https://www.researchgate.net/publication/331579591_Odour_Identification_Using_Machine_Learning_Techniques/citations

1. ICTS24632019-EC4030
239
International Conference on Technical Sciences (ICST2019)
March 201906–04
Odour Identification Using Machine Learning
Techniques
Ali M. Abdulshahed
Electrical & Electronic Engineering
department, Misurata University
Misurata, Libya
a.abdulshahed@eng.misuratau.edu.ly
Abstract—In recent years, the development of a simple, and
low-cost odour identification system using an electronic nose
has been the concern of many researchers. This work
investigates the abilities of machine learning techniques;
Adaptive Neuro-Fuzzy Inference System (ANFIS), Artificial
Neural Networks (ANNs), and Gaussian classifiers to identify
different airborne substances. Furthermore, its future trend,
perspectives and challenging problem are also mentioned. The
performances of the classifiers used in this study were
computed using four performance criteria: Root Mean Square
Error (RMSE), Nash-Sutcliffe Efficiency coefficient (NSE),
correlation coefficient (R) and also accuracy. According to the
results, it was found that Artificial intelligence (AI) classifiers
could be employed successfully in odour identification. In
addition, results showed that the ANFIS classifier outperforms
the other machine learning classifiers.
Keywords— machine learning, odour identification, artificial
neural networks, fuzzy logic, electronic nose
I. INTRODUCTION
An odour, especially an unpleasant one is caused by one
or more volatilized chemical compounds that are generally
found in low concentrations that humans can perceive by
their sense of smell. The identification of odour is an
important task for many applications, including the detection
and diagnosis in medicine, quality control in food-processing
chains, finding drugs and explosives, or the monitoring of
pollution levels in air [1]. One prominent example is drones
and mobile robots equipped with electronic noses conducting
tasks like a survey of farmland collecting necessary
information such as ambient and crop conditions [2]. Given
their high versatility to host multiple sensors while still being
compact and lightweight, odour identification systems has
demonstrated to be a promising technology to real-world gas
recognition and enormous commercial potential [3], which is
our main concern in this paper. Nowadays, autonomous
robots and drones are used in agriculture for increasing
efficiency, and especially reducing the cost of the scarce
human labor. They can be used to survey the farmland
collecting necessary information such as ambient and crop
conditions, soil fertility, pest and disease, etc. [4]. Electronic
noses are useful devices, which mimic the sense of human
being smell. These devices generally consist of an array of
sensors utilized to sense and distinguish odour in harsh
environment and at low cost. Recently developed techniques
have offered great potential for electronic noses to detect
different contaminations in foods by examining the pattern of
volatile compounds produced. Changes in the generated
fingerprint can be resulting from either, the appearance of
new chemical compounds or to variations in the quantity of
the original volatile compounds without changes in the
qualitative composition. The application of an electronic
nose can provide a fast and accurate means of sensing the
types of food contaminant origin such as microbiological,
chemical and physical with minimal efforts. The applications
of electronic noses used in the food industry have been
discussed in the review paper by Loutfi et al. [1]. The authors
indicated that there is a strong commonality between the
different application area in terms of the sensors used, and
the data processing algorithms applied. Generally different
types of classification approaches are used for odour
identification. The relationships between the system inputs
and outputs are not based on physical equations, but are
deduced through suitable experimental tests. A convenient
and common way of doing this is to use regression
classifiers. The most popular regression models include
multiple linear regression [5], principal component
regression [6], partial least squares [7] or Artificial Neural
Networks (ANNs) in the case of non-linear classifiers. Linear
regression is the simplest method to correlate measured
sensor's data with resulting output. A Least Squares (LS)
approach is used to obtain the coefficients that determine the
relationship between inputs and output without using any
physical equation. Although this method can provide
reasonable results for a given simple task, the sensing data
usually changes with the environment due to robot motion,
which introduces an error into the model [8]. The linear
regression classifier is also time-consuming and labour
intensive to design. To improve the above-mentioned
classifiers, one has to use a set of explaining variables and a
set of dependent variables. The set of explaining variables
contain the information from the chemical sensors
comprising an electronic nose device. The set of dependent
variables includes the values of odour intensity or hedonic
tone expressed in the verbal scale, which originate from a
group of assessors utilizing suitable olfactometer techniques
[9]. A task of the regression methods is to construct such a
model, which would allow quantitative evaluation of a
particular odour feature (odour intensity, hedonic tone) based
on the set of explaining variables [9]. Classification of odour
with an array of gas sensors is still a challenging task [1].
