Wireless �Structural Health Monitoring
ChenyangLuCyber-PhysicalSystemsLaboratory
AmericanSocietyforCivilEngineers2017ReportCardforAmerica'sInfrastructureØ Bridges C+
q Almost four in 10 are 50 years or older. q 56,007 (9.1%) bridges were structurally deficient.
q Backlog of bridge rehabilitation needs: $123 billion.q https://youtu.be/JjN1FwzbJaY
Ø Dams DØ Levees DØ Roads DØ … …Ø America's Infrastructure GPA: D+
2!
h+p://www.infrastructurereportcard.org
SmartCivilStructuresDevelop smart structures (with monitoring and control) to prevent…
3!
Freewaya:er1989SanFranciscoEarthquakeMinneapolisBridgeCollapse
StructuralHealthMonitoringCurrentPracCce
Ø Bridges: inspected manually once every two years.Ø Costly and time consuming.
4
Highway40ClosingforBooneBridgeInspecConMondayAugust10,2009If you'reheading toSt.Charles thisweekend,Highway40 isnotyourbest opNon. Westbound 40 from Long Road in St. Louis County toRoute94inSt.CharlesCountywillbeclosed(weatherpermiSng)whileworkcrewsinspecttheDanielBooneBridgeacrosstheMissouriRiver.The roadwill closeat5:30a.m.onAugust15andwon't reopenunNlsomeNmea:er9p.m.onAugust16.
WirelessStructuralHealthMonitoringØ Detect and localize damages to structuresØ Wireless sensor networks monitor at high spatiotemporal granularities
Ø Key Challengesq Computationally intensiveq Resource constraints
q Long-term monitoring
5
ExisCng(non-CPS)Approach
Ø Centralized: stream all data to base station for processing.q Too energy-consuming for long-term monitoring
Ø Example: Golden Gate Bridge project [Kim IPSN'07].q Nearly 1 day to collect enough data.
q Lifetime of 10 weeks w/4 x 6V lantern battery.
Ø Separate designs of sensor networks (cyber) and damage detection (physical).Ø Sensor networks focus on data transport.Ø Not concerned with method for damage detection.
6
DistributedArchitecture
Ø Dilemmaq Too much sensor data to stream to the base station
q Damage detection algorithms are too complex to run entirely on sensors
➪ Distributed Architectureq Perform part of computation on sensor nodesq Send (smaller) intermediate results to base station
q Complete computation at base station
7!
Raw
DataParNalResults
Cyber-PhysicalCo-design
Ø Employ damage detection approach amenable for distributed implementation in sensor networks.
Ø Optimally map damage detection algorithm onto distributed architecture.
8
Raw
DataParNalResults
OurSoluCon
q Physical: Damage Localization Assurance Criterion (DLAC) [Messina96]
q Identify structure’s natural frequencies based on vibration data.• “Signature” of structure’s health
q “Match” natural frequencies to structural models with damages.
q Cyber: optimally partition data flow between sensors and base station.q Minimize energy consumption
q Subject to resource constraints
9
(1)FFT
(2)PowerSpectrum
(3)CurveFiSng
(4)DLAC
DIntegers
HealthyModel DamagedLocaNon
2DFloats
DFloats
PFloats
D:#ofsamplesP:#ofnaturalfreq.(D»P)
DataFlowAnalysis DLACAlgorithm
10
(3a)CoefficientExtracNon
(3b)EquaNonSolving
5*PFloats
DataFlowAnalysis DLACAlgorithm
11
(4)DLAC
(1)FFT
(2)PowerSpectrum
(3)CurveFiSng
HealthyModel DamagedLocaNon
8192bytes
4096bytes
D:2048P:5Integer:2bytesFloat:4bytes
(3a)CoefficientExtracNon
(3b)EquaNonSolving
100bytes
20bytes
EffecNvecompressionraNoof204:1
4096bytes
ImplementaCon
Ø Platform: Imote2 + ITS400 sensor boardq 13 – 416 MHz XScale CPUq 32 MB ROM, 32 MB SDRAM
q CC2420 802.15.4-compliant radioq 3-axis accelerometer on sensor board
Ø Data collection and processing application written with TinyOS 1.1q 243 KB ROM, 71 KB RAM
12!
EvaluaCon:Truss
Ø 5.6m steel truss structure at UIUCØ 14 0.4m long bays, on 4 rigid supportsØ 11 Imote2s attached to frontal pane
13!
Wireless SensorTruss Frontal Panel
Fig 12. DLAC results for truss bay # 3
6.0 CONCLUSIONS
In this study a successful demonstration for an in-situ experimental validation of a
correlation-based decentralized damage detection technique using a wireless sensor network has
been performed. Structural damage was detected with sufficiently high correlation percentage in
two experimental structures independently of the damage hypothesis used in the sensitivity
matrix. On-board processing iMote2 capacities were exploited to reduce communication load
and make the application scalable within a wireless sensor network.
