Content uploaded by M.Z. Naser
Author content
All content in this area was uploaded by M.Z. Naser on Jan 29, 2021
Content may be subject to copyright.
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
1
MACHINE LEARNING-DRIVEN ASSESSMENT OF FIRE-INDUCED
1
CONCRETE SPALLING OF COLUMNS.
2
M.Z. Naser and Hadi Salehi
3
Biography:
4
M.Z. Naser is an assistant professor at the Glenn Department of Civil Engineering at Clemson
5
University. Dr. Naser is an active member in three ACI committees (216 – Fire Resistance and
6
Fire Protection of Structures, ACI 133 – Disaster Reconnaissance, and 440 – Fiber-Reinforced
7
Polymer Reinforcement). He is a registered professional engineer in the state of Michigan. His
8
research interests span over structural fire engineering, computational intelligence and
9
extraterrestrial construction. (email: mznaser@clemson.edu).
10
Hadi Salehi is a postdoctoral research fellow in the Department of Aerospace Engineering at
11
the University of Michigan, Ann Arbor, Michigan. Dr. Salehi received his MS and PhD in
12
Structural Engineering from Michigan State University. His research interests include data
13
analytics and machine learning for data-driven decision making and smart
14
infrastructure/aerospace monitoring. (email: hsalehi@umich.edu).
15
16
ABSTRACT
17
The past few years have witnessed the rise of serious research efforts directed towards
18
understanding fire-induced spalling in concrete. Despite these efforts, one continues to fall
19
short of arriving at a thorough examination of this phenomenon and of developing a modern
20
assessment tool capable of predicting the occurrence and intensity of spalling. Unlike other
21
works, this papers presents an approach that leverages a combination of machine learning (ML)
22
techniques; namely k-nearest neighbor (k-NN) and genetic programming (GP), to examine
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
2
spalling in fire-tested reinforced concrete (RC) columns. In this analysis, due diligence was
1
taken to examine 11 factors known to influence spalling and to identify those of highest impact
2
to be then used to develop a predictive tool. The outcome of this analysis shows that it is
3
possible to predict the occurrence of spalling (with a successful rate ranging 77%- 90%.)
4
through a simple, robust, and easy to use ML-driven tool.
5
6
Keywords: fire-induced spalling; pattern recognition; machine learning; artificial intelligence;
7
k-nearest neighbor; principal component analysis.
8
9
INTRODUCTION
10
The inert thermal diffusivity of concrete facilitates slow rise in cross-sectional temperature and
11
moderate loss in strength and stiffness of concrete structural members, thus permitting the use
12
of concrete materials without special treatment/handling in severe working conditions i.e.
13
power/chemical/nuclear plants.1 As such, it is a common practice not to externally fire-proof
14
concrete structures given that a sufficient cover to embedded reinforcement is provided.2 This
15
practice continues till this present day despite observations from recent fire tests as well as
16
post-fire on-scene investigations reporting the tendency of concrete to spall.3,4
17
Fire-induced spalling is generally defined as the breakage of concrete chunks or cover due to
18
thermally-induced effects.5 In the case that spalling occurs, the integrity of a concrete member
19
is adversely threatened on two fronts. The first is related to the fact that breakage of concrete
20
cover exposes steel (or in some cases fiber reinforced polymer (FRP)) reinforcement to direct
21
flames and heat – causing severe degradation in mechanical properties and development of
22
higher core (internal) temperatures in RC members. The second being; any loss in cross
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
3
sectional size would also lead to a reduction in load bearing capacity (i.e. moment capacity in
1
beams etc.). Unfortunately, the above two fronts are seldomly accounted for during the design
2
stage of a concrete structure and this negligence has been shown to trigger unexpected
3
failure/collapse mechanisms due to spalling in the event of a fire breakout.6,7
4
As such, a number of studies have alluded to the notion that without properly addressing the
5
phenomenon of fire-induced spalling, then an accurate evaluation of fire performance of
6
concrete structures may not be fully realized.8,9 The same studies also pointed out that the
7
complexity and randomness of spalling may limit the ability of practicing engineers to arrive
8
at optimal designs for RC structures. This is especially true for those of unique/complex
9
functionality such as mega buildings, bridges and tunnels where fire is considered a major
10
threat.10
11
On a parallel note, this incompetence in properly evaluating fire-induced spalling also hinders
12
ongoing standardization efforts aimed at promoting performance-based fire design solutions
13
and in a way handicaps future developments in this niche research area. It is unfortunate to note
14
that currently adopted (prescriptive) codal provisions falls short of providing an adequate
15
guidance to mitigate or to account for spalling.2,11 While such provisions do provide tabulated
16
listings that can be used to estimate the required concrete cover thickness to satisfy a given fire
17
rating, the same listings were not derived to account for the adverse effects of spalling. As such,
18
the application of these listings is limited in practical applications, as well as in scenarios
19
utilizing modern types of concretes i.e. high strength concrete (HSC) and ultra-high
20
performance concrete (UHPC).
21
Fire-induced spalling can potentially be classified under explosive (violent) spalling and
22
surficial spalling.12,13 Spalling is often explained through either: 1) development of pore
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
4
pressure build up facilitated by moisture migration in temperature range of 250-420°C; which
1
once exceeds the tensile strength of concrete leads to spalling, or 2) thermomechanical
2
processes (i.e. thermal dilation/shrinkage gradients) that occur within heated RC members.14–
3
16
4
However, it is interesting to note that there is enough experimental evidence to verify the
5
validity, as well as to contradict the principles of the aforementioned theories.17,18 For example,
6
Harmathy16 noted that spalling of concrete occurs early into fire exposure (within the first
7
25 minutes), yet tests carried out by Han et al.19 reported occurrence of spalling during later
8
stages (60-90 minutes) of fire exposure and in few incidents during the cooling phase as well.20
9
While the development of high thermal gradients (due to rapid heating) has been shown to
10
yield high risk of spalling, Noumowe et al.21 reported severe spalling in uniformly heated
11
concrete at a low rate equivalent to 0.5°C/min. On a separate front, Kalifa et al.22 noted the
12
positive effect of incorporating polypropylene fibres in minimizing spalling, however
13
Klingsch18 reported that incorporating such fibers did not positively reduce spalling in concrete.
