Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Evaluation of Factors Affecting Long-term Creep of Concrete Using

Machine Learning Regression Models


H. Daou, W. Raphael & F. Geara
Ecole Supérieure d’Ingénieurs de Beyrouth (ESIB), Saint-Joseph University, Lebanon

ABSTRACT: The major effects of creep in concrete bridge structures can be summarized in three categories:
camber and deflection, stress redistribution, and prestress loss. Creep affects the settings of bearings and the
size of sliding plates or laminated bearing pads as well as the size and setting of expansion joints. Creep also
influences the redistribution of forces in certain structures where the static system changes during construction,
and therefore, plays a major role in stress redistribution for composite construction. Moreover, creep affects the
amount of girder shortening due to the prestress and the corresponding loss of prestress, thereby also affects the
secondary moments in a prestressed girder bridge. Many factors including water-cement ratio, aggregate-ce-
ment ratio, cement content, type of cement, compressive strength, loading age, volume-surface ratio, tempera-
ture, relative humidity and stress may affect the long-term creep of concrete. This paper aims to investigate the
main factors affecting the long-term creep of concrete using machine learning. For that, multiple regression
models for predicting creep coefficient (Linear Model, Ridge Model, Lasso Model, Decision Tree Model, Bag-
ging Model, Random Forest Model, Generalized Boosting Model and Extreme Gradient Boosting Model) are
developed using the statistical software R. The creep tests used in this study are from the Northwestern Univer-
sity database. An evaluation of the regression models is performed by comparing the predicted value to the
experimental values, and the factors affecting creep are discussed. The results show that ensemble trees give a
more accurate prediction of the creep coefficient at long-term and that the factors have different degrees of
importance for the various regression models which must be well controlled to avoid the consequences of in-
accurate prediction of creep.

