Full length article
Quantitative recognition of
flammable and toxic gases with
arti
ficial neural network using metal oxide gas sensors in
embedded platform
B. Mondal
a,*, M.S. Meetei
a, J. Das
b, C. Roy Chaudhuri
c, H. Saha
daDepartment of Electronics and Communication Engineering, Tezpur University, Tezpur 784028, India bDepartment of Physics, Jadavpur University, Kolkata 700032, India
cDepartment of Electronics and Telecommunication Engineering, Indian Institute of Engineering Science and Technology, Howrah, India dCentre of Excellence in Green Energy and Sensors Systems, Indian Institute of Engineering Science and Technology, Kolkata 711103, India
a r t i c l e i n f o
Article history:
Received 15 November 2014 Received in revised form 19 December 2014 Accepted 31 December 2014 Available online 31 January 2015
Keywords: Neural network Reliable detection Sensor array
a b s t r a c t
Artificial Neural Network (ANN) based pattern recognition technique is used for ensuring the reliable evaluation of responses from an array of Zinc Oxide (ZnO) based sensors comprising of pure ZnO nano-rods and composites of ZnOeSnO2. All the sensors were fabricated in the lab. The paperfirst reports the development of an artificial neural network based model for successfully recognizing different con-centration of hydrogen, methane and carbon mono-oxide. Feed forward back propagation neural network was used for the classification of the gases at critical concentrations. The optimized ANN al-gorithm is then embedded in the microcontroller based circuit andfinally verified under lab conditions. © 2015 Karabuk University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
1. Introduction
Qualitative and quantitative detection offlammable and toxic gases like hydrogen (H2), methane (CH4) and carbon mono-oxide
(CO) in coal mines, petroleum industries and gas storage plants are of great importance for the safely of workers, health hazards and to avoid risk of accidental explosion. H2is a potential source of
energy of future and is used in fuel cell[1], H2engine cars, chemical
industry etc. H2 is a colorless, odorless and highly explosive gas
havingflammability in the range (4e75%) in air[2]. In coal mines presence of methane is always expected, which has caused most loss of life than any other gas. Methane in the range (5e14%) in air is flammable [2]. CO is also a colorless and tasteless gas which is believed to be the most dangerous gas in mine. CO in the range of 50 ppm is fatal and is alsoflammable and explosive[2]in higher range of concentrations.
Oxide based semiconductor sensors present advantages like high sensitivity to low concentration of gases, less fabrication cost, long term stability, low power consumption etc. which makes them
attractive material for the detection of toxic, hazardous and combustible gases. A wide range of sensors based on semi-conductor metal oxides including ZnO, TiO2, SnO2, WO3, In2O3, etc.
were used for the detection of gases like H2[3], CH4[4], LPG[5], CO [6]etc. Both ZnO and SnO2are n-type wide band semiconductors
which have attracted remarkable attention as highly sensitive sensor material. These material were expected to be superior ma-terial for sensing gases like H2[7,8], CO[9,10], and CH4[10,11]. In
spite of the high response, metal oxide semiconductor (MOx) based sensor presents a serious drawback of poor selectivity, which hin-ders their commercial application. Therefore, MOx sensors fall short when a precise detection of specific concentration of gaseous component is required or when they are used to discriminate a mixture of gas. Drift in the sensor output due to ageing, poor sta-bility or high sensitivity to humidity of the ambient environment are also serious drawbacks of metal oxide based sensors. Both drift and poor selectivity make the sensor output unreliable[12].
Many authors have suggested various ways to overcome the problem of cross-sensitivity to achieve selective sensor which may include, a) addition of noble catalytic metal to promote the reaction to specific gas[13], b) operating the sensor in dynamic mode[14], c) surface modification[15]and d) use of multi-compositional sensing film [16,17]. However, extremely selective sensors are very * Corresponding author. Tel.: þ91 03712 267007/8/9x5266.
E-mail address:[email protected](B. Mondal). Peer review under responsibility of Karabuk University.
