Elsevier

Measurement

Volume 165, 1 December 2020, 108141
Measurement

Ensemble machine learning model for corrosion initiation time estimation of embedded steel reinforced self-compacting concrete

https://doi.org/10.1016/j.measurement.2020.108141Get rights and content

Highlights

  • Corrosion initiation time of reinforced cement-limestone concrete is estimated.

  • Ensemble machine learning algorithm is used in corrosion initiation time estimation.

  • Random forest ensemble proved great in estimating corrosion initiation time.

  • Using random forest ensemble yielded a good prediction outcome (CC = 0.9867).

Abstract

Corrosion initiation time of embedded steel is an important service life parameter, which depends on concrete material make-up, exposure environment, and duration of exposure. Early and accurate determination of corrosion initiation time will aid in designing durable reinforced concrete, saves cost and time. This study leveraged on the power of ensemble machine learning by combining the performances of different models in estimating the corrosion initiation time of steel embedded in self compacted concrete using corrosion potential measurement. The concrete specimens were prepared with limestone powder as supplementary addition to Portland cement and was exposed to 5% sodium chloride in accordance with the requirements of ASTM C876 – 15 for 8 months. During the exposure, corrosion potential of the embedded steel was measured, and the recorded datasets were used in training five different machine learning models. With cement, limestone powder, coarse aggregate, fine aggregate, water and exposure period.as input variables, five different models were developed to estimate the corrosion initiation time (determined from the corrosion potential measurements) of the embedded steel. With respect to model predictive performance, the acquired results demonstrated that the random forest (RF) ensemble model amongst other trained models performed best with 85/15 dataset percentage split for the training and testing. RF ensemble performed best with CC and RMSE of 99.01% and 18.2747 mV for training, and 98.67% and 25.0298 mV for testing respectively. Hence, due to its superior and robust performance, this study proposes RF ensemble model in the estimation of corrosion initiation time of embedded steel in reinforced limestone-cement blend concrete.

Introduction

Structural degradation of concrete structures triggered by the corrosion of embedded steel remains the major durability problem faced by the construction industry [1], [2]. Steel corrosion is activated by either CO2 or chloride (Cl-) ion ingression into concrete, reacting with and destroying the alkaline hydrated cement products necessary to shield (passivate) the embedded steel. As corrosion sets in, the cross-section of the embedded steel is reduced, thereby losing its tensile and flexural strengths capacities [3]. This leads to structural cracks on the concrete that, in turn, reduces the structure's ability to carry loads. In comparison with carbonation-induced steel corrosion, according to reports, chloride ion induced corrosion is more detrimental and costlier to repair [4].

According to the 2013 NACE Impact Study, the cumulative cost of corrosion globally was estimated at US$2.5 trillion, which is almost 4% of the global gross domestic product (GDP). With the corrosion prevention best practices, a global savings of 15–35% cost of damage could be realized [5]. In some extreme reinforced concrete deteriorative state, the cost of repair is more than the initial cost of construction [6]. Therefore, the service life prediction of reinforced concrete (RC) structures based on corrosion of embedded steel is essential. The initiation time of corrosion for embedded steel helps in determining the service life of the RC structure [7]. It is time it takes for CO2 or Cl ions to journey from an RC-exposed environment to reach a threshold level on the surface of the embedded steel, compromises its passivity, and initiates active corrosion [6]. Time to corrosion initiation mainly determines the resistance of reinforced concrete exposed to the CO2 or Cl ion-laden environment.

For RC structures, the occurring electrochemical processes at the passive layer (a region between the concrete and the surface of the steel) along with reinforcement corrosion complicate understanding and measurement [8]. However, ASTM C876 [9] standard was developed as a guideline for monitoring reinforcement corrosion through empirical-based potential measurement. The embedded steel potential is readily measured through the determination of the voltage difference between steel immersed in the chloride-laden environment and a suitable reference electrode. The non-destructive assessment approach reveals the active or passive corrosion state of the embedded steel, which is used to estimate the time to the initiation of corrosion. A highly dense and low permeable concrete such as self-compacting concrete (SCC) is one of the proven ways to thwart the ingress of Cl ions into the concrete giving assurance of longer service life to RC structures. Many studies [10], [11], [12] have examined the different SCC corrosion resistance physiognomies and its associated durability characteristics and, as such, confirmed the superior performance of SCC compared to the conventional concrete.