7.0 ACKNOWLEDGMENT S
Funding for this research is provided in part by the National Science Foundation; grant NSF
NeTS-NOSS Grant CNS-0627126, by Washington University in St. Louis. Additionally, the
authors would like to thank Prof. Bill Spencer and Shin-Ae Jang for the use of and assistance
with the experimental truss.
8.0 REFERENCES
Clayton, E.H. (2002), “Development of an Experimental Model for the Study of Infrastructure
Preservation”, Proceedings of the National Conference on Undergraduate Research,
Whitewater, Wisconsin.
Clayton, E.H., Koh, B.H., Xing, G., Fok, C.L., Dyke, S.J. and Lu, C. (2005), “Damage
Detection and Correlation-based Localization Using Wireless Mote Sensors”, Proceedings
of ’05 The 13Th
Mediterranean Conference on Control and Automation, Limassol, Cyprus.
Clayton, E.H. (2006), “ Frequency Correlation-based Structural Health Monitoring with Smart
Wireless Sensors”, Master of Science Thesis, Washington University in St. Louis.
1234567891011120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
X = 3Y = 0.868
DLAC WS #32
Truss Central Bay Position
1234567891011120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
X = 3Y = 0.864
DLAC WS #45
Truss Central Bay Position
1234567891011120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
X = 3Y = 0.871
DLAC WS #67
Truss Central Bay Position
1234567891011120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
X = 3Y = 0.873
DLAC WS #28
Truss Central Bay Position
1234567891011120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
X = 3Y = 0.825
DLAC WS #35
Truss Central Bay Position
1234567891011120
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
X = 3Y = 0.865
DLAC WS #75
Truss Central Bay Position
Damage correctly localized to third bay!
EnergyConsumpCon
0 0.05 0.1 0.15 0.2 0.25
Decentralized
Centralized
EnergyconsumpCon(mAh)
Sampling
ComputaNon
CommunicaNon
EvaluaCon
14!
EnergyConsumpCon
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5
RawDataCollecNon
FFT
PowerSpectrum
CoefficientExtracNon
EquaNonSolving
EnergyConsumpCon(mAh)
Sampling
ComputaNon
CommunicaNon
EvaluaCon
15!
MemoryConsumpCon
0 50000 100000 150000 200000 250000
RAM
ROM
Size(bytes)
RawDataCollecNon
FFT
PowerSpectrum
CoefficientExtracNon
EquaNonSolving
EvaluaCon
16!
Whatwehavelearnedsofar
Ø Cyber-physical co-design of a distributed SHM system.q Reduces energy consumption by 71%
q Implemented on iMote2 using <1% of its memory
Ø Effectively localized damage on two physical structures.
Ø Demonstrated the promise of cyber-physical co-design.
17
G.Hackmann,F.Sun,N.Castaneda,C.LuandS.Dyke,AHolisNcApproachtoDecentralizedStructuralDamageLocalizaNonUsingWirelessSensorNetworks,RTSS2008.
HierarchicalDamageLocalizaCon
Ø The DLAC method employs no collaboration among sensors à limitations in SHM capabilities.q For example, cannot detect multiple damages.
Ø New hierarchical architecture for collaborative localization.q Embed processing into a hierarchical architecture
q Send (smaller!) partial results between layers of hierarchy
q Multi-level damage localization
Ø Demonstrate the generality of cyber-physical co-design
18!
Flexibility-basedMethods
Ø Structures flex slightly when a force is applied
Ø Structural weakening => decreased stiffnessØ Flexibility acts as a “signature” of the structure’s health
Ø Two flexibility-based methods of interest for our workq Beam-like structures: Angles-Between-String-and-Horizon
flexibility method (ASHFM) [Duan, J. Structural Engineering and Mechanics 09]
q Truss-like structures: Axial Strain flexibility method (ASFM) [Yan, J. Smart Structures and Systems 09]
19!
θ!
HierarchicalArchitecture1
Ø Sensors form groups
Ø Group membersq collect raw vibration dataq à power spectrum
Ø Group leadersq collect and correlate power
spectrum from children q à modal parameters (natural
frequencies + mode shapes)
20!
BaseStaCon
GroupLeader
GroupLeader
GroupMember
GroupMember
GroupMember
GroupMember
GroupMember
HierarchicalArchitecture2Ø Base station
q collects modal parameters from group leaders
q à structural flexibilityq compared against “baseline” collected when
structure was known to be healthy
Ø Differences in flexibility à localize damage
21!
DistributedDataFlow
22!
Sensing
FFT
PowerSpectrum
CrossSpectralDensity
SingularValueDecomposiNon
2Dints
Dfloats
Group Leader!