H O S T E D BY Contents lists available atScienceDirect
Engineering Science and Technology,
an International Journal
j o u rn a l h o m e p a g e : h t t p : / / w w w . e l s e v i e r . c o m / l o c a t e / j e s t c h
http://dx.doi.org/10.1016/j.jestch.2014.12.010
2215-0986/© 2015 Karabuk University. Production and hosting by Elsevier B.V. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/ licenses/by-nc-nd/4.0/).
expensive and operating the sensor in dynamic mode makes the measurement system complex. Although, the problem of cross-sensitivity can be addressed with any of the above mentioned ways, the troubles of reliability due to the drift in sensor output still persist. A potentially more cost effective way to perform such measurement is to use a sensor array coupled with a pattern clas-sification system [18e20]. The use of multiple sensors with an appropriately modeled pattern classifier is therefore expected to improve the reliability in the recognition of specific concentration of gaseous components. This will also improve the selectivity for a particular gas when there is a mixture of gases.
Various pattern analysis methods like principle component analysis (PCA) [21], partial least square analysis (PLS) [22], and artificial neural network[23]etc. can be used with gas sensor arrays for the classification and identification problem. Among these methods artificial neural network is reported to provide superior performance when quantification or multicomponent mixture classification is required [24]. Multilayer perceptron (MLP) and radial basis function (RBF) are the two commonly used neural network algorithms used for pattern recognition. However, the neuron connectivity in RBF is comparatively lower than MLP. This is the reason why the mean square error of an RBF network is less compared to an MLP one having the same number of neurons. Therefore the parameters for tuning the network are higher in MLP. MLP with nonlinear transfer function, like sigmoid function, allow the network to learn any inputeoutput relationship adjusting the weights and biases in the network by a gradient descent of tech-nique known as back propagation of errors. Many authors have reported the successful utilization of ANN in many applications including pattern classification, identification, decision making etc.
[25e28]. Multilayer perceptron neural network have been used by Joo et al. to identify breast nodule malignancy using sonographic images[26]. Lv et al. reported the detection of low concentration of formaldehyde in binary gas mixtures using micro gas sensor array and a neural network[27]. Lee et al. reported about a combustible gas recognition system including sensor array and neural network based pattern recognition[28].
Most of the commercially available gas recognition systems are PC based and hence are bulky. Embedded systems can be developed with a suitable training function which can automatically read the sensor input from the sensor array and produce the neural network output. This may overcome the drawback of PC based gas pattern recognizer, making them portable and convenient to the user. In this work, we report the development of a prototype toxic gas recognizer using sensor array and multilayer neural network. The sensors were based on pure ZnO nano-rods and ZnOeSnO2
com-posites which were fabricated in the lab. The sensors were exposed to H2, CH4 and CO in different concentration ranges. The data
generated from the sensors were used to train a multilayer neural network with back propagation learning algorithm, which can reasonably produce output to unseen input data. A process flow chart for realization of the ANN in microcontroller is also formu-lated and verified by downloading the network in the microcon-troller. The use of an array of sensors instead of one or two along with a suitably trained neural network can increase the reliability of the measurement.
2. Sensor fabrication
The sensor used in the experiment consists of pure oxide of zinc as well as composite oxide of zinc and tin, as the active gas sensing element. For the fabrication of sensors separate sols were prepared for pure ZnO and composite ZnOeSnO2 with polyvinyl (PVA) as
capping agent. The detail material synthesis and methodical char-acterization of the sensor materials were reported in our previous
communications[29,30]. In a typical procedure, sol for pure ZnO is prepared by mixing 0.6 g of zinc acetate di-hydrate [Zn(CH3COO)2$2H2O] with aqueous solution of PVA under
con-stant stirring. The 1.0 M sodium hydroxide (NaOH) solutions was then mixed drop wise to the above solution until a pH value reached 7.0. The solution was then stirred at 80C for 1 h. Sol for ZnOeSnO2 sensor is prepared by mixing appropriate amount of
zinc acetate dehydrate (Zn(CH3COO)2$2H2O) and stannic chloride
(SnCl4$5H2O) with PVA. 1.0 M NaOH solution was used as reducing
agent until pH of the solution becomes 7.0. Thefinal solution was stirred at 110e120C for 1 h. Finally, the resulting precursors were
spin-casted on the cleaned<100> oriented Si/SiO2substrates
fol-lowed by annealing at 550C for 30 min in a tube furnace. 3. Measurement techniques
A total of six numbers of sensors were used for the collection of data generated from the sensors when exposed to three different gases namely hydrogen, methane and carbon monoxide. The sen-sors were exposed to each test gas in different concentrations (0.1, 0.3, 0.5 and 1.0%) separately. The sensor types and their operating conditions are given in Table 1. The detail arrangement of the sensor measurement set up and sensor structure were reported elsewhere [29,30]. Typically, the sensors were placed inside a closed test chamber with suitable arrangement for gas inlet and outlet. Measured amount of test gas was fed to the test chamber using Alicat®massflow controllers (MFC). The test chamber was fitted with heating arrangement that would maintain a uniform temperature (with an accuracy of±3C) inside the chamber. The
sensors were designed to operate in a resistive mode andFig. 1
shows a model of the sensor measuring circuit. PdeAg contact electrodes were made on the top of the sensing layer by e-beam evaporation technique in order to provide electrical connection. An electrometer (Agilent U1253B) was used to monitor the variation of sensor resistance. The resistance variation of the sensor is con-verted into voltage output by the measurement circuit.