1 INTRODUCTION control the quality of the structural elements. Unfor-


tunately, these long-term effects are not easy to un-
Creep is one of the most significant time-dependent derstand or to model and predict (Raphael et al.
behaviors of concrete, which forms the necessary ba- 2012).
sis for precise prediction and evaluation of the me- Creep is a complex phenomenon that is affected by
chanical response of concrete structures under sus- several factors such as concrete mixtures (i.e., water-
tained loads. In concrete bridges, the major effects of cement ratio, compressive strength, type of cement,
creep can be summarized in three categories: camber and admixture) (Chunping et al. 2019, Daou &
and deflection, stress redistribution, and prestress Rapahel 2021, Daou & Raphael 2020, Kamen et al.
loss. Creep affects the settings of bearings and the 2009, Shen et al. 2017), environmental conditions
size of sliding plates or laminated bearing pads as (i.e., temperature, and relative humidity) (Briffaut et
well as the size and setting of expansion joints. Creep al. 2012, Theiner et al. 2017, Zheng et al. 2020), and
also influences the redistribution of forces in certain loading schemes (i.e., age at loading, and load) (Shen
structures where the static system changes during et al. 2020, Su et al. 2017). These tests provided val-
construction, and therefore, plays a major role in uable data for understanding the creep behavior of
stress redistribution for composite construction concrete. However, an accurate prediction remains a
(Balant & Panula 1980, Bazant et al. 2011, Daou et challenge because of the high sensitivity of creep to
al. 2019, Raphael et al. 2018). Therefore, the effects wide ranges of parameters and the interplay between
of creep are needed to be calculated and considered to these parameters. This study provides a Machine
learning (ML) solution to predict creep coefficient at
long-term and evaluate the factors affecting it. As a
tributary of artificial intelligence, supervised ML has
been considered as a promising data-driven approach
to build robust models for predicting various proper-
ties and behaviors of heterogeneous materials.
Machine learning (ML) algorithms are ubiquitous
with their easy implementation over time. ML models
are often used to find trends and patterns in given ob-
servations. For quantitative data, regression models
are used for finding patterns. In recent years, many
researchers in the field of construction materials have
been implementing ML models in predicting com-
Figure 1. Flowchart showing the implementation of the problem
pressive strength (Marani et al. 2020), elastic modu-
lus (Han et al. 2020), shear capacity (Solhmirzaei et
al. 2020, Vu & Hoang 2015), flexural strength (Kang 2.1 Data collection
et al. 2021), cracking propagation (Bayar & Bilir The data was retrieved from the Northwestern Uni-
2019), and drying shrinkage (Bal & Buyle-bodin versity (NU) database, which was assembled during
2013). ML models were also used in predicting the the period 2010-2013 under the support of the U.S.
crack condition of highway pavements (Inkoom et al. Department of Transportation (Hubler et al. 2015).
2019), detecting bridges defects (Zhu et al. 2019), as-
This database is constituted of approximately 1433
sessing bridge condition (Liu & Zhang 2020), pre- creep tests performed using different concrete mix
dicting crack classification in roads (Jo & Jadidi compositions and under various environmental con-
2019), and predicting fatigue damage of highway sus- ditions. Each test or experiment has many readings at
pension bridge hangers (Deng et al. 2020). Ensemble different times. This study aims to study the signifi-
learning is a machine learning paradigm, which im- cance of factors affecting creep at long-term; there-
proves stability and fits the base learner models such fore 212 experiments conducted under different con-
as decision trees, support vector machines and neural
ditions for specimens without admixtures are
network models by aggregating their outputs to pro- considered from the NU database, and the creep coef-
duce a single decision. Ensemble tree machine learn- ficient is calculated after 3000 days of loading.
ing models have proved useful for solving poorly un- Each observation has ten attributes namely aggregate-
derstood and complex problems (Dietterich 2000b). cement ratio, water-cement ratio, cement type, ce-
In this study, ML models (i.e., linear, ridge, lasso, de- ment content (kg/m3), compressive strength at the age
cision tree, bagging, random forest, generalized of 28 days (MPa), concrete age at loading (days), vol-
boosting, and extreme gradient boosting models) are ume-surface ratio, sustained stress over the compres-
established to learn the underlying patterns behind sive strength at loading age (MPa), temperature (°C),