The high nonlinearity nature of embedded steel corrosion due to complex features of reinforced concrete structures makes corrosion properties prediction difficult due to the lack of theoretical basis for some phenomena [13]. Most of the models used for prediction adopt empirical equations in calibrating experimental data where empirical constants are required but difficult to obtain. This is due to the inherently complex relationship between concrete's mixture proportions and the properties to be modeled [14]. An advanced tool can adequately monitor the data complexities and deliver accurate results.

The adoption of machine learning over the years in modeling real and complex civil engineering problems was built on its ability to capture interrelationships between input and output data [15], [16]. The potential of machine learning to predict and discover patterns in materials properties and characteristics has been reported [17]. Machine learning tools have been deployed in many research areas [18], [19], [20], [21], [22], [23] and specifically in areas such as crack propagation investigation in concrete [24], durability assessment and safety monitoring [25], corrosion initiation time studies [26], chloride diffusion [19], autogenous shrinkage in concrete [27], slope reliability analysis [28] and so on. In corrosion-related studies, Taffese et al. [15] reviewed the machine learning capabilities in durability and service life assessment, focusing on the adequacy and applicability of models compared to the empirical models. Jui-Sheng Chou et al. [29] developed a single, ensemble and hybrid models from four machine learning techniques including artificial neural networks (ANNs), classification and regression tree (CART), linear regression (LR) and support vector regression/machines (SVR/SVM), to estimate corrosion of embedded steel in RC and marine corrosion rate of carbon steel. The hybrid meta-heuristic regression model performed better than other considered models (single and ensemble) in the estimation of pitting corrosion risk and marine corrosion rate. Sadowski and Nikoo [30] used an ANN in combination with the imperialist competitive algorithm (ICA) model to predict the corrosion current density. The inputs of the model were temperature, AC resistivity over and away from embedded steel, and DC resistivity over the embedded steel. When the results from the ICA-ANN model were compared with the ones from genetic algorithm, it was found that ICA-ANN performed better in predicting corrosion current density of the embedded steel bar. Volker and Gino Ebell [31] creatively deployed machine learning inspired data fusion approach based on logistic regression algorithms for corrosion recognition in order to enhance the non-destructive corrosion testing of RC structure. There are other parameters used to define the corrosion properties such as time to initiation of corrosion, which are yet to be modeled using machine learning.

The ability to predict corrosion initiation time of embedded steel in reinforced concrete structures is essential for (i) planning of proper maintenance of the existing structure and optimized cost of repair works for failed concrete, (ii) designing new structures, and (iii) sustainability assessments of new cement materials [26]. With some computational intelligence models developed to predict the initiation time of steel corrosion, many models [7], [29], [30], [32], [33], [34], [35], [36] focused on the prediction or estimation of parameters to qualify the level of embedded steel corrosion in RC structures. The closest to this study are the works of Jiang et al. [37], where steel corrosion initiation time of sewer RC is predicted and used to estimate the corrosion rate. Similarly, Hodhod and Ahmed [38] predicted the corrosion initiation time of slag concrete using ANN. The corrosion properties of embedded steel in RC structure depend on the exposure and the mixture proportion of ingredients, especially the cementitious materials. The ingredients containing the binder affect the concrete's durability performance. The time to initiation of reinforcement corrosion influences the service life studies of RC structure. The primary aim of this paper is to assess and compare the performances of different ML techniques to propose an optimal model in predicting the corrosion initiation time of the embedded steel. The datasets for the study were generated from the experimental investigation of the effect of incremental addition of limestone powder to cement (CLSP) on corrosion initiation time of embedded reinforcement in SCC. To develop the model, CLSP percentage addition and exposure periods are the descriptors (inputs), and corrosion potential is the target (output) for both the training and testing stages.

