In-Flight Fatigue Crack Monitoring in Aircraft Using Acoustic Emission and Neural Networks Eric. v. K. Hill Samuel G. Vaughn Christopher L. Rovik.

In-Flight Fatigue Crack In-Flight Fatigue Crack Monitoring in AircraftMonitoring in Aircraft

Using Acoustic Emission and Neural NetworksUsing Acoustic Emission and Neural Networks

Eric. v. K. HillEric. v. K. Hill

Samuel G. VaughnSamuel G. Vaughn

Christopher L. RovikChristopher L. Rovik

Fatigue Crack MonitoringFatigue Crack Monitoring

Background Background Test Setup & ProcedureTest Setup & Procedure Acoustic EmissionAcoustic Emission Neural NetworksNeural Networks ResultsResults Conclusions & RecommendationsConclusions & Recommendations

BackgroundBackground

Fatigue CrackingFatigue Cracking

Brittle failure in a normally ductile material Brittle failure in a normally ductile material due to cyclic loads below yield stressdue to cyclic loads below yield stress

Plastic deformation plus cyclic loads leads Plastic deformation plus cyclic loads leads to strain hardening, then fatigue crackingto strain hardening, then fatigue cracking

Small cyclic loads can cause significant Small cyclic loads can cause significant damage over timedamage over time

Notable Fatigue FailuresNotable Fatigue Failures

1988 Aloha Airlines flight: a piece of a B-737 1988 Aloha Airlines flight: a piece of a B-737 fuselage tore off during flight due to fuselage tore off during flight due to corrosion/fatigue crackingcorrosion/fatigue cracking

Recent F-15E fuselage longeron failure behind Recent F-15E fuselage longeron failure behind cockpit – grounding of fleetcockpit – grounding of fleet

Aging aircraft are progressively accumulating Aging aircraft are progressively accumulating fatigue damagefatigue damage

This leads to costly mandatory inspections and This leads to costly mandatory inspections and parts replacement at “safe” intervalsparts replacement at “safe” intervals

In-Flight Acoustic Emission In-Flight Acoustic Emission ApplicationsApplications

Military: KC-135, C-5A, F-105Military: KC-135, C-5A, F-105

Commercial: N/ACommercial: N/A Civil: Civil: Piper PA-28 CadetPiper PA-28 Cadet & & Cessna T-303 Cessna T-303

CrusaderCrusader

Goal:Goal: In-flight fatigue crack detection systems In-flight fatigue crack detection systems promote maintenance schemes based on promote maintenance schemes based on replacement for causereplacement for cause rather than rather than replacement at replacement at conservatively calculated intervalsconservatively calculated intervals using linear using linear elastic fracture mechanics.elastic fracture mechanics.

Relevant M.S. ThesesRelevant M.S. Theses 1994 A.F. de Almeida: Neural Network Detection of 1994 A.F. de Almeida: Neural Network Detection of

Fatigue Crack Growth in Riveted Joints Using Fatigue Crack Growth in Riveted Joints Using Acoustic Acoustic EmissionEmission

1995 W.P. Thornton: Classification of Acoustic Emission 1995 W.P. Thornton: Classification of Acoustic Emission Signals from an Aluminum Pressure Vessel Using Signals from an Aluminum Pressure Vessel Using

a a Self-Organizing MapSelf-Organizing Map 1996 M.L. Marsden: Detection of Fatigue Crack Growth 1996 M.L. Marsden: Detection of Fatigue Crack Growth

in a Simulated Aircraft Fuselagein a Simulated Aircraft Fuselage 1998 S.G. Vaughn III: In-Flight Fatigue Crack Monitoring 1998 S.G. Vaughn III: In-Flight Fatigue Crack Monitoring

of an Aircraft Engine Cowlingof an Aircraft Engine Cowling 1998 C.L. Rovik: Classification of In-Flight Fatigue 1998 C.L. Rovik: Classification of In-Flight Fatigue

Cracks in Aircraft Structures Using Acoustic Cracks in Aircraft Structures Using Acoustic Emission Emission and Neural Networksand Neural Networks

Test Setup & ProcedureTest Setup & Procedure

Testbed 1: Engine CowlingTestbed 1: Engine Cowling

In-Flight Test Set-UpIn-Flight Test Set-Up

In-Flight Test Set-UpIn-Flight Test Set-Up

4 acoustic emission transducers 4 acoustic emission transducers symmetrically mounted on engine symmetrically mounted on engine cowlingcowling

