Key Engineering Materials Submitted: 2018-10-24

ISSN: 1662-9795, Vol. 792, pp 166-169 Accepted: 2018-11-08

doi:10.4028/www.scientific.net/KEM.792.166 Online: 2018-12-21
© 2018 Trans Tech Publications, Switzerland

Adapting ANNs in SONREB Test to Estimate Concrete Compressive

Strength
Yu-Ren Wang1,a, Loan T.Q. Ngo1,b, Yi-Fang Shih1,c,
Yen-Ling Lu1,d and Yi-Ming Chen1,e
1
Department of Civil Engineering, National Kaohsiung University of Science and Technology,
415 Chien-Kung Road, Kaohsiung, Taiwan 807
a
[email protected], [email protected], [email protected],
d
[email protected], [email protected]

Keywords: Non-Destructive Testing; Ultrasonic Pulse Velocity Test; Rebound Hammer Test; ANNs;
SONREB; Concrete Compressive Strength.

Abstract. SONREB method is a non-destructive testing (NDT) method for estimating the concrete
compressive strength. It is conducted by combining two popular NDT methods: ultrasonic pulse
velocity (UPV) test and rebound hammer (RH) test. Several researches have been attempted to find
the correlation of the different testing method data with actual compressive strength. This research
proposes a new Artificial Intelligence based approach, Artificial Neural Networks (ANNs), to
estimate the concrete compressive strength using the UPV and RH test data. Data from a total of 315
cylinder concrete samples are collected to develop and validate the ANFIS prediction model. The
model prediction results are compared with actual compressive strength using mean absolute
percentage error (MAPE). With the adaption of ANFIS, the estimation error of SONREB test can be
reduced to 5.98% (measured by MAPE).

Introduction
For the construction industry, it is important to examine the quality of concrete materials after they
are casted in place. Especially, it is crucial to make sure the concrete compressive strength meet the
requirements. The most accurate way to obtain the concrete compressive strength is to conduct
destructive compressive tests. During construction, random concrete samples are collected from
different batches of concrete mix and then sent to the lab for destructive tests. However, it is not
possible to collect these samples after the construction is completed. As a result, non-destructive
(NDT) methods are implemented to estimate concrete compressive strength for existing structures.
Two of the most popular NDT methods to estimate compressive tests are ultrasonic pulse velocity
(UPV) tests and rebound hammer (RH) tests. Nevertheless, the estimations obtained from these two
tests are not very accurate. For these two individual tests, researches have shown that about 10%
~20% variance is observed between the estimation and actual compressive strength. For this reason,
some researches attempted to combine the measurement data from these two NDT methods (UPV
and RH). The combining method (ultrasonic pulse velocity test plus rebound hammer test) is named
SONREB method. Results from these researches showed that the variance is reduced to around 10%.
Most of these researches applied regression (linear or non-linear) analysis to find the relationship
between the NDT measurement data and actual compressive strength. This research intends to
improve the SONREB concrete compressive strength estimation accuracy with artificial intelligence
(AI) based techniques, Artificial Neural Networks (ANNs). ANNs are one of the popular AI
techniques and ANNs concepts are inspired by the biological neural networks of human brains.
ANNs have been successfully adopted in many fields, such as computer vision, speech recognition,
machine translation, social network filtering, playing board and video games, financing, engineering
and medical diagnosis. To collect experimental data, the researchers collaborated with a local
construction material testing lab. Both UPV and RH methods are conducted using cylinder concrete
samples. Destructive tests are also conducted to obtain actual compressive strength of these samples.
Data from a total of 315 test samples are collected to develop and validate the ANNs prediction
model. The results are presented in the paper and suggestions are also provided.

