Nitrate and Sulfate Estimations in Water Sources Using ... - IEEE Xplore

0 downloads 0 Views 2MB Size Report
Nitrate and Sulfate Estimations in Water Sources. Using a Planar Electromagnetic Sensor Array and Artificial Neural Network Method. Alif Syarafi Mohamad Nor, ...
IEEE SENSORS JOURNAL, VOL. 15, NO. 1, JANUARY 2015

497

Nitrate and Sulfate Estimations in Water Sources Using a Planar Electromagnetic Sensor Array and Artificial Neural Network Method Alif Syarafi Mohamad Nor, Student Member, IEEE, Mahdi Faramarzi, Mohd Amri Md Yunus, Member, IEEE, and Sallehuddin Ibrahim

Abstract— The primary advantages of planar electromagnetic sensors can be listed as low cost, convenient, suitable for in situ measurement systems, rapid reaction, and highly durable. In this paper, the outputs of a planar electromagnetic sensors array were observed and analyzed after testing it with different types of water samples at different concentrations. The output parameters were derived to decompose by wavelet transform. The energy and mean features of decomposed signals were extracted and used as inputs for an artificial neural network (ANN) model. The analysis model was targeted to classify the amount of nitrate and sulfate contamination in water. Nitrates and sulfate samples in the form of KNO3 and K2 SO4 , each having different concentrations between 5 and 114 mg dissolved in 1 L of distilled water, were used. Furthermore, the analysis model was tested with seven sets of mixed KNO3 and K2 SO4 water samples. A three-layer multilayer perceptron is used as a classifier. It is understood from the results that the model can detect the presence of nitrate and sulfate added in distilled water and is capable of distinguishing the concentration level in the presence of other types of contamination with a root mean square error (RMSE) of 0.0132. The validity of the ANN model was verified by removing the ANN model in estimating the water contamination, where the RMSE rose to 0.0977. The system and approach presented in this paper have the potential to be used as a useful low-cost tool for water source monitoring. Index Terms— Artificial neural network, wavelet transform, planar electromagnetic sensors array, feature extraction, and water contamination.

I. I NTRODUCTION

A

GRICULTURAL activities, farm animals, and industrial areas are the primary causes for the increase of impurities in water resources [1], [2]. These impurities or foreign substances can cause a variety of diseases that are dangerous to

Manuscript received June 30, 2014; revised August 11, 2014; accepted August 11, 2014. Date of publication August 15, 2014; date of current version November 11, 2014. This work was supported in part by the Exploratory Research Scheme Grant through the Ministry of Higher Education Research Malaysia, (Vote No. 4L039), in part by the ScienceFund through the Ministry of Science, Technology and Innovation Malaysia (Vote No. 4S098), and in part by the Universiti Teknologi Malaysia, Research University Grant Scheme (Vote No. 08J88). The associate editor coordinating the review of this paper and approving it for publication was Dr. Richard T. Kouzes. (Corresponding author: Mohd Amri Md Yunus.) The authors are with the Department of Control and Mechatronic Engineering, Faculty of Electrical Engineering, Infocomm Research Alliance, Universiti Teknologi Malaysia, Johor Bahru 81310, Malaysia (e-mail: [email protected]; [email protected]; [email protected]; [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JSEN.2014.2347996

humans. Detection of pollutants, commonly known as nitrates and sulfates, that exist in natural water resources, are the key points in this study. Both of these contaminants could have a negative impact on human health. The nitrate level in water sources is around 20 mg/L, but excessive intake of nitrate for adults and infants will affect the function of platelets [3], [4] and lead to fatalities [5] respectively, as it limits the ability of each red blood cell to carry oxygen. Sulfate can change the taste of water to bitter if the concentration of sulfate in the water is 200 mg/L [6]. Apart from that, the pH value of water depends on the concentration of sulfate in the water. The higher the concentration of sulfate, the lower the pH [7], which could encourage more pollutant to be dissolved [8]. Current nitrate and sulfate detection techniques such as those using electrochemical sensors [9] and biosensors [10] need the preparation of extra reagents, while other detection methods such as spectrometry [11], and ion chromatography [12], [13] need a lab-based setup which requires a tedious working step. The laboratory equipment used is usually huge, expensive, or difficult to carry. In addition, companies such as Life Technologies, MoboSens, and Lenntech have introduced many devices that are used to monitor the nitrate and sulfate concentration in water. However, the detection method used by Life Technology needs the preparation of extra reagents while an expensive strong ionic resin is used by Lenntech. From considering these disadvantages, this research proposes the use of a low-cost sensor array based on a planar electromagnetic sensor array to estimate different concentration levels of nitrate and sulfate. Planar electromagnetic sensors are mainly used to determine the near-to-the-surface properties, such as dielectrics, permeability and conductivity, as reported in [14]. Planar electromagnetic sensors have been widely used in many applications such as food safety [15] and bacterial content detection [16], where the sensors are sensitive to the different magnetic susceptibilities and dielectric properties of each material [17], [18]. In this paper, series connection of meander and interdigital sensors is used for the detection of water contamination. Previous work on a single planar electromagnetic sensor utilized Independent Component Analysis (ICA) as an estimation tool [19], [20]. This tool was powerful in reducing the dimension of the analysis, but the ICA-based analysis was time-consuming as the measurement was repeated many times to obtain a sufficient amount of data before estimation