2 transducers monitoring crack growth 2 transducers monitoring crack growth and the other 2 recording the noiseand the other 2 recording the noise

3 Flights with 5 particular maneuvers 3 Flights with 5 particular maneuvers monitored on each flightmonitored on each flight

Testbed 2: Vertical TailTestbed 2: Vertical Tail

Cessna Crusader N106ER

Equipment Setup (Flight)Equipment Setup (Flight)

Equipment Setup (Lab)Equipment Setup (Lab)

Acoustic EmissionAcoustic Emission

Acoustic Emission (AE)Acoustic Emission (AE)

Definition:Definition:

The transient elastic waves generated by The transient elastic waves generated by the rapid release of energy within a material the rapid release of energy within a material due to flaw growth mechanismsdue to flaw growth mechanisms

AE Signal (Voltage vs. Time) AE Signal (Voltage vs. Time) Waveform ParametersWaveform Parameters

AE Sources & CharacteristicsAE Sources & Characteristics

Fatigue CrackingFatigue Cracking Plastic DeformationPlastic Deformation Mechanical Noise (Rubbing & Rivet Fretting)Mechanical Noise (Rubbing & Rivet Fretting)

AE SourceAE Source DurationDuration AmplitudeAmplitude

Fatigue CrackingFatigue Cracking ShortShort HighHigh

Plastic DeformationPlastic Deformation ShortShort Low-MediumLow-Medium

Mechanical NoiseMechanical Noise LongLong Low-MediumLow-Medium

AE Duration vs. Amplitude PlotAE Duration vs. Amplitude Plot

Source LocationSource Location

Source Location PlotSource Location Plot

Finite Element Analysis (FEA) Finite Element Analysis (FEA) AnalysisAnalysis

Data AcquisitionData Acquisition

AE source (e.g., fatigue crack) emits acoustic AE source (e.g., fatigue crack) emits acoustic emission energy in the form of stress wavesemission energy in the form of stress waves

Piezoelectric crystal within AE transducer senses Piezoelectric crystal within AE transducer senses the signalthe signal

AE signal amplified and transmitted to a computer AE signal amplified and transmitted to a computer where its waveform quantification parameters are where its waveform quantification parameters are digitized and storeddigitized and stored

Records signals in the frequency range 100 kHz to Records signals in the frequency range 100 kHz to 1 MHz1 MHz

Neural NetworksNeural Networks

Classification Neural NetworkClassification Neural Network

Kohonen Self-Organizing Map (SOM) neural Kohonen Self-Organizing Map (SOM) neural network uses mathematical processes to classify network uses mathematical processes to classify “things” based on a set of inputs: six AE “things” based on a set of inputs: six AE quantification parameters (amplitude, duration, quantification parameters (amplitude, duration, counts, energy, rise time, and counts-to-peak) counts, energy, rise time, and counts-to-peak)

SOM Neural Network ArchitectureSOM Neural Network Architecture

SOM Data ProcessingSOM Data Processing

Two primary steps in implementing a Two primary steps in implementing a Kohonen SOM neural network:Kohonen SOM neural network:

Training the SOM – sample of dataTraining the SOM – sample of data Testing the SOM – remainder of dataTesting the SOM – remainder of data

Training the SOMTraining the SOM

Create a training fileCreate a training file 5 steps to training:5 steps to training: 1.1. Randomly set weights between 0 and 1Randomly set weights between 0 and 1 2.2. Introduce first input vector ( 6 signal Introduce first input vector ( 6 signal

parameters for AE hit)parameters for AE hit) 3.3. Find minimal planar distance between the Find minimal planar distance between the

input vector and Kohonen neuronsinput vector and Kohonen neurons 4.4. Identify the neuron with the minimal distanceIdentify the neuron with the minimal distance 5.5. Adjust/update the weights Adjust/update the weights

Testing the SOMTesting the SOM

Create testing fileCreate testing file Pass test file through the trained neural Pass test file through the trained neural

network and it will be classifiednetwork and it will be classified

ResultsResults

Anticipated ResultsAnticipated Results

Neural network classifies lab test data into 3 Neural network classifies lab test data into 3 categories: fatigue cracking, plastic deformation, categories: fatigue cracking, plastic deformation, and rubbing (mechanical noise)and rubbing (mechanical noise)

Trained neural network classifies the entire lab Trained neural network classifies the entire lab test file with a high degree of accuracytest file with a high degree of accuracy