212 creep test results at long-term (i.e., 3000 days). and relative humidity (%). All the attributes are quan-
The creep tests are selected from the Northwestern titative variables except the cement type. In the da-
University database, which consists of experiments taset, the cement type may be R (normal hardening),
conducted with different mixture proportions and un- RS (rapid hardening), or SL (slow hardening).
der various environmental conditions. Then, the anal-
ysis of significant factors on concrete creep is done. 2.2 Data preprocessing
The study aims to find out the importance of each fac- The attributes in the NU database have different
tor for each of the regression models in order to figure ranges. For example, the water-cement ratio ranges
out the most affecting factors on creep of concrete to between 0.3 and 0.8 while relative humidity varies
control them well and avoid the consequences of in- from 20% to 100%. Therefore, the data should be pre-
accurate prediction of creep. processed. The quantitative variables are standard-
ized using a z-score to easily calculate the most robust
2 METHODOLOGY model as shown in Equation 3.
𝑥−𝑥̅
In order to know the best ML algorithm, which fits 𝑧= 𝜎
(1)
better the given data, different ML regression models
or regressors are used in this study. Therefore, multi- where x = original feature vector, 𝑥̅ = mean of the fea-
ple regressors are performed to select the best regres- ture vector x, and σ = standard deviation.
sor which is the best fit for the calculation of the con- 2.3 Model building
crete creep coefficient. Figure 1 shows the process of
the implementation of the problem. Different ML regression models were used in this
study for the prediction of the creep coefficient,
namely Linear Model (Engel 2019), Ridge Model
(Hoerl & Kennard 1970), Lasso Model (Tibshirani 2.3.6 Random Forest model
1996), Decision Tree Model (Breiman et al. 1984), Random forests are a modification of bagged decision
Bagging Model (Breiman 1996), Random Forest trees that build a large collection of de-correlated
Model (Breiman 2001), Generalized Boosting Model trees to further improve predictive performance. Ran-
(Friedman 2001) and Extreme Gradient Boosting dom forests are a combination of tree predictors such
Model (Chen & Guestrin 2016). that each tree depends on the values of a random vec-
tor sampled independently and with the same distri-
2.3.1 Linear model bution for all trees (Breiman 2001).
The linear model is the most general algorithm based
on supervised learning for machine learning predic- 2.3.7 Generalized Boosting model
tion. Linear regression technique identifies a linear Generalized boosting models are an extremely popu-
relationship between xi (input variables) and y (output lar machine learning algorithm, and have proven suc-
variable) as follows: cessful across many fields. There is a combination be-
tween decision tree algorithms and boosting methods.
𝑦 = 𝛼0 + 𝛼1 𝑥1 + 𝛼2 𝑥2 + ⋯ + 𝛼𝑛 𝑥𝑛 (2)
The boosting framework was generalized to regres-
where y = dependent variable value, xi = independent sion problems by relating boosting algorithm to im-
variable value, and αi = estimated coefficients. portant statistical concepts like loss functions
(Friedman 2001).
2.3.2 Ridge model
Ridge regression model is an extension of the linear 2.3.8 Extreme Gradient Boosting model
regression where it controls the estimated coefficients Extreme gradient boosting is an efficient and scalable
using the square regulation (L2-norm) as follows: variant of the gradient boosting machine. Gradient
boosted trees use regression trees in a sequential
𝑝 2 𝑝
𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒(∑𝑛𝑖=1(𝑦𝑖 𝛼0 − ∑𝑗=1 𝛼𝑗 𝑥𝑖𝑗 ) + 𝜆 ∑𝑗=1 𝛼𝑗2 ) (3) learning process as weak learners. Extreme gradient
boosting yields a direct interpretation of boosting
where n = number of observations, p = number of pre- methods from the point of view of numerical optimi-
dictors or features, and λ = tuning parameter (Hoerl zation in function space and generalize them by al-
& Kennard 1970). lowing optimization of an arbitrary loss function
(Chen & Guestrin 2016).
2.3.3 Lasso model
Lasso regression is similar to the ridge regression but 2.4 Model evaluation
the control of the estimated coefficients is made using The mean absolute error (MAE) and root mean square
the least absolute regulation (L1-norm) as follows error (RMSE) are the metrics used for the evaluation
(Tibshirani 1996): of the models.
MAE computes the mean of the absolute difference
𝑝 2 𝑝
𝑚𝑖𝑛𝑖𝑚𝑖𝑧𝑒(∑𝑛𝑖=1(𝑦𝑖 𝛼0 − ∑𝑗=1 𝛼𝑗 𝑥𝑖𝑗 ) + 𝜆 ∑𝑗=1|𝛼𝑖 |) (4) between the actual and the predicted creep coefficient
at long-term. MAE is defined as:
2.3.4 Decision tree model 1
𝑀𝐴𝐸 = 𝑛 ∑𝑛𝑖=1|𝑦𝑖 − 𝑦̂𝑖 | (5)
The decision tree can handle both numeric and cate-
gorical data. Tree-based models are a class of nonpar- where n = the total number of observations, yi = the
