Automation & Robotics Research Institute (ARRI)The University of Texas at Arlington

F.L. Lewis, Fellow IEEE, Fellow IFAC, Fellow UK InstMCMoncrief-O’Donnell Endowed Chair

Head, Controls & Sensors Group

http://ARRI.uta.edu/acslewis@uta.edu

Structural Health Monitoring

Rytter’s Levels:

Worden et al. 2009

Machine Learning

Methods of Learningsupervisedunsupervisedreinforcement

classification

SHM

defect

Interrogate (active or passive)

Change in propertyMaterialSample

sensors

Property Changes:visualacoustictemperature, pressurethermal conductance propertiesmagnetic propsglobal props.- Modal- stiffness, vibration freqslocal props – local vibration freqs, impedance

strain, stressforce causes AE stress waves

wave propagation properties –scattered, reflected, freq content

Model-basedData-based DSP

filter, preprocess, detrendFeature extraction

Decision-makingBayesNNfuzzyrule-based

Diagnosis & PrognosisDetectionClassificationLocalizationAssessmentPrediction

Fault Types

CompositesMatrix (resin) crackDelaminationFibre breaks

MetalsMaterial Defects

CorrosionCrackFatigue

System Defects Rivet FailureSurface Ice

Detection Methods

Vibration (LF- wavelength >> plate thickness)global- LF - changes in structural props- stiffness, vibr. freqs. local- HF – changes in res. freqs., impedance

SonicUltrasound (HF- wavelength << thickness)Acoustic Emission (mid freq.- wavelength ~ thickness)Wave Propagation (single freq waves)

Strain, Stress

X-RayVisualThermalMagnetic

Electromagnetic spectrumhttp://imagers.gsfc.nasa.gov/ems/waves3.html

Sound

Frequency in Hz

Wavelength (STP at sea level)

20 200 2,000 20,000

Infrasound Ultrasound

Humans DogsElephants

100,000

Cats

BatsDolphins

200,0005

50m 10m 1m 10cm 1cm 1mm

Sound

Frequency in Hz

Wavelength (STP at sea level)

20 200 2,000 20,000

Infrasound Ultrasound

Humans DogsElephants

100,000

Cats

BatsDolphins

200,0005

50m 10m 1m 10cm 1cm 1mm

Acoustic Spectrum

Sensor Modalities

Overlap in freqs!

Transmission depends on the medium

Sensors Based on Physical Transduction Principles

Mechanical SensorsPiezoresistive Effect converts an applied strain to a change in resistance Piezoelectric Effect converts an applied stress (force) to a potential difference. PZTCapacitive Sensors convert displacement (force) into change in capacitance

Magnetic and Electromagnetic Sensors do not require direct physical contactHall Effect. Magnetic field applied perpendicular to current flow causes induced voltageMagnetic Field Sensors detect metallic objectsEddy Current Sensors use magnetic probe coils to detect defects in metallic structures

F.L. Lewis, “Wireless Sensor Networks,” in Smart Environments: Technologies, Protocols, Applications, Chapter 2, ed. D.J. Cook and S.K. Das, Wiley, New York, 2005.

Thermal Sensors measure temperature or heat fluxThermo-Mechanical Transduction. Heat causes thermal expansionThermoresistive Effects. Resistance R changes with temperature TThermocouples. Junctions of two different metals at different temperatures causes current flowResonant Temperature Sensors. Temp change in some materials causes a change in resonant frequency

Optical Transducers. Convert various properties to lightOptical fiber interferometers and gratings– changes in length (strain), temp. cause changes in phaseOptical fiber accelerometers based on time of flight

Acoustic SensorsUltrasound. High Frequency. Can penetrate structures. Reflected and scattered from defectsAcoustic Wave Sensors

surface acoustic wave (SAW), thickness-shear mode (TSM), flexural plate wave (FPW), or acoustic plate mode (APM)

PiezoelectricSensor- develops a voltage difference across two of its faces when compressed Actuator- physically changes shape when an external electric field is applied

PyroelectricHeat Sensor- Develops a voltage difference across two of its faces when it experiences a temperature change.

TEMP. COMPENSATION

Ferroelectric-Has a spontaneous electric polarization (electric dipole) which can be reversed in the presence of an electric field.

