You are on page 1of 20

Probabilistic model for rebar-

concrete bond failure mode


prediction considering corrosion

Benjamin (Ahmad) Soraghi, Ph.D. Candidate, as481@zips.uakron.edu


Qindan Huang, Associate Professor, qhuang@uakron.edu
Department of Civil Engineering, University of Akron
 Background and motivation
 Research objectives
 Methodology
o Experimental
• Crack pattern
o Analytical
 Results and conclusions
Background and motivations

 Corrosion of infrastructure (bridges, dams, and buildings) has a


significant impact on safety, economic and financial health of structures
 Direct cost of corrosion for highway bridges is $13.6 B
Cracking
Corrosion in a jetty structure
Load carrying capacity
 Corrosion Spalling
Failure mode
Deterioration in bond
between rebar-concrete
Bond stress, τ(s)

τmax
Pull-out τu,split

Pull-Out
Splitting
τf

Slip (s)
Splitting s1,split s1 s2 s3
Background and motivations (cont.)

 Bonding between rebar and concrete is critical to ensure the reliable performance of RC structures, and it is
affected by
• Concrete cover • Rebar geometry
• Transverse reinforcement • Concrete properties

 Bond behavior is also recognized to be fail either in pull-out or splitting failure, which is important in
determine the performance of RC structures

 Pull-out failure:
Due to shearing of concrete between ribs, when there is
adequate concrete cover Pull-out

 Splitting failure:
Due to lack of cover or confinement to attain full pull-out
strength Splitting
Research objectives

Since,
 Bond failure mode changes structural behavior
 Existing available deterministic models (e.g., ACI and FIB-CEB criteria) are not accurate enough for
failure mode prediction
 No probabilistic formulation is available to predict bond failure mode.

Research goal of this study:


To develop a prediction model for bond failure mode considering various factors including
corrosion
Methodology

Experimental Analytical

 Specimen design  Failure mode classifications


 Specimen preparation, curing, • Logistic
and corroding • Lasso logistic
 Experiment set-up and Loading • Discriminant
• Decision tree
• k-Nearest
• Naïve Bayes
• Random Forest
• Support Vector Machines
Methodology
Experimental (Specimen Design)

 Design parameters for beam-end specimens


o Concrete strength, f’c
• f‘c (30, 40, and 50 MPa)
o Ratio of cover size to diameter of rebar C/d
• C/d (1.5 ∼ 4.8)
o Corrosion level Q
• 12 intact (0% corrosion) & 22 corroded (5% ∼ 20%)
o Loading type (monotonic or cyclic)
• 18 monotonic & 26 cyclic
o Confinement provided by transverse reinforcement, Ktr
Beam-end specimen configuration based on ASTM
A944-10 • 22 w/ reinforcement Ktr = (3.68 ∼ 5.89)
• 22 w/o reinforcement Ktr = 0
Methodology
Specimen preparation, curing, and corroding

The process of constructing the specimens


are:
 Making rectangular same-sized forms
 Making specified hole with respect to
rebar sizes (#5, #6, and #8)
 Making specified cover sizes
 Implementing bond regions (3.5”, 4.5”,
and 6”) using PVC pipes as bond brake
 Adding the reinforcements
o Parallel rebar
o Longitudinal rebar
o Transverse stirrups
 Concrete casting
Methodology Challenges:
 Time dependent corrosion
Specimen preparation, o Using the technique “accelerated
curing, and corroding
corrosion” to expedite corrosion
o Using power supplies
 Predicting the corrosion amount
 Providing enough humidity
 Drop of current as corrosion occurs and rust
appears around the test bar
o Designing a parallel circuit to keep the
voltage constant
o Installing light switches to be able to
monitor, track and control current
without blocking it from other
specimens that need to corrode more
 Accumulation of carrion in one side of the
test bar as the direction of the current is from
one side
o Installing light switches to the front and
back of each specimen to control the
current from both direction
Methodology  Type of loading to simulate seismic
loading of an earthquake, especially
Experiment set-up and
when corrosion is present
loading
o Monotonic
o Cyclic
 failure of a concrete specimen
Cracking pattern
o Pull-out
o Splitting
 Cracking is a random inevitable phenomenon
 No matter if the concrete is reinforced, what is cover size, what is the
strength of the concrete, in presence or absence of Q
 Understanding and recognizing these cracks and the way they will
affect the structure is of great importance.
testing 132 specimens, showed a
pattern forms on the concrete
associated with the failure mode of
the member.

 Splitting failure
o Starts from the center of
the rebar
o Go through the cover
o Breaking bond make the
concrete detached from
itself
 Pullout failure
o Cracks surrounding the
main rebar
o Attached to rebar like a
circular shape forming
some kind of
confinement even after
failure of the bond
Splitting failure pattern

PULL-OUT FAILURE PATTERN


Methodology (Analytical)

Failure mode
classifications
 Logistic  Classification algorithms are used for Bond stress, τ(s)

