A Programmable Sensor Network Based Structural Health Monitoring System

1
A Programmable Sensor Network Based Structural
Health Monitoring System

• Krishna Kant Chintalapudi
• Embedded Networking Laboratory,
• University of Southern California, Los Angeles,
  USA

2
Agenda

• Whats the talk about ?
• Whats structural health monitoring (SHM)?
• SHM techniques and their impact on sensor network
  design
• Architecture design for sensor network based SHM
• A prototype implementation and deployment
• What next?

3
Whats the talk about?

• A programmable sensor network based system for
  structural health monitoring
• What are the requirements of SHM applications?
• How do we architect a sensor network system to
  satisfy these requirements?
• A prototype and its performance

4
Agenda

• Whats the talk about ?
• Whats structural health monitoring (SHM)?
• SHM techniques and their impact on sensor network
  design
• Architecture design for sensor network based SHM
• A prototype implementation and deployment
• What next?

5
What Is Structural Health Monitoring (SHM)?

• Structural integrity assessment for buildings,
  bridges, offshore rigs, vehicles, aerospace
  structures etc.

• Goals of SHM are
• damage detection is there damage?
• damage localization where is the damage?
• damage quantification how severe?
• damage prognosis future prediction

6
How Are Damages Caused?

• Extreme stress leading to fatigue in elements
• several freeway bridges today bear traffic far
  exceeding tolerance levels they were originally
  designed to bear.

• Rusting and degradation of material properties
• leads to change in stress distribution and
  overloading of certain elements more than others

• Continuous vibrations/cyclic stresses in the
  structure
• waves shaking offshore oil-rigs, gales shaking
  bridges.

• Catastrophes (earthquakes)

7
How Do Damages Evolve?

• Most damages start as tiny cracks caused by
  metal fatigue (microns-mm).

• If unattended the cracks creep and grow in size
  leading to deterioration of the material.

• If unchecked, it eventually results in an
  unpredictable, sudden and catastrophic failure.

• SHM techniques focus on detection and
  localization of damages as early as possible.

8
SHM Today

• Today SHM is carried out by
• collecting sensor data from several locations in
  the structure and analyzing it on a high end
  platform
• periodic (bi-annual) human inspections
  (visual/using portable devices),
• expensive and dedicated data-acquisition systems
  (for structures where monitoring is critical) .

• SHM suffers from
• human error and inaccessibility of locations
  within the structure
• expensive labor (for inspection), cabling and
  installation (for data-acquisition systems)
• possibility of catastrophic failure between
  inspections

9
Agenda

• Whats the talk about ?
• Whats structural health monitoring (SHM)?
• SHM techniques and their impact on sensor network
  design
• Architecture design for sensor network based SHM
• A prototype implementation and deployment
• What next?

10
Local vs. Global Techniques
GLOBAL
LOCAL

• Detect tiny cracks (mm/cm) and small corroded
  patches.

• Target larger damages e.g. undermined cables,
  braces or columns

• Use sophisticated imaging techniques 250KHz
  ultrasound, x-ray, thermal, magnetic etc.

• Use accelerometers to collect structural
  response.

• Detect structural damages in the entire structure

• Can detect damages within a few inches of the
  equipment

11
Feasibility of Local SHM Techniques

• They are expensive, require a lot of power and
  bulky
• Demand extremely dense deployments

• Local SHM techniques are not amenable to sensor
  network deployments

• So let us focus on global schemes henceforth

12
Ambient vs. Forced Excitation
AMBIENT
FORCED

• Rely on ambient sources (wind, passing vehicles,
  earthquakes)

• Rely on induced excitation (impact hammer,
  rotating mass etc.)

• Unpredictable in nature and timing

• Pre-meditated and precise.

• Much higher signal-to-noise ratio.

• Low signal-to-noise ratio.

• Require continuous monitoring hard to implement
  duty cycles.

• Amenable to extremely low duty cycle functioning.

13
Recall Our Goal

• We want a system that SHM engineers can program
  not experts in TinyOS
• We explore existing SHM schemes to find what SHM
  engineers want?
• We design our system based on requirements of SHM
  schemes.

14
What SHM Engineers want?

• Structural integrity assessment for buildings,
  bridges, offshore rigs, vehicles, aerospace
  structures etc.

• Today SHM engineers want
• damage detection is there damage?
• damage localization where is the damage?
• damage quantification how severe?
• damage prognosis future prediction

15
Structural Dynamics 101
Structures are no different from strings!!
16
Structural Dynamics 101

• Structural response is the spatio-temporal
  deformation induced in the structure.
• The dynamics of a structure are often expressed
  as,

• The impulse response is given by

• vl are mode shapes normalized structural
  deformation patterns
• are modal/resonant frequencies of the
  structure
• are the amplitude and phase of the
  mode induced in the structure

17
Structural Dynamics 101

• mode shapes and frequencies are fundamental to
  the structure
• material properties, geometry and assemblage of
  elements
• depend on both the sensing and
  actuating locations
• mode are global phenomena may span the entire
  structure

18
How Does Damage Affect Modes?

• Modal (resonant) frequencies and mode shapes
  change

• Modal frequencies decrease

• Break in symmetry of the structure may lead to
  splitting of overlapping modes and cause extra
  modes to appear

• Non-linearities may introduce new modes.

