Measurement 67 (2015) 71–77

Contents lists available at ScienceDirect

Measurement
journal homepage: [Link]/locate/measurement

Indirect measure of thermal conductivity of rocks through

adaptive neuro-fuzzy inference system and multivariate
regression analysis
Mohsen Hajihassani a, Aminaton Marto b,⇑, Nima Khezri b, Roohollah Kalatehjari b
a
Construction Research Alliance, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia
b
Faculty of Civil Engineering, Dept. of Geotechnics and Transportation, Universiti Teknologi Malaysia, 81310 UTM Skudai, Johor, Malaysia

a r t i c l e i n f o a b s t r a c t

Article history: Thermal conductivity is an important property of rocks which is considered for energy-
Received 29 September 2014 efficient building construction. This paper is aimed to predict the thermal conductivity
Received in revised form 26 January 2015 of rocks utilizing the adaptive neuro-fuzzy inference system (ANFIS) and multivariate
Accepted 2 February 2015
regression (MVR) analysis. In this regard, 44 datasets including the most effective para-
Available online 21 February 2015
meters on thermal conductivity of rocks were collected from the literature. The physico-
mechanical properties of rocks including uniaxial compressive strength, P-wave velocity,
Keywords:
bulk density and porosity were used to develop the predictive models. The correlation of
Thermal conductivity
Rock properties
determination equal to 0.99 and 0.95 were obtained by ANFIS and MVR models respective-
Energy-efficient ly. The obtained results suggest that the ANFIS model outperforms the MVR model and is
Adaptive neuro-fuzzy inference system an applicable tool to predict thermal conductivity of rocks with high degree of accuracy.
Ó 2015 Elsevier Ltd. All rights reserved.

1. Introduction stresses are induced by changing in the temperature which

cause thermal strains in rocks. The TC of rocks is influenced
Thermal conductivity (TC) of rocks is an important part by the rock type, moisture content, P-wave velocity and
of their properties especially when the energy conserva- porosity. In Addition, mineral composition and crystalliza-
tion is considered. The energy usage in a building is tion of a rock influence its TC [5–7].
reduced when natural rocks with low TC are used in build- The TC of a construction material is usually determined
ing construction [1,2]. Therefore, considering this property through the hot plate steady-state method [8]. The TC of a
of rocks for the countries that excessively use heating or rock is obtained by measuring the temperature gradients
cooling systems is essential. Three different parameters and heat input [9]. The TC of a rock might be obtained
including thermal diffusivity (TD), heat capacity (HC) as using the laboratory experiments or in the field under in-
well as TC describe the rock thermal properties [3]. TD situ condition [10] and can be determined through the
shows the areas of rock which have been influenced by Fourier’s law as follows:
the heat in each second. HC shows the rock capacity to heat
accumulation where TC reveals the heat transmission Qh
k¼ ð1Þ
capability of rocks. According to Verma [4], thermal ADTt

⇑ Corresponding author. Tel.: +60 127165310; fax: +60 75566157. in which k represents the TC (W/m K), Q is the heat flow, h
E-mail addresses: mohsen_hajihassani@[Link] (M. Hajihassani),
is rock thickness in the direction of heating (m) and A is the
aminaton@[Link], aminaton.marto09@[Link] (A. Marto), nima. area of heated surface (m2). DT is temperature difference
khezri.n@[Link] (N. Khezri), kalatehjary@[Link] (R. Kalatehjari). between surface of material (°C) and t is heating time (h).

[Link]
0263-2241/Ó 2015 Elsevier Ltd. All rights reserved.
72 M. Hajihassani et al. / Measurement 67 (2015) 71–77

