Bridge & Structure Laboratory University of Tokyo 1 Structural health monitoring using Imote2 Tomonori Nagayama Assistant Professor University of Tokyo.

Bridge & Structure LaboratoryUniversity of Tokyo

1

Structural health monitoring using Imote2 Tomonori NagayamaAssistant Professor University of Tokyo

07/10/2009


2

Wireless sensor components -functionality- The Imote2 has promising features. But not all the

functionalities needed in SHM are provided in OS/HW.

Hardware

OS

Middleware

SHM applications

RF

CP

UM

emory

Pow

er

Sensor/

actuator

sensing

Following functionalities are provided as middleware services. Users can utilize them to assemble their own SHM applications Time Synchronization Synchronizaed sensing Reliable communication Efficient data aggregation Others

timesync data aggregation

comm/networking


3

Synchronization basics

Node synchronization Nodes exchange packet and estimate local clock offsets

Time synchronization protocols Reference Broadcast Synchronization (RBS), Timing-sync

Protocol for Sensor Network (TPSN), Flooding Time Synchronization Protocol (FTSP)

t1

t2

t3

Node1 clock

Node2 clock

Node3 clock

T2

T3 +T3

+T2

Global time


4

Time synchronization middleware Based on Flooding Time Synchronization Protocol

(FTSP)

By cascading, this synchronization works on a multihop network

Send packet

Append time stampt1

Obtain reception timet2

Global time = local time + t1-t2+t3

t3

Concept


5

Time synchronization accuracy check Timestamps of receivers are examined

Send packet

Concept Time synchronization

t3Get

globaltime

Getglobaltime

Getglobaltime

Getglobaltime

Getglobaltime

Getglobaltime

Repeat n times

Time synchronization

t3


6

Synchronized Sensing accuracy check

Synchronization accuracy

…

BeaconReply global time

Time synchronization error < 150 Time synchronization error < 150 s. Mostly < 20s. Mostly < 20ss

Difference in returned global time stamps

50 100 150 200 250 300

-80

-40

0

40

80

Repetition

Tim

e (

s)


7

t2+T

t3+T

Synchronization basics -drift-

Drift Due to difference in clock speed of each node, difference among

local times changes (almost linearly) Synchronization error accumulates as time passes after the last

synchronization unless appropriate compensation is performed.

t1

t2

t3

Node1 clock

Node2 clock

Node3 clock

T2

T3 +T3

+T2

Global time

+T3+T

+T2+T

t1+T


8

Time synchronization drift check Difference among local clocks (T)are

examined

Send packet

Time synchronization

t3

Concept

GetT2

GetT3

GetT4

Repeat n times

Get T2+T1

GetT3+T1

GetT4+T1

Get T2+T2

GetT3+T2

GetT4+T2


9

Drift estimation

…

BeaconReply offset

• is almost constant over timeis almost constant over time• Difference in clock rates can be as large as 50 Difference in clock rates can be as large as 50 s/ss/s

0 40 80 120 160-4000

0

4000

8000

Time (s ec)

Drif

t(

s)

Clock drift

T

(s

)

However, time synchronization of However, time synchronization of the nodes does not provide the nodes does not provide

synchronized sensing.synchronized sensing.

However, time synchronization of However, time synchronization of the nodes does not provide the nodes does not provide

synchronized sensing.synchronized sensing.


10

Toward synchronized sensing EVEN If a command to start sensing is

issued at the same time, the execution timing is different

Sampling timing has individual difference

node1

“Start sensing”

node2

node3

Sampling timing time

Actual start t1

t2

t3

t1 != t2 != t3


11

Two approaches for synchronized sensing

Strict HW control of sampling timing Sampling has high priority than

other tasks. No need for post processing Other tasks are delayed.