All rights reserved. No part of contents of this paper may be reproduced or transmitted in any form or by any means without the written permission of Trans
Tech Publications, www.scientific.net. (#112433240, Iowa State University, Ames, USA-25/01/19,07:27:46)
 Key Engineering Materials Vol. 792 167

Literature References
SONREB Method. The most popular non-destructive tests (NDT) to measure the concrete
compressive strength are Ultrasonic Pulse Velocity (UPV) test and Rebound Hammer (RH) test. UPV
tests are conducted by passing an ultrasonic pulse wave through the concrete, as shown in Fig. 1. The
distance and passing time are measured and the velocity is calculated. Higher velocities indicate good
quality and continuity of the concrete material, and the concrete compressive strength should be
higher. Rebound hammers (Schmidt hammer in this study) are used to estimate concrete strength by
referencing surface hardness and penetration resistance, as shown in Fig. 2. The test hammer hits the
concrete with a defined energy and the hammer records the rebound by measuring the height of the
bounce. Its rebound related to the hardness of the concrete and the rebound value can be used to
determine the compressive strength. Softer material will absorb more of the impact energy and the
rebound distance will be less.

Fig. 1. UPV test Fig. 2. RH Test

One of the major drawbacks of rebound hammer test is that it only measures the surface part of the
testing object. Combining rebound hammer test with ultrasonic pulse velocity test, known as
SONREB test, the concrete compressive strength estimation accuracy can be improved [1]. Linear
regression models with 95% prediction intervals are developed to estimate the concrete compressive
strength using the data from UPV and RH tests [2]. The results have shown that worst results are
obtained if only RH test data are used to predict the concrete strength. The models produce better
estimation results with UPV test data. With both the RH and UPV test data, know as SONREB
method, the prediction models produce the best concrete compressive strength estimations with an
average R2 of 0.95 [2]. Similar results are observed from other researches that RONREB method
produces better prediction results when comparing to individual RH or UPV test [3,4,5,6].
Artificial Neural Networks. Artificial neural networks (ANNs) are one of the main machine
learning tools used by many researchers to make predictions (categorization or regression). They are
inspired by the biological neural connections in the human brain. A collection of inter-connected
processing units or nodes named artificial neurons (perceptron) replicates the neural structure in the
human brain. Each artificial neuron has weighted inputs, transfer function and one output. Although a
single neuron can perform certain simple information processing functions, the power of neural
computations comes from connecting neurons in a network [6, 7]. A single neuron can perform
certain simple information processing functions, but the power of neural computations comes from
layers of inter-connected neurons. A typical 3-layer neural network is shown in Fig. 3, with one input
layer, one hidden layer and one output layer.
168 Engineering and Innovative Materials VII

Fig. 3. Artificial Neural Networks

ANNs are successfully applied in many research fields to make predictions. Topçu and Sarıdemir
[8] have successfully adapted ANNs to predict concrete compressive strength using inputs such as
day, Portland cement, water, sand, crushed stone, high range water reducing agent and fly ash. Their
results showed that ANNs models are able to produce good compressive strength prediction results
with an R2 of 0.99. As a result, this research also adapts ANNs to build the prediction model for the
SONREB method to estimate concrete compressive strength.

Model Development and Result Analysis

The researchers collaborate with a local construction material testing lab to collect the SONREB
experiment data. In a period of one year, a total of 315 concrete cylinder samples (φ15 cm × 30cm)
are tested and relative information is collected. SilverSchmidt N-Type rebound hammer is used to
conduct the rebound hammer test following the ASTM C805 standards. Proceq TICO ultrasonic
tester is used to conduct ultrasonic pulse velocity tests following the ASTM C597-02 standards. After
the UPV and RH tests, destructive compressive strength tests are conducted using HT-8391 hydraulic
concrete compression testing equipment to obtain the actual compressive strength. The rebound
number (Q) and ultrasonic pulse velocity (V) measured in the experiments are set as the input
variables for the ANNs model and the actual compressive strength is set as the output variable.
Among the 315 samples, around 80% (252) are selected as the training dataset for the prediction
models. The remaining 63 samples are set as the testing dataset to validate the prediction model. For
comparison purpose, linear regression analysis is first conducted using the training dataset. The linear
regression equation obtained from the training dataset is shown in Eq. 1.