1530-437X © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

498

IEEE SENSORS JOURNAL, VOL. 15, NO. 1, JANUARY 2015

of the water contamination could be carried out. Due to the mentioned disadvantages, this work is motivated to reduce the total number of measurement cycles by introducing a planar electromagnetic sensors array and to apply a more accurate method to classify the water sources contamination based on Artificial Neural Networks (ANN). ANN is a computational model inspired by natural neurons which attempt to model the information processing capabilities of nervous systems [21]. This concept, which is a datadriven technique, was brought out by McCulloch and Pitts in 1943 and has been developed in recent decades by many researchers [22]. ANNs can be categorized into Feed Forward Neural Networks (FFNN) and Recurrent Neural Networks (RNN); both have advantages and disadvantages and can be used based on the problem properties. Traditional FFNN has various architectures and the most popular one is the multilayer perceptron (MLP). MLPs have been used in a broad assortment of problems in which it is important to apply rule-based scheduling e.g. in machine learning [23], pattern recognition [24], nonlinear control [25], and water quality forecasting [26], [27]. The MLP network consists of an input layer, hidden layer(s) of neurons and an output layer. The basic concept of the MLP for function approximation can be defined as: J X j W j k + bk (1) yk = j =1   I (2) Ui Wi j + b j xj = f i=1

where x is the activation value of hidden neurons, b is a bias value of neurons, W is a connection weight between neuron i and j , J , and I are the number of hidden neurons and inputs, respectively, y is the output and f is an activation function. In this research, the output responses of a planar electromagnetic sensors array were analyzed and classified by a powerful classification method, namely Artificial Neural Networks (ANN). In order to reduce the dimensions of the input data to suit the application of ANNs with the aim of classifying water contamination, Wavelet Transform (WT) is applied to extract signal features. A three-layer MLP neural network was trained, validated, and tested by the different water contamination sources. The BP algorithm is used to train the MLP Neural Network, using MATLAB software. II. D ESCRIPTION OF THE P LANAR E LECTROMAGNETIC S ENSOR A RRAY Early versions of planar electromagnetic sensors consisted of both meander-type and interdigital-type sensors [28]. The overall dimensions of the meander sensor are 20 mm × 20 mm with five square loops. The distance between any outer loop and the neighboring inner loop is 0.5 mm, as shown in Fig. 1(a). The interdigital sensor employs a conventional configuration having consecutive negative and positive electrodes placed apart at a fixed distance. The positive and negative electrodes’ widths are 0.5 mm and 1 mm respectively, as shown in Fig. 1(a). Fig. 1(b) shows a grounded backplane deposited on the other side of the interdigital sensor. The criterion for selecting the distance between interdigital electrodes

Fig. 1. (a) Schematic diagram of the sensor: top layer, (b) schematic diagram of sensor: bottom layer, and (c) planar electromagnetic sensors with star array configuration.