14
The effect of non-uniform heating was reported to cause spalling in RC columns tested by Raut
15
and Kodur23, while the same effect was shown not to cause any spalling in tests carried out by
16
Xu and Wu24. Another aspect that needs to be remembered is that the majority of the above
17
works investigated fire-induced spalling either through experimentation, or theoretical
18
derivation/numerical simulation.17,25–27 As such, these are confined to specific concrete
19
mixtures and testing set-ups and hence are hard to replicate and lack thorough verification due
20
to the lack of sample duplication.
21
Building upon the collective findings of past and recent works, combined with the fact that
22
fire-induced spalling is believed to be triggered by a complex chain of reactions involving
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
5
multi-dimensional parameters, this study aims at leveraging modern techniques in order to
1
examine the influence of geometric, material and loading features on the susceptibility of RC
2
columns to spall under fire conditions. Due to the randomness of spalling, understanding this
3
phenomenon can better be attained via an ML-driven perception. Unlike other works in which
4
AI-derivatives were primarily used to optimize concrete mix proportions28–30 or predict
5
properties of concrete31–34, this study applies contemporary approaches that falls under artificial
6
intelligence (AI) and machine learning (ML) i.e. pattern recognition (PR) and genetic
7
programing (GP) to identify the hidden relations that govern both the occurrence as well as
8
magnitude of fire-induced spalling. The identified critical parameters are then used to develop
9
a simple and easy-to-apply GP-powered assessment tool that can accurately predict propensity
10
and magnitude of spalling in fire-exposed RC columns. This tool indirectly takes into account
11
how material properties of concrete and reinforcement vary with temperatures, is freely
12
available, can be continuously upgraded and hence is attractive for both researchers and
13
practitioners.
14
15
RESEARCH SIGNIFICANCE
16
The use of ML has been steadily rising over the past few decades; thanks in part to the rapid
17
advancements in computing and data analytics. Due to the complex nature of fire-induced
18
spalling, this paper explores the potential of ML in examining the susceptibility of reinforced
19
concrete columns to spalling. This work leverages two distinct ML algorithms, i.e. PR and GP,
20
to showcase the merit in adopting ML as a modern technique, in parallel to traditional methods
21
e.g. testing and simulation. Adopting ML is expected to open the door towards research
22
opportunities encompassing performance-based fire design of concrete structures.
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
6
AN OVERVIEW TO FIRE-INDUCED SPALLING PHENOMENON
1
This study recognizes the multidimensionality of spalling as highlighted by past and recent
2
studies14–18,25,26,35–40 and in order to accommodate a collective view into this phenomenon,
3
spalling is said to be governed by factors that can be grouped under three categories: 1) material
4
characteristics, 2) geometric configurations, and 3) loading conditions. These factors are
5
further examined in the following subsections and are also summarized in Table 1. It should
6
be stressed that a more inclusive review on other key factors, including; cement type, degree
7
of pore saturation, fire cooling phase, with regard to spalling is spared for brevity and can be
8
found elsewhere.41–45
9
10
Material characteristics
11
The material characteristics that primarily govern spalling behavior comprises of concrete mix
12
components such as aggregate and cement type, additives (fibers, superplasticizers etc.), and
13
water/binder ratio. In the case of aggregates, carbonate aggregate often delivers a
14
comparatively better resistance to heat effects than other types of aggregates (e.g. silicate). This
15
can be attributed to the capability of carbonate aggregates to develop an endothermic reaction
16
at temperatures close to 700°C which reduces temperature rise and slows down strength
17
degradation.1 In the case of additives, incorporating steel fibers (~1.75% by weight) or
18
polypropylene fibers (~0.15% of volume) seems to minimize the extent of fire-induced
19
spalling.36,46 Using silica fume or limestone fillers is expected to increase spalling occurrence
20
as these lower permeability and limit vapor release. On a similar line, concretes with high
21
moisture content or water/cement ratio or those of low permeability also tend to spall due to
22
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
7
developing high pore pressure. Thermal incompatibility between mix components may also
1
promote spalling.5,38
2
3
Geometric size and reinforcement configuration
4
The geometric configuration of concrete components can also affect its susceptibility to fire-
5
induced spalling. For instance, edged, as opposed to round, components attract higher
6
magnitudes of heat via bi-lateral heat transmission which may facilitate spalling.5 Hertz38 and
7
Kanéma et al.47 reported that members of bigger sizes have increased tendency to spall as they:
8
1) hold larger amounts of moisture, and 2) can develop sharp thermal and pressure gradients
9
across their cross sections. The configuration of embedded steel reinforcement is another
10
governing factor to the phenomenon of spalling. In general, columns incorporating closely
11
space hooked ties (bent at 135°) do not seem to spall as much as columns with traditional or
12
spaced-out ties.44
13
14
Loading conditions
15
The arrangement and magnitude of loading and heating regimes also affect spalling behavior
16
of RC members. While the existence of axial or eccentric forces put a loaded member under a
17
constant state of compression; which in a way limits cracking development, stressed members
18
may also become susceptible to spalling due to the amplifying effects of pore pressure.5 From
19
a thermal point of view, a fire with a rapid and intense heating rate has the potential to thermally
20
shock concrete material and to develop large thermal gradients, causing high thermal stresses
21
and non-uniform expansion within exterior and interior layers of concrete thus promoting
22
spalling.
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
8
PATTERN RECOGNITION AND GENETIC PROGRAMMING MODEL
1
RATIONAL AND DATABASE DEVELOPMENT
2
This work formulates the following hypothesis, “if observations on fire-induced spalling are
3
collected from independent fire tests, then it is possible to intelligently tie such observations to
4
factors governing the phenomenon of spalling through a systematic analysis”. Given the
5
complexity and high dimensionality of successfully completing such analysis, a decision can
6
be justifiably taken to utilize ML as an exploratory engine. ML-based techniques mimic
7
human-like reasoning process in order to resolve phenomena that may not be truly understood
8
by means of conventional methods or may necessitate resource-intensive experimentations or
9
specialized computing software/workstations. Of these techniques, pattern recognition (PR)
10
and genetic programing (GP) are of interest to this work and hence are applied herein. These
11
techniques have been widely used during the last decade in various structural and fire
12
engineering applications.35,48
13
The followed investigation philosophy starts by collecting information on spalling observations
14
from standard fire tests. Thus, a thorough analysis of published fire tests49–55, together with
15
recommendations of notable works5,38,44,56, was carried out to pinpoint the critical parameters
16
that influence spalling as outlined earlier. These parameters include: 1) concrete type, fc, 2)
17
cross sectional size, W, 3) boundary conditions, BC, 4) tie spacing, S, 5) stirrup configuration,
18
SC, 6) steel reinforcement ratio, r, 7) aggregate type, A, 8) fiber type, f, 9) humidity, H, 10)
19
magnitude, P, and 11) eccentricity of applied loading, e.