ametric algorithms that work by partitioning the fea- actual value of the creep coefficient for the observa-
ture space into many smaller and non-overlapping tion i, and 𝑦̂𝑖 = the predicted value of the creep coef-
regions with similar response values using a set of ficient for the observation i.
splitting rules (Breiman et al. 1984). RMSE computes the root mean square of the differ-
ence between the predicted creep coefficient and the
2.3.5 Bagging model actual creep coefficient for all the observations in the
Bagging, also called bootstrap aggregating, is one of dataset. RMSE is defined as:
the first ensemble algorithms machine learning prac-
(𝑦 2
titioners learn and is designed to improve the accu- ̂−𝑦𝑖)
𝑅𝑀𝑆𝐸 = √∑𝑛𝑖=1 𝑖
𝑛
(6)
racy and stability of classification and regression al-
gorithms. Bagging prediction models fit multiple 2.5 Feature importance
versions of a prediction model and then ensemble
them into an aggregate prediction (Breiman 1996). The feature importance is studied for the different ML
Bagging works especially well for unstable, high var- regression models or regressors. Different regressors
iance base learners algorithms whose predicted out- have different importance values for the given fea-
put undergoes a major change in response to small tures. Each model is analyzed to find out which fea-
changes in the training data (Dietterich 2000a). tures are the most significant.
3 RESULTS AND DISCUSSION gradient boosting model are the best performing algo-
rithms. Ensemble learners always tend to outperform
The dataset used in this study is retrieved from the the conventional machine learning algorithms and, in
Northwestern University database and includes 212 this study too, ensemble learners have shown good
readings of creep coefficient from experiments per- performance for predicting the output (i.e., creep co-
formed using different concrete mix compositions efficient). However, the bagging model is the best
and under various environmental conditions. The model to predict the creep coefficient at long-term in
creep coefficient is calculated at 3000 days using the the given dataset in terms of MAE and RMSE. In fact,
interpolation and extrapolation equations (Eqs 1-2). the bagging model works well on decision tree mod-
The dataset is randomly divided into a training set els where it highly improves regression performance
(159 readings) and a test set (53 readings). Different and overfitting, and decreases the prediction variance.
machine learning regression models are used to per- As shown in Table 1, there is an improvement in pre-
form the prediction of the concrete creep coefficient diction accuracy of approximately 32% and 43% in
at long-term and carried out using the statistical soft- MAE and RMSE, respectively, between bagging and
ware R. These models include linear model, ridge decision tree models.
model, lasso model, elastic net model, decision tree The importance of the given factors for different re-
model, bagging model, random forest model, gener- gression models is calculated and shown in Figures 2-
alized boosting model, and extreme gradient boosting 9.
model. The evaluation of models was done based on The rows index in the Figures 2-9 represents the fea-
the MAE and RMSE. The mathematical expressions tures as shown in Table 2.
for MAE and RMSE are given in Equations 5 and 6.
The linear model predicts the creep coefficient in the Table 2. Index and feature name
function of the ten variables. The ridge model adds a Index Features (Factors)
square regulation (L2-norm) to the cost function w/c Water-cement ratio
while the lasso model uses the absolute regulation a/c Aggregate-cement ratio
(L1-norm). The decision tree model creates partitions c Cement content
in predictors so that the creep coefficient can be pre- V/S Volume-surface ratio
dicted based on partitions between the ten input vari- RH Relative humidity
ables. The bagging model fits multiple versions of a t0 Loading age
prediction model and then ensembles them into an ag- fcm Compressive strength
gregate prediction. The random forest model injects load Sustained load over com-
more randomness in the tree-growing process by ran- pressive strength at loading
domly selecting a bootstrap sample to train on and a T Temperature
random sample of features to use at each split. The R, RS and SL Type of cement
generalized boosting model iteratively improves re-
gression trees, where each model in the sequence
slightly improves the performance of the previous
model. This improvement is done primarily by focus-
ing on the rows of the training data where the previ-
ous tree had the largest errors or residuals. The ex-
treme gradient boosting model increases speed and
performance by introducing regularization parame-
ters to reduce overfitting. The performance of all
models is summarized in Table 1.