in electromagnetic sensors is based on the wavelength of an interdigital sensor, where the strength of the output signal can be manipulated by changing the distance between the electrodes [29]. An alternating 10 V peak-to-peak sinusoidal wave signal from a function generator is supplied to the sensors. The meander-type sensor will produce magnetic fields, while the interdigital-type sensor will produce electric fields. Hence, the combination of both fields allows the formation of an electromagnetic field. The electromagnetic field interacts with the material under test. So, the resultant electromagnetic field is altered and consequently the impedance of the sensor is changed. This project proposes a set of planar electromagnetic sensors arranged in a star array. Basically, the star sensor array consists of three identical planar electromagnetic sensors. The star sensor array shows that S2 and S3 are placed 45° and -45° from S2 , respectively. S2 and S3 are separated by an angle of 90° as shown in Fig. 1(c), which shows the photo of the fabricated star sensor array. An electrical equivalent circuit of the star sensor array is shown in Fig. 2. The sensor is connected to a function generator where Rg is the output resistance with a nominal value of 50 . R1 denotes the series surface mount resistor connected to sensor 1 (S1 ), as shown in Fig. 1(c). Hence, current I3-1 can be calculated from  (3) I3−1 = V3−1 R1 where I3−1 and V3−1 are the rms value of current through the sensor and voltage across R1 respectively. The absolute total impedance for sensor S1, Z1 is given by Z 1 = V1 /I3−1

(4)

MOHAMAD NOR et al.: NITRATE AND SULFATE ESTIMATIONS IN WATER SOURCES

499

Fig. 3. Fig. 2.

Equivalent circuit for star sensor array. TABLE I

MEMS T ECHNOLOGY A PPLICATIONS FOR WATER M ONITORING

where V1 is the rms value of the input voltage signal. Consider that θ1 is the phase difference between V1 (t) with V3−1 (t) in degree, taking V3−1 (t) or I3−1 as a reference. The total impedance of sensor S1 , Z 1 , can be written as:  (5) Z 1 = V1  θ1 I3−1  0◦ The same method of calculation can be used to calculate the impedance for both sensors S2 and S3 by using Equations (3) to (5). The total number of sensors in the array is not limited to three and could be increased. However, only three sensors were used in this prototype model in view of the limitation of the four input channels of the oscilloscope. The total number of sensors could be increased by adding another oscilloscope, but an extra oscilloscope is not effective during the prototype phase. Sensors-based micro electromechanical system (MEMS) technology can be applied in the near future to minimize the size of the sensor. The related environmental monitoring application using interdigital and a combination of planar meander or interdigital elements based on MEMS is reported in Table I. III. E XPERIMENTAL S ETUP AND R ESULTS The experimental setup is shown in Fig. 3, where the setup has a frequency waveform generator which generates a standard sinusoidal waveform of 10 V peak-to-peak and is set as the input signal for the sensors. A retort stand was used to hold the star sensor array, which was partially immersed in the

Experimental setup for determining nitrate and sulphate levels.

water sample. An oscilloscope was interfaced to a PC where the output signals and the sensor’s impedance was calculated using the LabVIEW software. The measurements were done at a frequency range between 1 kHz and 10 MHz. Although the current setup is still bulky and takes time and effort to carry, it is still in the prototype stage and the system can be reduced to a small-scale portable device, that is compact and portable, as demonstrated in [31]. Before the experiment was conducted, a sensor coating with the product Wattyl Killrust Incralac was sprayed on the sensors in order to form an acrylic resin-based protective coating. The effect of the samples on the sensors’ impedance was recorded. Water samples consisting of both nitrate and sulfate samples at different levels of concentration were prepared in milligram per liter (mg/L). The nitrate samples were prepared from potassium nitrate while the sulfate samples were prepared from potassium sulfate. Both of these compounds were dissolved in distilled water at an appropriate weight to obtain the respective concentration levels. In order to achieve greater accuracy, the weight of the respective compounds was measured using an electronic beam balance with an accuracy of up to 0.1 mg. During the experiment the data were saved automatically into the Excel file via the LabVIEW software. The total impedance of the sensor array versus frequency tested with different concentrations of nitrate and sulfate is illustrated in Fig. 4(a) and 4(b), respectively. Based on Fig. 4, it can be seen that for both nitrate and sulfate the total impedance decreases as the total concentration is increased. The decreased impedance showed that more current was allowed to pass through the sensor due to the increased conductivity of the water sample. In addition, different patterns could be associated with different types of contaminants and concentrations, as shown in Fig. 4. The next section will discuss how the chemical differentiation can be augmented. IV. A NALYSIS M ODEL BASED ON A RTIFICIAL N EURAL N ETWORKS A. Description of the ANN Model Structure An ANN demonstrates a complex, nonlinear relationship between pertinent inputs and outputs with many adjustable parameters, e.g. weights and biases. Such parameters must be optimized during the training procedure. If the data set for training is enough and meets the training procedure, the

500

IEEE SENSORS JOURNAL, VOL. 15, NO. 1, JANUARY 2015

Fig. 5.