20
This compiled observations are then put into a database. This database is examined using PR
21
to identify most critical parameters (out of all 11 collected parameters) that govern spalling
22
phenomenon. Once identified, then GP is applied to generate simple expressions that can be
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
9
used to predict the occurrence, as well as expected magnitude of spalling (see Fig. 1). These
1
expressions are finally encoded into a simple assessment tool. One can see that carrying out
2
the proposed ML analysis is quite dissimilar to traditional analytical/simulation methods as
3
these require inputting appropriate temperature-dependent constitutive models and
4
development of thermal and structural models.
5
It should be noted that it is due to the complex nature of fire testing, and lack of instruments
6
capable of surviving harsh temperatures or measure the magnitude of fire-induced spalling, this
7
phenomenon continues to be reported qualitatively (i.e. binary notation – spalling/no spalling
8
or descriptive notation – no, minor, major spalling) without being actually measured. This work
9
maintains the same notation in order to show the validity of the proposed framework. In more
10
details, if a fire test reported that a particular column underwent “minor spalling”, then this
11
column is also labeled to undergo “minor spalling”. This approach was followed since the
12
majority of selected fire tests did not report specific information pertaining to spalling
13
magnitude (i.e. average spalling depths or max spalling depths). A future work is currently in
14
its early stages to develop quantitative measurements of spalling through ML.
15
16
Pattern recognition (PR)
17
PR thrives to learns patterns hidden in varying dimensions of observations as to establish a
18
relation between input parameters and expected output(s).57,58 Among the different PR
19
techniques, k-nearest neighbors (k-NN) has been widely used in the field of damage detection
20
and condition assessment, and hence is applied herein.59 In general, k-NN is a non-parametric
21
classification algorithm belonging to the instance-based learning methods. Further, the k-NN
22
algorithm: 1) makes no assumption about the data distribution, thus yielding a flexible decision
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
10
boundary with minimum learning process and higher accuracy, 2) tends to be robust even with
1
noisy training data, 3) has the ability to learn complex concepts by local approximations, and
2
4) is easily implemented and fully tractable. As such, k-NN is used to analyze spalling-related
3
data collected from 87 fire tests (as 18 columns from the originally compiled database lacked
4
essential details on some input parameters and hence were deemed unsuitable for analysis).
5
The k-NN classification algorithm is formulated by assuming the pair (xi, φ(xi)) which denotes
6
the feature vector xi and its corresponding label φ(xi); where i=1, 2,…, n and φ{1, 2, …, m}
7
where n and m are the number of training feature vectors and the number of classes,
8
respectively. Considering xi as an arbitrary feature vector, the distance between this feature and
9
feature vector xj is calculated by:
10
),(),( ji xxfjid =
(1)
11
where is a distance function that can be defined as:
12
−−= )()(),( ji
T
jiji xxxxxxf
(2)
13
Equation (2) is the generalized distance, and for the case of = I it denotes the Euclidean
14
distance (T herein denotes the transpose function). The distance vector D(i) is defined by
15
Equation (3):
16
},...,2,1,,...,2,1|),({)( trai ntest njnijidiD ===
(3)
17
The D(i) vector is arranged in an increasing order (Dn(i),) and the k-nearest vote vector is
18
defined by using the first K elements as follows
19
))})(((),...,1)((({ KiDiDV nn
=
(4)
20
The classification is then performed by determining the k-nearest vote vector V. In this regard,
21
the test feature xi is classified to the class that has the most votes in V. In this procedure, the 11
22
parameters identified above e.g., compressive strength, width, steel ratio, etc., were considered
23
),( ji xxf
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
11
as pattern features. Thus, each pattern was represented with 11 features (z1, z2, z3,…, z11) i.e.,
1
the dimension of the PR problem herein is 11. To implement the k-NN algorithm, data set was
2
classified to two and three classes for binary and multiple spalling classification, respectively.
3
For the case of binary spalling classification, the dataset was split to ‘class N’ denoting patterns
4
with no spalling and ‘class S’ representing patterns of spalling. For the case of multiple spalling
5
classification, the dataset was classified to three classes, where ‘class N’ presents no spalling,
6
‘class MN’ represent patterns due to minor spalling, and class ‘MJ’ denotes patterns as a result
7
of major spalling. The dataset for the k-NN analysis was randomly classified into three subsets;
8
namely, training, validation, and testing. The training set was used to train the classifier, while
9
the validation set was used to compute the optimal k for the k-NN classifier. The best models
10
were selected based on their performance on the validation data. Performance of the classifier
11
with optimal k was then investigated on the test set.
12
13
Genetic programming (GP)
14
Fundamentally, GP follows the Darwinian philosophy of “survival of the fittest” to develop
15
solution candidates with high arrive at predictive capabilities. In this method, a population
16
comprising of candidate solutions is first randomly generated through arithmetic operators and
17
mathematical functions i.e. addition (+), trigonometric functions (i.e. tangent) etc.60 Suitable
18
solutions are then manipulated via operations such as mutation (randomly changing a the layout
19
of candidate) and/or crossover (combining two, or more, candidate solutions to get an improved
20
solution) – see Fig. 2. The theory and application of GP into fire-based problems have been
21
thoroughly documented in companion works and hence is avoided herein for brevity.35,61,62
22
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
12
Development of database
1
As discussed in a previous section, the proposed methodology requires compiling data on
2
spalling from fire tests. This section provides a brief description to such selected tests wherein
3
full details can be found at their respective references.14–18,25,26,35–40 For a start, the National
4
Research Council of Canada (NRCC) carried out numerous tests on columns made of different
5
concretes (normal, high, fiber-reinforced concrete) and various features (shape: square,
6
rectangular, and circular; cross-sectional size: 203 mm – 406 mm; ratio of longitudinal steel
7
rebars: 2-4%; aggregates: carbonate, siliceous and lightweight etc.).26,40,50 These tests proved
8
to be very valuable from the point of this work.