PZT - Lead zirconate titanate

X-rayVisualCoherent OpticsFiber optics – no EMI, lightweight, low noise, high BW

InterferometryFiber Bragg Grating - FBG

Thermography - IRMagnetics

Eddy currenta coil induces eddy currents in a conductive sampledefects cause change in the impedance of the sample

Interrogation / Interaction Modalities

Group 1

Ultrasound HF (5 MHz)wavelength << thicknessshort prop. Distancea single excit. freq.

Acoustic Emission AE Mid freq (100 kHz-2 MHz)wavelength ~ thicknesssubstantial prop. dist. (a few m.)multiple excitation freqs. (e.g. white noise)

Lamb Wave is AE with a single excit. freq.

Group 2- SonicSound propagation depends on the mediumDefects scatter or reflect sound waves

Group 3- Vibration

LF (20 Hz – 20 kHz) Global- changes in physical props. Modal parameters (stiffness, res. freq.)

HF (100 kHz – 1 MHz) Local- changes in frequency contentEM impedance

PZT sensors

Monitor vibration signature- pattern recognition

Wavelength >> thickness

Group 4- Strain, Stress

Force on defect causes AE stress waves

wikipediaLamb Waves

Elastic waves that propagate in a solid thin plate

2-D Wave equation

solutions split into two sets of waves-symmetric & antisymmetric

Irradiate entire thicknessPropagate substantial distances

S0 Extensional mode

A0 Flexural mode

wavelength ~ thickness

s0 Scattered and reflected by crack

a0 Detects delamination

A major challenge and skill in the use of Lamb waves for ultrasonic testing is the generation of specific modes at specific frequencies that will propagate well and give clean return "echoes". This requires careful control of the excitation and identification of the correct waves.

Eddy-coil EM actuatordischarges capacitor through a coil, induces pulsed magnetic field in conductive sample,generates a forceneeds 1-10 J

Passive vibration monitoring-in-flight aircraft vibration freqs are very lowsuccessful in a boat hull monitoring application

Actuation / Interrogation

Active vs. Passive

PZT actuator

Active -

Actuator interrogation signalsBurst sinusoids

kk-N

has DFT

Square window DFT swamps out the signal DFT

time freq

Square window

Hamming window

Hann window

FFT of burst sinusoids with: time freq

Chaotic Interrogator Actuation Todd et al. 2009

Actuation by Lorenz Signal

Signals received atsensors

Sensor 1 Sensor 2

Actuator and Sensor Locations

Based on Physical ModelsFEA or dynamics model

Based on Engineering KnowledgeTodd et al. 2009FBG sensors on Boat waterjet

Worden & Manson 2009CR = actuatorCi= PZT sensors

3 sensor networks based onknown damage regions

Ihn & Chang 2004Sensors along a rivet joint andBuilt into a composite skin

c.f. human nervous system proprioceptors

Deploy at Hot SpotsDeploy over a Large Area-

limits the frequencies and interrogation methods

No methodical procedures

Aircraft wing

Features

Time momentsFrequency domain properties-

Resonant freqs, sidebandsPower content in specific frequency bandsTransmissibilities

Strain, stress

Transmissibility from sensor j to sensor i

( )( )( )

iij

j

PSDTPSD

ωωω

=

Do DFT for sensor i signal

actuatorSensor i

Do DFT for sensor j signal

Sensor j

( )ijT ω

Time-Varying Frequency ContentShort Time Fourier Transform – Windowed DFT

time

frequency

Must select window lengthMust use good window w(n) - Hamming, Hann

WaveletDoes not need windowMulti-resolution analysis

Hilbert-Huang Transform (HHT)

2 ( 1)( 1)/

( 1)( , ) ( ) ( )

tj k n N

n t NX k t x n w t n e π− − −

= − −

= −∑STFTSpectrogram

Wavelet xformScaleogram

Basic or mother wavelettime

freq

Fault detection & Identification

Model-Basedmake physical model using FEA or physics-based methodsdetermine comparison metriclook for departures of real measured data from the model

Data-Basedbased on moments, freq response, or statisticsestablish normal operating limits basedestablish abnormality thresholdsdepartures indicate faults

Both methods look for departures from the normthis means Statistical Pattern Recognitionpreprocessing of data, filter, detrendoutlier rejection

• Variations of available empirical and deterministic fatigue crack propagation models are based on Paris’ formula:

Where:α = instantaneous length of dominant crackΝ = running cyclesCo, n = material dependent constantsΔК = range of stress intensity factor over one loading cycle

( )no KCdNda

Δ=

e.g. Deterministic Crack Propagation Modelse.g. Deterministic Crack Propagation Models

Physical Modeling

Dr. George Vachtsevanoshttp://icsl.gatech.edu/icsl

Ihn and Chang 2004

Lamb Waves

Must focus on ONE frequency

Nascent freq

Where is the right Lamb wave?