The goal of this work is to train a machine learning classifier
that allows a mobile robot or flying drone to be discriminate
between different airborne substances, for instance, Acetone
and Propanol (see Fig. 1). The next section first gives a short
introduction to machine learning systems, and then
concentrates on methods for obtaining classifiers from data.
These approaches are commonly referred to as machine
learning techniques since they take decisions without being
explicitly programmed to perform a particular task. Within

2. ICTS24632019-EC4030
240
this section, only architectures of artificial neural network
and neuro-fuzzy technique is considered.
FIG. 1 THE PROPOSED BLOCK DIAGRAM.
II. MACHINE LEARNING
A machine learning system is a system that can make
decisions which would be considered intelligent if made by a
human being. Machine Learning (ML) is becoming more
popular and particularly amenable to modelling complex
systems, because it has demonstrated superior classification
ability compared to traditional methods [10, 11]. In this
work, the aim has been to present a description and analysis
of the ML systems that will be used throughout this
classification task. This section first gives a short
introduction to artificial neural networks and fuzzy systems,
and then concentrates on methods for obtaining classifiers
from data. These approaches are commonly referred to as
neuro-fuzzy techniques since they exploit a link between
fuzzy systems and neural networks. Within this section, only
one architecture of neuro-fuzzy techniques is considered, the
so called an Adaptive Neuro-Fuzzy Inference System
(ANFIS).
A. Artificial neural network
Artificial neural network as a form of ML is a data-driven
approach. It is designed in a way that mimics the behaviour
of biological neural network. A typical artificial neural
network has an input layer, one or more hidden layers, and
an output layer. The neurons in the hidden layer, which are
connected to the neurons in the input and output layers by
adaptable weights, enable the ANN to compute complex
associations between the input and output variables [12] (see
Fig. 2). The inputs of each neuron in the hidden and output
layers are summed and the resulting summation is processed
by an activation function [12]. Training the classifier is the
process of determining the adjustable weights and it is
similar to the process of determining the coefficients of a
regression model by least squares approach. The weights are
initially selected randomly and an optimisation algorithm is
then used to find the weights that minimise the differences
between the model-calculated and the target outputs [13].
Across the whole classification procedure, no physical
equation is used. To find the relationship between inputs and
outputs of a complex system, ANN techniques have drawn
more attention rather than statistical techniques, and produce
results without requiring a detailed mechanistic description
of the phenomena that is governing the system. There are
different ANN architectures to building classifiers, Back-
Propagation (BP) artificial neural network has proved to be a
suitable nonlinear classification method [14]. One of the
major advantages of ANNs is efficient handling of highly
non-linear relationships in data.
FIG. 2 THE STRUCTURE OF ASSOCIATED NETWORK CLASSIFIER.
B. Fuzzy Logic and Fuzzy Systems
The concept of Fuzzy Logic (FL) was pioneered by
Zadeh [15, 16] and was introduced not as a control
methodology, but as a way of processing data by allowing
partial set membership rather than a crisp set membership or
non-membership. In fuzzy logic, the membership function is
a curve that defines how each point in the input space is
mapped to a degree of membership between 0 and 1.
Classical logic needs a deep understanding of a system’s
exact physical equations and precise crisp values. Fuzzy
logic demonstrates an alternative way of thinking, which
allows complex modelling using a higher level of abstraction
created particularly from human knowledge and experience.