Resampling based approach Sensing starts at the approximately

same time ． Resampling based on accurate

timestamping Less requirement on HW Timestamp + Resampling + linear

interpolation -> VERY accurate synchronized sensing is realized

Strict HW control

HW control

Resample


12

Resampling basics

fsfs11

L MLL MML MLL MML MLL MML MLL MM

L MLL MML MLL MML MLL MM

fsfstargettarget

upsample filter downsample

To eliminate aliasing components

Resampling without distortion in signal


13

Combination of resampling and linear interpolation

L MLL MML MLL MML MLL MML MLL MM

L MLL MML MLL MML MLL MM

upsample filter downsample

What if we need data at these timing ? Linear interpolation


14

Cross spectrum

Synchronized sensing accuracy checkAccuracy of synchronization among signals

Cross spectral densities among sensors have almost flat phase meaning accurately synchronized signals

1 degree at 100Hz 1/360/100 = 28s

synchronization error

1x t

2 1x t x t t

Fourier transform 1X 2 1 expX X i t *

1 2X X t


15

packet

Reliable communication Redundant packet

transmission Packet retransmission:

Same packets are transmitted more than once

Erasure codelost packets can be reconstructed

To transfer Send

13 15 21 13

To transfer Send

13 15 21

Received

13 15 21

Reconstruct13 15 21

13 15 21 13 15 21x xPacket loss

13 15 21 13+15+21xPacket loss

Received Reconstruct

However burst loss may happen, then ?

x x

x x

15 21 13 15 21

13 15 21 13+15+21

13


16

Reliable communication Acknowledgement based approach

To transfer

13 15 21

Send

13

ACK

15 21

13 15 21

Reconstruct Received

13 15 21

15

ACK ACK

Reliable but slow to transfer a large amount of data


17

Reliable communication

Acknowledgement based approach: fewer ACK packets

To transfer

13 15 21 …16

Send

13 21

13 15 21 …16

Reconstruct

13 15 21

15

ACK15 is

missing

Reliable and fast to transfer a large amount of data

16

16

…

15

15

All received


18

Efficient data aggregation

Application specific knowledge is utilized to efficiently perform data aggregation

Data

Efficient data aggregation

Information

Application specificknowledge


19

Application specific knowledge -Natural Excitation Technique-

Definition:Estimate:

iR Correlation function

ref ref refx x x x x xMR CR KR 0

0 i i refR E x t x t

*

1

1ˆ dn

xy i iid

G X Yn T

i i refR E x t x t

1 ˆˆxy xyR G F

（ Cross Spectrum Density estimation )

Natural Excitation Technique (NExT)“Correlation functions satisfies EOM for free vibration”

Data compression through averaging

1/20 － 1/10(nd = 10-20)

ix

t t t t Mx Cx Kx f

Measurement

Subsequently decomposed into modal vibrations


20

Node 1x1

Node 2x2

Node 3x3

Node 4x4

Node nsxns

. . .Node 5x5

1 1ˆ

i iE x t x t R i =1,2,…,ns

Centralized data aggregationCorrelation function estimation

Requires signals from 2 nodes 2 approaches

Centralized implementation O(N·nd·ns)

Distributed implementation

*

1

1ˆ dn

xy i iid

G X Yn T

1 ˆˆxy xyR G F

1d sN n n Transmission


21

Packet transfer Broadcast and unicast

Broadcast: 1-to-”others in the range”

Unicast: 1-to-1 Specify the destination by

node ID Basically broadcast, but

others ignore.

Unicast

Broadcast


22

Correlation function estimation Requires signals from 2 nodes 2 approaches

Centralized implementation O(N·nd·ns)

Distributed implementation O( N(nd+ns))

Node 1x1

Node 2x2

Node 3x3

Node 4x4

Node nsxns

. . .Node 5x5

1 1 11ˆE x t x t R

12R̂ 13R̂ 14R̂ 15R̂ 1ˆ

snR

Distributed Data Aggregation

*

1

1ˆ dn

xy i iid

G X Yn T

/ 2 1d sN n N n Transmission

1 ˆˆxy xyR G F

Data transfer requirement is a primary Data transfer requirement is a primary factor for power consumption. factor for power consumption.