The thermal conductivity of construction materials has the non-linear relationships between input and output
been investigated by several researchers. Özkahraman data [20–22].
et al. [2] conducted several laboratory tests to obtain the To explain the modeling procedure by ANFIS, it is
TC of various rocks. They investigated the relationships supposed that the FIS under consideration consists of two
between the physical and mechanical characteristics of inputs (x, y) and one output (f) and the rule base includes
rocks including porosity, P-wave velocity, compressive two fuzzy rule set ‘‘if-then’’ as bellow [17]:
strength and density. Finally, they resulted that P-wave
velocity yields the best correlation among all characteris- Rule I: if x is A1 and y is B1, then f1 = p1x + q1y + r1
tics and proposed a relationship between this property Rule II: if x is A2 and y is B2, then f2 = p2x + q2y + r2
and TC. Yasßar et al. [1] collected various rock samples from
different part of Turkey and conducted several laboratory in which pi, qi, and ri are the consequent parameters to
tests on the collected samples. Finally, using the statistical be settled. According to Jang [17] and Jang et al. [19], an
analysis, they established several relationships to assess ANFIS with five layers and two rules can be explained as
the TC of rocks using their mechanical and physical follows:
properties. Layer I: Every node i in layer I produces a membership
Recently, artificial intelligence (AI) methods have been grade of a linguistic label. For example, the node func-
widely used in many fields of science and engineering tion of the ith node is:
[11–14]. This is due to the ability of these methods for gen- 1
Q 1i ¼ lAi ðxÞ ¼  2 bi ð2Þ
erating the non-linear relationships between input and
xv i
output parameters. Singh et al. [15] and Khandelwal [16] 1þ r1
employed artificial intelligence (AI) approaches to predict
TC of rocks using simple rock parameters. They concluded in which Q 1i and x are the membership function and
that these methods are able to predict TC of rocks with input to node i respectively. Ai is the linguistic label
good accuracy. related to node i and r1, vi, bi are parameters that make
This paper presents an ANFIS model to predict TC of changes in the form of the membership functions. The
rocks based on their physical and mechanical properties. existing parameters in this layer are related to the pre-
In this regard, a database consisting of 44 dataset was mise part, as in Fig. 1(a).
collected from the literature [1,2,15]. For the sake of com- Layer II: Each node in layer II computes the firing
parison, the predicted results by proposed ANFIS model strength of each rule through multiplication:
were compared with those of multivariate regression Q 2i ¼ wi ¼ lAi ðxÞ  lBi ðyÞ i ¼ 1; 2 ð3Þ
(MVR) analysis.
Layer III: The ratio of firing strength of the ith rule to the
sum of firing strengths of all rule is obtained in this
layer.
wi
2. Adaptive neuro fuzzy inference system Q 3i ¼ W i ¼ P2 i ¼ 1; 2 ð4Þ
j¼1 wj
Adaptive neuro-fuzzy inference system (ANFIS) devel-
Layer IV: Every node i in this layer is a node function
oped by Jang [17] based on the Takagi–Sugeno [18] fuzzy
whereas Wi is the output of layer III. Parameters of this
inference system (FIS). ANFIS is a universal predictor with
layer are related to consequent part.
the capability to approximate any real continuous func-
Q 4i ¼ W i f i ¼ W i ðpi x þ qi y þ ri Þ ð5Þ
tions [19]. ANFIS works based on the construction of a
set of if-then fuzzy rules with proper membership func- Layer V: The incoming signals are summed in this layer
tions to produce the required output data. In general, a and form the overall output. P X
FIS is generated based on five functioning blocks including: wi f
several if-then fuzzy rules, a database to define the mem- Q 5i ¼ Overall Output ¼ W if i ¼ P i ð6Þ
wi
bership functions, a decision-making element to conduct
the inference operations on the rules, a fuzzification inter-
face to convert the inputs utilizing linguistic values and 3. Prediction of thermal conductivity of rocks through
finally, a defuzzification interface to convert the fuzzy ANFIS
results into an output.
ANFIS integrates the philosophies of artificial neural An ANFIS-based model was developed to predict TC of
networks (ANNs) and FIS and therefore, potentially pre- rocks. This model utilized four input parameters includ-
sents all benefits of them in a unique framework. ing uniaxial compressive strength, P-wave velocity, bulk
Through the hybrid learning, ANFIS is able to evaluate density and porosity. These parameters are the most
the relationships between inputs and target data by influential factors on TC of rocks. The input and output
determining the optimum distribution of membership parameters utilized in this model and range of them
functions. Fig. 1 shows a basic ANFIS architecture. are tabulated in Table 1. Fig. 2 shows the frequency dis-
According to this figure, ANFIS architecture consists of tribution of utilized TC values in the proposed ANFIS-
two parts including premise and consequent parts. based model. The input datasets were randomly divided
ANFIS has extensively used in various applications of into two categories: 80% as the training datasets (35
science and engineering based on its ability to predict datasets) for learning procedure and 20% as the testing
 M. Hajihassani et al. / Measurement 67 (2015) 71–77 73

Fig. 1. (a) Sugeno fuzzy model with two rules and (b) equivalent ANFIS architecture [17].