Fuzzy logic allows formulating this knowledge in a
subjective way which is mapped into exact crisp ranges. In
classic set theory, elements either completely belong to a set
or are completely excluded from it. The process of
expressing the mapping from inputs to an output using fuzzy
logic is named the Fuzzy Inference System (FIS) [17]. The
particular structure of the fuzzy model, can be classified into:
(i) Fuzzy linguistic model (Mamdani model) [18] (ii) Fuzzy
relational model [19] (iii) Takagi-Sugeno (T-S) fuzzy model
[20]. A main distinction can be made between the Mamdani
model, which has fuzzy propositions in both antecedents and
consequents of the rules, and the T-S model, where the
consequent is a crisp function of the input variables, rather
than a fuzzy proposition [21]. Fuzzy relational models can be
regarded as a generalisation of Mamdani model, allowing
one particular antecedent proposition to be associated with
several different consequent propositions via a fuzzy relation
[22]. In the literature, it can be clearly seen that the Mamdani
model structure demonstrates several advantages. It provides
a natural framework to include expert human knowledge in
the form of linguistic fuzzy “if-then” rules. This knowledge
can be easily gathered with rules that describe the relation
between system input-output [21]. Moreover, Mamdani
model provides a flexible means to formulate knowledge,
while at the same time it remains interpretable, as long as a
proper design is developed. However, although Mamdani
model possesses several advantages, it also comes with some
weaknesses. One of the main drawbacks is the lack of
accuracy when modelling some high-dimensional, complex
systems. This is due to the limitation of human cognitive
ability of codifying these complex systems. Therefore,
during the last few years much of the research developed in
fuzzy logic modelling focused on increasing the accuracy as
much as possible, giving little attention to the interpretability
of the resultant model. Hence, the T-S fuzzy models played a
pivotal role in the contemporary research. These models are
relatively easy to identify, and their structure can be readily

3. ICTS24632019-EC4030
241
calibrated. As discussed above, fuzzy logic is a useful
modelling technique for assessing ambiguous complex
processes such as odour identification. However, its
applicability needs further evaluation with experimental data.
Several hybrid methods have been introduced in the artificial
intelligence field including a neuro-fuzzy technique. Within
this work only one architecture of neuro-fuzzy techniques is
considered, the so called an adaptive neuro-fuzzy inference
system.
C. Adaptive Neuro-Fuzzy Inference System
The Adaptive Neuro-Fuzzy Inference System (ANFIS),
was first introduced by Jang [17]. According to Jang, ANFIS
is a neural network that is functionally the same as a Takagi-
Sugeno type inference model. The ANFIS is a hybrid
intelligent system that takes advantages of both ANN and
fuzzy logic theory in a single system. By employing the
ANN technique to update the parameters of the Takagi-
Sugeno type inference model, the ANFIS is given the ability
to learn from training data, the same as ANN. The solutions
mapped out onto a Fuzzy Inference System (FIS) can
therefore be described in linguistic terms. In order to explain
the concept of ANFIS structure, five distinct layers are used
to describe the structure of an ANFIS classifier. The first
layer in the ANFIS structure is the fuzzification layer; the
second layer performs the rule base layer; the third layer
performs the normalization of membership functions (MFs);
the fourth and fifth layers are the defuzzification and
summation layers, respectively. More information about the
ANFIS structure is given in [17]. Fig. 3 shows basic structure
of the ANFIS with two inputs. Adaptive Neuro-Fuzzy
Inference System. ANFIS classifier design consists of two
sections: constructing and training. Construction involves
selecting the input variables, input space partitioning,
choosing the number/type of MFs for inputs, generating
fuzzy rules, premise and conclusion parts of fuzzy rules and
selecting initial parameters for MFs. Training data patterns
should first be generated to build an ANFIS classifier. These
data patterns consist of ANFIS classifier inputs and the
desired output. However, the size of the input-output data
pattern is very crucial when the generation of data is a costly
affair. Construction of the ANFIS classifier requires the
division of the input-output data into rule patches. This can
be achieved by using a number of methods such as grid