Table 1. MAE and RMSE for the different models.


Model MAE RMSE
Linear model 0.47 0.58
Ridge model 0.47 0.57
Lasso model 0.48 0.57
Decision tree model 0.41 0.60
Bagging model 0.28 0.34
Random forest model 0.30 0.35
Generalized boosting model 0.28 0.37
Extreme gradient boosting model 0.30 0.42

As shown in Table 1, the bagging model, random for-


est model, generalized boosting model and extreme
Figure 2. Feature importance for the linear regression model
Figure 3. Feature importance for the ridge regression model Figure 5. Feature importance for the decision tree model

Figure 4. Feature importance for the lasso regression model Figure 6. Feature importance for the bagging model
Figure 7. Feature importance for the random forest model Figure 9. Feature importance for the extreme gradient boosting
model

As shown in Figures 2-9, the factors have different


degrees of importance for the various regression mod-
els. For the linear, ridge and lasso model, relative hu-
midity is the most influential factor on creep coeffi-
cient. For the decision tree and ensemble trees
models, the sustained load is the most factor affecting
the long-term creep coefficient. Therefore, creep de-
formations are strongly influenced by the intensity of
the applied sustained load.
Based on the best prediction model (i.e., bagging
model), the concrete mixture proportions including
the aggregate-cement ratio (a/c), cement content (c)
and water-cement ratio (w/c) have a significant im-
pact on creep. Therefore, these factors must be well
predicted and identified during the design phase and
should necessarily be controlled by a quality control
on the worksite, because the creep depends largely on
them.