Distributed implementation has an Distributed implementation has an advantage advantage

Ex)Ex)

NN = 1024, = 1024, nndd=20, =20, nnss = 15= 15

Centralized implementation Centralized implementation 286,720 286,720Distributed implementation Distributed implementation 27,648 27,648A reduction factor of 10.4A reduction factor of 10.4

Data transfer requirement is a primary Data transfer requirement is a primary factor for power consumption. factor for power consumption.

Distributed implementation has an Distributed implementation has an advantage advantage

Ex)Ex)

NN = 1024, = 1024, nndd=20, =20, nnss = 15= 15

Centralized implementation Centralized implementation 286,720 286,720Distributed implementation Distributed implementation 27,648 27,648A reduction factor of 10.4A reduction factor of 10.4


23

Integration of middleware services into applications Example1 Distributed Computing Strategy

for SHM Example2 Railroad bridge vibration

monitoring


24

Distributed Computing Strategy for SHM

Sensing

NExT

SDLV

DCS logic

Cluster formation

ERA

（常時微動計測を仮定）1. Vibration measurement ambient vibration measurement2. Modal analysis in each cluster

OutPut: Natural frequency, Mode shape, A,C matrices

Method: NExT, ERA

3. Damage assessment in each communityOutput: Damage locationMethod: Stochastic damage locating vector

4. Synthetic judgment among cluster headsOutput: Damage locationMethod: DCS logic

DCS flow chart


25

DCS implementation middleware

Reliable communication Synchronized sensing Efficient data

aggregation

Numerical library FFT SVD Eigensolver sort

Static stress analysis (a part of damage detection)

All the tasks are predefined. Once parameters are injected to the network, the Imote2s autonomously perform damage identification.


26

Experimental Verification Ten Imote2s, 3 clusters

autonomously monitor the 3D truss scale model

Longitudinal & vertical measurements

Damage simulated by an element with a small cross-section is localized by Imote2s

53% cross section reduction


27

The SDLV method

The damaged element 8 has small stress indicating damage.

0

0.2

0.4

0.6

0.8

1

3 4 5 6 7 8 9 1011 1112 1314 151617 18197 8 9 10 111213 1415Element ID

No

rma

lize

d a

ccu

mu

late

d s

tre

ss

Threshold0.3

Damaged element


28

Railroad bridge monitoring application

Node wakeup based on train schedule Extract time traffic vibration Data processing

Cluster head

Leaf node

Leaf node

Leaf node

Detailed modal analysis of viaducts from ambient vibration is Detailed modal analysis of viaducts from ambient vibration is non-trivial exploit traffic vibration ⇒ non-trivial exploit traffic vibration ⇒

After train passageAfter train passage natural frequenciesnatural frequencies linear damagelinear damage

During train passagDuring train passagee ：：Vibration amplitude Vibration amplitude abnormal abnormal vibrationvibrationCoherence functionCoherence function non-linearitynon-linearity

Report to BS

Amplitude levelcoherence function

Modal identification

Signal extraction

Sensing

Wakeup


29

Service-Oriented Architecture Service-Oriented ArchitectureService-Oriented Architecture (SOA) in the SHM

Toolkit simplifies SHM software development Applications are comprised of manageable, modular

servicesservices that exchange data in a common format The middleware frameworkmiddleware framework connects the services by

providing communication and coordination

SDLV

Numerical ServicesApplication Services Foundation Services SHM Application


30

SHM Toolkit Contents Foundation services

Universal sensing Time synchronization Reliable communication Numerical library

Application services Correlation function estimation (CFE) Eigensystem Realization Algorithm (ERA) Stochastic Damage Load Vector (SDLV) Stochastic Subspace Identification (SSI) Synchronized sensing

Test applications, tools and utilities Radio & antenna testing Data acquisition (local and remote) Test applications for each component of the toolkit


31

http://shm.cs.uiuc.edu

Bridge & Structure Laboratory University of Tokyo 1 Structural health monitoring using Imote2 Tomonori Nagayama Assistant Professor University of Tokyo.

Documents

time difference

reception time t2 global

time stamp t1

global time stamps

clock rates

ss clock drift t s

clock speed

local clocks t