19
Some practical aspects

• Modal frequencies are typically in the range of
  few tens of Hz
• Real structures are often heavily damped and
  decay within a second
• Most SHM engineers prefer 10 times oversampling
• Sampling rates desired are around 200-500Hz for
  most structures.

20
Literature Review Damage Detection

• Literature is very vast

• Model the structural response using ARMA/AR
  based linear predictors and look for a
  significant change in coefficients.

• Look for shifts/changes in modal frequencies
  through spectral analysis.

• Look for changes in mode shapes.

• Use non-linear techniques such as neural
  networks.

21
Literature Review Damage Localization

• Significantly more challenging and still a very
  hot research topic.

• Time domain methods, model structure as a LTI
  system
• try to solve for A,B,C and D using response from
  all sensors
• compute stiffness of elements using A, B, C and D
  
• loss of stiffness indicates damage in an element

.
22
Damage Localization Techniques

• Frequency domain - estimate mode shapes using
  structural response from all sensors and use mode
  shapes to estimate stiffness of members
• ERA (Eigenvalue Realization Algorithm) perform
  SVD on the Hankel matrix
• y is the impulse response
  
  
• 
  
  vector
• Select modes corresponding to the high singular
  values to forma reduced order system, and
  calculate the modal vector matrix V using,

23
Whats common to SHM schemes?

• Inherently Centralized Global nature of modes
  naturally leads to centralized algorithms for
  detection and localization.
• Can leverage local computation Almost none of
  the schemes uses data in its raw form
• ARMA/AR models need coefficients
• Modal frequency based schemes need to use the
  estimated spectrum
• Compute these quantities locally and transmit
  instead of raw data.
• 40 ARMA coefficients instead of 5000 samples
  (over 99 savings!!!)
• Little or no collaboration/aggregation most
  algorithms do not require inter-node
  collaboration (eg SVD is hard to decentralize)

24
How many sensors would a typical structure need?

• Strategies for deploying sensors
• Deploy a tri-axial sensor at the end of every
  member (damage localization/member)
• Divide the structure into sections and deploy a
  tri-axial sensor at every corner (damage
  localization/section)
• Number of sensors determines the granularity of
  localization (per floor? Per column?)
• A real structure can have several 100s of
  members/sections
• Local computation is absolutely critical

25
What are the requirements of SHM schemes?

• High data rates 100 sensor will generate a few
  Mbps of data
• Reliable Delivery SHM algorithms do not
  tolerate sample losses
• Time Synchronization - Required by most schemes
  
• error in time-synchronization manifests as phase
  error in modes
• error , the higher the modal
  frequency the more accuracy one needs
• For 1 error in a 20Hz mode, an accuracy of
  about 100
• Local computation data acquisition system
  based solutions will not scale

26
Agenda

• Whats the talk about ?
• Whats structural health monitoring (SHM)?
• SHM techniques and their impact on sensor network
  design
• Architecture design for a programmable sensor
  network based SHM system
• A prototype implementation and deployment
• What next?

27
Recall Our Goal

• We want a system that SHM engineers can program
  in Matlab/C
• An SHM engineer should be able to write and test
  variety of algorithms without having to
  re-program the motes
• The system should be evolvable a if better mote
  platform come, the SHM engineer should not need
  to rewrite his code

28
Typical operation of an SHM system

• Sensors collect noise unless the structure is
  shaking!!!
• Ambient Schemes rely on significant event
  (heavy wind, passing truck)
• Forced Schemes rely on actuators (impact
  hammers)
• Structural Response lasts a few seconds!!!
• Sensors sleep unless an event occurs or the
  users requests actuators to test
• Sleep --- test/significant event ---- collect
  data and locally process --- transmit to central
  location --- sleep (wake once a day/ once a few
  hrs)
• SHM systems will be Triggered Systems

29
Architecture Design Decisions

• Two-level Hierarchy A higher more endowed
  layer is required to manage the aggregate data
  rates generated by the motes.
• Isolate Application code from mote code Mote
  class devices provide a generic task interface
  but no application specific code
• getSamples(startTime, noSamples, sampFreq, axis)
• getFFTSamples(startTime,noSamples,sampFreq,axis,ff
  tSize)
• actuateStructure(startTime,type, parameters)
• conveyed to motes as tasking packets by
  gateway-class devices

30
What does code isolation buy us?

• Reusability Application programmers can use
  the generic task interface and write many
  different SHM applications.
• Basic SHM library functions can me provided on
  motes fft, auto-correlation, ARMA coefficient
  estimation, spectral estimation etc.
• Evolvability If a new mote comes along with
  greater processing power, just add new
  functionality, no need to rewrite application.
• Gateway class nodes translate C/Maltab
  application code into mote tasking commands

31
Agenda

• Whats the talk about ?
• Whats structural health monitoring (SHM)?
• SHM techniques and their impact on sensor network
  design
• Architecture design for a programmable sensor
  network based SHM system
• A prototype implementation and deployment
• What next?