Table 1
Summary of measured parameters in the predictive models.

Parameter Type Unit Symbol Min Max Mean St. deviation

Uniaxial compressive strength Input MPa UCS 22.27 120.80 64.54 22.60
P-wave velocity Input m/s PWV 2745 6300 4553.30 1037.78
Bulk density Input kg/m3 BD 1108 2970 2541.75 325.34
Porosity Input % P 0.86 36 4.59 6.55
Thermal conductivity at 300 K Output W/m K TC 0.35 3.01 1.43 0.76

Fig. 2. Frequency distribution of utilized TC in ANFIS-based model.

Fig. 3. Relationship between UCS and TC.

74 M. Hajihassani et al. / Measurement 67 (2015) 71–77

Fig. 4. Relationship between PWV and TC.

Fig. 5. Relationship between BD and TC.

Fig. 6. Relationship between P and TC.

Table 3
Characterizations of the proposed ANFIS model.

ANFIS parameter type Value

Type of membership function Gaussian
Number of membership functions 5
Table 2
Number of nodes 1297
Performance of trained ANFIS models.
Number of linear parameters 625
Number of 2 3 4 5 6 7 Number of nonlinear parameters 40
membership function Total number of parameters 665
Number of training data pairs 35
RMSE 0.140 0.123 0.095 0.078 0.124 0.167
Number of checking data pairs 9
R2 0.976 0.981 0.986 0.990 0.976 0.957
Number of fuzzy rules 625
 M. Hajihassani et al. / Measurement 67 (2015) 71–77 75

Fig. 7. Membership functions assigned for the input parameters.

Table 4 The relationships between the input parameters and TC

Analysis of variance. of rocks are shown in Figs. 3–6. According to these figures,
df SS MS F Significance F the input parameters are largely related to the thermal
Regression 4 21.81778 5.45445 194.90013 9.09965E25
conductivity of rocks. PWV, BD and UCS parameters are
Residual 38 1.063462 0.027986 directly proportional to the TC of rocks. In fact, increasing
Total 42 22.88125 in the value of these parameters might cause to increase
the value of TC. However, there is an inverse relationship
between the P and TC values. This means that increasing
Table 5 the P of rocks might cause to decrease the TC values.
Statistical information for TC predictive model. To determine the optimum ANFIS architecture, several
models were constructed and evaluated (see Table 2). It
Input Coefficients St. error t-value p-value
was found that the model with five membership functions
Constant 0.06072 0.52356 0.11599 0.90828
yields superior results among other models concerning the
PWV 0.00037 0.00004 9.70670 0.00000
P 0.01653 0.00851 1.94360 0.05938 root mean squared error (RMSE) and correlation of deter-
BD 0.00058 0.00021 2.75920 0.00886 mination (R2). In modeling process, the Gaussian member-
UCS 0.02053 0.00220 9.33574 0.00000 ship function which is the most common membership
function in fuzzy systems was utilized. The characteriza-
tions of the proposed ANFIS-based model are tabulated in
datasets (9 datasets) to evaluate the accuracy of the Table 3. The linguistic variables for the input parameters
model. were assigned as very low (VL), low (L), medium (M), high

Fig. 8. Correlation between measured and predicted values of TC by MVR analysis.

76 M. Hajihassani et al. / Measurement 67 (2015) 71–77

Fig. 9. Correlation between measured and predicted values of TC by ANFIS.

Table 6 The graph of predicted TC values using MVR approach

Performance indices of predictive models. against the measured TC values for all 44 datasets is
Predictive model R2 Adj. R2 RMSE VAF (%) illustrated in Fig. 8. As shown in this figure, the R2 value
ANFIS 0.9894 0.9891 0.0779 98.92 equal to 0.9535 reveal the reliability of the MVR method
MVR 0.9535 0.9524 0.1599 95.42 to predict TC of rocks. The R2 can be calculated as Eq. (8).
Fig. 9 shows the predicted TC values by performing
ANFIS model plotted against the measured TC values. The
(H) and very high (VH). Fig. 7 shows the membership func- R2 value of 0.9894 suggests the superiority of the proposed
tions assigned for the input parameters in the proposed ANFIS approach in predicting TC of rocks.
ANFIS model. PN  2
yi  y0i
R2 ¼ 1  i¼1
PN 02 ð8Þ
PN 2 y
i¼1 yi 
i¼1 i
4. Multivariate regression analysis N