 Lasso logistic categorical outputs while the regression


algorithms are used for numerical
 Discriminant
predictors τmax
 k-Nearest  The predictors in this research are failure τu,split
 Naïve Bayes mode, pull-out failure and splitting failure Pull-Out
 Support Vector  Those classification algorithms have been Splitting
Machines adopted which are relevant to the objective τf
 Decision tree of this paper
Slip (s)
 Random Forest s1,split s1 s2 s3
Methodology Logistic classification Lasso classification
 Advantages  Advantages
o Suitable for categorical o Suitable for categorical
Failure mode
responses responses
classifications
(Cont.) o Explicit formulation o Explicit formulation
o Practical application for o Practical application for
engineering purposes engineering purposes
 Disadvantages o Constraint the coefficients
o Dependent on the residuals  Disadvantages
o Not efficient for large o Dependent on the cross
samples validation folds

 
N
l (β)   y j βT x j  log 1  exp  x j β    
N p

j 1
l (β)    y j β X j  ln  (1  exp  X j β      | i
T
|
coefficient j 1  i 0 
s
constraint

Probability of failure Y being pull-out exp   0   i xi 


p(x )  Pr(Y  1 | X  x ) 
1  exp   0    i xi 
Probability of failure
Model intercept Model coefficients
Vector of input variables
Methodology Discriminant classification K-nearest neighbors Naïve Bayes classification
 Advantages  Advantages  Advantages
o Suitable for o Non parametric o Robust to isolate
Failure mode categorical method noise points
classifications responses o Simple o Robust to
(Cont.) o Uses the Gaussian o No training time irrelevant attributes
distribution o Handle missing
 Disadvantages
 Disadvantages values
o Memory intensive
o Complex  Disadvantages
o Estimation is slow
distribution o Loss of accuracy
formulation o Relates on k
(assuming class
o Not good for conditional
insufficient independence)
observations
Prior probability Class m

Conditional
probability
Neighbor k
Methodology Support Vector Machine Decision trees Random forest
 Advantages classification classification
Analytical
 Effective on large  Advantages  Advantages
Failure mode sample size o Easy to o Producing accurate
classifications
 Uses a subset of understand classifier
(Cont.)
training points in o Easy to generate  Disadvantages
decision function rules o Overfitting for
 Disadvantages  Disadvantages noisy classification
 Not efficient for o Overfitting o Large number of
data including o Classifying by trees slow down
noises rectangular the algorithm
 Non probabilistic partitioning

Positive real constants

Hyper-plane line
Individual decision trees
Methodology  Model Accuracy
 y  p(x j ) 
2
SSE  j
o Sum of squared of error, SSE
Actual failure
Predicted failure

o Hit/Miss method which is suitable when the response is a


categorical variable
Correct predicted
p(x)  50%  Pull-out nTS  nTN
p(x) < 50%  Splitting PCD 
nTS  nTN  n FP
Total predicted

 Model Development
o Conducting a preliminary study to select potential xi.
o Variables that show the potential influences on the failure mode
(Y) are: C/d, Ktr, Q, and MC, where MC is a dummy variable and
defined as xi
0 for specimens under monotonic loading
MC   Ktr, C/d, Q, MC      
1 for specimens under cyclic loading
KtrC/d, C/dQ, QMC, KtrMC, C/dMC, KtrQ

o KtrC/dQ, C/dQMC, KtrQMC, KtrC/dMC


Examining interaction terms
KtrC/dQMC
Methodology

 Model Selection
o A full model with a model size of 15
o Applying model selection to the full model to eliminate variables
that do not have contribution to the model prediction
o Removing models having p-values larger than 10% and VIF larger
than 10
o Measuring model quality using R-sq, Adj-R-sq, AIC, BIC

Model coefficients for the logistic regression model Model coefficients for lasso logistic regression

β0 β1 β2 β3 β0 β1 β2 β3
Model coefficients Model coefficients
(Intercept) (MC) (Ktr·C/d) (C/d·Q) (Intercept) (C/d) (Ktr·C/d) (MC·Q)
Mean -3.06 2.41 0.15 9.43 Mean -3.56 0.70 0.10 53.35
Standard deviation 1.13 1.07 0.08 4.59 Standard deviation 0.41 0.10 0.01 6.18
Coefficient of variance -0.37 0.44 0.53 0.49 Coefficient of variance -0.12 0.14 0.10 0.12
Model β0 β1 β2 β3 SSE PCD
Results Coefficients
& Logistic (Intercept) (MC) (Ktr· C/d) (C/d· Q) 2.03 88%
Conclusions
Lasso logistic (Intercept) (C/d) (Ktr· C/d) (MC· Q) 1.5 94%

k-Nearest Discriminant Naïve Bayes SVM Decision Tree Random Forest


PCD 74% 84% 66% 74% 89% 87%
1 1

Prediction
True

 Logistic and Lasso comparison


Probability of Failure Mode
Probability of Failure Mode

0.5 0.5 o 1, 2, and 3 show that the likelihood of being pull-
out increases of the associated term
Prediction
True o both models indicate that the failure mode leans to
0 0

1 1
pull-out failure when more confinement (e.g., the,
Prediction plot based on logistic regression model Prediction
True C/d, and transverse stirrups) is provided, or Q
increases, or structure subjected to cyclic loading
Probability of Failure Mode
Probability of Failure Mode

0.5 0.5
o The predictions using lasso logistic classification are
closer to actual failure mode compared to logistic
0
Prediction
True
0
classification

Pull-out failure
Prediction plot based on lasso logistic Splitting
regressionfailure
model

You might also like