Classification of Taste using a Neural Network: A Case Study in Mineral Water and Drinking Water Classification TEO JAU SHYA 1, MOHD NOOR AHMAD 1, MUHAMMAD SUZURI HITAM and ALI YEON SHAKAFF 1 School of Chemistry School of Industrial Technology School of Electrical and Electronic Engineering University Sains Malaysia 118, Penang MALAYSIA Abstract: -A prototype of disposable one-time measurement taste sensor has been developed. The taste sensor consists of an array of non-selective sensors is fabricated using screen printing technology. A multi-layer feed forward neural network is employed to classify the taste of the tested medium. In this paper, the mineral water and drinking water was chosen to evaluate the classification ability of the prototype. Experiments were carried out on variety of a commercially available mineral water and drinking water. The classification results were analyzed by using several algorithms namely standard back-propagation algorithm, Conjugate Gradient algorithm, Quasi- Newton algorithm and Lavenbert-Marquardt algorithm. Results show that a standard backpropagation algorithm could not converged. However the other learning algorithms could perform the classification task with 1% accuracy. Key-Words: Neural network, classification, disposable taste sensor, back-propagation, Conjugate Gradient, Quasi-Newton and Lavenbert-Marquardt algorithm. 1 Introduction An expert tester has a unique ability to classify taste with high accuracy. It is anticipated that if this unique ability of human expert to perform taste classification could be carried out by an artificial taste sensory system, a human taste expert is no longer needed to perform such difficult job. Artificial taste sensory system or sometimes referred to as electronic tongue system [1], should be able to substitute biological tongue in some obvious fields such as quality checking of foodstuffs. In addition, it could increase the range of tasting media, i.e., from the poisonous food into the non-eatable food, [] and []. In this paper, a study towards the development of a simple but reliable, one-shot disposable taste sensor, which is capable of identifying and classifying predetermined taste is presented. In other words, the prototype is able to identify and classify a set of selected pattern samples in such a way that each samples must be assigned to one of a set of pre-defined classes. A mineral water and drinking water were selected because their tastes are quite subtle and hence difficult for a human to discriminate between the two. Natural mineral water is defined as ground water which is obtained for human consumption from subterranean water-bearing strait through a spring, well, bore or other exit, with or without the addition of carbon dioxide. Whereas, the drinking water is a potable water or treated potable water, other than natural mineral water, that is hermetically sealed in bottles or other packages and is intended for human consumption []. It should be noted that the market price for both types of water is very much difference. Without proper quality control, there is no assurance that the water contained in the bottle is the same as what is stated on the outside label. The conventional analytical methods, such as capillary
electrophoresis require long and complicated processes. Therefore, a more suitable, economical and faster method is required. Human Taste System and the Artificial Taste System The basic stimulus for taste arises from the contact of substances with our receptors, i.e., taste buds, located throughout our tongue. The chemical interaction at the receptors leads to chemical change which generates a neural impulse []. The impulses are transmitted along the nerve fiber into the brain that leads to taste perception [6]. The perception of taste is acquired through a learning process. The design of the artificial taste sensory system is based on biological principles of human taste sensory system. The taste sensor acts as a receptor and the neural network plays a similar role to human brain; to recognize the taste. Fig.1 illustrates the relation between the human taste recognition system and the artificial taste recognition system. An Artificial Taste System In this project, the artificial taste sensing system consists of an array of non-selective disposable screenprinted sensors which is interfaced to a personal computer via a Multi-Interface Unit. The disposable strips of the taste sensors were designed and fabricated at the University Sains Malaysia with the help from Screen Technology Corp., Malaysia. Fig. shows the top view of a miniature disposable taste sensor fabricated by screen-printing technology. Each of the electrodes was printed in arrays of eight tracks of working electrodes and a track printed with Ag/AgCl, as the reference electrode. The sensor array is later deposited with lipid membranes. The lipid materials used are similar to the one reported by Toko et al, [] and []. Taste recognition Taste recognition Brain Nerve fibers Receptors (taste buds) Human taste system Neural net Data acquisitions An array of nonselective screenprinted sensors Artificial taste system Fig.1 Relationship between human taste system and the artificial taste system. Fig. Disposable taste sensor. The responses of all the eight sensors are based on the measurement of electrical properties of the chemical substances in the media. The principle is based on the fact that a large number of different compounds contribute to defining a measured taste; the chemical sensor array of the taste sensor provides an output sample pattern due to the potential difference, which represents a combination of all the components [7]. Fabrication of Taste Sensor Screen-printing technology is a technique whereby the screens allow ink or paste to be applied into a substrate with a squeegee in a particular size, shape and sequence of the print, as illustrated in Fig.. The open pattern in the screen defines the pattern that will be printed on the substrate. Each of the screen-printed electrodes was printed in arrays of eight tracks of working electrodes