4. Design of ANN in matlab®
A feed forward multilayered neural network (Fig. 2) trained by a gradient descent technique, also called back propagation of errors, was used to process the data arising from the sensors to investigate the feasibility of discrimination of gases. The network was designed in the Neural Network toolbox in MATLAB®with six input neuron equal to the number of sensors used, two hidden layers with ten and twelve neurons in the hidden layer 1 and 2, respectively, and thirteen output nodes corresponding to a specific concentration of a specific gas, as shown inTable 2. Sigmoid function was used as the activation function for the hidden and the output neurons. Training of the network was performed using supervised back propagation learning algorithm which wasfirst initialized with random weights and biases. During the training process, weights and biases of the network were iteratively adjusted in order to minimize the
Table 1
Sensor type and operation temperature.
Sensor Sensing material Operating temperature (C)
S1 ZnO 150 S2 200 S3 250 S4 ZnOeSnO2 150 S5 200 S6 250
performance function of the network. The default performance function for feed-forward networks is mean square error (MSE), which is the average squared error between the network output and the target output and is defined as[31],
MSE¼1
N
XN
i¼1
ðTi YiÞ2 (1)
where T is the target and Y is the network output.
In this study 260 data samples were used out of which 70% data were used as training data, 20% as validation and 10% as testing data. The performance of a feed forward MLP neural network is significantly affected by the network parameters like number of hidden layers, number of neurons in each hidden layer, training style and activation function used. The weight matrix for the neu-rons was initially chosen randomly and the training algorithm used was Resilent Backpropagation (RP).
4.1. Optimization of neural network structure
MLP with a single hidden layer and nonlinear transfer function like sigmoid function can learn any non-linear relationship. How-ever, it has been reported earlier by various authors that the addition of a second hidden layer improves the efficiency of the network to a great extent and also drastically reduces the total number of hidden neurons[32e34]. A network with two hidden layers was chosen in the present case and the number of neurons in each layer is varied in search of best performance. The network gave the lowest MSE (0.005) when the number of neurons in hid-den layer 1 and layer 2 was ten and twelve, respectively.Fig. 3plots the MSE of the network for different combination of neurons in the two hidden layers.
4.2. Optimization of activation function
Two sigmoid (Logsig and tansig) and one linear (pureline) activation function were studied in this work. These three functions are defined in eq.(2)e(4) [35]. A set of 27 combinations of these activation functions, as shown inTable 3, for the two hidden and one output layer are investigated for achieving best network per-formance. The network performance for the various combinations of these functions is presented inFig. 4. For the given set of data one can observe fromFig. 4that the network gave the best performance when‘logsig’ activation function is used for both the hidden layers and the output layer (corresponding to combination code 8 in
Table 3) LogsigðnÞ ¼ 1 1þ en (2) TansigðnÞ ¼ 2 1þ e2n 1 (3) PurelinðnÞ ¼ n (4)
5. Implementation of ANN in microcontroller
The optimized ANN is implemented in an Arduino due micro-controller board based on the Atmel (SAM3X8E ARM Cortex-M3 CPU) 32-bit ARM core microcontroller. It has 54 digital input/ output pins, 12 analog inputs, a 84 MHz clock, 2 DAC (digital to Fig. 1. Schematic model of the circuit diagram.
Fig. 2. Structure of neural network. Table 2
Test gas and sensors designated output.
Test gas Concentration (%) Designated output
No gas 1000000000000 H2 0.1 0100000000000 0.3 0010000000000 0.5 0001000000000 1 0000100000000 CH4 0.1 0000010000000 0.3 0000001000000 0.5 0000000100000 1 0000000010000 CO 0.1 0000000001000 0.3 0000000000100 0.5 0000000000010 1 0000000000001
analog), 96 Kbytes of SRAM and 512 Kbytes offlash memory and can be powered by a USB connector or with an external power supply.