Flexibility
Base Station!
Dmatrices
Group Member!
DfloatsPnaturalfrequencies+
modeshapes
D:#ofsamplesP:#ofnaturalfreq.
(D»P)
EnhancedDistributedDataFlow
23!
Sensing
FFT
PowerSpectrum
2Dints
Dfloats
PeakPicking
Dfloats
CrossSpectralDensity
SingularValueDecomposiNon
Group Leader!
Flexibility
Base Station!
Pmatrices
Pnaturalfrequencies+modeshapesPfloats
Group Member!
D:#ofsamplesP:#ofnaturalfreq.
(D»P)
MulC-LevelDamageLocalizaConØ Only a handful of sensors are needed to detect damageØ As more sensors are added, localization gets more preciseØ Save energy by exploiting localized nature of flexibility-based approach
24!
ImplementaCon
Ø Hardware: Imote2 + ITS400 sensor boardq 13 – 416 MHz PXA271 XScale CPUq 32 MB ROM, 32 MB SDRAM
q CC2420 802.15.4-compliant radioq 3-axis accelerometer on sensor board
Ø Software platformq TinyOS 1.1 operating systemq UIUC’s ISHM toolsuite used for sensing,
reliable communication, and time sync
25!
EvaluaCon:Truss
Ø Simulation of 5.6 m, 14-member steel truss structure at UIUC
Ø Simulated sensor data generated in MATLAB and injected into live application using “fake” sensor driverq Intact data set: no damages
q Damaged data set: three members reduced on left side of truss, four on right side
26!
EvaluaCon:Truss
Ø Level 1: nine sensors at uniform points along truss’s length
27!
0 2 4 6 80
1
2
3
4x 10
−6
Element Number
Da
ma
ge
In
dic
ato
r
Damage identified on left half
Damage identified on right half
EvaluaCon:Truss
Ø Level 2: move all nine sensors to respective halves (emulate higher density)
28!
1 2 3 4 5 6 7 8 9 101112130
1
2
3
Element Number
Da
ma
ge
In
dic
ato
r
1 2 3 4 5 6 7 8 9 101112130
1
2
3
Element Number
Da
ma
ge
In
dic
ato
r
Damage localized correctly to all seven members
EvaluaCon:Truss
Ø Codesigned architecture reduces communication latency from estimated 87s to 0.21s
Ø 78.9% of energy attributable to synchronization and sensing
Ø Compare to theoretical energy supply of 20,250 J (3x 1.5 V, 1250 mAh AAA batteries)
29!
GroupMember
SynchronizaNon 12.1JSensing 23.0JComputaNon 9.28J
CommunicaNon 0.08J
GroupLeader
SynchronizaNon 16.2J
Sensing 21.2J
ComputaNon 8.52J
CommunicaNon 0.76J
Full-ScaleTruss
30!
Image source: Zhuoxiong Sun, Purdue University!
TestResults:Full-ScaleTruss
Ø Two levels of damage localization
Ø Level 1: localized�damage to bay 9
Ø Level 2: localized�damage to element 42
31!
2 3 4 5 6 7 8 910 20 3132 420
0.2
0.4
0.6
0.8
1
1.2
1.4
1.6
1.8
2x 10−6
Truss Element Number
AS F
lexi
bilit
y D
amag
e In
dica
tor
Summary
Ø Cyber-physical co-design for wireless structural health monitoringq Distribute flexibility-based damage localization methods in a
hierarchical architectureq Multi-level search strategy only activates sensors in area of
interest; many sensors remain asleep
Ø Localize damage to a simulated truss and a real full-size truss with low energy consumption
Ø Long-term goal: a general cyber-physical co-design approach to integrated sensing and control
32!
G.Hackmann,W.Guo,G.Yan,Z.Sun,C.LuandS.Dyke,Cyber-PhysicalCodesignofDistributedStructuralHealthMonitoringwithWirelessSensorNetworks,IEEETransacNonsonParallelandDistributedSystems,25(1):63-72,January2014.
ReflecCon:TradiConalMethodologyØ Localize damages on structures using wireless sensors.
Ø Traditional: separate network and civil engineeringq Cyber: Wireless network streams all data to a base station
q Physical: Base station runs damage localization algorithm
Ø Clean separation of concern, but ineffective
q Streaming raw data consumes too much energy
CPSCo-Design
Ø Get hands dirtyØ Understand the data flow of damage localization
Ø But still employ clean abstraction and methodologyØ Optimal data flow embedding in a network
Ø Highly effectiveØ Reduces energy consumption by 71% [RTSS'08]
34!
1. Design a damage localization method suitable for distributed processing. 2. Model the data flow. 3. Optimally embed the data flow in a sensor network.