In order to control the capacity performance of the

Multivariate regression (MVR) analysis is performed by predictive models, other performance indices including,
considering more than one independent variable to adjusted R2 (Adj. R2), root mean squared error (RMSR)
achieve the best-fit equation between input and output and variance account for (VAF) between measured and
variables. In general, the aim of MVR analysis is to predicted values of these predictive models were obtained.
determine a relationship between the input (independent The aforementioned indices can be calculated as follows:
variables) and output (dependent variable). The form of
MVR equation is y = b1x1 + b2x2 + bnxn + c, in which {b1, b2, ð1  R2 ÞðN  1Þ
Adj:R2 ¼ 1  ð9Þ
. . ., bn} are the coefficients of regression and parameter c NP1
is a constant value of y. As long as independent variables rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
have acceptable correlation with dependent variable, they 1 XN
RMSE ¼ ðy  y0 Þ2 ð10Þ
can be utilized as inputs in predictive models [23]. N i¼1

In this study, MVR analysis was conducted to establish a  

relationship between input and output parameters. The varðy  y0 Þ
VAF ¼ 1  ð11Þ
statistical software package MICROSOFT EXCEL 2010 was varðyÞ
used for MVR analysis. The obtained equation for predic-
in which y and y0 represent the measured and predicted
tion of TC is shown in Eq. (7). More details on the statistical
values, respectively. P is the number of predictors and N
information of the proposed equation can be seen in Tables
is the total data number. The model will be excellent if
4 and 5.
the RMSE is zero and VAF is 100. Results of performance
indices of the proposed predictive models for all 44
TC ¼ 0:00037PWV  0:01653P  0:00058BD
datasets are tabulated in Table 6. Comparison between
þ 0:02053UCS  0:06072 ð7Þ the performance indices for the proposed ANFIS and MVR
models reveals that the ANFIS model produces more accu-
5. Results and discussion rate prediction compared to the MVR model. This is due to
the fact that ANFIS is a non-linear approach while MVR is a
In this paper, application of ANFIS and MVR approaches linear method. The low value of RMSE equal to 0.0779 and
for prediction of TC of rocks were described. In this regard, high value of VAF equal to 98.92% confirm the superiority
the collected datasets was utilized to conduct the ANFIS of the proposed ANFIS model in predicting TC of rocks.
and statistical analyses. Subsequently, a comparison was The ANFIS model displays several benefits, as it combines
conducted between the measured and predicted values the advantages of the artificial neural networks and fuzzy
obtained by the proposed ANFIS and MVR predictive inference system to demonstrate a high prediction capa-
models. city in nonlinear engineering problems.
 M. Hajihassani et al. / Measurement 67 (2015) 71–77 77