Section snippets

Linear regression model

Linear regression (LR) finds the relationship between the target and one or more predictors. Simple linear regression is basically a technique used to establish statistical but not deterministic relationships between two continuous variables. One of the variables is the predictor or independent and the other is response or dependent variable. The correlation between the dependent variables and independent variables can be said to be deterministic, if there is a 100% accuracy expressed by one of

Performance criteria for the ML models

The performance of the developed models were assessed using some statistical tools, which are coefficient of correlation (CC), mean absolute error (MAE) and root mean square error (RMSE) and their equations are labelled below as Eqs. (16), (17), (18).

Developed from Eqs. (16), (17), (18), root mean square error (RMSE), mean absolute error (MAE) and coefficient of correlation (CC) are all used to evaluate the performances of the developed models [60].RMSE=1ni=1nei2MAE=1ni=1neiCC=i=1nCpi(exp)-Cp

Materials and mixture designs

A comprehensive corrosion study was conducted on embedded steel in self-compacting RC prepared using a blend of ordinary Portland cement and limestone powder. The experimental investigation details about the extraction of the datasets are shown in the coming segments.

A short introduction of the materials used in the development of the self-compacting concrete is discussed in the following segments.

In the experimental study, ASTM C 150 Type I Portland cement with specific gravity (SG) of 3.15

Data collection

After the concrete was water-cured for the required 28 days, the lollipop concrete specimens were exposed to 5% NaCl solution for concrete resistance evaluation following the ASTM C876-15 as shown in Fig. 2. The standard is an empirical-based procedure of corrosion potential measurements used as guidance for reinforcement corrosion monitoring. The open-circuit corrosion potentials of the SCC specimens were measured for 8 months using a high impedance voltmeter. The reference electrode and the

Preliminary dataset preparation and evaluation

The checking of outliers and extreme values returned an after-check result of no outliers and extreme values in the corrosion datasets. After the check, the dataset was randomized, as mentioned in Section 7.3.

Effect of data transformation on ML algorithm selection

In order to understand the effect of data transformation on the performances of the developed models, three operations (attribute-selection, normalization and standardization) were applied on the datasets using the same hyperparameters for the respective models. This is imperative to

Concluding remarks

The inherent complexity and nonlinearity in the behavior of Portland cement concrete is already known, a blend of limestone with cement brings additional complexities to the concrete behavior. As a result, there is a huge nonlinear relationship between the concrete mixture parameters and the durability performance (corrosion initiation time for this study) of the developed concrete. Hence, there is need for computationally intelligent techniques with high predictive performance to learn these

CRediT authorship contribution statement

Babatunde Abiodun Salami: Conceptualization, Methodology, Software, Investigation, Data curation, Writing - original draft, Writing - review & editing, Visualization, Supervision, Project administration. Syed Masiur Rahman: Supervision, Writing - review & editing, Validation. Tajudeen Adeyinka Oyehan: Methodology, Visualization, Software. Mohammed Maslehuddin: Resources, Project administration, Writing - review & editing. Salah U. Al Dulaijan: Project administration, Writing - review & editing.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgement

The authors express their sincere appreciation and gratefully acknowledge the support received from the Civil and Environmental Department and Center of Engineering Research (CER), King Fahd University of Petroleum & Minerals, Dhahran, Saudi Arabia. Our appreciation also goes to Dr Sunday O. Olatunji of Imam Abdulrahman Bin Faisal University, Dammam, Saudi Arabia and Dr Fatai Anifowose of Saudi Aramco, Saudi Arabia for their guidance.

References (73)

  • W.Z. Taffese et al.

    Machine learning for durability and service-life assessment of reinforced concrete structures: Recent advances and future directions

    Autom. Constr.