partitioning, subtractive clustering method and fuzzy c-
means (FCM) [23]. According to Jang [17], grid partition is
only suitable for problems with a small number of input
variables (e.g. fewer than 6). A classifier with three inputs
with three fuzzy sets per input produces a complete rule set
of 27 rules, whereas a classifier with six inputs requires 729
(36) rules. Clearly standard ANFIS classifiers are practically
limited to low dimensional modelling. It is important to note
that an effective partition of the input space can decrease the
number of rules and thus increase the speed in both learning
and application phases. In order to obtain a small number of
fuzzy rules, a fuzzy rule generation technique that integrates
ANFIS with FCM clustering will be applied in this paper,
where the FCM is used to systematically create the fuzzy
MFs and fuzzy rules base for ANFIS. In addition, it helps to
determine the initial parameters of the fuzzy classifier. This
is important because an initial value, which is very close to
the final value, will eventually result in the quick
convergence of the classifier towards its final value during
the training process. In order to maximise the classifier
performance, a learning procedure is followed to refine the
classifier parameters. In the training section, the membership
function parameters are able to change through the learning
process. The adjustment of these parameters is assisted by a
supervised learning of the input-output dataset that are given
to the classifier as training data. Different learning
techniques can be used, such as a hybrid-learning algorithm
combining the least squares method, and the gradient descent
method is adopted to solve this training problem.
FIG. 3 BASIC STRUCTURE OF ANFIS CLASSIFIER.
III. EXPERIMENTAL WORK
Decision making is carried out in four stages as follows:
(i) collect the dataset, (ii) train the classifier using training
dataset (iii) testing the resulting classifier with new unseen
dataset, which are not used during training stage, (iv) identify
the best classifier structure based on statistical performance
criteria values. The performance criteria equations will be
given in next section.
A. Performance evaluation of various classifiers
Once a classifier has been trained, it is necessary to check
the classification quality of the resulting classifier and to
assess the parameter accuracy. This will give the confidence
behind the classifier, and tell the designer if he needs to
revise the training process. This procedure is called model
validation, which consists of several steps. The first test is to
examine whether the obtained classifier can classify the
experimental dataset that has been used for the training
process. Otherwise, there is clearly something wrong in the
training procedure, and it has to be modified and repeated.
Cross validation is used to examine the performance of the
classifier, to check its generalization capability. Therefore,
enough dataset must be available and divided these into two
subsets, one for training stage (and afterward direct
validation), and the other for cross validation. The
performances of the classifiers used in this work were
computed using four performance criteria: Root Mean
Square Error (RMSE), Nash-Sutcliffe Efficiency coefficient
(NSE), correlation coefficient (R), and also accuracy.
B. ANFIS classifier Development
Extensive simulations were conducted to determine the
optimal structure of the ANFIS classifier through various
experiments. The optimal number of MFs was determined by
assigning different values to the number of clusters (nc)
(equal to number of MFs) for the ANFIS classifier. Too few
MFs will not allow an ANFIS classifier to be mapped well.
However, too many MFs will increase the difficulty of
training and will lead to over-fitting or memorising

4. ICTS24632019-EC4030
242
undesirable inputs such as noise. The classification errors
were measured separately for each classifier using the root
mean square error (RMSE) index with the testing dataset. An
example of selecting the optimal structure for the ANFIS
classifier is presented as follows: In this classification
method, the optimal size of the ANFIS classifier was
determined. Different numbers of epochs were selected for
each classifier because the training process only needs to be
carried out until the errors converge. It was found that the
ANFIS classifier with three (nc=3) clusters exhibited the
lowest RMSE value (1.8) for the testing dataset.
Consequently, this ANFIS classifier with 3 rules was
considered to be the optimal.