6. Conclusion [6] A. Hasan, Optimising insulation thickness for buildings using life
cycle cost, Appl. Energy 63 (1999) 115–124.
[7] H. Çanakci, R. Demirboğa, M. Burhan Karakoç, O. Sßirin, Thermal
Two predictive model utilizing ANFIS and MVR conductivity of limestone from Gaziantep (Turkey), Build. Environ.
approaches were developed for prediction of thermal con- 42 (4) (2007) 1777–1782.
[8] British Standards Institute, Determining Thermal Insulating
ductivity of rocks. In this regard, numbers of 44 datasets
Properties, Guarded hot Plate Method. Part 2: Sect. 2.1 BS 874, 1986.
were collected from literature and utilized to develop pre- [9] ASTM C 1045-90, Practice for the Calculation of Thermal
dictive models. For this purpose, the most influential para- Transmission Properties from Steady State Heat Flux
Measurements, American Society for Testing and Materials,
meters on thermal conductivity of rocks including uniaxial
Pennsylvania, 1990.
compressive strength, bulk density, P-wave velocity and [10] C. Clauser, E. Huenges, Thermal conductivity of rocks and minerals,
porosity were utilized in development of predictive mod- AGU Ref. Shelf 3 (1995) 105–126.
els. To determine the optimum ANFIS architecture, several [11] M. Hajihassani, D. Jahed Armaghani, H. Sohaei, E. Tonnizam
Mohamad, A. Marto, Prediction of airblast-overpressure induced by
models were examined. Among these models, a model blasting using a hybrid artificial neural network and particle swarm
with five Gaussian membership functions found to be the optimization, Appl. Acoust. 80 (2014) 57–67.
best model in terms of RMSE. Finally, an ANFIS model [12] R. Rooki, Estimation of pressure loss of Herschel–Bulkley
drilling fluids during horizontal annulus using artificial neural
was proposed to predict thermal conductivity of rocks. network, J. Disper. Sci. Technol. (2014). DOI: 0.1080/01932691.
Furthermore, through the statistical analysis, a MVR model 2014.904793.
was established utilizing the same inputs and outputs in [13] [Link] Armaghani, M. Hajihassani, E.T. Mohamad, A. Marto, S.A.
Noorani, Blasting-induced flyrock and ground vibration prediction
the ANFIS modeling. The results demonstrated that the through an expert artificial neural network based on particle swarm
ANFIS model gives higher performance compared to MVR optimization, Arab J. Geosci. (2013), [Link]
results. However, the MVR models with acceptable accura- s12517-013-1174-0.
[14] M. Fazavi, S.M. Hosseini, M. Arabloo, M. Shokrollahi, M. Nouri-
cy might be used for the preliminary estimation. Taleghani, M. Amani, Applying a smart technique for accurate
determination of flowing oil–water pressure gradient in horizontal
Acknowledgements pipelines, J. Disper. Sci. Technol. 35 (6) (2014) 882–888.
[15] T.N. Singh, S. Sinha, V.K. Singh, Prediction of thermal conductivity of
rock through physico-mechanical properties, Build. Environ. 42 (1)
The authors would like to extend their appreciation to (2007) 146–155.
the Ministry of Science, Technology and Innovation, [16] M. Khandelwal, Prediction of thermal conductivity of rocks by soft
Malaysia through the provision of research Grant No. 03- computing, Int. J. Earth Sci. 100 (2011) 1383–1389.
[17] R.J.S. Jang, Anfis: adaptive-network-based fuzzy inference system,
01-06-SF1185 (UTM 4S072) to undertake the research. IEEE Trans. Syst. Man Cybern. 23 (1993) 665–685.
Acknowledgement is also conveyed to the Universiti [18] T. Takagi, M. Sugeno, Fuzzy identification of systems and its
Teknologi Malaysia for providing the required facilities applications to modeling and control, IEEE Trans. Syst. Man
Cybern. 15 (1985) 116–132.
which made this research possible. [19] R.J.S. Jang, C.T. Sun, E. Mizutani, Neuro-Fuzzy and Soft Computing,
PTR Prentice Hall, 1997.
References [20] M. Rezaei, A. Majdi, M. Monjezi, An intelligent approach to predict
unconfined compressive strength of rock surrounding access tunnels
in longwall coal mining, Neural Comput. Appl. 24 (1) (2014) 233–
[1] E. Yasßar, Y. Erdoğan, H. Güneyli, Determination of the thermal
241.
conductivity from physico-mechanical properties, Bull. Eng. Geol.
[21] H. Jalalifar, S. Mojedifar, A.A. Sahebi, H. Nezamabadi-Pour,
Environ. 67 (2) (2008) 219–225.
Application of the adaptive neuro-fuzzy inference system for
[2] H.T. Özkahraman, R. Selver, E.C. Isßık, Determination of the thermal
prediction of a rock engineering classification system, Comput.
conductivity of rock from P-wave velocity, Int. J. Rock Mech. Min. Sci.
Geotech. 38 (6) (2011) 783–790.
41 (4) (2004) 703–708.
[22] I. Yilmaz, G. Yuksek, Prediction of the strength and elasticity
[3] I.H. Gul, A. Maqsood, Thermophysical properties of diorites along
modulus of gypsum using multiple regression, ANN, and ANFIS
with the prediction of thermal conductivity from porosity and
models, Int. J. Rock Mech. Min. Sci. 46 (4) (2009) 803–810.
density data, Int. J. Thermophys. 27 (2) (2006) 614–626.
[23] N. Ceryan, U. Okkan, A. Kesimal, Prediction of unconfined
[4] B.P. Verma, Rock Mechanics for Engineers, Khanna Publishers, New
compressive strength of carbonate rocks using artificial neural
Delhi, Khanna, 2000.
networks, Environ. Earth Sci. 68 (3) (2012) 807–819.
[5] F.P. Incropera, D.P. Dewitt, Fundamentals of Heat and Mass Transfer,
Wiley, New York, 1990.