Predicting the slump of industrially produced concrete using machine learning: A multiclass classification approach
Introduction
Concrete is one of the most used cement-based materials, a composite that contributes 40–45% of the global cement consumption [1]. Its workability is considered a vital performance and quality control criterion for fresh concrete. Workability indicates the ease of mixing, adjusting, compacting, and placing fresh concrete into a desirable shape without losing homogeneity and performance [2,3]. Depending on the placing conditions, different concrete structures require different workability, ranging from 10 mm to 220 mm [4]. Concrete is mainly delivered to the job sites as ready-mix concrete using agitating trucks. The concrete trucks will be rejected if the concrete does not meet the required slump value. Such rejections could significantly impact the construction project, costing extra resources, time, and money associated with construction reworks and the dumping of wasted concrete [1,5]. Workability is traditionally assessed by a slump test, which is time-consuming and costly, involving repetitive and manual operations in preparing a slump cone and measuring it with ruler apparatus.
In recent years, analytical and computational methods are widely used in material science and engineering. Machine learning (ML) has been applied in materials design and discovery, material behavior modeling, and the prediction and optimization of material performance [[6], [7], [8]]. Materials investigated using ML include metals, alloys, composites, ceramics, polymers, and biomaterials. Numerous previous studies have used ML-based soft computing techniques to explore the performance of concrete. Concrete performance is investigated and assessed in terms of compressive, tensile, flexural and shear strengths, modulus of rupture and elasticity, hardened density, slump, tortuosity, permeability, carbonation depth, surface chloride concentration, and thermal expansion coefficient [7,[9], [10], [11], [12]]. For example, DeRousseau et al. [7], Cook et al. [11], and Zhang et al. [13] predicted concrete compressive strength and compared the performance of ML methods, including multiple linear regression (MLR), polynomial regression, artificial neural networks (ANN), support vector machine (SVM), random forest (RF), and extreme gradient boosting (xgboost). Nguyen et al. [14] used a heuristic algorithm to estimate compressive strength of normal and high performance concrete. Asteris et al. [15,16] used ANN to predict the compressive strength of concrete masonry structures. In another study, Asteris et al. [17] used ANN genetic programming (GP) to predict the compressive strength of metakaolin-based mortars. Liao et al. [18] predicted the ultimate axial load of concrete-filled steel tubes. Nilsen et al. [19] used the RF model to predict different concrete properties such as the coefficient of thermal expansion, compressive strength, modulus of elasticity and rupture, and splitting tensile strength. In particular, ML has proven to be a robust, economical, and sustainable way to evaluate and optimize concrete performance and control concrete quality.
Previous ML-based studies in concrete slump assessment may be classified into two groups. The first group used numerical data related to the concrete mix of laboratory-produced concrete. Yeh [20,21] and Öztaş et al. [22] applied ANN to model slump as a function of the quantities of concrete constituents, water to binder ratio, and superplasticizer to binder ratio. Cihan [16,23], Amlashi et al. [24], and Moayedi et al. [25] used multiple ML tools of ANN, multivariate adaptive regression spline (MARS), M5 model tree (M5Tree), RF, bagging, fuzzy logic, and SVM to predict slump. The second group used image-based data and applied computer vision technologies to predict slump at the actual construction site [26]. Kim et al. [27], Tuan et al. [26] and Ding et al. [28] analyzed slump images by deep learning tools such as convolutional neural networks (CNN), region-based CNN, fast region-based CNN, and long short-term memory (LSTM). These studies shed light on the fresh concrete performance assessment and quality control by predicting and assessing the slump, generating many valuable results that have improved the practice of the concrete industry. However, there were shortcomings. Studies in group one were mainly based on laboratory concrete with a small number of data points (30–250), failing to account for the uncertainties due to the varying conditions in concrete design, production, transportation, mixing, and placement [13,29]. Studies in group two used field concrete data. Nonetheless, the computer vision models were not that robust as they were computationally extensive and resource-consuming, necessitating repetitive equipment operations and the learning of image data in three-dimensional red, green, and blue metrics.
Our study intends to improve the quality control and performance assessment of fresh concrete by predicting the slump using an ML-based classification approach. The novelties and contributions of this study are summarized as follows: First, we applied ML methods to predict the concrete field slump. Such a computational approach is beneficial over empirical testing procedures and saves time, money, and material. Second, we used a large number of 3359 observations of industrially produced concrete to model and predict the concrete slump. Using actual field data instead of laboratory data, as in most previous studies, better reflects the on-site conditions and gives more reliable results [13,29]. Third, we proposed a multiclass classification methodology that is more effective than previously proposed computer vision-based methodologies. The classification approach is robust and more computationally efficient without requiring high-dimensional image-based data. It is applicable at the batching plant just after preparing the concrete mix without consuming the resources necessary for computer vision-based testing and setup (e.g., setup of cameras). Lastly, this study followed a comprehensive methodology and provided practical guidance on using the multiclass classification method. We explored the mainstream ML methods, compared them against a wide range of performance evaluation metrics, and drew insights from data analytics and visualization.