The required data for implementing the ANN in microcontroller are the weight matrix, biases and mean and standard deviation of the input data set. The mean and standard deviation of the input data are required for data normalization. The input data are normalized using the following relation[36]
Xn¼ ðXi MiÞ=Sd (5)
In this equation Xnis the normalized data, Xiis the input data and
Mi, Sdare mean and standard deviation of the given set of data.
The weight and bias matrices are obtained after training the network in MATLAB®environment. For the network under study (six input neuron, ten and twelve neuron in hidden layer one and two respectively and thirteen output neuron), we have [10 6] weight matrix for thefirst hidden layer, [12 10] weight matrix for the second hidden layer and [13 12] weight matrix for the output layer. In addition, we also have three bias matrices for the two hidden layers and one output layer.
The processflow chart for the implementation of the ANN in microcontroller is shown inFig. 5. The weights and biases of the trained network connections were first stored in the microcon-troller memory. The output of the first hidden neurons is then computed using the following expression(6)e(8),
Pj ¼ flogsig
Xn
i¼1
Wji$Xiþ bj (6)
where i¼ 1e6 is the number of input nodes and j ¼ 1e10 is the number of neurons in thefirst hidden layer.
The output of the second hidden layer is computed using the following expression
Qk ¼ flogsig
Xn
j¼1
Wkj$Pjþ bk (7)
where k¼ 1e12, is the number of neurons in the second hidden layer.
Andfinally the output of the output neurons is computed using the following expression
Ol ¼ flogsig
Xn
k¼1
Wlk$Qkþ bl (8)
where l¼ 1e13, is the number of output neurons.
This course is coded in Arduino software environment“Arduino 1.5.2” based on Cþþ programming language. The code is then converted into a HEX file before downloading into the microcontroller.
Fig. 3. Optimization of number of neurons in the two hidden layers.
Table 3
Combination of activation function.
Activation function combination code
Activation function
First hidden layer Second hidden layer Output layer
1 Purelin Purelin Purelin
2 Purelin Purelin Logsig
3 Purelin Purelin Tansig
4 Purelin Logsig Purelin
5 Purelin Tansig Purelin
6 Logsig Purelin Purelin
7 Tansig Purelin Purelin
8 Logsig Logsig Logsig
9 Logsig Logsig Purelin
10 Logsig Logsig Tansig
11 Logsig Purelin Logsig
12 Logsig Tansig Logsig
13 Purelin Logsig Logsig
14 Tansig Logsig Logsig
15 Tansig Tansig Purelin
16 Tansig Tansig Logsig
17 Tansig Purelin Tansig
18 Tansig Logsig Tansig
19 Purelin Tansig Tansig
20 Logsig Tansig Tansig
21 Logsig Tansig Purelin
22 Logsig Purelin Tansig
23 Tansig Logsig Purelin
24 Tansig Purelin Logsig
25 Purelin Logsig Tansig
26 Purelin Tansig Purelin
27 Tansig Tansig Tansig
Fig. 5. Flowchart for ANN implementation.
6. Results and discussion
After successful completion of the learning process, the pro-posed network produced very high recognition rate. To verify the performance of the system, we plot inFig. 6, the regression values for training (a), validation (b), testing (c) and also for the whole set (d) of data. The regression values give the relationship between the outputs of the network with that of the targets.Fig. 6indicates that the output is closelyfit with the target with an accuracy of ~94% during training, ~96% during validation and ~95% during testing. The overall accuracy of the network was 94.6%.
The network was successfully implemented in embedded plat-form. The network is implemented in the microcontroller using 185 KB in the flash memory and the execution time was nearly 100 s. In order to verify the performance of the recognizer, the network in both the environment (MATLAB®and microcontroller) was fed with the same inputs. Approximately 98% of the output from both the system found to match perfectly.