9
Another testing program was carried out by Hass.55 In their tests, RC columns of two sizes:
10
200×200 mm2 and 300×300 mm2 that were reinforced with two sizes of reinforcement
11
(diameter of 14 and 20 mm) were tested. Buch and Sharma63 tested 11 RC columns (six of
12
which were normal strength concrete columns (NSC) and five were made of HSC). All columns
13
were 3.15 m in height, had a square cross-section of 300×300 mm2, and were reinforced with
14
longitudinal and transverse reinforcements with an average yield tensile strength of 491.5-
15
499.5 MPa. These columns were tested to explore the effect of loading arrangement (i.e.
16
eccentricity) on spalling. Shah and Sharma17, Myllymaki and Lie52, Rodrigues et al.54 also
17
conducted fire resistance experiments on RC columns and varied restraint conditions, concrete
18
type, loading magnitude etc. Observations from all of the above tests were organized into a
19
database which can be accessed online.64
20
21
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
13
VALIDATION OF METHODOLOGY AND DEVELOPMENT OF SPALLING
1
ASSESSMENT TOOL
2
The developed database was then input into Matlab simulation environment for analysis. The
3
compiled tests were randomly shuffled in order to maintain an unbiased ML analysis. At the
4
beginning of the analysis, the initial classification accuracy of the PR analysis was low (i.e.,
5
around 60%) primarily due to the higher dimensionality of spalling phenomenon (i.e. 11
6
features/dimensions). Thus, incorporating ML optimization techniques was deemed necessary.
7
A data compressing technique is used to reduce the dimensions of feature space through finding
8
principal components (i.e. those of maximum variance for the dataset). This technique is
9
referred to as principal component analysis (PCA).65 Using this technique, the original input
10
feature vector (z) is projected on the first two principal components for single and multiple
11
spalling classification (see Fig. 3). As can be seen from this figure, the defined classes for all
12
cases (e.g., binary spalling and multiple spalling) overlap even using first two principal
13
components z=[z1,z2]t, thus resulting in a low classification accuracy.
14
Since these preliminary results indicate the necessity of using feature selection techniques to
15
enhance classification accuracy, sequential forward selection (SFS) and sequential backward
16
selection (SBS) were applied. The SFS feature selection method starts with an empty set of
17
features and adds the best feature z+ sequentially (from the set of full features) which gives the
18
highest value for the objective function J(Xk + z+). On the other hand, SBS feature selection
19
method starts with the full set of features and removes the worst feature z− sequentially that
20
gives the lowest value for the objective function J(Xk + z
−
) – see Fig. 4 and Fig. 5. It is worth
21
noting that k-fold cross validation technique was also used to prevent overfitting of the k-NN
22
algorithm. In this study, 10-fold cross validation (i.e. assuming k=10) was considered.
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
14
To visualize the outcome of PCA-based PR analysis, a confusion matrix that contains
1
information on actual and predicted classes is employed as the main metric to assess
2
performance of the carried out methodology. The performance of the damage detection model
3
with k-NN method was thus measured using the detection performance rate defined in Equation
4
(5). MATLAB was utilized for implementing k-NN algorithm and to compute confusion
5
matrices.
6
Damage Detection Accuracy = Number of patterns correctly classified /Total number of patterns (5)
7
Accordingly, the best feature vectors were selected by the SFS and SBS algorithms for binary
8
and multiple spalling classifications. To obtain the best classification performance, the
9
optimum value of k, i.e., number of neighbors for the k-NN algorithm, was determined through
10
computing the classification accuracy on training, validation and test data. It should be noted
11
that different combinations in terms of size of data subsets used for training, validation, and
12
testing were considered in this study for binary and multiple classification, and the performance
13
of the k-NN method was evaluated based on each combination. Yet, the presented results are
14
based on the combinations for which the accuracy of the k-NN was highest. On this basis, for
15
the scenario of binary spalling, 50% of data was used for classification and 10-fold cross
16
validation and 50% of data was used for testing. On the other hand, for the case of multiple
17
spalling classification, 60% of data was used for training and 10-fold cross validation and the
18
remaining 40% of dataset was used for testing the k-NN classifier.
19
The k-NN algorithm, without feature selection techniques, was initially used to determine the
20
classification accuracy on test data for both binary and multiple spalling classification (see
21
Table 2 and Table 3, respectively). These tables show that the optimal number of k for the
22
case of binary and multiple spalling classification was found as 5 and 12, respectively. Still,
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
15
results of k-NN classification with different number of neighbors using original set of features
1
and SFS and SBS feature selection methods are also plotted and presented in Fig. 6, where
2
classification results confirm the optimal number of k noted above.
3
In the next phase of analysis, the best features for the case of binary spalling were selected as
4
compressive strength, eccentricity, and humidity, as well as width, stirrup spacing, eccentricity,
5
humidity, and load level using SFS and SBS methods, respectively. For the multiple spalling
6
scenario, best features using SFS algorithm were chosen as compressive strength and additive
7
fibers type, whereas SBS algorithm resulted in similar features as for the case of binary spalling.
8
As noted above, a confusion matrix containing information regarding actual and predicted
9
classes/patterns was used to explore the performance of the k-NN algorithm. The diagonal
10
entries of the confusion matrix denote the spalling cases that are correctly classified. Also,
11
entries in the off-diagonal cells represent the spalling cases that are misclassified. Confusion
12
matrices on test data based on optimal k for both binary and multiple spalling with original
13
feature sets and SFS and SBS feature selection algorithms are presented in Table 4 to Table 6,
14
from which it can clearly be seen that using feature selection methods reduced classification
15
accuracy error from 29% to 25%. This is while, classification error for class S (denoting
16
spalling occurs) decreased from 32% to 23% using selected features, confirming the
17
importance of feature selection techniques in this study.