C. ANN classifier development
In order to assess the ability of the ANFIS classifier
relative to that of a neural network classifier, an ANN
classifier was constructed using the same input variables to
the ANFIS. It is worth noting that the range of the training
data must be representative of the entire operating conditions
of the system in order to overcome the problem of
extrapolation error. Usually ANN classifiers have three
layers: Input, hidden and output layer. Although, for
common engineering problems, one hidden layer is sufficient
for model training, two or more hidden layers may be needed
for other applications. An ANN classifier with three layers
was used in this study: the input layer has 2 input variables
and the output layer has one neuron (the classifier output).
Selection of the number of neurons in the hidden layer is
important for finding a suitable ANN classifier structure.
Although increasing the neuron numbers in the hidden layer,
may help to improve the neural network performance,
however, the possibility of over-fitting may increase.
Furthermore, a large number of hidden neurons can increase
classifier training time. In this work, the minimum RMSE is
determined by changing the number of hidden neurons.
Therefore, after a series of experiments to find the best
architecture, an ANN classifier with 10 neurons in the hidden
layer was constructed to discriminate between two possible
airborne substances, namely Acetone and Propanol.
D. Results and Discussion
In this work, the use of ANFIS, ANN, Quadratic
Gaussian classifier QG and linear Gaussian classifier LG, for
discriminate between two possible airborne substances,
namely Acetone and Propanol, was described and compared.
The final classifiers being trained and validated in the
training stage have been verified further by a new separate
dataset, not used during training stage. The confusion matrix
results using ANFIS, ANN, QG and LG classifiers are
shown in Fig. 4, Fig. 5, Fig. 6 and Fig. 7 respectively. The
performance of each of the four classifiers is presented and
compared in Table 1, where the four classifiers are trained
using the same training dataset and validated by the same
testing dataset. According to the discriminative results and
evaluation criteria values in Table 1, it is very clear that the
ANFIS classifier has a smaller RMSE, higher efficiency
coefficient NSE=0.86, higher accuracy 96.6% and higher
correlation coefficient (R) contrasting with the ANN, QG
and LG classifiers. The ANN classifier performed better than
the QG and LG classifiers for discriminate between to
possible airborne substances. It can be also observed from
Table 1 that the classifiers developed using the artificial
intelligence techniques outperformed the Gaussian
classifiers. The Gaussian classifiers are known for their
simplicity and have less complexity, compared with other
non-parametric classifiers. However, due to the nonlinearity
of the problem under consideration, the Gaussian classifiers
may not provide a satisfactory result. In order to avoid the
tedious trial and error approach, AI algorithm can be used in
order to improve the performance of the classifier. However,
the ANN classifier does improve the classification accuracy
to higher than 95%, the number of ANN model parameters is
high. Furthermore, it is worth noting that these classifiers
(i.e., ANN classifier) need a proper optimisation to
discriminate effectively. For instance, the ANN classifier
needs 10 neurons in the hidden layer, which was difficult to
optimise. Therefore, the ANFIS classifier is a good classifier
choice for discriminate between to possible airborne
substances, namely Acetone and Propanol with the benefit of
fewer rules.
TABLE 1. PERFORMANCE CALCULATION OF THE USED CLASSIFIERS.
Classifier
Performance indices
R RMSE NSE Accuracy
ANFIS 0.92 0.18 0.86 96.6%
ANN 0.89 0.20 0.82 91.8%
Quadratic 0.55 0.42 0.25 82.0%
Linear 0.53 0.42 0.22 81.0%
As described above, each ML technique has its own
limitations. The fusion of two or three of these techniques
will continue to be one of the trends in the area of odour
identification. Another trend in ML applications is likely to
be the fusion of ML and hard computing. The fusion of ML
and hard computing should be able to provide innovative
solutions to the problems with high-performance, cost-
effective, and reliable computing systems. In order to
improve the results, the following stages can be considered:
pre-processing the data (e.g. data normalization), feature
extraction and classification. Furthermore, success in
obtaining a reliable and robust classifier depends heavily on
the choice of the domain used for construction and training
purposes. For feature extraction stage, the wavelet transform
can be used. Use of the wavelet transformation technique
will give the results in both the frequency domain and time
domain, so that we can extract valuable features that can
reflect the occurrence of a certain action.