Use s0 for cracksUse a0 for composite delamination

Group velocity Dispersion

From FEA for the specific material G

roup

vel

ocity

Frequency

s0 arrives before a0 below 600kHz

Time

Freq

uenc

y

s0 a0

Optimal sensor location wrt crack

Ihn and Chang 2004Smart Suitcase

Acellent Technol. Smart patch

Rivet failure

Cracks0

Debonda0

Composite faults

Acellent Smart Layer

a0

s0

debond

crack

Dam

age

inde

xD

amag

e in

dex

Tomography- imaging by sections using wave energy2D or 3D images

x-ray CTgammaelectron

Reconstruction algorithmsfiltered back projectioniterative reconstruction

ART- algebraic reconstruction techniques(Kaczmarcz algorithm)

wikipediaCardiac CT scan

s0 energy- compute energy up to MAXIMUM PeakUse RMS value for tomographic reconstruction

Sticky gum to hold sensors?

Split plate into a uniform gridMount sensors at grid points400 sensors!

Freq = 500 kHz

Uniform angular sampling of plate with few sensors

Improved sensor placement

Wide band Lamb wavesExcitation – rectangular impulse 10 microsec wideExcites Lamb waves covering a broad frequency spectrum

Used Kohonen NN to classify damage

Compensate for Propagation- amplitude A with distance x

Hickman et al. 1991

LF vibration 1-5 kHz

Defects cause energy redistribution in freq. spectrogram

FEA

Rivet removal and cracks both lower the HF content

Select Features?

24” sq. plate, 0.08” thick

Eddy coilActuator1-10 J

Screwsaroundedges

Sensor placement determined experimentally!

Compute features for each sensor

Best Features:Energy distribution- Power in specific freq bands

3-4PZT sensors

NN classification?

Features-Energy distribution- Power in specific freq bands

1-1.5 kHz

1.5-3 kHz

Rivet failuresCracksicing

Power in freq range 1-1.5 kHz

Pow

er in

freq

rang

e 1.

5-3

kHz

Aircraft Monitoring Wing cuff

1. Damage detection

EDS

EDS= Electrodynamic shakerLF EM Vibration at 1-2 kHzCompute DFT

4 PZT sensors

Transmissibility from sensor j to sensor i

( )( )( )

iij

j

PSDTPSD

ωωω

=

1 2

3

4

panel

Damaged panels

Features = power in specific freq bands

Classification and departure detectionNNclustering (K-means, NN)outlier analysis using norm distance measure

Training data Set and Validation Set

Unsupervised learning for fault detection

Band 1 Band 2

No fault

crack

2. Damage Location

Network of sensors

Damage = remove panels

CR = PZT actuatorCi = PZT sensors

NN MLP classifier – competitive learning

Supervised learning for fault classification or fault location

Actuator and sensor locations based onKNOWN possible fault locations

4 networks with 1 actuator and 3 sensors

Fiber Bragg Grating – FBG

Interrogate length scales in the mm rangeNo EMILightweightCan be directly photo written into silica fiber using UVEmbed inside composites

2r nTλ =n= fiber core model index, T= grating period

Axial compression or tension changes T – can measure strain

Boat Hull Monitoring – passive wave excitationJoint Degradation – active excitation - EDS

about 0.1-0.3 nm

A-K= 11 sensor arrays56 sensors in allRosette= 3-D sensor?

Boat Hull Monitoring

Passive excitation

Sensor placementhull monitoringwaterjet monitoring

Discrete wavelet transformSet scale factor equal to 2j

Feature selectionSelect specific levels only

No defect

Defect

1024

516256

128

1996 IEEE Ultrasonics Symp.

Changes in time delay and freq due to physical quantity y(t)

Interrogation freqs 434MHz = 5-10 m RFID range2.5 GHz = 1-2 m RFID range

SAW + RFID

Gate this part = 1-2 micro sec

Sensitive to y(t) = temp., displacementy(t)= strain, force, accel. needs proper packaging

Haiying Huang, ME Dept, UTA

Passive induction coupling-remote interrogation

Res fre

q f 01

Res freq f10

Strain causesRes freq shift

Patch Antenna

Crack Monitoring Haiying Huang, ME Dept, UTA

Res fre

q f 01

Res freq f10