Fig. 4. Total impedance of different concentration levels for: (a) nitrate and (b) sulfate.

ANN should be able to realize the output of “unseen” input data [32]. The most applicable ANN is the multilayer perceptron (MLP), which is a feed forward neural network. When the MLP neural network weights and biases are initialized, the network is ready for training. The most popular algorithm for training MLP is the Back Propagation (BP) algorithm, which was proposed by Rumelhart et al. in 1986 [33]. In BP the training process needs a set of data vectors as inputs P and target outputs T . Conventional BP uses a gradient descent algorithm to reduce the Mean Square Error (MSE) of output data, during the epochs, gradually. The MSE of network output O and target output T can be defined as: 1 N (Ti − Oi ) (6) E= i=1 N where, N is the number of samples. B. Stopping Criteria During the training step, the training set is used to update weights and the validation set is applied to avoid over-fitting. The learning step is continued until the stopping criterion is satisfied. The stopping criterion is typically an exact number of iterations (epochs), an arbitrary level of low error (error level is problem-dependent), or an epoch in which validation error increased.

Two level Wavelet Transform decomposition.

1) Wavelet Transform: Wavelet transform uses basis functions called small wavelets with limited duration to represent other functions based on the following formula:  ∞ t −b 1 ) f (t)dt (7) Ψ ∗( F(a, b) = √ a a −∞ where F is the Continuous Wavelet Transform (CWT) of f(t) at scale a and translation b and Ψ is called the mother wavelet. The translation factor locates the main position, while the scaled factor of wavelets allows the signal to be analyzed at different scales. Equation (7) is an extended version of a wavelet, but in the case of discrete signals a Discrete Wavelet Transform (DWT) is needed for analysis and synthesis on the original signals. To decompose the signals at different scales, the DWT uses a series of high-pass and low-pass filters with different cut-off frequencies for coarse approximation and detailed information, respectively. The original signal is first passed through two different half-band filters, a high-pass and a low- pass. The output response of the filters is down-sampled by two; therefore, half of the samples are eliminated. This procedure continues on the result of the high-pass filter until a pre-determined level. The whole procedure for two levels is illustrated in Fig. 5. 2) Feature Extraction: Feature extraction plays an important role in classification problems due to dimension reduction. There is no direct method for finding proper features and it must be done by trial and error. Therefore, in this paper, two different features were extracted; energy and mean by using the following formulas: Energy(x) =

N 

| (x)(n) |2

(8)

N 1  (x)(n) N

(9)

n=1

C. Input Variables and Data Processing The data set includes signals of water contamination quality measured by three electromagnetic sensors over different frequencies from 1 kHz to 10 MHz. Raw signals are not adequate to feed neural networks because of the huge number of samples. The dimensions of the input data for the neural network must be reduced to an acceptable number. This procedure is called feature extraction. These signals have non-stationary behavior and in such case a Wavelet Transform (WT) can be used for decomposition and feature extraction.

Mean(x) =

n=1

where N is the number of samples and x is the sample. These features are used for neural network inputs. V. R ESULTS ON THE C ALCULATION OF C ONTAMINATION AND D ISCUSSIONS A. Classes of Water Samples The classes comprise three groups of contamination, which are potassium nitrate (KNO3 ), potassium sulfate (K2 SO4 ), and

MOHAMAD NOR et al.: NITRATE AND SULFATE ESTIMATIONS IN WATER SOURCES

Fig. 6.

501

Second derivative for three classes of samples (logarithm scale).

Fig. 7. Four successive details and approximation for second derivative (%Z  ): (a) Original signal of class 1, (b), (c), (d) and (e) details at levels 1,2,3, and 4, respectively, and (f) approximation at level 4.