18
Similarly, confusion matrices on test data for multiple spalling classification demonstrate that
19
classification accuracy is also increased using SFS and SBS methods. As can be seen from
20
Table 7 to Table 9, the total k-NN classification error decreased from 47% (in the case of
21
original features) to 39% (with SBS algorithm), i.e., classification accuracy increased from
22
53% to 61%. It is noted that although the maximum accuracy achieved using SBS algorithm is
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
16
not particularly high, such accuracy is still acceptable given the size of dataset and complexity
1
of the spalling phenomenon. The outcome of this analysis is also in agreement with
2
experimental studies indicating that factors such as compressive strength, width, stirrup
3
spacing, eccentricity, humidity, and load level have significant effect on the extent of spalling.
4
In fact, the above SBS-arrived feature selections led to highest classification accuracy in term
5
of prediction of spalling. While these features are known to govern spalling, however the
6
relation/magnitude of their governance was not quantified till now.
7
Then, these six features are input into the developed GP model to arrive at simple expressions
8
that can be used to predict the occurrence as well as intensity of spalling (SP) by substituting
9
the values of the governing parameters. Similar to PR analysis, two sets of expressions are
10
developed; the first for No spalling/Spalling and the second for No spalling/Minor
11
spalling/Major spalling classification. These expressions are listed in Table 10 which also
12
shows their coefficient of determination (R2) metric. The same table also lists the range of
13
limitations and applicability of these expressions.
14
It can be seen from above table that there is a strong correlation between predicted and
15
measured data points and that the GP-derived expressions succeeded in capturing spalling
16
occurrence. For simplicity, these derived expressions are used to develop a dedicated spalling
17
assessment tool (see Fig. 7) to enable fellow researchers/engineers from examining and using
18
such tool in a plug-and-play mode without needing to re-conduct the ML analysis shown herein.
19
This tool will be made available at the authors’ website.
20
As noted earlier, fire tests on RC columns are very scare and limited. While the compiled
21
database presented herein is the most comprehensive database developed up to date, we cannot
22
but acknowledge that having additional data points could potentially lead to better training of
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
17
the developed ML model and, thus improving its prediction capability. One of the contributions
1
of this paper is to show that even using a limited data set, the proposed AI-based model is
2
capable to predict the spalling for reinforced concrete columns.
3
While the developed database herein accounts for various independent parameters, this
4
database can be further improved by adding outcome of other studies as well as future fire tests
5
(of standard fire or design fire nature). It is expected from future tests to give due consideration
6
to AI-based modeling, training, and validation – this accommodation can be through testing
7
replicates specimens and specimens of varying sizes and configurations; a feature that lacks in
8
those testing programs discussed earlier. As such, the developed fire assessment tool is
9
anticipated to undergo a series of improvements and calibrations in the near future. For example,
10
efforts at the moment are being taken to enable manual and automatic updating of the
11
developed tool via a procedure that allows a centralized repository to harvest data from users
12
to evolve the developed tool such that all users will have the ability to use the most up-to-date
13
tool at all times. Future editions of this tool are expected to be able to propose solutions and
14
spalling mitigation strategies as to aid designers into arriving at safe and optimal designs of RC
15
structures for standard and design fire conditions.
16
17
CONCLUSIONS
18
This paper integrates PR and GP techniques to examine the phenomenon of fire-induced
19
spalling in RC columns. Based on the analysis carried out in this study, the main factors that
20
affect this phenomenon were shown to be compressive strength, width, stirrup spacing,
21
eccentricity, humidity, and load level. As such, these factors were used to develop a user-
22
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
18
friendly fire assessment tool capable of predicting the occupancy of spalling in RC columns.
1
The following conclusions could also be drawn from the results of this study:
2
• There is a need to incorporate ML techniques to develop modern computational
3
methods that can comprehend fire behavior of concrete structures. These approaches
4
can conveniently be developed through PR and GP.
5
• Results confirm that PR using SBS feature selection technique can be effectively used
6
to predict spalling with an accuracy reaching 77%, even with limited dataset. Further,
7
GP-based analysis is also capable of successfully identifying fire-induced spalling with
8
>90% accuracy.
9
• Fire-induced spalling is a complex phenomenon that is not fully understood yet. Future
10
works are expected to deepen the knowledge and understanding on tendency of
11
concrete to spall.
12
• It is noted that the performance of the proposed AI-based model can be notably
13
improved by increasing the number of data set (number of fire tests) used in training
14
and testing the developed model. In fact, results obtained based on the limited number
15
of dataset used in this study further showcase the acceptable performance of the
16
developed AI-based model.
17
REFERENCES
18
1. Kodur, V. “Properties of concrete at elevated temperatures,” International Scholarly
19
Research Notices: Civil Engineering, 2014.
20
2. ACI216.1. “Standard method for determining fire resistance of concrete and masonry
21
construction assemblies,” American Concrete Institute (ACI) - ACI 216, 2014.
22
3. Meacham, B., Park, H., van Straalen, Ij., et al. “Fire and Collapse, Faculty of
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
19
Architecture Building, Delft University of Technology: Data Collection and
1
Preliminary Analyses,” 2009.
2
4. Mueller, K. A., Kurama, Y. C., and McGinnis, M. J. “Out-of-plane behavior of two
3
reinforced concrete bearing walls under fire: A full-scale experimental investigation,”
4
ACI Structural Journal, 2014.
5
5. Khoury, G. A. “Effect of fire on concrete and concrete structures,” Progress in
6
Structural Engineering and Materials, V. 2, No. 4, 2000, pp. 429–47.
7
6. Peris-Sayol, G., Paya-Zaforteza, I., Balasch-Parisi, S., et al. “Detailed Analysis of the
8
Causes of Bridge Fires and Their Associated Damage Levels,” Journal of Performance
9
of Constructed Facilities, 2017.
10
7. Meacham, B., Engelhardt, M., and Kodur, V. “Collection of Data on Fire and
11
Collapse, Faculty of Architecture Building.” NSF Engineering Research and
12
Innovation Conference. Honolulu, Hawaii , 2009.
13
8. Missemer, L., Ouedraogo, E., Malecot, Y., et al. “Fire spalling of ultra-high
14
performance concrete: From a global analysis to microstructure investigations,”
15
Cement and Concrete Research, V. 115, 2019, pp. 207–19.