In-flight data verifies fatigue crack growth between In-flight data verifies fatigue crack growth between Channels 1 & 2 on Piper Cadet cowling Channels 1 & 2 on Piper Cadet cowling

Fatigue crack growth activity associated with Fatigue crack growth activity associated with stressful maneuvers on Cessna Crusader vertical stressful maneuvers on Cessna Crusader vertical tailtail

Lab Test ConfigurationLab Test Configuration

Lab Test ResultsLab Test Results

Over twenty AE files were recorded during Over twenty AE files were recorded during the lab fatigue teststhe lab fatigue tests

File twenty: 3 minutes 30 seconds in length; File twenty: 3 minutes 30 seconds in length; recorded fatigue cracking for the last minuterecorded fatigue cracking for the last minute

Duration vs. Amplitude plot of file twenty Duration vs. Amplitude plot of file twenty shows good separation between failure shows good separation between failure mechanismsmechanisms

Duration vs. Amplitude (File 20)Duration vs. Amplitude (File 20)

ATPOST Filtering LimitsATPOST Filtering Limits

AE sources filtered into individual files:AE sources filtered into individual files:

Mechanism Duration (µs) Amplitude (dB)Cracking 0-6,000 65-100Rubbing 6,000-32,000 30-70Plst. Def. 0-6,000 30-65

Training the SOMTraining the SOM

100 hits each of fatigue cracking, plastic 100 hits each of fatigue cracking, plastic deformation, and rubbing for a total of 300 deformation, and rubbing for a total of 300 hits were used for traininghits were used for training

Trained neural network tested 99% accurate Trained neural network tested 99% accurate when testing the remaining 70,000+ hitswhen testing the remaining 70,000+ hits

One column by three row (1x3) matrix One column by three row (1x3) matrix Kohonen classification layer gave the most Kohonen classification layer gave the most concise output concise output

Piper PA-28 Cadet Engine Piper PA-28 Cadet Engine Cowling ResultsCowling Results

In-Flight Test Set-UpIn-Flight Test Set-Up

Piper Cadet In-Flight Data SOM Piper Cadet In-Flight Data SOM Output (Channels 1&2)Output (Channels 1&2)

25-Sept-97(Chan. 1&2)Manuever Total hits Crk hits % Rub hits % Plst. hits %

Taxi 16373 771 4.7% 502 3.1% 15099 92.2%Take-Off 16734 496 3.0% 648 3.9% 15229 91.0%Climb Out 234 0 0.0% 233 99.6% 1 0.4%

Steady Level Flight 132 0 0.0% 132 100.0% 0 0.0%Final/Touch and Go 287 1 0.3% 285 99.3% 1 0.3%

10-Oct-97(Chan. 1&2)Manuever Total hits Crk hits % Rub hits % Plst. hits %

Taxi 16374 1674 10.2% 104 0.6% 14594 89.1%Take-Off 12336 753 6.1% 278 2.3% 11304 91.6%Climb Out 1086 0 0.0% 1085 99.9% 1 0.1%

Steady Level Flight 2174 189 8.7% 1830 84.2% 156 7.2%Final/Landing 16373 3004 18.3% 1334 8.1% 12030 73.5%

Fatigue Crack (Channels 1&2)Fatigue Crack (Channels 1&2)

Piper Cadet In-Flight Data SOM Piper Cadet In-Flight Data SOM Output (Channels 3&4)Output (Channels 3&4)

10-Oct-97(Chan. 3&4)Manuever Total hits Crk hits % Rub hits % Plst. hits %

Taxi 16373 1323 8.1% 3 0.0% 15036 91.8%Take-Off 2257 195 8.6% 514 22.8% 1548 68.6%Climb Out ----- ----- ----- ----- ----- ----- -----

Steady Level Flight 1301 0 0.0% 1301 100.0% 0 0.0%Final/Landing 16373 3212 19.6% 2940 18.0% 10220 62.4%

7-Oct-97(Chan. 3&4)Manuever Total hits Crk hits % Rub hits % Plst. hits %

Taxi 16374 2942 18.0% 20 0.1% 13410 81.9%Take-Off 982 167 17.0% 527 53.7% 287 29.2%Climb Out 1366 0 0.0% 1366 100.0% 0 0.0%

Steady Level Flight ----- ----- ----- ----- ----- ----- -----Final/Landing 16373 3261 19.9% 1205 7.4% 11906 72.7%

Fatigue Crack (Channels 3&4)Fatigue Crack (Channels 3&4)