    (2017)
  • N.-D. Hoang

    Estimating punching shear capacity of steel fibre reinforced concrete slabs using sequential piecewise multiple linear regression and artificial neural network

    Measurement

    (2019)
  • N.-D. Hoang et al.

    Prediction of chloride diffusion in cement mortar using Multi-Gene Genetic Programming and Multivariate Adaptive Regression Splines

    Measurement

    (2017)
  • A. Bagheri et al.

    The use of machine learning in boron-based geopolymers: Function approximation of compressive strength by ANN and GP

    Measurement

    (2019)
  • T. Gupta et al.

    Prediction of mechanical properties of rubberised concrete exposed to elevated temperature using ANN

    Measurement

    (2019)
  • J. Jonbi et al.

    Modeling the water absorption and compressive strength of geopolymer paving block: An empirical approach

    Measurement

    (2020)
  • A. Müsevitoğlu et al.

    Experimental and analytical investigation of chemical anchors’s behaviour under axial tensile

    Measurement

    (2020)
  • G. Bayar et al.

    A novel study for the estimation of crack propagation in concrete using machine learning algorithms

    Constr. Build. Mater.

    (2019)
  • Y. Yu et al.

    Machine learning aided durability and safety analyses on cementitious composites and structures

    Int. J. Mech. Sci.

    (2019)
  • U.M. Angst

    Predicting the time to corrosion initiation in reinforced concrete structures exposed to chlorides

    Cem. Concr. Res.

    (2019)
  • J. Liu et al.

    Prediction of autogenous shrinkage of concretes by support vector machine

    Int. J. Pavement Res. Technol.

    (2016)
  • H. Zhao

    Slope reliability analysis using a support vector machine

    Comput. Geotech.

    (2008)
  • J.-S. Chou et al.

    The use of artificial intelligence combiners for modeling steel pitting risk and corrosion rate

    Eng. Appl. Artif. Intell.

    (2017)
  • Y.F. Wen et al.

    Corrosion rate prediction of 3C steel under different seawater environment by using support vector regression

    Corros. Sci.

    (2009)
  • G. Jiang et al.

    Predicting concrete corrosion of sewers using artificial neural network

    Water Res.

    (2016)
  • O.A. Hodhod et al.

    Modeling the corrosion initiation time of slag concrete using the artificial neural network

    HBRC J.

    (2014)
  • H. Liang et al.

    Improved Estimation in Multiple Linear Regression Models with Measurement Error and General Constraint

    J. Multivar. Anal.

    (2009)
  • J.-S. Chou et al.

    Machine learning in concrete strength simulations: Multi-nation data analytics

    Constr. Build. Mater.

    (2014)
  • M.R. Moghaddasi et al.

    ICA-ANN, ANN and multiple regression models for prediction of surface settlement caused by tunneling

    Tunn. Undergr. Space Technol.

    (2018)
  • M. Sarıdemir

    Prediction of compressive strength of concretes containing metakaolin and silica fume by artificial neural networks

    Adv. Eng. Softw.

    (2009)
  • M.A. Kewalramani et al.

    Concrete compressive strength prediction using ultrasonic pulse velocity through artificial neural networks

    Autom. Constr.

    (2006)
  • F. Altun et al.

    Predicting the compressive strength of steel fiber added lightweight concrete using neural network

    Comput. Mater. Sci.

    (2008)
  • M. Sarıdemir

    Predicting the compressive strength of mortars containing metakaolin by artificial neural networks and fuzzy logic

    Adv. Eng. Softw.

    (2009)
  • F. Demir

    Prediction of elastic modulus of normal and high strength concrete by artificial neural networks

    Constr. Build. Mater.

    (2008)
  • M.Y. Mansour et al.

    Predicting the shear strength of reinforced concrete beams using artificial neural networks

    Eng. Struct.

    (2004)
  • K.O. Akande et al.

    Investigating the effect of correlation-based feature selection on the performance of support vector machines in reservoir characterization

    J. Nat. Gas Sci. Eng.

    (2015)
  • Cited by (62)

    View all citing articles on Scopus
    View full text