Analysis of FWD Data from the National Airport Pavement Test Facility NAPTF: Neural NetworkBased Alg

1
Analysis of FWD Data from the National Airport
Pavement Test Facility (NAPTF) Neural
Network-Based Algorithms vs. Traditional Approach
Dr. Halil Ceylan, Assistant Professor Dr.
Kasthurirangan Gopalakrishnan, Postdoctoral
Research Associate Alper Guclu, Graduate Research
Assistant Mustafa Birkan Bayrak, Graduate
Research Assistant Iowa State University
Pavement Research Team

• 2005 FWD USERS GROUP CONFERENCE
• Austin, Texas October 15-18, 2005

2
FAAs NAPTF

• Generate full-scale test data for performance
  analysis of pavement subjected to New Generation
  Aircraft
• Facility was dedicated on April 1999
• First series of tests (CC1) between Sep. 1999 and
  July 2001
• Boeing 777 gear trafficking (North side)
• Boeing 747 gear trafficking (South side)

3
NAPTF - Plan View
4
NAPTF Flexible Test Sections


5
FWD Tests at NAPTF

• 9-kip FWD
• Document uniformity
• Prior to traffic testing
• Diameter 12-inch
• Pulse width 27-30 msec
• 7 geophones
• (D0 _at_ 0 to D5 _at_ 60)

6
FWD Test Locations
Same layout for each test section
STATIONS
7
Layer Moduli Backcalculation

• Required for mechanistic pavement analysis and
  design
• Good indicator of pavement layer condition
• Particularly appealing for characterizing
  subgrade soils
• AI Design Manual MS-1 recommends ERi as input for
  ELP analysis

8
Layer Moduli Backcalculation

• BAKFAA
• WESDEF
• ANN prediction models

9
Backcalculation - BAKFAA

• FAA Airport Technology Branch
• Based on LEAF ELP
• Moduli are adjusted to minimize the RMS of the
  differences between FWD sensor measurements and
  the JULEA-computed deflection basin
• Stiff layer (E 1 x 106 psi, ? 0.5)
• At 10 ft. for the medium-strength sections
• At 12 ft. for the low-strength sections

10
Backcalculation - BAKFAA
11
Backcalculation - WESDEF

• Originally developed by Van Cauwelaert et al.
  (1989)
• Based on WESLEA ELP
• Utilizes an iterative procedure to obtain a set
  of moduli that, when used in WESLEA will produce
  deflections similar to the measured values
• Ability to backcalculate moduli values using
  deflections with depth (e.g., MDD)

12
Backcalculation - WESDEF
13
Backcalculation ANN Approach

• Rapid, real-time moduli backcalculation
• Unbound granular layer non-linearity at NAPTF
• ILLI-PAVE (incorporates stress-dependent material
  models)
• Multi-layer, feed-forward network trained to
  approximate the FWD backcalculation function

14
ANN Models Input Ranges
15
Subgrade Model
Thompson and Elliott (1985)
?


-

?

• M

K
K
K
(
lt K
)
when
d
2
R
d
1
3
2

-
-
?
?

• M

K
K
K
(
)
gt K
when
d
2
R
d
1
4
2
K3
Resilient Modulus,
where ?d ?1 - ?3
K4
K2
K1
Deviator Stress,
16
Unbound Aggregate Base Model
450
400
K-? model
350
Uzan model
Crushed Stone
300
n
250
q
Resilient Moduli, MR (MPa)

K
M
200
R
p
o
150
K
K
3
2
Sand
s
q
100
d

K
M
1
R
p
p
50
o
o
0
0
1000
200
400
600
800
Bulk Stress, q (kPa)
17
Mechanistic Based Pavement Design
Structural Model
Structural Model
Structural Model
INPUTS
INPUTS
INPUTS
Pavement Responses
Materials Characterization
s, e, D
s, e, D
s, e, D
Paving Materials
Paving Materials
Paving Materials
Soils, Aggregates
Soils, Aggregates
Soils, Aggregates
Performance Prediction
Traffic
Traffic
Traffic
Transfer Functions
Transfer Functions
Transfer Functions
Climate
Climate
Climate
Design Reliability
Design Reliability
Design Reliability
Final Design
Rehabilitation
Pavement Distress
Evaluation
Evaluation
Evaluation
18
ANN Architecture

• 28,500 data Training
  Phase
• 1,500 data Testing
  Phase
• ANN
  Eac Model
• 8 - 60 - 60 - 1
• ANN
  Eri Model

Do-D12-D24-D36-D48-D60 Hac - Hbase
19
Eac Predictions (ANN vs. ILLI-PAVE)
20
ANN Model Eac Prediction Curve
21
Eri Predictions (ANN vs. ILLI-PAVE)
22
Results MFCAC Modulus
23
Results MFCAC Modulus
24
Results MFSAC Modulus
25
Results MFSAC Modulus
26
Results LFCAC Modulus
27
Results LFCAC Modulus
28
Results LFSAC Modulus
29
Results LFSAC Modulus
30
Results MFCSubgrade Modulus
31
Results MFCSubgrade Modulus
32
Results MFSSubgrade Modulus
33
Results MFSSubgrade Modulus
34
Results LFCSubgrade Modulus
35
Results LFCSubgrade Modulus
36
Results LFSSubgrade Modulus
37
Results LFSSubgrade Modulus
38
Backcalculation Results
39
Subgrade Lab Resilient Modulus
Gopalakrishnan (2004)
40
Subgrade Lab Resilient Modulus
Gopalakrishnan (2004)
41
Summary Findings

• Material non-linearity at NAPTF
• Need for a suitable structural model for analysis
  and backcalculation
• Major benefits of ANN approach over conventional
  approach
• Ease of analysis
• Real-time results
• Ability to calculate critical pavement responses
  directly from the FWD deflection data MEPDG
  integration

42
Summary Findings

• The predicted AC moduli (ANN) is consistent with
  the results from the WESDEF and BAKFAA.
• The predicted subgrade moduli (ANN) is lower than
  the results from the WESDEF and BAKFAA.
• ANN subgrade moduli predictions agree well with
  the laboratory resilient modulus test results

43
Advantages of ANN Approach

• Backcalculation of stress-dependent pavement
  layer properties in real time
• Backcalculation of critical pavement responses
  directly from the FWD deflection basins in real
  time
• Prediction of critical pavement responses and
  pavement surface deflections for a given pavement
  system forward analysis

44
Advantages of ANN Approach

• Pavement engineers designers provided with the
  state-of-the-art finite element (ILLI-PAVE 2000,
  GT-PAVE, etc.) solutions
• No complex input, no seed moduli, no moduli range
  requirements
• No large computer resources needed
• Computation performed practically in real time
• (50,000 analyses in 1 sec.!)

45
THANK YOU!
Questions ?