4 CONCLUSION

The study of algorithms for predicting creep of the


concrete is of high importance. Since the design of
bridges depends on the time-dependent properties of
concrete, the factors affecting creep should be studied
Figure 8. Feature importance for the generalized boosting model very well. This study builds eight machine learning
models (i.e., linear model, ridge model. lasso model,
decision tree model, bagging model, random forest
model, generalized boosting model, and extreme gra-
dient boosting model) for predicting creep coefficient
at long-term. Ten factors including the water-cement
ratio, aggregate-cement ratio, cement content, type of
cement, loading age, compressive strength, volume-
surface ratio, sustained load, temperature and relative
humidity were considered as input variables or fea- Learning 24: 123–140.
tures. The dataset contains 212 readings of creep co- Breiman, L. 2001. Random forests. Machine
efficient at 3000 days. Learning 45: 5–32.
The dataset was preprocessed using the z-score and Breiman, L. Friedman, J. Stone, C. J. & Olshen, R. A.
was randomly divided into a training set and a test set. 1984. Classification and Regression Trees,
Chapman and Hall/CRC.
Then, the machine learning models were built to pre- Briffaut, M. Benboudjema, F. Torrenti, J. & Nahas,
dict the creep coefficient at 3000 days using the sta- G. 2012. Concrete early age basic creep:
tistical software R. The models were evaluated based Experiments and test of rheological modelling
on the mean absolute value (MAE) and root mean approaches. Construction and Building
square error (RMSE). The ensemble trees algorithms Materials 36: 373–380.
showed a good performance for predicting the long- Chen, T. & Guestrin, C. 2016. XGBoost: A Scalable
term creep coefficient in terms of MAE and RMSE. Tree Boosting System. In Proceedings of the
Moreover, the bagging model was the best model for 22nd ACM SIGKDD International Conference
predicting the long-term creep coefficient. The results on Knowledge Discovery and Data Mining, New
showed that there is a prediction accuracy improve- York, NY, USA: Association for Computing
ment of approximately 32% and 43% in MAE and Machinery, pp. 785–794.
RMSE, respectively, between bagging and decision Chunping, G. Yicong, W. Fan, G. … Jin, C. 2019.
Early age tensile creep of high performance
tree models, which shows that the bagging model concrete containing mineral admixtures:
works well on decision tree models as it significantly Experiments and modeling. Construction and
improves the regression performance and overfitting, Building Materials 197: 766–777.
and reduces the prediction variance. Daou, H. Abou Salha, W. Raphael, W. &
The analysis of significant factors showed that the Chateauneuf, A. 2019. Explanation of the
factors have different degrees of importance for the collapse of Terminal 2E at Roissy–CDG Airport
various regression models. For the linear, ridge and by nonlinear deterministic and reliability
lasso model, relative humidity is the most influential analyses. Case Studies in Construction
factor on creep coefficient. For the decision tree and Materials 10. doi:10.1016/j.cscm.2019.e00222
ensemble trees models, the sustained load is the most Daou, H. & Rapahel, W. 2021. Identifying the Weight
factor affecting the long-term creep coefficient. How- of Factors Affecting Creep of Concrete Using
ever, according to the best predicting model, load, Factorial ANOVA. Key Engineering Materials
872: 21–25.
mixture proportions (i.e., aggregate-cement ratio, ce- Daou, H. & Raphael, W. 2020. Investigation of the
ment content and water-cement ratio) in addition to Significance of Factors Affecting Long-Term
compressive strength have major effects on concrete Creep of Concrete using Screening Designs. IOP
creep. Therefore, and in order to avoid the conse- Conference Series: Materials Science and
quences of inaccurate prediction of creep, these fac- Engineering 809. doi:10.1088/1757-
tors should be highlighted and taken into account in 899X/809/1/012016
all models predicting creep. Moreover, the variables Deng, Y. Zhang, M. Feng, D. & Li, A. 2020.
which have the biggest influence on creep should nec- Predicting fatigue damage of highway
essarily be controlled by a quality control on the suspension bridge hangers using weigh-in-
worksite, because the creep depends on it largely. motion data and machine learning. Structure and
Infrastructure Engineering: Maintenance,
5 REFERENCES Management, Life-Cycle Design and
Performance 1–16.
Dietterich, T. 2000a. An Experimental Comparison of
Bal, L. & Buyle-bodin, F. 2013. Artificial neural Three Methods for Constructing Ensembles of
network for predicting drying shrinkage of Decision Trees: Bagging, Boosting, and
concrete. Construction and Building Materials Randomization. Machine Learning 40: 139–
38: 248–254. 157.
Balant, Z. P. & Panula, L. 1980. Creep and Shrinkage Dietterich, T. 2000b. Ensemble Methods in Machine
Characterisation for Analyzing Prestressed Learning. International Workshop on Multiple
Concrete Structures. PCI Journal 25(3): 86– Classifier Systems 1–15.
