-world data
+ Low cost of non-destructive experiments
- Absence of measurements related to real dike failures
- High cost of destructive experiments
8A. PyaytUrbanFlood
Knowledge about anomaliesPhysical modellingReal measurements
+ Real-world data
+ Low cost of non-destructive experiments
- Absence of measurements related to real dike failures
- High cost of destructive experiments
+ Cheaper than field experiments
- Model adequacy
9A. PyaytUrbanFlood
Knowledge about anomaliesPhysical modellingReal measurements
+ Real-world data
+ Low cost of non-destructive experiments
- Absence of measurements related to real dike failures
- High cost of destructive experiments
+ Cheaper than field experiments
- Model adequacy
10A. PyaytUrbanFlood
Measurements with anomalies
Normal
Abnormal
Normal
Normal
Abnormal
Failure
Nondestructive experiment,not stable dike
Destructive experiment, real dike failure
Zeeland dijkStammerdijk Ijkdijk
11A. PyaytUrbanFlood
Pre-processing
1) Wavelet denoising
2) Spectrum Singular Analysis (SSA )
3) Hodrick-Prescott filter
4) L1 trend filtering
5) Moving average1
0
1 N
i i ll
x yN
−
−=
= ∑
12 2
1 11 2
min ( ) [( ) ( )] ,T T
i i i i i ii t
y x x x x x smoothing parameterλ λ−
+ −= =
− + − − − −∑ ∑
21
1 2
min ( ) ,T T
i i i ii i
y x x x smoothing parameterλ λ−= =
− + − −∑ ∑
DWT Hard thresholding iDWTy xCoefficients
SSA1
N
nn
y c=
=∑ Take only c which corresponds to maximal eigenvalues
y – measurementsx – estimation of a signal
y x
The main idea : to apply methods, which have minimal number of adjustable parameters and require minimal information about signal
Logical groups
Pre-processing
Feature Extraction
Cl1
Cl2
Clk
Decision support
12A. PyaytUrbanFlood
Data pre-processing
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
x 104
0
20
40
60
80
time
valu
es
1.292 1.294 1.296 1.298 1.3 1.302 1.304 1.306 1.308 1.31 1.312
x 104
12
14
16
18
20
22
24
26
28
Smoothing of 'jumps'
time
valu
es
noisy signalwaveletl1hp filterSSAMAsignal
1.44 1.46 1.48 1.5 1.52 1.54 1.56
x 104
22
24
26
28
30
32
34
36
Smoothing of outliers
time
valu
es
noisy signalwaveletl1hp filterSSAMAsignal
Synthesized data
1.180.500.490.240.23RMSE
MASSAhpl1Wavelet
13A. PyaytUrbanFlood
One-side classification
--ABNORMAL BEHAVIOUR-- NORMAL BEHAVIOUR
Logical groups
Pre-processing
Feature Extraction
Cl1
Cl2
Clk
Decision support1) Hypercube
3) Neural Clouds2) Parzen window
14A. PyaytUrbanFlood
Virtual Dike
ffcstab τφσ −+= tan
fτ
fσ
- cohesion, [Pa]
- friction angle, [grad]
- shear stress, [Pa]
- effective normal stress, [Pa]
φ
Mohr-Coulomb plasticity model:
c
stab - stability factor, [Pa] (<0 in case of plastic deformations )
, where
9 m
0 m
Livedijk, Eemshaven, the Netherlands VVK1
Slide 14
VVK1 LIVEDIKEValeria Krzhizhanovskaya; 3-11-2011
15A. PyaytUrbanFlood
Macro-instability Modelling
Distribution of stability factor in the dike at different loading steps:
Beginning of plastic deformations
critical state, water is at 6.6 m above ref. level
Plastic deformations
16A. PyaytUrbanFlood
Simulated sensor data
17A. PyaytUrbanFlood
Anomaly detection in artificial data
0 500 1000 15000
1
2x 10
-6 First principal strain
0 500 1000 15001.5
2
2.5x 10
-5 X deformation
0 500 1000 15000
0.5
1Confidence value of dike normal behaviour
0 500 1000 15000
0.5
1Confidence value of dike normal behaviour
0 500 1000 1500-1
0
1
2x 10
4Stability factor
Detection of anomaly
“Local failure”
460 1123
18A. PyaytUrbanFlood
Variant of combination of the AI component and the Virtual Dike
Committee
Cl1
X1
X2
-5 -4 -3 -2 -1 0 1 2 3 4 5
x 10-3
-4
-2
0
2
4
6x 10
-3
Cl2
Clk
Decision support
19A. PyaytUrbanFlood
AI component implementation
JMS consumer
Java
Mes
sage
Ser
vice
(JM
S)
AI component
Self monitoring
XML reader
Data analysis
XML writerAnySense
Sensor cabinet
DSS, other components
JMS producer
WebDashBoard
Web browser
Dike measurements Confidence values of normal behaviour AI component state
Logical groups
Pre-processing
Feature Extraction
Cl1
Cl2
Clk
Decision support
20A. PyaytUrbanFlood
AI component as part of the UrbanFlood EWS
JMS consumer
Java
Mes
sage
Ser
vice
(JM
S)
AI component
Self monitoring
XML reader
Data analysis
XML writerAnySense
Sensor cabinet
DSS, other components
JMS producer
WebDashBoard
Web browser
Dike measurements Confidence values of normal behaviour AI component state
Joint UrbanFlood & SSG4Env International Monitoring and FloodSafety Workshop,Amsterdam, the Netherlands, Nov 2010
21A. PyaytUrbanFlood
Summary
� Current results◦ Detection of anomalies� Real dikes
� Stammerdijk (published)
� Zeelandijk (to be published)
� Livedijk-based generated data by the Virtual Dike
� Next steps◦ Ijkdijk modelling and data analysis
◦ AI component� application of parametric methods
22A. PyaytUrbanFlood
Acknowledgements
� This work is supported by the EC FP7 project UrbanFlood, grant N 248767
� Siemens LLC Corporate Technology, Russia for financial support
� Alert Solutions (particularly, Erik Peters), WaterNet (particularly, Rob van Putten) for providing data