Beruflich Dokumente
Kultur Dokumente
1 Introduction
D. Dutta
Civil Engineering, Manipal Institute of Technology, Manipal 576104, India
e-mail: diprodutta96@gmail.com
S. V. Barai (&)
Department of Civil Engineering, Indian Institute of Technology Kharagpur,
Kharagpur 721302, West Bengal, India
e-mail: skbarai@civil.iitkgp.ernet.in
dry particles of the aggregates need to be removed completely after the addition of
water-to-concrete mix. However, in practice, fully compacted concrete still contains
1% of air gaps [3].
Along with the basic components of conventional concrete (Portland cement,
fine and coarse aggregates and water), HPC incorporates supplementary cementi-
tious ingredients like fly ash, silica fume, blast furnace slag and usually a chemical
admixture, such as superplasticizer. These additional ingredients make HPC a
complex conglomerate, and predicting its strength becomes difficult.
Compressive strength can also be predicted using regression analyses and sev-
eral methods which can be collectively classified as machine learning (ML). One of
the popular machine learning approaches is Artificial Neural Networks (ANN).
The ANN model is inspired by the biological neural network. It utilises several
computing elements to parallel process the input data. These elements, called
artificial neurons, are arranged in layers and interact with each other via weighted
connections. Backpropagation algorithms are popularly used to solve problems of
ANN. Backpropagation minimises the error using a gradient descent method. Yeh
[5] compared regression and ANN for predicting compressive strength of the
concrete. He found that the ANN model performs better than regression.
There is a need to search and develop better machine learning models apart from
well-reported ANN models. This study explores ANN further, along with another
model of ML called the Instance-Based Learning (IBL) or k-Nearest Neighbours
(k-NN).
In the present study, the main objective is to compare the machine learning
models—ANN and IBL—for predicting compressive strength of the concrete. This
study presents—ANN based two back propagation algorithms and the IBL model
for integral values of k ranging from 1 to 6.
Machine learning (ML), a subset of artificial intelligence, has evolved from the
study of pattern recognition. Two ML subfields have been studied. Yeh [5] com-
pared regression and ANN for predicting compressive strength of the concrete. He
found that the ANN model performs better than regression.
Instance-based learning (IBL) is a learning method which compares new
problem instances with the instances that the programme has been trained. Its k-
nearest training examples (k-nearest neighbours or k-NN) are utilised to deliver the
output, and the value is determined as the mean of the output values of its k-nearest
neighbours.
The contribution of the neighbours has weights assigned to them. This allows the
output to be more similar to the nearer neighbours than the distant ones.
Artificial neural networks (ANN) is learning method used to solve problems
requiring nonlinear and parallel processing through a mesh of simple computing
elements called artificial neurons. Each neuron can have R inputs (x) and each input is
Prediction of Compressive Strength of Concrete: Machine … 505
weighted (w). There is a bias with a constant input of 1. The sum of the weighted inputs
and the bias passes through the transfer function (f) and generates an output (a).
!
X
R
a ¼ f bþ wi xi
1
The data was collected from the machine learning repository of the University of
California, Irvine [5]. It provided 1030 samples. The output variable is the com-
pressive concrete strength (MPa) which depends on eight input variables:
1. Cement (kg/m3),
2. Fly ash (kg/m3),
3. Blast furnace slag (kg/m3),
4. Water (kg/m3),
5. Superplasticizer (kg/m3),
6. Coarse aggregate (kg/m3),
7. Fine aggregate (kg/m3)
8. Age at the time of testing (days).
506 D. Dutta and S. V. Barai
The set includes data from several sources. This verifies the reliability of the data
[2]. Only, the 28-day samples have been taken for the study. The summary of the
425 samples is shown in Table 1.
The data set was randomised before each iteration. 43 samples ( 10%) were
kept aside for testing, while the remaining was used for training the algorithm. The
same sets of randomised data were utilised to conduct the experiment across the
three algorithms to ensure a perfect comparison between the three.
4 Model Development
The weighting scheme used in the analysis gave each neighbour a weight, which
was the inverse of the Euclidean distance between the simulation point and the
neighbour. The analysis was run for 1-NN through 6-NN models. The average of
the absolute error in the prediction of the output had been calculated and a graph of
predicted versus expected values of the output was generated.
The feedforward neural network is trained using seven inputs in the input layer and
one output in the output layer. Training is the process of presenting testing data to
the network and has it to modify its weights to accurately produce the strength
when presented with the mix proportions. The performance function in the net-
works is mean squared normalised error (MSE). The parameters for the training of
the network are as follows:
For LM,
• No. of hidden layers = 1,
• No. of hidden units = 7,
• Initial momentum (mu) = 0.001,
Prediction of Compressive Strength of Concrete: Machine … 507
5 Model Performance
The absolute error percentage (AEP) was calculated for each testing data point and
subsequently, the mean of that was calculated (MAEP). The MAEP thus produced
by each algorithm (IBL, LM and GDX) was statistically analysed using hold-out
and cross-validation. These give an estimate of the accuracy of the performance of
the models in a practical setting [6].