16
9. Sanjayan, G., and Stocks, L. J. “Spalling of high-strength silica fume concrete in fire,”
17
ACI Materials Journal, 1993.
18
10. Garlock, M., Paya-Zaforteza, I., Kodur, V., et al. “Fire hazard in bridges: Review,
19
assessment and repair strategies,” Engineering Structures, 2012.
20
11. BSI., and European Committee for Standardization. “Design of concrete structures -
21
Part 1-2: General rules - Structural fire design,” 2004.
22
12. Kodur, V., and Naser, M. “Structural Fire Engineering,” 1st edition, McGraw Hill
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
20
Professional, 2020, 480 pp.
1
13. Hertz, K. D. “Reinforcement data for fire safety design,” Magazine of Concrete
2
Research, 2004.
3
14. Bažant, Z. P., Kaplan, M. F., and Bazant, Z. P. “Concrete at High Temperatures:
4
Material Properties and Mathematical Models,” 1996.
5
15. Ulm, F.-J., Coussy, O., and Bažant, Z. P. “The ‘Chunnel’ Fire. I: Chemoplastic
6
Softening in Rapidly Heated Concrete,” Journal of Engineering Mechanics, V. 125,
7
No. 3, 1999, pp. 272–82.
8
16. Harmathy, T. Z. “Effect of mositure on the fire endurance of building elements,”
9
ASTM special technical publication, 1965.
10
17. Shah, A. H., and Sharma, U. K. “Fire resistance and spalling performance of confined
11
concrete columns,” Construction and Building Materials, V. 156, 2017, pp. 161–74.
12
18. Klingsch, E. W. H. “Explosive spalling of concrete in fire,” Tese de doutorado, 2014.
13
19. Han, L. H., Tan, Q. H., and Song, T. Y. “Fire performance of steel reinforced concrete
14
columns,” Journal of Structural Engineering (United States), 2015.
15
20. Agrawal, A., and Kodur, V. “Residual response of fire-damaged high-strength
16
concrete beams,” Fire and Materials, 2019.
17
21. Noumowé, A., Carré, H., Daoud, A., et al. “High-Strength Self-Compacting Concrete
18
Exposed to Fire Test,” Journal of Materials in Civil Engineering, V. 18, No. 6, 2006,
19
pp. 754–8.
20
22. Kalifa, P., Chéné, G., and Gallé, C. “High-temperature behaviour of HPC with
21
polypropylene fibres - From spalling to microstructure,” Cement and Concrete
22
Research, V. 31, No. 10, 2001, pp. 1487–99.
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
21
23. Raut, N. K., and Kodur, V. K. R. “Response of High-Strength Concrete Columns
1
under Design Fire Exposure,” Journal of Structural Engineering, V. 137, No. 1, 2011,
2
pp. 69–79.
3
24. Xu, Y., and Wu, B. “Fire resistance of reinforced concrete columns with L-, T-, and +-
4
shaped cross-sections,” Fire Safety Journal, V. 44, No. 6, 2009, pp. 869–80.
5
25. Kodur, V., McGrath, R., Leroux, P., et al. “Experimental studies for evaluating the fire
6
endurance of high-strength concrete columns,” 2005.
7
26. Kodur, V. R., Mcgrath, R. C., Latour, J. C., et al. “Experimental Studies for Evaluating
8
the Fire Endurance of High- Strength Concrete Columns,” 2000.
9
27. Dwaikat, M. B., and Kodur, V. K. R. “Response of Restrained Concrete Beams under
10
Design Fire Exposure,” Journal of Structural Engineering, 2009.
11
28. Ji, T., Lin, T., and Lin, X. “A concrete mix proportion design algorithm based on
12
artificial neural networks,” Cement and Concrete Research, V. 36, No. 7, 2006, pp.
13
1399–408.
14
29. Nehdi, M., El Chabib, H., and El Naggar, M. H. “Predicting performance of self-
15
compacting concrete mixtures using artificial neural networks,” ACI Materials
16
Journal, 2001.
17
30. Rafiei, M. H., Khushefati, W. H., Demirboga, R., et al. “Novel approach for concrete
18
mixture design using neural dynamics model and virtual lab concept,” ACI Materials
19
Journal, V. 114, No. 1, 2017.
20
31. Keshtegar, B., Bagheri, M., and Yaseen, Z. M. “Shear strength of steel fiber-
21
unconfined reinforced concrete beam simulation: Application of novel intelligent
22
model,” Composite Structures, 2019.
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
22
32. Ashrafian, A., Shokri, F., Taheri Amiri, M. J., et al. “Compressive strength of Foamed
1
Cellular Lightweight Concrete simulation: New development of hybrid artificial
2
intelligence model,” Construction and Building Materials, 2020.
3
33. Golafshani, E. M., and Behnood, A. “Estimating the optimal mix design of silica fume
4
concrete using biogeography-based programming,” Cement and Concrete Composites,
5
V. 96, 2019, pp. 95–105.
6
34. Rafiei, M. H., Khushefati, W. H., Demirboga, R., et al. “Neural network, machine
7
learning, and evolutionary approaches for concrete material characterization,” ACI
8
Materials Journal, 2016.
9
35. Naser, M. Z. “Heuristic machine cognition to predict fire-induced spalling and fire
10
resistance of concrete structures,” Automation in Construction, V. 106, 2019, p.
11
102916.
12
36. Liu, J.-C. C., Tan, K. H., and Yao, Y. “A new perspective on nature of fire-induced
13
spalling in concrete,” Construction and Building Materials, V. 184, 2018, pp. 581–90.
14
37. Phan, L. T., and Carino, N. J. “Fire Performance of High Strength Concrete: Research
15
Needs.” Advanced Technology in Structural Engineering. Reston, VA, American
16
Society of Civil Engineers, 2000. pp. 1–8.
17
38. Hertz, K. D. D. “Limits of spalling of fire-exposed concrete,” Fire Safety Journal, V.
18
38, No. 2, 2003, pp. 103–16.
19
39. Song, T.-Y., Han, L.-H., and Tao, Z. “Structural Behavior of SRC Beam-to-Column
20
Joints Subjected to Simulated Fire Including Cooling Phase,” Journal of Structural
21
Engineering, V. 141, No. 9, 2015.