0 1
0
1
1628
58.5%
33
1.2%
98.0%
2.0%
62
2.2%
1059
38.1%
94.5%
5.5%
96.3%
3.7%
97.0%
3.0%
96.6%
3.4%
Target Class
OutputClass
Confusion Matrix
FIG. 4 CONFUSION MATRIX FOR THE CLASSIFICATION OF ACETONE AND
PROPANOL WITH ANFIS

5. ICTS24632019-EC4030
243
0 1
0
1
1618
58.2%
43
1.5%
97.4%
2.6%
73
2.6%
1048
37.7%
93.5%
6.5%
95.7%
4.3%
96.1%
3.9%
95.8%
4.2%
Target Class
OutputClass
Confusion Matrix
FIG. 5 CONFUSION MATRIX FOR THE CLASSIFICATION OF ACETONE AND
PROPANOL WITH ANN.
0 1
0
1
1470
52.8%
191
6.9%
88.5%
11.5%
311
11.2%
810
29.1%
72.3%
27.7%
82.5%
17.5%
80.9%
19.1%
82.0%
18.0%
Target Class
OutputClass
Confusion Matrix
FIG. 6 CONFUSION MATRIX FOR THE CLASSIFICATION OF ACETONE AND
PROPANOL WITH QG.
0 1
0
1
1492
53.6%
169
6.1%
89.8%
10.2%
353
12.7%
768
27.6%
68.5%
31.5%
80.9%
19.1%
82.0%
18.0%
81.2%
18.8%
Target Class
OutputClass
Confusion Matrix
FIG. 7 CONFUSION MATRIX FOR THE CLASSIFICATION OF ACETONE AND
PROPANOL WITH LG.
IV. CONCLUSIONS
Odour identification remains one of the important
problems of the modern systems. In this paper, ANFIS,
ANN, QG and LG, classifiers were utilized to discriminate
between two possible airborne substances, namely Acetone
and Propanol. The results show that the ANFIS classifier
could be a powerful tool for a discrimination task, which is
difficult to achieve using conventional methods, especially
for the highly non-separable problem. The proposed
methodology has its capacity for fast learning from
experimental data and linguistic knowledge, and the ability
to provide a simple, transparent and robust classifier. This is
a new attempt to discriminate between two possible airborne
substances using ANFIS and other machine learning
classifiers. There is still large room for enhancement of these
classifiers by including more explaining variables into
consideration, and try different hybrid machine learning tools
to optimize the model architecture for odour identification.
REFERENCES
[1] A. Loutfi, S. Coradeschi, G. K. Mani, P. Shankar, and J.
B. B. Rayappan, "Electronic noses for food quality: A review,"
Journal of Food Engineering, vol. 144, pp. 103-111, 2015.
[2] H. Fan, V. H. Bennetts, E. Schaffernicht, and A. J.
Lilienthal, "A cluster analysis approach based on exploiting
density peaks for gas discrimination with electronic noses in open
environments," Sensors and Actuators B: Chemical, vol. 259, pp.
183-203, 2018.
[3] J. G. Monroy and J. Gonzalez-Jimenez, "Gas
classification in motion: An experimental analysis," Sensors and
Actuators B: Chemical, vol. 240, pp. 1205-1215, 2017.
[4] T. Pobkrut and T. Kerdcharoen, "Soil sensing survey
robots based on electronic nose," in Control, Automation and
Systems (ICCAS), 2014 14th International Conference on, 2014,
pp. 1604-1609.