122. Engel, G. 2019. Machine Learning to Approximate
Bayar, G. & Bilir, T. 2019. A novel study for the Solutions of Ordinary Differential Equations :
estimation of crack propagation in concrete Neural Networks vs . Linear Regressors. In
using machine learning algorithms.
International Conference on Computational
Construction and Building Materials 215: 670– Science, Springer, pp. 169–177.
685. Friedman, J. H. 2001. Greedy Function
Bazant, Z. P. Hubler, M. H. & Yu, Q. 2011.
Approximation: A Gradient Boosting Machine.
Pervasiveness of Excessive Segmental Bridge The Annals of Statistics 29(5): 1169–1232.
Deflections : Wake-Up Call for Creep. ACI Hamed, E. 2015. Non-linear creep effects in concrete
Structural Journal 108(6): 766–774. under uniaxial compression. Magazine of
Breiman, L. 1996. Bagging predictors. Machine Concrete Research 67(16): 876–884.
Han, T. Siddique, A. Khayat, K. Huang, J. & Kumar, Experimental investigations on early-age tensile
A. 2020. An ensemble machine learning creep of internally cured high strength concrete
approach for prediction and optimization of under different initial stress / strength ratios.
modulus of elasticity of recycled aggregate Construction and Building Materials 265:
concrete. Construction and Building Materials 120313.
244: 118271. Solhmirzaei, R. Salehi, H. Kodur, V. & Naser, M. Z.
Harinadha Reddy, D. & Ramaswamy, A. 2018. 2020. Machine learning framework for
Experimental and numerical modeling of creep predicting failure mode and shear capacity of
in different types of concrete. Heliyon 4(7). ultra high performance concrete beams.
doi:10.1016/j.heliyon.2018.e00698 Engineering Structures 224: 111221.
Hoerl, A. E. & Kennard, R. W. 1970. Ridge Su, L. Wang, Y. Mei, S. & Li, P. 2017. Experimental
Regression: Biased Estimation for investigation on the fundamental behavior of
Nonorthogonal Problems. Technometrics 12(1): concrete creep. Construction and Building
55–67. Materials 152: 250–258.
Hubler, M. H. Wendner, R. & Bažant, Z. P. 2015. Theiner, Y. Drexel, M. Neuner, M. & Hofstetter, G.
Comprehensive Database for Concrete Creep 2017. Comprehensive study of concrete creep ,
and Shrinkage: Analysis and Recommendations shrinkage , and water content evolution under
for Testing and Recording. ACI Materials sealed and drying conditions. Strain (53): 1–13.
Journal 112(4). doi:10.14359/51687453 Tibshirani, R. 1996. Regression shrinkage and
Inkoom, S. Sobanjo, J. Barbu, A. & Niu, X. 2019. Selection via the Lasso. Journal of the Royal
Prediction of the crack condition of highway Statistical Society. Series B (Methodological)
pavements using machine learning models. 267–288.
Structure and Infrastructure Engineering: Vu, D. & Hoang, N. 2015. Punching shear capacity
Maintenance, Management, Life-Cycle Design estimation of FRP- reinforced concrete slabs
and Performance 0(0): 1–14. using a hybrid machine learning approach.
Jo, J. & Jadidi, Z. 2019. A high precision crack Structure and Infrastructure Engineering:
classification system using multi-layered image Maintenance, Management, Life-Cycle Design
processing and deep belief learning. Structure and Performance.
and Infrastructure Engineering: Maintenance, doi:10.1080/15732479.2015.1086386
Management, Life-Cycle Design and Zheng, Z. Hu, D. Liu, P. Sha, F. Liu, L. & Yu, Z.
Performance 1–9. 2020. Considering the effect of the randomness
Kamen, A. Denarie, E. Sadouki, H. & Bruhwiler, E. of concrete strength and relative humidity on
2009. UHPFRC tensile creep at early age. concrete creep. Structural Concrete 1–15.
Materials and Structures 42: 113–122. Zhu, J. Zhang, C. Qi, H. & Lu, Z. 2019. Vision-based
Kang, M. Yoo, D. & Gupta, R. 2021. Machine defects detection for bridges using transfer
learning-based prediction for compressive and learning and convolutional neural networks.
flexural strengths of steel fiber-reinforced Structure and Infrastructure Engineering:
concrete. Construction and Building Materials Maintenance, Management, Life-Cycle Design
266: 121117. and Performance 0(0): 1–13.
Liu, H. & Zhang, Y. 2020. Bridge condition rating
data modeling using deep learning algorithm.
Structure and Infrastructure Engineering:
Maintenance, Management, Life-Cycle Design
and Performance 1–14.
Marani, A. Jamali, A. & Nehdi, M. L. 2020.
Predicting Ultra-High-Performance Concrete
Compressive Strength Using Tabular
Generative. Materials 13: 4757.
Raphael, W. Faddoul, R. Geara, F. & Chateauneuf, A.
2012. Improvements to the Eurocode 2
shrinkage model for concrete using a large
experimental database. Structural Concrete
13(3): 174–181.
Raphael, W. Zgheib, E. & Chateauneuf, A. 2018.
Experimental investigations and sensitivity
analysis to explain the large creep of concrete
deformations in the bridge of Cheviré. Case
Studies in Construction Materials 9: e00176.
Shen, D. Jiang, J. Wang, W. Shen, J. & Jiang, G.
2017. Tensile creep and cracking resistance of
concrete with different water-to-cement ratios at
early age. Construction and Building Materials
146: 410–418.
Shen, D. Li, C. Liu, C. Li, M. & Kang, J. 2020.

You might also like