Transcript www.auravector.com

Acoustic Emission Fatigue Life Prediction in Bridge Steel
Eric v. K. Hill, Andrej Korcak, Jamil Suleman and Fady F. Barsoum
I-35W Mississippi River Bridge in Minneapolis
Collapsed due to Fatigue of Gussett Plate
OBJECTIVES
 Monitor acoustic emission (AE) activity during
fatigue testing of notched tensile and I-beam
specimens
 Use Kohonen self-organizing map (SOM) neural
network to classify AE data from plastic
deformation (ahead of crack tip), plane strain and
plane stress fracture in notched specimens
 Employ back propagation neural network (BPNN)
to predict fatigue lives from first (0-25%) and third
(50-75%) quarter AE amplitude histogram data
 Goal: Predict fatigue or cyclic life to failure with
worst case error within ±15%
APPROACH/TECHNICAL CHALLENGES

Monitor AE from cyclic fatigue crack growth in ten
12x1x0.25 inch tensile specimens and ten S4x7.7
I-beams, all made from notched A572-G50 steel
Train BPNN on 0-25% and 50-75% AE data for 7
tensile specimens and test (predict fatigue life) on
remaining 3; train on 6 I-beams and test on 4
ACCOMPLISHMENTS/RESULTS
• SOM successfully classified AE fatigue cracking
data into plastic deformation, plane strain and
plane stress fracture mechanisms
• BPNN worst case errors for 0-25% and 50-75% AE
data, respectively: tensile specimens, -19.4% and

Cause of Failure: Fatigue of Undersized
Gussett Plate
Fatigue Testing of Notched A572-G50 Bridge Steel
Tensile Specimens in MTS Machine
Fatigue Specimen
with AE Transducers
Plane Stress vs. Plane
Strain Fracture
Failure Mechanism Occurrences
as a Function of Time
Pocket AE Analyzer
AE Waveform Parameters
SOM Failure Mechanism Classifier
I-Beam Fatigue Testing Using MTS Hydraulic Actuator

10 Transversely loaded I-Beams (S4x7.7)

3 Point bending

1.0 Hz, 0-3,800 lbf @ center of 112 inch span

A572-G50 bridge steel

0.10 inch deep V-notch on bottom flange
V-Notch on Bottom Flange along with
Extensometer and AE Transducer
I-Beam in 3-Point Bending with MTS Hydraulic Actuator
Applying 1 Hz Cyclic Loading (0-3,800 lbf)
Fatigue Crack Failure through Bottom
Flange then through Web Section
Back Propagation Neural Network (BPNN)


NeuralWorks Professional/II Plus software

71 input neurons for AE amplitude
histogram data and 1 for actual fatigue
life

Variable number of hidden layer
processing elements or neurons

Output neuron for predicted fatigue life
or cycles to failure
10 notched I-beams used for high cycle fatigue
(HCF ≥ 10,000 cycles) life prediction

0-25% and 50-75% AE amplitude
histograms used as BPNN input

Train on AE data from 6 beams; test
(predict fatigue life) on AE data from
remaining 4 beams
0-25%
50-75%
50-75%
Data
Worst
Case
Error
BPNN Parameters
Trained HCF BPNN
# of Hidden Layer Neurons
3
Hidden Layer Learning Coefficient
0.08
Output Layer Learning Coefficient
0.05
Transition Point
1000
Learning Coefficient Ratio
0.5
F' Offset
0.1
Momentum
0.4
Learning Rule
Norm-Cum-Delta
Transfer Function
TanH
Sample
Beam 10
Beam 12
Beam 13
Beam 14
Fatigue Life Predicted % Error
19188
19550
1.9
15833
16485
4.1
15710
16419
4.5
16084
16362
1.7