[5] J. Gebicki, B. Szulczynski, and M. Kaminski,
"Determination of authenticity of brand perfume using electronic
nose prototypes," Measurement Science and Technology, vol. 26,
p. 125103, 2015.
[6] H. Zhang, J. Wang, and S. Ye, "Predictions of acidity,
soluble solids and firmness of pear using electronic nose
technique," Journal of Food Engineering, vol. 86, pp. 370-378,
2008.
[7] J. Nicolas, C. Cerisier, J. Delva, and A.-C. Romain,
"Potential of a network of electronic noses to assess in real time the
odour annoyance in the environment of a compost facility," in
Chemical engineering transactions: NOSE2012 International
Conference on Environmental Odour, 2012, pp. 133-138.
[8] A. M. Abdulshahed, A. P. Longstaff, S. Fletcher, and A.
Myers, "Thermal error modelling of machine tools based on
ANFIS with fuzzy c-means clustering using a thermal imaging
camera," Applied Mathematical Modelling, vol. 39, pp. 1837-1852,
2015.
[9] B. Szulczyński, K. Armiński, J. Namieśnik, and J.
Gębicki, "Determination of Odour Interactions in Gaseous
Mixtures Using Electronic Nose Methods with Artificial Neural
Networks," Sensors, vol. 18, p. 519, 2018.
[10] J. B. Mitchell, "Machine learning methods in
chemoinformatics," Wiley Interdisciplinary Reviews:
Computational Molecular Science, vol. 4, pp. 468-481, 2014.
[11] N. M. Nasrabadi, "Pattern recognition and machine
learning," Journal of electronic imaging, vol. 16, p. 049901, 2007.

6. ICTS24632019-EC4030
244
[12] Y. Nagata and K. H. Chu, "Optimization of a
fermentation medium using neural networks and genetic
algorithms," Biotechnology letters, vol. 25, pp. 1837-1842, 2003.
[13] N. Nasr, H. Hafez, M. H. El Naggar, and G. Nakhla,
"Application of artificial neural networks for modeling of
biohydrogen production," International Journal of Hydrogen
Energy, vol. 38, pp. 3189-3195, 2013.
[14] D. S. O. Correa, D. F. Sciotti, M. G. Prado, D. O. Sales,
D. F. Wolf, and F. S. Osorio, "Mobile robots navigation in indoor
environments using kinect sensor," in 2012 Second Brazilian
Conference on Critical Embedded Systems, 2012, pp. 36-41.
[15] L. A. Zadeh, "Fuzzy sets," Information and Control, vol.
8, pp. 338-353, 1965.
[16] L. A. Zadeh, The concept of a linguistic variable and its
application to approximate reasoning: Springer, 1974.
[17] J. S. R. Jang, "ANFIS: Adaptive-network-based fuzzy
inference system," Systems, Man and Cybernetics, IEEE
Transactions on, vol. 23, pp. 665-685, 1993.
[18] E. H. Mamdani, "Application of fuzzy logic to
approximate reasoning using linguistic synthesis," Computers,
IEEE Transactions on, vol. 100, pp. 1182-1191, 1977.
[19] J. Casillas, O. Cordón, F. H. Triguero, and L.
Magdalena, Interpretability issues in fuzzy modeling vol. 128:
Springer, 2003.
[20] T. Takagi and M. Sugeno, "Fuzzy identification of
systems and its applications to modeling and control," Systems,
Man and Cybernetics, IEEE Transactions on, pp. 116-132, 1985.
[21] O. Cordón, "A historical review of evolutionary learning
methods for Mamdani-type fuzzy rule-based systems: Designing
interpretable genetic fuzzy systems," International journal of
approximate reasoning, vol. 52, pp. 894-913, 2011.
[22] J. Abonyi, Fuzzy Model Identification: Springer, 2003.
[23] S. Guillaume, "Designing fuzzy inference systems from
data: An interpretability-oriented review," Fuzzy Systems, IEEE
Transactions on, vol. 9, pp. 426-443, 2001.