potassium nitrate combined with potassium sulfate (KNO3 + K2 SO4 ) solutions at different concentrations between 5 mg/L and 115 mg/L. Each group contains different classes based on the amount of contamination. In total, there are 19 classes, where each class has 36 sets. The advantage of the star array sensor is proved here, where only 12 measurements are needed to obtain 36 sets instead of 36 measurements being required using the settings in the previous work. This is due to the number of sensors available on the star array sensor. Table II summarizes the water solutions for the reference sets. KNO3 and K2 SO4 were chosen for the experiment because they are two of the most common chemical compounds used for making inorganic fertilizer. B. Derivative of Impedance Sensitivity The signals that are used by the ANN are based on the sensitivity of the sensor, which is the impedance across the series resistor as governed by the following equation:  %Z = (Z sample − Z Dist illed W at er) Z Dist illed W at er ×100% (10) where %Z is the impedance sensitivity, Z sample is the impedance of the sensor when the star sensors array is immersed in the water sample and Z Dist illed W at er is the impedance of the sensor when the star sensors array is immersed in the distilled water. Applying the ANN with the signal from %Z is not desirable due to the difficulty in differentiating each of the signals, which have almost the same concentration of contamination level and type. Hence, the normalized second derivative is applied to overcome the problem and, furthermore, to eliminate the unwanted baseline level. The second derivative for the impedance sensitivity is given by: %Z  = d 2 (%Z )/d f 2

(11)

where %Z  is the normalized second derivative of impedance sensitivity and %Z is the impedance sensitivity. Fig. 6 shows the normalized second derivative of impedance sensitivity for three different classes. In Fig. 6 all the spectra have the same maximum peak at about 11 kHz, but they are significantly different in the rest of the frequency range, i.e., below and above 11 kHz. It is clearly seen that the second derivative signals can be used to differentiate between different chemical species. C. Wavelet Transform Implementation Different types of mother wavelets can be found in the literature, such as Haar, Daubechies, Coiflet and Symmlet wavelets. In this paper, the very popular Haar wavelet is used to decompose signals into four levels, as shown in Fig. 7. It can be seen from Figs. 7 (b) to (e) that the active parts of the details are moving towards the higher frequencies. This shows that the details at each level can provide different information on the original signal (normalized second derivative). D. Input Space for Neural Network The objective of water contamination classification is to demonstrate the effectiveness of the proposed feature selection (energy and mean) of the second derivative %Z  . For this purpose, the feature vectors in each class of Table II are used as the input of the neural network. There are 19 contamination classes of KNO3 and K2 SO4 and a combination of both, as shown in Table II. Twelve different measurements were performed for each class and the output of the three sensors was gathered for analysis. This means that there were 36 sets for each class (12 measurements for each sensor). The energy and mean of approximation

502

IEEE SENSORS JOURNAL, VOL. 15, NO. 1, JANUARY 2015

TABLE II WATER S AMPLE C LASSES

Fig. 9.

MLP Neural Network Architecture (6-6-19).

Fig. 10. MSE result of increasing the number of hidden neurons in three-layer MLP.

The total size of data is 6 × 684, where 6 is the dimension of the features and 684 is the total number of samples in the 19 classes (each class has 36 samples). Sets of each class are separated into three data sets: training 70%, validation 10% and test 20%. E. MLP Neural Network Architecture

Fig. 8. Two-dimensional input vectors of 19 classes: (a) energy of detail in level 2 and mean of approximation at level 4, and (b) energy of detail in level 4 and energy of approximation at level 4.

in level 4 and also the energy of details in all levels were calculated. As shown in Fig. 9, these 6 features were used as the input vector for the ANN. Fig. 8(a) and 8(b) show the typical two-dimensional inputs for the ANN. These input vectors must be classified; their features are shown in two dimensions for their graphical representation. As an example, from Fig. 8(a), the class in the upper right section in this dimension is situated far from the other classes, so this class can be distinctively classified by the ANN, but Fig. 8(b) shows that all clusters of the classes are overlapped and it is difficult for any existing classification methods to perform data classification using only this dimension (energy of detail in level 4 and energy of approximation in level 4).

A typical three-layer MLP with inputs and activation function is shown in Fig. 9. The optimum architecture of threelayers MLP is obtained by trial and error. A network with six neurons in the hidden layer is used as an initial guess. Fig. 9 illustrates that sigmoid functions and linear functions are applied in the hidden layer and output layer, respectively. F. Neural Network Performance To achieve the optimal parameter of the neural network, a mean square error of the training and validation sets was determined based on Eq. (6). Fig. 10 shows that the error of the training data decreased when the number of neurons in the hidden layer exceeded around 20. After this point, there is a small decrease in training error, but additional neurons increase the complexity of the network. A compromise between error reduction and network complexity was made by using 25 neurons in the hidden layer. Fig. 10 also shows some fluctuation with 14, 16, 25, and 34 neurons, which happened because neural networks become

MOHAMAD NOR et al.: NITRATE AND SULFATE ESTIMATIONS IN WATER SOURCES

503

VI. C ONCLUSION

Fig. 11.