Output ObservationsOutput Observations Fatigue crack growth unexpectedly Fatigue crack growth unexpectedly

detected on both sides of the aircraft detected on both sides of the aircraft cowlingcowling

Inspection revealed cracking between Inspection revealed cracking between Channels 3 - 4 as well as 1 - 2Channels 3 - 4 as well as 1 - 2

Cracking in the engine cowling occurred Cracking in the engine cowling occurred predominantly during ground predominantly during ground operations: taxi, take-off, and final operations: taxi, take-off, and final approach/landingapproach/landing

Cessna T-303 Crusader Cessna T-303 Crusader Vertical Tail ResultsVertical Tail Results

Equipment Setup (Flight)Equipment Setup (Flight)

Cessna Crusader In-Flight Data Cessna Crusader In-Flight Data SOM OutputSOM Output

Maneuver Crack Events Crack % PD Events PD % MN Events MN %Taxi 0 0.0% 963 93.0% 73 7.0%

Takeoff 69 14.4% 250 52.3% 159 33.3%Flight 277 23.3% 515 43.3% 398 33.4%Flight 1 0.1% 782 66.1% 400 33.8%Flight 7 3.5% 104 52.5% 87 43.9%

Dutch Roll 10 3.9% 113 44.5% 131 51.6%Roll 40 5.1% 383 48.7% 364 46.3%Roll 212 71.9% 52 17.6% 31 10.5%

Dutch Roll 204 73.9% 42 15.2% 30 10.9%Flight 472 64.1% 160 21.7% 104 14.1%

Landing 11 2.3% 331 69.0% 138 28.8%Taxi 0 0.0% 895 93.7% 60 6.3%

Conclusions & Conclusions & RecommendationsRecommendations

ConclusionsConclusions SOM trained successfully to classify fatigue cracking, SOM trained successfully to classify fatigue cracking,

plastic deformation, and rubbing noisesplastic deformation, and rubbing noises Fatigue crack growth successfully detected in-flight Fatigue crack growth successfully detected in-flight

from both engine cowling of the Piper PA-28 Cadet from both engine cowling of the Piper PA-28 Cadet and vertical tail of the Cessna T-303 Crusader using and vertical tail of the Cessna T-303 Crusader using AE parameter dataAE parameter data

Engine cowling fatigue cracking occurred mostly Engine cowling fatigue cracking occurred mostly during ground-based operations while vertical tail during ground-based operations while vertical tail fatigue cracking occurred predominantly in-flight, fatigue cracking occurred predominantly in-flight, especially during rolls and Dutch rollsespecially during rolls and Dutch rolls

In-flight crack detection systems should help to In-flight crack detection systems should help to minimize maintenance costs and extend the service minimize maintenance costs and extend the service lives of aging aircraft.lives of aging aircraft.

Problem: Data Overlap in AE Problem: Data Overlap in AE Parameter PlotsParameter Plots

Fatigue Crack WaveformFatigue Crack Waveform

-100

-80

-60

-40

-20

0

20

40

60

80

100

-50.0 50.0 150.0 250.0 350.0 450.0

Duration (s)

Am

plit

ude

(dB

)

Plastic Deformation WaveformPlastic Deformation Waveform

-100

-80

-60

-40

-20

0

20

40

60

80

100

-50.0 50.0 150.0 250.0 350.0 450.0

Duration (s)

Am

plit

ude

(dB

)

Source Location PlotSource Location Plot

Second hump indicates two Lamb wave types:Second hump indicates two Lamb wave types:• Symmetric (sSymmetric (s00) – plane strain crack) – plane strain crack• Antisymmetric (aAntisymmetric (a00) – plane stress crack) – plane stress crack

Tearing Cracks Mode III

Tensile CracksMode I

Mixed Mode Cracks

Average Frequency vs RA Average Frequency vs RA Value PlotValue Plot

RecommendationsRecommendations

Problem: AE parameter data overlapProblem: AE parameter data overlap

Possible solutions:Possible solutions:• Frequency analysis of waveforms using Fast Fourier or Frequency analysis of waveforms using Fast Fourier or

Wavelet TransformsWavelet Transforms• Symmetric and antisymmetric Lamb waves separated Symmetric and antisymmetric Lamb waves separated

using average frequency (fusing average frequency (favgavg = C/D) vs. RA parameter = C/D) vs. RA parameter

(RA = RT/A) plot(RA = RT/A) plot• High fidelity broadband AE sensors needed for High fidelity broadband AE sensors needed for

frequency analysis and Lamb wave identificationfrequency analysis and Lamb wave identification