7. Conclusion
Using an array of six, moderately selective, ZnO and ZnO-SnO2
based sensors, the proposed pattern recognition methods allow us to construct the simple pattern classifier, which is able to recognize and classify hydrogen, methane and carbon oxide gas samples of different concentrations with good reliability and accuracy. Multi-layer perceptron is used for gas recognition architecture. The optimized network consists of six input neurons equal to the number of sensors, two hidden layers with ten and twelve neurons in hidden layer-1 and hidden layer-2, respectively, and thirteen output neurons. The network is fed with 260 experimentally ob-tained data sets out of which 70% of the data were used for training, 20% were used for validation and 10% for testing. The network gave an overall accuracy of 94.6% when trained with RP algorithm. A processflow chart for the realization of ANN in microcontroller is then formulated which is later written in Arduino software envi-ronment and downloaded in the microcontroller to develop a prototype gas recognition system. The developed system can ef fi-ciently produce reasonable output when fed with input data. References
[1] N. Sammes, Fuel Cell Technology: Reaching toward Commercialization, Springer.
[2] http://www.mathesongas.com.
[3] A. Qurashi, M. Faiz, N. Tabet, M.W. Alam, Low temperature synthesis of hex-agonal ZnO nanorods and their hydrogen sensing properties, Superlattice Microstruct. 50 (2011) 173e180.
[4] P. Bhattacharyya, P.K. Basu, H. Saha, S. Basu, Fast response methane sensor using nanocrystalline zinc oxide thinfilms derived by solegel method, Sens. Actuat. B 124 (2007) 62.
[5] C.S. Prajapati, S.N. Pandey, P.P. Sahay, Sensing of LPG with nanostructured zinc oxide thinfilm grown by spray pyrolysis technique, Physica B 406 (2011) 2684e2688.
[6] R.K. Joshi, Q. Hu, F. Alvi, N. Joshi, A. Kumar, Au decorated zinc oxide nanowires for CO sensing, J. Phys. Chem. C 113 (2009) 16199e16202.
[7] O. Lupan, V.V. Ursaki, G. Chai, L. Chow, G.A. Emelchenko, I.M. Tiginyanu, A.N. Gruzintsev, A.N. Redkin, Selective hydrogen gas nanosensor using indi-vidual ZnO nanowire with fast response at room temperature, Sens. Actuat. B 144 (2010) 56e66.
[8] S.F. Bamsaoud, S.B. Rane, R.N. Karekar, R.C. Aiyer, Nano particulate SnO2based
resistivefilms as a hydrogen and acetone vapour sensor, Sens. Actuat. B 153 (2011) 382e391.
[9] S. Chang, T. Hsueh, I-Cherng Chen, B. Huang, Highly sensitive ZnO nanowire CO sensors with the adsorption of Au nanoparticles, Nanotechnology 19 (2008) 175502.
[10] M.A. Andio, P.N. Browning, P.A. Morris, S.A. Akbar, Comparison of gas sensor performance of SnO2 nano-structures on microhotplate platform, Sens.
Actuat. B 165 (2012) 13e18.
[11] P. Bhattacharyya, P.K. Basu, B. Mondal, H. Saha, A low power MEMS gas sensor based on nanocrystalline ZnO thinfilms for sensing methane, Microelectron. Reliab. 48 (2008) 1772e1779.
[12] G. Korotcenkov, Gas response control through structural and chemical modification of metal oxide films: state of the art and approaches, Sens. Actuat. B 107 (2005) 209e232.
[13] W. Lim, J.S. Wright, B.P. Gila, S.J. Pearton, F. Ren, Wei-Ta Lai, Li-Chyong Chen, Ming-Shien Hu, K. Chen, Selective-hydrogen sensing at room temperature with Pt-coated InN nanobelts, Appl. Phys. Lett. 93 (2008) 202109.
[14] W. Maziarz, T. Pisarkiewicz, Gas sensors in a dynamic operation mode, Meas. Sci. Technol. 19 (2008) 05520.
[15] P.K. Basu, S.K. Jana, H. Saha, S. Basu, Low temperature methane sensing by electrochemically grown and surface modified ZnO thin films, Sens. Actuat. B 135 (2008) 81e88.
[16] X. Song, Z. Wang, Y. Liu, C. Wang, L. Li, A highly sensitive ethanol sensor based on mesoporous ZnOeSnO2nanofibers, Nanotechnology 20 (2009) 075501.
[17] H. Woo, C.W. Na, Il-Doo Kim, J. Lee, Highly sensitive and selective trime-thylamine sensor using one-dimensional ZnOeCr2O3hetero-nanostructures,
Nanotechnology 23 (2012) 245501.