22
40. Kodur, V., Cheng, F., Wang, T., et al. “Fire resistance of high-performance concrete
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
23
columns,” National Research Council Canada, 2001, p. 136.
1
41. Maluk, C., Bisby, L., and Terrasi, G. P. “Effects of polypropylene fibre type and dose
2
on the propensity for heat-induced concrete spalling,” Engineering Structures, 2017.
3
42. Zhang, Y., Zeiml, M., Maier, M., et al. “Fast assessing spalling risk of tunnel linings
4
under RABT fire: From a coupled thermo-hydro-chemo-mechanical model towards an
5
estimation method,” Engineering Structures, V. 142, 2017, pp. 1–19.
6
43. Shakya, A. M., and Kodur, V. K. R. “Response of precast prestressed concrete
7
hollowcore slabs under fire conditions,” Engineering Structures, V. 87, 2015, pp. 126–
8
38.
9
44. Kodur, V. K. R. “Innovative strategies for enhancing fire performance of high-strength
10
concrete structures,” Advances in Structural Engineering, V. 21, No. 11, 2018, pp.
11
1723–32.
12
45. Ali, F., Nadjai, A., Silcock, G., et al. “Outcomes of a major research on fire resistance
13
of concrete columns,” Fire Safety Journal, V. 39, No. 6, 2004, pp. 433–45.
14
46. Heo, Y.-S., Sanjayan, J. G., Han, C.-G., et al. “Synergistic effect of combined fibers
15
for spalling protection of concrete in fire,” Cement and Concrete Research, V. 40, No.
16
10, 2010, pp. 1547–54.
17
47. Kanéma, M., Pliya, P., Noumowé, A., et al. “Spalling, Thermal, and Hydrous Behavior
18
of Ordinary and High-Strength Concrete Subjected to Elevated Temperature,” Journal
19
of Materials in Civil Engineering, V. 23, No. 7, 2011, pp. 921–30.
20
48. Salehi, H., and Burgueño, R. “Emerging artificial intelligence methods in structural
21
engineering,” v. vol. 171, Elsevier, 2018.
22
49. Hou, Q., Liu, L., Zhen, L., et al. “A novel projection nonparallel support vector
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
24
machine for pattern classification,” Engineering Applications of Artificial Intelligence,
1
V. 75, 2018, pp. 64–75.
2
50. Lie, T., and Woollerton, J. “Fire Resistance of Reinforced Concrete Columns - NRC
3
Publications Archive - National Research Council Canada,” 1988.
4
51. Kodur, V., and McGrath, R. “Fire endurance of high strength concrete columns,” Fire
5
Technology, V. 39, 2003, pp. 73–87.
6
52. Myllymaki, J., and Lie, T. “Fire Resistance Test of a Square Reinforced Concrete
7
Column,” National Research Council Canada, 1991, p. 36.
8
53. Dotreppe, J.-C., Franssen, J.-M., Bruls, A., et al. “Experimental research on the
9
determination of the main parameters affecting the behaviour of reinforced concrete
10
columns under fire conditions,” Magazine of Concrete Research, 1997.
11
54. Rodrigues, J. P. C., Laím, L., and Correia, A. M. “Behaviour of fiber reinforced
12
concrete columns in fire,” Composite Structures, V. 92, 2010, pp. 1263–8.
13
55. Hass, R. “Practical rules for the design of reinforced concrete and composite columns
14
submitted to fire,” v. vol. 69, T.U. Braunschweig, 1986.
15
56. Schneider, U. “Concrete at high temperatures - A general review,” Fire Safety Journal,
16
V. 13, No. 1, 1988, pp. 55–68.
17
57. Salehi, H., Das, S., Chakrabartty, S., et al. “Damage identification in aircraft structures
18
with self-powered sensing technology: A machine learning approach,” Structural
19
Control and Health Monitoring, 2018.
20
58. Salehi, H., Biswas, S., and Burgueño, R. “Data interpretation framework integrating
21
machine learning and pattern recognition for self-powered data-driven damage
22
identification with harvested energy variations,” Engineering Applications of Artificial
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
25
Intelligence, 2019.
1
59. Denœux, T. “A k-nearest neighbor classification rule based on Dempster-Shafer
2
theory,” Studies in Fuzziness and Soft Computing, 2008.
3
60. Koza, J. R. “A genetic approach to finding a controller to back up a tractor-trailer
4
truck,” Proceedings of the 1992 American Control Conference, 1992, pp. 2307–11.
5
61. Mousavi, S. M., Aminian, P., Gandomi, A. H., et al. “A new predictive model for
6
compressive strength of HPC using gene expression programming,” Advances in
7
Engineering Software, 2012.
8
62. Naser, M. Z. “Autonomous Fire Resistance Evaluation,” ASCE Jounral of Structural
9
Engineering, 2020.
10
63. Buch, S. H., and Sharma, U. K. “Fire Resistance of Eccentrically Loaded Reinforced
11
Concrete Columns,” Fire Technology, 2019, pp. 1–36.
12
64. Naser, M. “Database for Fire-induced spalling in reinforced concrete (RC) columns,”
13
2020.
14
65. Abdi, H., and Williams, L. J. “Principal component analysis,” Wiley Interdisciplinary
15
Reviews: Computational Statistics, V. 2, No. 4, 2010, pp. 433–59.
16
17
18
19
20
21
22
23
24
25
26
27
28
29
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
26
TABLES AND FIGURES
1
List of Tables:
2
Table 1 – Factors affecting occurrence and magnitude of fire-induced phenomenon
3
Table 2 – K-NN classification accuracy (binary spalling)
4
Table 3 – K-NN classification accuracy (multiple spalling)
5
Table 4 – Confusion matrix for binary spalling (without feature selection)
6
Table 5 – Confusion matrix for binary spalling (SFS feature selection)
7
Table 6 – Confusion matrix for binary spalling (SBS feature selection)
8
Table 7 – Confusion matrix for multiple spalling (without feature selection)
9
Table 8 – Confusion matrix for multiple spalling (SFS feature selection)
10
Table 9 – Confusion matrix for multiple spalling (SBS feature selection)
11
Table 10 – GP-derived expressions to be used to evaluate fire response of RC columns and
12
statistics
13
List of Figures:
14
Fig. 1 – Framework of proposed methodology.