MLP training procedure.

TABLE III P ERFORMANCE OF MLP N EURAL N ETWORK

stuck in local minima. The three-layer MLP was trained in batch mode by the LM algorithm and a learning rate of 0.001 was used. The objective function for training the MLP is mean square error based on Eq. (6). Fig. 11 shows how the error reduced during the training procedure. The iteration at which the validation performance reached a minimum was 55. The training continued for 6 more iterations before the training stopped. Fig. 11 also does not show any over-fitting in the training data because errors in the test and validation sets follow the training error. Table III shows the performance of the optimal threelayer MLP. As is obvious in Table III, the root mean square error (RMSE) for all data sets is 0.0132. It can be seen that the number of incorrect samples is 9 out of 684, which means that the error is very low and that our feature extraction method is suitable for the input of the MLP neural network. Table III also shows the result of classification without the ANN model. In this procedure without application of the ANN, the same data inputs were used. However, all the training and validation data was set as reference data. Similarly, classification of the test data was carried out by calculating the root mean square error (RMSE) between each test sample and all the reference data. Each test sample data class to which it belongs, is determined based on the minimum RMSE. From Table III, the total RMSE for the ANN is 0.0132, which was calculated from both the training and test results, while the total RMSE for the classification without ANN is 0.0977, which was calculated from the test data. This result shows that the model with ANN is more accurate, and the RMSE has been reduced to almost 7 times smaller. The MLP neural network architecture provides a learning process for the model to comprehend the important features and characteristics of the inputs before the classification process take place.

This paper has succesfully demonstrated the application of an ANN with the planar electromagnetic sensor array in estimating nitrate and sulfate contamination in water sources. From the characteristics of the sensor array, it was found that the impedance of the sensor array decreased when the concentration of the chemical was increased. The distinct differentiation between different amounts and types of water samples can be seen from the second derivative signals. The method for classifying different water samples was carefully constructed by taking into account the proper types and form of inputs for the ANN, MLP parameters of the ANN, and the limitations of the ANN. The dimensions of the input signals of the MLP were reduced using DWT. To this end, a Haar wavelet on four levels was used. Continuing with the DWT, the energy and mean of the approximation and details are extracted to form the MLP inputs. A three-layer MLP was trained and tested using these inputs to classify 19 different levels and combination of water contamination by nitrate and sulfate. The optimal number of 25 hidden neurons of the neural network was determined based on the minimum mean square error achieved when 25 hidden neurons were applied in the structures of the ANN during the training and validation procedure. The total measurement time to perform the estimation of nitrate and sulfate contamination was significantly reduced when the planar electromagnetic sensors array was used. Furthermore, the results show that in the case of nitrate and sulfate water pollution, the ANN method can estimate nitrate and sulfate contamination with an RMSE of 0.0132. The importance of the ANN method in the model of estimating nitrate and sulfate contaminations was demonstrated when the ANN was removed and the results showed that the RMSE rose to 0.0977. This is because the ANN demonstrates the capability to relate inputs and outputs with many adjustable parameters, e.g. weights and biases which can be perceived as prior information. However, another important point is that the parameters must be optimized during the training procedure. R EFERENCES [1] E. Molina-Navarro, D. Trolle, S. Martínez-Pérez, A. Sastre-Merlín, and E. Jeppesen, “Hydrological and water quality impact assessment of a Mediterranean limno-reservoir under climate change and land use management scenarios,” J. Hydrol., vol. 509, pp. 354–366, Feb. 2014. [2] H. Hayzoun et al., “Impact of rapid urbanisation and industrialisation on river sediment metal contamination,” Environ. Monitor. Assessment, vol. 186, no. 5, pp. 2851–2865, 2014. [3] J. W. Park, B. Piknova, P. L. Huang, C. T. Noguchi, and A. N. Schechter, “Effect of blood nitrite and nitrate levels on murine platelet function,” PLoS ONE, vol. 8, no. 2, p. e55699, 2013. [4] S. Lidder and A. J. Webb, “Vascular effects of dietary nitrate (as found in green leafy vegetables and beetroot) via the nitrate-nitrite-nitric oxide pathway,” Brit. J. Clin. Pharmacol., vol. 75, pp. 677-696, 2013. [5] N. S. Bryan and H. van Grinsven, “The role of nitrate in human health,” Adv. Agronomy, vol. 119, pp. 153–182, Jan. 2013. [6] I. A. Raja, M. Y. Khan, N. A. Khan, M. R. Wani, and A. A. Bhat, “Assessment of some metals in the drinking water of Dal Lake Kashmir,” Nature Sci., vol. 11, no. 3, pp. 63–64, 2013. [7] G. T. Carling et al., “Relationships of surface water, pore water, and sediment chemistry in wetlands adjacent to Great Salt Lake, Utah, and potential impacts on plant community health,” Sci. Total Environ., vol. 443, pp. 798–811, Jan. 2013.