[18] V.D. Lecce, A. Amato, R. Dario, C. Martines, Air quality control for health care centres, the application of an intelligent distributed system, 2009 IEEE Workshop on Environmental, Energy, and Structural Monitoring Systems, EESMS-2009 Crema (Italy). pp. 27e30.
[19] I. Concina, M. Falasconi, E. Gobbi, F. Bianchi, M. Musci, M. Mattarozzi, M. Pardo, A. Mangia, M. Careri, G. Sberveglieri, Early detection of microbial contamination in processed tomatoes by electronic nose, Food Control 20 (2009) 873e880.
[20] R.F. Machado, D. Laskowski, O. Deffenderfer, T. Burch, S. Zheng, P.J. Mazzone, T. Mekhail, C. Jennings, J.K. Stoller, J. Pyle, J. Duncan, R.A. Dweik, S.C. Erzurum, Detection of lung cancer by sensor array analyses of exhaled breath, Am. J. Respir. Crit. Care Med. 171 (2005) 1286e1291.
[21] J.W. Gardner, Detection of vapours and odours from a multisensor array using pattern recognition part principal component and cluster analysis, Sens. Actuat. B 4 (1991) 109e115.
[22] M. Barker, W. Rayens, Partial least squares for discrimination, J. Chemometrics 17 (2003) 166e173.
[23] D.S. Lee, H.Y. Jung, J.W. Lim, M. Lee, S.W. Ban, J.S. Huh, D.D. Lee, Explosive gas recognition system using thickfilm sensor array and neural network, Sens. Actuat. B 71 (2000) 90e98.
[24] I. Sayago, M. del C. Horrillo, S. Baluk, M. Aleixandre, M.J. Fernandez, L. Ares, M. García, J.P. Santos, J. Gutierrez, Detection of toxic gases by a tin oxide multisensor, IEEE Sens. J. 2 (5) (2005).
[25] J.I. Civera, E.G. Breijo, N.L. Miro, L.G. Sanchez, J.G. Baixauli, I.R. Gil, R.M. Peris, M.A. Fillol, Artificial neural network onto eight bit microcontroller for secchi depth calculation, Sens. Actuat. B 156 (2011) 132e139.
[26] S. Joo, W.K. Moon, H.C. Kim, Computer-aided diagnosis of solid breast nod-ules on ultrasound with digital image processing and artificial neural network, in: Proceedings of the 26th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC '04), 2004, pp. 1397e1400.
[27] P. Lv, Z. Tang, G. Wei, J. Yu, Z. Huang, Recognizing indoor formaldehyde in binary gas mixtures with a micro gas sensor array and a neural network, Meas. Sci. Technol. 18 (2007) 1997e3004.
[28] D. Lee, S.M. Lee, D. Lee, Micor gas sensor array with neural network for recognizing combustible leakage gases, IEEE Sens. J. 5 (3) (2005).
[29] B. Mondal, B. Basumatary, J. Das, C. Roychaudhury, H. Saha, N. Mukherjee, ZnOeSnO2based composite type gas sensor for selective hydrogen sensing,
Sens. Actuat. B 194 (2014) 389e396.
[30] B. Mondal, L. Dutta, C. Roychaudhury, D. Mohanta, N. Mukherjee, H. Saha, Effect of annealing temperature on the morphology and sensitivity of the zinc oxide nanorods based methane senor, Acta Metall. Sin., English Lett. 27 (4) (2014) 593e600.
[31] http://in.mathworks.com/help/matlab/.
[32] V. Kurkova, Kolmogorov's theorem and multilayer neural networks, Neural Netw. 5 (1992) 501e506.
[33] G.B. Huang, Learning capability and storage capacity of two-hidden-layer feedforward networks, IEEE Trans. Neural Netw. 14 (2003) 274e281. [34] D. Stathakis, How many hidden layers and nodes? Int. J. Rem. Sens. 30 (8)
(2009) 2133e2147.
[35] M. Dorofki, A.H. Elshafie, O. Jaafar, O.A. Karim, S. Mastura, Comparison of artificial neural network transfer functions abilities to simulate extreme runoff data, in: International Conference on Environment, Energy and Biotechnology, IPCBEE, vol. 33, IACSIT Press, Singapore, 2012.
[36] S. Sumathi, P. Surekha, Computational Intelligence Paradigms: Theory and Applications Using MATLAB, CRC Press, Inc., 2010, ISBN 978-1-4398-0902-0.