Section snippets
Data used in this study
A contractor and concrete producer in China provided the raw data for this study. The company supplies a wide range of concrete composites for various building and infrastructure projects throughout China and Southeast Asia. The concrete design and batching comply with standards such as British Standards (BS), British Standard European Norm (BS EN), and the American Society for Testing and Materials (ASTM) [[30], [31], [32], [33], [34]]. A total of 3599 datapoints related to 24 concrete mix
Seven slump classification models
Seven well-known classification models are employed to predict the slump class, which embody linear, non-linear, and ensemble learning. The following subsections provide a concise overview of these classification models and their associated hyperparameters to avoid exhaustive non-trivial debate. Readers are recommended to read Hastie et al. [39] and Mohri et al. [40] if an in-depth understanding is desired.
Data statistics
Key data statistics such as mean, median, and standard deviation are evaluated to describe the central tendency and variability of the data across the slump classes (see Table 1). Also, Fig. 2 shows the percentages of datapoints in each of the eight slump classes. It is seen that there is a considerable difference in the percentages between the classes. For example, the percentage of the 100 mm slump is 32.2% (1082 out of 3359 datapoints) and the 135 mm class is only 0.5% (18 out of 3359
Overall ML process
The ML modeling and analysis scripts are developed in the R programming language, a robust programming environment for data mining and modeling [37,38]. Table 2 enlists the R-libraries, and the corresponding R-functions for each ML model. Some details of this ML modeling and analysis process are discussed in the following sections.
Model evaluations metrics
The data used in this study is highly imbalanced, with a considerable difference in the frequencies of the slump classes. Performing multiclass classification on imbalanced data necessitates a credible performance evaluation. The following metrics are used in the evaluation: accuracy, Kappa, Matthews correlation coefficient (MCC), logLoss, receiver operating characteristic (ROC) curve, precision-recall (PR) curve, and the area under the curve (AUC) [57,58,70,71]. The metrics values
Performance comparison of the ML classification models
The seven ML classification models are trained and validated according to the procedures discussed in Section 5. This section compares the performance of the seven models in the training and validation stages and identifies the best ones for predicting the slump class.
Observations and discussions
The first observation is that nonlinear classification models perform much better than linear ones. The reason is twofold. First, linear models cannot deal with the concrete data's nonlinearity due to varying design and production conditions and the complex interactions between the properties of concrete and its constituents. Second, the data is highly imbalanced. For example, there are 1082 datapoints for the 100 mm slump class while only 35 datapoints for the 50 mm slump class. It is known
Conclusions
This study used ML-based multiclass classification approach to predict the slump of industrially produced concrete. In contrast to previous studies using laboratory data of fewer than 300 datapoints, this work is based on 3359 observations of industrially produced concrete for a wide range of construction projects. Seven classification models are explored, taking mix proportions as the input variables to predict slump into one of eight classes. Moreover, the ML models are evaluated against
Credit author statement
Xueqing Zhang: Conceptualization, Methodology, Reviewing and Editing, Supervision. Muhammad Zeshan Akber: Writing - Original Draft, Formal Analysis, Visualization. Wei Zheng: Resources, Investigation, Data Curation.
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.
References (74)
- et al.
Waste generation from the production of ready-mixed concrete
Waste Manag.
(2019) - et al.
Fresh and mechanical properties of concrete made of binary substitution of millet husk ash and wheat straw ash for cement and fine aggregate
J. Mater. Res. Technol.
(2021) - et al.
Materials discovery and design using machine learning
J. Mater.
(2017) - et al.
A comparison of machine learning methods for predicting the compressive strength of field-placed concrete
Construct. Build. Mater.
(2019) - et al.
Modelling and optimization of microhardness of electroless Ni-P-TiO2composite coating based on machine learning approaches and RSM
J. Mater. Res. Technol.
(2021) - et al.
Machine learning prediction of mechanical properties of concrete: critical review
Construct. Build. Mater.
(2020) - et al.
Estimating strength properties of geopolymer self-compacting concrete using machine learning techniques
J. Mater. Res. Technol.
(2020) - et al.
Prediction of seven-day compressive strength of field concrete
Construct. Build. Mater.
(2021) - et al.
Heuristic algorithm-based semi-empirical formulas for estimating the compressive strength of the normal and high performance concrete
Construct. Build. Mater.
(2021) - et al.
Soft computing-based models for the prediction of masonry compressive strength
Eng. Struct.
(2021)