504

[8] V. Lenoble et al., “Distribution and chemical speciation of arsenic and heavy metals in highly contaminated waters used for health care purposes (Srebrenica, Bosnia and Herzegovina),” Sci. Total Environ., vol. 443, pp. 420–428, Jan. 2013. [9] Z. Wang et al., “Electrochemical detection of nitrite based on difference of surface charge of self-assembled monolayers,” Int. J. Electrochem. Sci., vol. 9, pp. 1139–1145, Jan. 2014. [10] A. Amine and G. Palleschi, “Phosphate, nitrate, and sulfate biosensors,” Anal. Lett., vol. 37, no. 1, pp. 1–19, 2004. [11] B. Schnetger and C. Lehners, “Determination of nitrate plus nitrite in small volume marine water samples using vanadium(III)chloride as a reduction agent,” Marine Chem., vol. 160, pp. 91–98, Mar. 2014. [12] J. Liu, X. N. Ji, Q. K. Ren, S. S. Ai, L. J. Wan, and D. J. Bian, “Determination of nitrate in water by HPLC,” Appl. Mech. Mater., vol. 448, pp. 406–408, Oct. 2014. [13] E. L. C. Silveira, L. B. de Caland, and M. Tubino, “Simultaneous quantitative analysis of the acetate, formate, chloride, phosphate and sulfate anions in biodiesel by ion chromatography,” Fuel, vol. 124, pp. 97–101, May 2014. [14] N. Goldfine et al., “MWM-array eddy current sensors for detection of cracks in regions with fretting damage,” ASNT Mater. Eval., vol. 60, no. 7, pp. 870–877, 2002. [15] A. I. Zia et al., “MEMS based impedimetric sensing of phthalates,” in Proc. IEEE Int. Instrum. Meas. Technol. Conf. (I2MTC), May 2013, pp. 855–860. [16] A. R. M. Syaifudin, P. L. Yu, S. C. Mukhopadhyay, C. P. Gooneratne, and J. Kosel, “Analysis of different coating thickness on new type of planar interdigital sensors for endotoxin detection,” in Proc. IEEE Int. Instrum. Meas. Technol. Conf. (I2MTC), May 2013, pp. 1496–1501. [17] H. Müller, T. von Dobeneck, W. Nehmiz, and K. Hamer, “Nearsurface electromagnetic, rock magnetic, and geochemical fingerprinting of submarine freshwater seepage at Eckernförde Bay (SW Baltic Sea),” Geo-Marine Lett., vol. 31, no. 2, pp. 123–140, 2011. [18] O. Korostynska, A. Mason, and A. Al-Shamma’a, “Proof-of-concept microwave sensor on flexible substrate for real-time water composition analysis,” in Proc. 6th Int. Conf. Sensing Technol. (ICST), Dec. 2012, pp. 547–550. [19] M. A. M. Yunus, S. C. Mukhopadhyay, and A. Punchihewa, “Application of independent component analysis for estimating nitrate contamination in natural water sources using planar electromagnetic sensor,” in Proc. 5th Int. Conf. Sensing Technol. (ICST), Dec. 2011, pp. 538–543. [20] M. A. M. Yunus, S. C. Mukhopadhyay, and S. Ibrahim, “Planar electromagnetic sensor based estimation of nitrate contamination in water sources using independent component analysis,” IEEE Sensors J., vol. 12, no. 6, pp. 2024–2034, Jun. 2012. [21] J. Jiang, P. Trundle, and J. Ren, “Medical image analysis with artificial neural networks,” Comput. Med. Imag. Graph., vol. 34, no. 8, pp. 617–631, 2010. [22] H. R. Maier and G. C. Dandy, “Neural networks for the prediction and forecasting of water resources variables: A review of modelling issues and applications,” Environ. Model. Softw., vol. 15, no. 1, pp. 101–124, 2000. [23] P. Graff, F. Feroz, M. P. Hobson, and A. Lasenby, “Neural networks for astronomical data analysis and Bayesian inference,” in Proc. IEEE 13th Int. Conf. Data Mining Workshops (ICDMW), Dec. 2013, pp. 16–23. [24] J. K. Basu, D. Bhattacharyya, and T.-H. Kim, “Use of artificial neural network in pattern recognition,” Int. J. Softw. Eng. Appl., vol. 4, no. 2, pp. 23–34, 2010. [25] Q. Zhou, P. Shi, H. Liu, and S. Xu, “Neural-network-based decentralized adaptive output-feedback control for large-scale stochastic nonlinear systems,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 42, no. 6, pp. 1608–1619, Dec. 2012. [26] H.-G. Han, Q.-L. Chen, and J.-F. Qiao, “An efficient self-organizing RBF neural network for water quality prediction,” Neural Netw., vol. 24, no. 7, pp. 717–725, 2011. [27] P. Banerjee, V. Singh, K. Chatttopadhyay, P. Chandra, and B. Singh, “Artificial neural network model as a potential alternative for groundwater salinity forecasting,” J. Hydrol., vol. 398, nos. 3–4, pp. 212–220, 2011. [28] M. A. M. Yunus and S. C. Mukhopadhyay, “Novel planar electromagnetic sensors for detection of nitrates and contamination in natural water sources,” IEEE Sensors J., vol. 11, no. 6, pp. 1440–1447, Jun. 2011. [29] M. A. M. Yunus, V. Kasturi, S. C. Mukhopadhyay, and G. S. Gupta, “Sheep skin property estimation using a low-cost planar sensor,” in Proc. IEEE Instrum. Meas. Technol. Conf. (I2MTC), May 2009, pp. 482–486.