In hold-out (HO), the data was divided randomly into training and testing sets
consisting of 90% and 10% of the entire data. The errors in the output of the testing
set were used for the analysis. Ten iterations of each k-NN, LM and GDX were
evaluated using HO.
Cross-validation (CV) is preferred for prediction models [6]. In k-fold
cross-validation, the data set is equally split into k subsets, each of which is used as
the test set while the others (k-1) are used together as the training set. The average
error of the k trials is calculated. A 10-fold cross-validation has been used for the
analysis.
6 Results
1-NN through 6-NN models were run and the experiment was iterated 10 times and
analysed. Table 2 shows the results of the hold-out experiments. The 2-NN model
has shown the best performance and its summary is as follows:
508 D. Dutta and S. V. Barai
Mean ¼ 8:5%
S:D: ¼ 1:55
Minimum ¼ 5:81%
Maximum ¼ 10:87%
1st Quartile ¼ 7:60%
Median ¼ 8:25%
3rd Quartile ¼ 10:08%
The 10-fold cross-validation analysis on the k-NN model too shows that the
2-NN algorithm performs the best. Table 3 shows evaluation of the analysis, and its
summary is presented as follows:
Mean ¼ 9:32%
S:D: ¼ 0:78
Minimum ¼ 8:52%
Maximum ¼ 11:49%
1st Quartile ¼ 8:99%
Median ¼ 9:17%
3rd Quartile ¼ 9:33%
A scatter plotted graph of the expected values versus the predicted values of the
fifth experiment of 2-NN using hold-out is shown in Fig. 1.
Fig. 1 Scatter plot of expected versus predicted values for Exp. 5 of hold-out for 2-NN
Table 4 Arithmetic mean and standard deviation of ANN models using hold-out
LM GDX
MAEP 15.18 20.18
SD 2.36 6.93
A hold-out analysis was performed on the artificial neural network models. Table 4
shows the AEPs and MAEP as calculated from the LM and GDX algorithms. The
summary of the findings are as follows:
For LM,
Mean ¼ 15:18%
S:D: ¼ 2:36
Minimum ¼ 11:39%
Maximum ¼ 18:67%
1st Quartile ¼ 13:80%
Median ¼ 15:54%
3rd Quartile ¼ 16:78%
510 D. Dutta and S. V. Barai
For GDX,
Mean ¼ 20:18%
S:D: ¼ 6:93
Minimum ¼ 12:89%
Maximum ¼ 38:39%
1st Quartile ¼ 16:84%
Median ¼ 18:62%
3rd Quartile ¼ 20:49%
Tables 5 and 6 show the evaluation of the 10-fold analysis of the fifth experi-
ment of neural network method using LM and GDX, respectively. Their summaries
are as follows:
Prediction of Compressive Strength of Concrete: Machine … 511
For LM,
Mean ¼ 14:2%
S:D: ¼ 1:2
Minimum ¼ 12:83%
Maximum ¼ 16:33%
1st Quartile ¼ 13:1%
Median ¼ 14:08%
3rd Quartile ¼ 14:75%
For GDX,
Mean ¼ 19:45%
S:D: ¼ 4:19
Minimum ¼ 14:52%
Maximum ¼ 30:29%
1st Quartile ¼ 17:42%
Median ¼ 18:62%
3rd Quartile ¼ 20:56%
Scatter plotted graphs of the expected values versus the predicted values of the
seventh experiment of LM and GDX using hold-out are shown in Figs. 2 and 3,
respectively.
7 Comparative Study
The analysis shows that the ML algorithms can predict the compressive strength of
28-day hardened concrete with varying accuracy. A comparison of the data in
Tables 2 and 4 is shown in Fig. 4.
Figure 4 shows that IBL performs considerably better than either of the neural
network algorithms. It follows that the 2-NN model performs slightly better than the
other k-NN models with an MAEP of 8.50%. The 2-NN model also has the lowest
SD at 1.55.
8 Conclusions
HPC is a complex material and that makes the modelling and predicting its
behaviour difficult. The study was an attempt to demonstrate the possibilities of
using machine learning to predict the compressive strength of concrete. The ini-
tialisation values were not optimised, and hence there is a scope for further study of
the ANN models. Optimising the initial weights and constants may give better
results.
It must be noted that none of the explored methods is applicable to extrapolation.
Hence, the minimum and maximum of the training set must be selected carefully.
Also, more training instances will produce more accurate predictions. Hence, larger
training data sets should be utilised for modelling.
Based on the comparison of the models, the following can be concluded:
• Instance-Based Learning (IBL) is a preferred choice for the prediction of the
compressive strength with the seven parameters.
• The neural network models are comparatively easier to train and hence easier to
simulate the predictions, but their performance is relatively poor.
• The Levenberg–Marquardt (LM) algorithm can be used for approximations
where the tolerance for error is high (preferably above 15%).
• It is advisable to not use the variable-learning-rate-with-momentum gradient
descent algorithm (GDX) for the prediction.
References