32
We have a prototype
function shifts getModalShiftsFromBuilding()
create a group for sensors gidSensors
NetSHMCreateGroup(1,2,3,4) create a group
for actuators gidActuators NetSHMCreateGroup(5
) actuate after 22 seconds NetSHMCmdActuate(gid
Actuators,22) collect structural response
starting 20 seconds from now, 4000 samples at
200Hz,along x-axis only, samples
NetSHMCmdGetSamples(gidSensors,20,200,1,4)
find modal frequencies modes
findModes(samples) read original modes load
OriginalModes shifts findModalFreqShifts(modes,
OriginalModes)

• A complete SHM test
• Matlab API
• Matlab functions implemented as wrappers over C
  functions
• Platform MicaZ and starGates

33
The Stacks
34
The API

• Groups Every task is addressed to a group of
  sensors/actuators
• Create, AddNodes, DeleteNodes, ClearGroup etc
• Create returns a handle to the group
• Tasks task(groupId, parameters)
• getSamples, getFFTSamples, getXCorrSamples,
  getModalFreqs, actuate etc.

35
Mote Tasking Library

• Translates API commands into command packets to
  motes
• Uses TimeSynch Module to translate global time
  to sensor network time
• Dispatches command packets using the Reliability
  Layer
• Delivers results to applications according to
  API specifications
• A collection of C and Matlab Mex files

36
Reliability Layer

• Transactional Delivery Application expects
  results asynchronously
• Application issues a task
• Mote Tasking library breaks it up into commands
• Opens a connection to Reliability layer and
  sends command packet
• Reliability layer keeps connection open and
  forwards result packets to Mote Tasking Lib
• Mote Tasking Library aggregates results and
  returns to applications
• Takes care of out of order delivery
• Can handle several applications simultaneously
• OneShot Delivery Application does not expect
  any results (e.g. Actuate)

37
Time Synchronization

• Use FTSP
• Small modifications for compatibility with our
  code
• We use 28.8Khz timer and get accuracy to a few
  100micro-sec
• All motes are synchronized to a single mote

38
Routing

• Does not require any-to-any routing
• starGates to motes
• mote to starGates
• starGates to starGates
• both communication end points are never motes
• Routing Modules used
• starGate to starGate - a distance vector
  routing scheme, also passes on routes to motes
• motes to starGates CENS Extensible Sensing
  System,
• starGates to motes each node periodically
  transmits list of nodes in its sub-tree to its
• parent, the
  parent keeps a pointer on the reverse path
• We are still investigating better choices for
  routing

39
Sensing Hardware

• MDA400 vibration cards from Crossbow
• high quality low power vibration sensing
• 16-bit samples, on board storage (64k)
• 0-20000Hz sensing
• 4 simultaneous channels
• driven by a micaZ
• Accelerometers
• high sensitivity (1v/g)
• low noise
• Actuators
• off-the shelf door latch devices
• motor control board interfaced to micaZ

40
Deployment
Seismic Test Structure
Scaled Building Model
41
Damage Detection and Localizationon scaled model

• Building Details
• 48 inches high, 4 floors, 60 lbs
• Floors 1/2 x 12 x 18 aluminum plates
• steel 1/2 x 1/8 inch steel columns
• 5.5 lb/inch spring braces
• 4 actuators on the top floor
• 8 motes, 2/floor, dual axis, 200Hz, 2 starGates
• 4 Test Cases
• braces from floor 4 removed
• braces from floor 3 removed
• braces from floor 2 removed
• braces from floor 2 and 4 removed

42
Performance Analysis on Seismic Test Structure

• Structure details
• Full scale imitation of a hospital ceiling (28
  by 48)
• electric lights, drop ceiling, water pipes, fire
  sprinklers
• 55,000 lb actuator, 10 inch stroke, manually
  operated right now
• 15 micaZ motes, 2 starGates, 200Hz
• Latency and robustness to failure
• One starGate carrying most motes killed
• all samples recovered
• 3000 samples in about 5 minutes

43
Agenda

• Whats the talk about ?
• Whats structural health monitoring (SHM)?
• SHM techniques and their impact on sensor network
  design
• Architecture design for a programmable sensor
  network based SHM system
• A prototype implementation and deployment
• What next?

44
What next?

• Develop schemes that allow aggressive local
  computation
• for damage localization.
• Remotely actuate the Seismic Test Structure
• Developing local actuators for the Seismic Test
  Structure
• Damage Detection and Localization on the Seismic
  Test Structure
• Experiments on real bridges and structures with
  large scale deployments