IEEE SENSORS JOURNAL, VOL. 15, NO. 1, JANUARY 2015

[30] J. Garcia-Canton, A. Merlos, and A. Baldi, “High-quality factor electrolyte insulator silicon capacitor for wireless chemical sensing,” IEEE Electron Device Lett., vol. 28, no. 1, pp. 27–29, Jan. 2007. [31] M. A. M. Yunus, G. R. Mendez, and S. C. Mukhopadhyay, “Development of a low cost system for nitrate and contamination detections in natural water supply based on a planar electromagnetic sensor,” in Proc. IEEE Instrumen. Meas. Technol. Conf. (I2MTC), May 2011, pp. 1–6. [32] S. Palani, S.-Y. Liong, and P. Tkalich, “An ANN application for water quality forecasting,” Marine Pollution Bull., vol. 56, no. 9, pp. 1586–1597, 2008. [33] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, Learning Representations by Back-Propagating Errors. Cambridge, MA, USA: MIT Press, 1988.

A. S. M. Nor (S’14) received the bachelor’s degree in electrical engineering from Universiti Teknologi Malaysia, Johor Bahru, Malaysia, where he is currently pursuing the master’s degree in electrical engineering at the Department of Control and Mechatronic Engineering. His research area of interest is planar electromagnetic sensors array applications in water contamination estimation.

M. Faramarzi received the B.Eng. degree in communication system from Shahid Bahonar University, Kerman, Iran, in 2004, and the M.Sc. degree in control engineering from Gonabad Azad University, Gonabad, Iran, in 2010. He is currently pursuing the Ph.D. degree in electrical engineering at the Department of Control and Mechatronic Engineering, Universiti Teknologi Malaysia, Johor Bahru, Malaysia. His current research areas of interest are application of artificial neural network and fuzzy system application in ultrasonic process tomography.

M. A. M. Yunus (S’08–M’14) received the B.Eng. and M.Eng. degrees from Universiti Teknologi Malaysia (UTM), Johor Bahru, Malaysia, in 2002 and 2005, respectively, and the Ph.D. degree from Massey University, Palmerston North, New Zealand, in 2011, all in electrical engineering. He is currently a Senior Lecturer with the Department of Control and Mechatronics Engineering at UTM. His research interests include process tomography, planar electromagnetic sensors, and wireless sensing technology.

S. Ibrahim received the Ph.D. degree in process tomography from Sheffield Hallam University, Sheffield, U.K., in 2000. He is currently an Associate Professor with the Department of Control and Instrumentation Engineering, Faculty of Electrical Engineering, Universiti Teknologi Malaysia, Johor Bahru, Malaysia. He has also presented many papers in various conferences. His current research interest is in the field of instrumentation, sensors, and tomography.