15
Fig. 2 – Typical architecture of a GP model.
16
Fig. 3 – Data projection on to the first two principal components: (a) Single spalling
17
classification, and (b) Multiple spalling classification.
18
Fig. 4 –SFS feature selection algorithm.
19
Fig. 5 –SBS feature selection algorithm.
20
Fig. 6 – Classification accuracy with k-NN algorithm for different number of k: (a) Binary
21
spalling classification, (b) Multiple spalling classification.
22
Fig. 7 – Graphical interface of developed fire assessment tool.
23
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
27
Table 1– Factors affecting occurrence and magnitude of fire-induced phenomenon
1
May lead to
spalling
May minimize spalling
Mechanisms/Remarks
Material characteristics
Silicate/quartzite
aggregate
-
• Due to a change in the quartzite
phase at 573°C.
-
Carbonate aggregates
• Lowers temperature rise
• Slows down strength degradation
-
Polypropylene fibers
• Melts at 160–170°C and thus create
additional pores
-
Steel fibers
• Improves tensile strength of
concrete
Silica fume, high
content of cement
and limestone
fillers
-
• Dense microstructure
• Low permeability
High moisture
content
-
• Facilitates development of increased
pore pressure
Geometric
configurations
Sharp edges
-
• Attracts heat through bi-directional
transmission
Larger size
-
• Holds higher amounts of moisture
• Facilitates large thermal and pore
pressure gradient
-
Bent ties/close tie
spacing
• Improves resistance to pore
pressure
Loading
conditions
Axial loading/fixed
restraint conditions
-
• Continuous compression state
• Inhibits development of cracks
Eccentric loading
-
• Develops two states of stress
Rapid/intense
heating
-
• Causes thermal shock (leading to
high thermal stresses and non-
uniform expansion)
2
3
Table 2 – K-NN classification accuracy (binary spalling)
4
5
6
7
8
9
Number of Neighbors (k)
Classification Accuracy (%)
2
64%
3
71%
4
69%
5
72%
6
65%
7
64%
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
28
Table 3 – K-NN classification accuracy (multiple spalling)
1
2
3
4
5
6
7
Table 4 – Confusion matrix for binary spalling (without feature selection)
8
True Classes
Predicted Classes
True
Sum
Class N
Class S
Class N
8
11
19
Class S
2
23
25
Sum
10
34
44
Error (%)
0.20
0.32
0.29
9
10
Table 5 – Confusion matrix for binary spalling (SFS feature selection)
11
True Classes
Predicted Classes
True
Sum
Class N
Class S
Class N
13
6
19
Class S
5
20
25
Sum
18
26
44
Error (%)
0.28
0.23
0.25
12
13
14
15
Number of Neighbors (k)
Classification Accuracy (%)
5
45%
7
50%
10
53%
12
58%
15
52%
17
50%
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
29
Table 6 – Confusion matrix for binary spalling (SBS feature selection)
1
True Classes
Predicted Classes
True
Sum
Class N
Class S
Class N
10
9
19
Class S
2
23
25
Sum
12
32
44
Error (%)
0.17
0.28
0.25
2
Table 7 – Confusion matrix for multiple spalling (without feature selection)
3
4
5
6
7
8
9
Table 8 – Confusion matrix for multiple spalling (SFS feature selection)
10
11
12
13
14
15
16
17
18
19
True Classes
Predicted Classes
True Sum
N
MN
MJ
N
6
4
7
17
MN
2
3
1
6
MJ
2
1
10
13
Sum
10
8
18
36
Error (%)
0.40
0.62
0.44
0.47
True Classes
Predicted Classes
True Sum
N
MN
MJ
N
14
0
3
17
MN
3
1
2
6
MJ
7
0
6
13
Sum
24
1
11
36
Error (%)
0.42
0.00
0.45
0.42
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
30
Table 9 – Confusion matrix for multiple spalling (SBS feature selection)
1
2
3
4
5
6
7
Table 10 – GP-derived expressions to be used to evaluate fire response of RC columns
8
and statistics
9
Remarks
Derived expressions
R2
Binary
classification
No spalling = 0
Spalling = 1
95.1
Multi-classification
No spalling = 1
96.7
Minor spalling =
1
97.2
Major spalling = 1
94.6
Range of applicability
fc= 23.8-138 MPa
e= 0-40 mm
S = 50-406 mm
W = 203-406 mm
P = 0-5373 kN
H = 5-99%
10
11
12
13
14
True Classes
Predicted Classes
True Sum
N
MN
MJ
N
15
0
2
17
MN
4
1
1
6
MJ
5
2
6
13
Sum
24
3
9
36
Error (%)
0.37
0.67
0.33
0.39
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
31
1
Fig. 1– Framework of proposed methodology.
2
3
4
Fig. 2– Typical architecture of a GP model.
5
Step 1: Collect data points
Step 2: Carry out PR analysis
Step 3: Apply GP
Step 4: Develop assessment tool to predict
fire-induced spalling
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
32
1
Fig. 3– Data projection on to the first two principal components: (a) Single spalling
2
classification, and (b) Multiple spalling classification.
3
4
Fig. 4–SFS feature selection algorithm
5
6
7
Fig. 5–SBS feature selection algorithm
8
9
(b)
(a)
Algorithm 1 SFS Feature Selection
1. Start with empty feature set
2. Select the next best feature
3. Update
4. Return to step 2
Algorithm 2 SBS Feature Selection
1. Start with empty feature set
2. Remove the worst feature
3. Update
4. Return to step 2
This is a preprint draft. The published article can be found at: https://doi.org/10.14359/51728120
Please cite this paper as:
Naser M.Z., Salehi, H. (2020). “Machine Learning-Driven Assessment of Fire-Induced Concrete Spalling of
Columns.” ACI Materials Journal. https://doi.org/10.14359/51728120.
33
1
Fig. 6 – Classification accuracy with k-NN algorithm for different number of k: (a)
2
Binary spalling classification, (b) Multiple spalling classification.
3
4
5
Fig. 7 – Graphical user interface of developed fire assessment tool.
6