fc = -852.581+13.207×Q + 0.179×V (1)

Data from the 63 samples are applied into Eq. 1 to obtain the compressive strength estimations.
The estimations are compared with the actual compressive strength to examine prediction accuracy.
The prediction accuracy is measured by mean absolute percentage error (MAPE), as shown in Eq. 2.
The MAPE for the linear regression model is 7.16%
�𝐴𝐴𝑡𝑡 − 𝐹𝐹𝑡𝑡 �
∑𝑛𝑛
𝑡𝑡=1 𝐴𝐴𝑡𝑡
𝑀𝑀𝑀𝑀𝑀𝑀𝑀𝑀 = (2)
𝑛𝑛

where At is the actual value and Ft is the forecast value

The ANNs model is implemented with NeuroSolutions 6.0 software. The rebound number (Q) and
ultrasonic pulse velocity (V) data from the 252 training samples are set as the input variables and the
actual compressive strength data obtained from these samples are set as the target variable. ANNs
prediction models with one and two hidden layers are developed. In addition, the researchers vary the
 Key Engineering Materials Vol. 792 169

number of neurons in each hidden layer to find the best ANNs prediction model with lowest MAPE.
The results showed that ANNs model with one hidden layer and 8 neurons in the hidden layer, (2-8-1
network formulation), yields the lowest training error, with MAPE of 5.32% and R2 of 0.9536. For
validation purpose, data from the 63 testing dataset are input into the best ANNs training model
obtained. These data are unseen (or new) to the ANNs model and are used to test if the model can still
yield good estimations. The MAPE obtained from the testing dataset is 5.98%, which is only slightly
higher than the training MAPE, 5.32%. This indicates that the ANNs model can yield consistent
predictions.

Summary
Previous researches have shown that the SONREB method can yield better concrete compressive
strength estimations comparing to individual ultrasonic pulse velocity test or rebound hammer test.
This research adopted artificial neural networks to develop concrete compressive strength models in
hope to further improve the SONREB method estimations. Experiment data from a total of 315
concrete cylinder test samples are collected to develop and validate the proposed ANNs model. The
results showed that the ANNs model developed by the researchers has successfully improved the
SONREB method. The best training ANNs model obtained yields only 5.98% of MAPE using a
randomly chosen testing dataset of 63 samples.

References
[1] L. Nobile and M. Bonagura: Accuracy of non-destructive evaluation of concrete compression
strength, The 12th International Conference of the Slovenian Society for Non-Destructive
Testing, Portorož, Slovenia (2013).
[2] H. Y. Qasrawi: Concrete strength by combined nondestructive methods Simply and reliably
predicted, Cement and Concrete Research, Vol. 30, No. 5 (2000), p. 739-746.
[3] A. D. Ambrisi, M. T. Cristofaro and M. De Stefano: Predictive Models for Evaluating Concrete
Compressive Strength in Existing Buildings, The 14th World Conference on Earthquake
Engineering (2008), Beijing, China.
[4] S. Hannachi and M.N. Guetteche: Application of the Combined Method for Evaluating the
Compressive Strength of Concrete on Site, Open Journal of Civil Engineering, Vol. 2 (2012), p.
16-21.
[5] R. Pucinotti: Reinforced concrete structure: non destructive in situ strength assessment of
concrete, Construction and Building Materials, Vol. 75 (2015), p. 331-341.
[6] S. Biondi, and E. Candigliota: In situ tests for seismic assessment of RC structures. Beijing, 14th
World Conference on Earthquake Engineering (2008).
[7] J. Zupan and J. Gasteiger: Neural Networks: A new method for solving chemical problems or just
a passing phase? Anal. Chim. Acta 248 (1991), p. 1-30
[8] J. M. Zurada: Introduction to Artificial Neural System, PWS, Boston (1992)
[9] I. B. Topcu and M. Sarıdemir: Prediction of compressive strength of concrete containing fly ash
using artificial neural networks and fuzzy logic, Computational Materials Science, Vol. 41
(2008), p.305-311.