and a track printed with Ag/AgCl, as the reference electrode. The electrodes were manufactured by Screen Technology Corporation, Malaysia. The final step of the fabrication is the deposition of lipid membranes with a dispenser into the working electrodes of the array. Squeegee Screen Printing Direction Printing Ink channel 1 channel channel channel channel channel 6 channel 7 channel 8 Output Signal Substrate (polyester) Fig.: Screen printing of conductive paste. Taste Classification Each complex liquid presented to the disposable taste sensor produces a specific signature or a pattern characteristic of the solution. By presenting many different chemical solutions to the sensors, a database of signatures of the liquids were build up and later used for the neural network training purpose. In this project, a fully connected backpropagation neural network is employed. The multilayer feedforward neural network was trained in batch mode with four types of learning algorithms, namely, Levenberg-Marquardt, Quasi-Newton Algorithms, Conjugate-Gradient Methods and Gradient Descent [8]. The network is also trained with different hidden neuron sizes and various learning-rate to determine the optimal parameter of the network. Fig. illustrates a multi-layer feedforward neural network structure with eight input nodes and a single output which represent the eight sensor input channels and classification of the type of water tested, respectively. The hyperbolic tangent functions were used in the hidden layer and a linear function is used in the input and output layer. Input Layer Hidden Layer Fig.: The neural network structure. 6 Results and Discussions In this study, 8 samples of mineral water and drinking water that are commercially available in the Malaysia's market were obtained. Tests were conducted using each sample and were repeated 8 times each. A total of 76 measurements were recorded. The data were separated into three categories; training, validation and testing in the ratio of 6%, % and %, respectively. In all the experiments conducted, the network were set as such that the output value of the network will classify mineral water when the network response value is equal and more than., otherwise it will classify as drinking water. In all of the experiment employing the advance training algorithm, it was found that 1% identification success rate were achieved when using the testing samples. 6.1 Training algorithms As expected, the neural network performs differently when different learning algorithms were employed. Fig. shows the performance of the four different training algorithms during the training session when the network were set to hidden neurons at the learning
rate set to.1. Fig. (a) and Fig. (b) showed that there is a significant difference in the performance of learning between a standard backpropagation algorithm and the other three training algorithms namely the Conjugate Gradient algorithm, Quasi-Newton algorithm and Lavenberg-Marquardt algorithm. Fig. (a) indicates that the Levenberg-Marquardt learning algorithm provides fastest convergence as compared two the other two. However, it was found that the standard gradient descent algorithm could not converged even the network training was continued for 1, epochs. Training-Blue Goal-Black Training Curves varying learning algorithm with No.hidden neurons = & learning rate=.7 Conjugate Gradient Algorithms Quasi-Newton Algorithms.6 Levenberg-Marquardt M ea n- sq ua re d err or -.1 1 1 1 Performance is., Goal is 1e- 1 1 1-1 -.....1 Fig. (a): Advance algorithm. 1-6 1 1 Epochs Fig. (b): Standard back propagation algorithm. 6. The Effect of Hidden Neuron's Size and Learning Rate Parameter The number of adjustable parameters in ANN model is directly linked to the learning rate and the number of neurons in the hidden layer. As the number of training sample is low, it is recommended to use few hidden nodes as possible to avoid over-fitting and to make the models robust [9]. To determine the optimal number of hidden neurons and learning-rate in this study, the smallest number of hidden neurons with the learning rate parameter that yield convergence to a minimum in training and the corresponding error with the least number of epochs were chosen. However, to be certain of the results, for each algorithm, the best parameters and classification result is chosen based on the results from the validation set that is monitored during the training process. The size of the hidden neuron is varied from to at different learning-rate with the range from.9 to.1. Once a rough estimation of the suitable parameter value could be estimated, the process is repeated. From the initial experiment, it is found that the number of hidden neuron within the range of two to ten neurons provides better convergence. By using this number of hidden neuron, the learning rate value is varied between in the range from.1 to.9. The experiments were repeated 1 times with a minimum of epochs and the results are the average of these trials. Training curves varying the number of hidden neurons at learning rate of. and.1 are shown in Fig. 6(a) and 6(b), respectively. It can be observed that most of the plots show the similar trends; the numbers of hidden neurons at thirty and fifty results in a much slower convergence than the rest. It could also be observed that most of the networks converged at the average of sixth epoch suggests that the number of hidden neurons does not affected the convergence of the network. Fig. 7 shows an overall best learning curve. From Fig. 7, it appears that the optimal learningrate parameter is about.7 and the optimal number of hidden neurons is. 7 Conclusions The response of an array of disposable taste sensor fabricated using screen-printing technology is capable of identifies and classifies commercially available drinking water and mineral water by an aids of fully connected feed-forward neural network.
Combinations of different learning algorithms, numbers of hidden neurons and learning-rate parameters are investigated to observe their effect on network convergence properties. Each combination is trained with the same set of 76 sample data and the results of the experiment are compared directly. The optimal learning-rate parameter is found to be around.7 and the optimal number of hidden neurons four neurons. By using the above parameter, it is found out that 1% classification success rate could be obtained.. Mean-squared error Training Curves varying at diff no. hidden neurons & at learning rate =. 7 6 8 1 Mean-squared error 1.6 1. 1. 1.8.6.. BestTraining Curves hidden neurons with lr=.9 hidden neurons with lr=.7 hidden neurons with lr=.7 hidden neurons with lr=.7 8 hidden neurons with lr=.1 1 hidden neurons with lr=.7 -. 1 6 7 8 9 1 Fig. 7: The best learning curves. 1 Mean-squared error 1 6 8 1 1 1 16 Training Curves varying at diff no. hidden neurons & at learning rate =.1 9 8 7 8 1 6 6 8 1 1 1 Fig. 6(b) Training curves varying number of hidden neurons at learning rate.1. Acknowledgement The authors gratefully acknowledge the research grants provided by University Sains Malaysia, Penang as well as the Ministry of Science, Technology and the Environment Malaysia that have resulted in this article. References: [1] Yu.G.Vlasov, A.V. Legin, A.M. Rudnitskaya, A.D Amico, C.Di Natale, Electronic tongue new analytical tool for liquid analysis on the basis of non-specific sensors and methods of pattern recognition, Sensors and Actuators B, Vol.6,, pp.-6. [] Robert Koncki, Lukasz Tymecki, Elzbieta Zwierkowska, Stanislaw Glab, Screen-printed copper ion-selective electrodes, Fresenius J Anal Chem, Vol.67,, pp. 9-9. [] C.A. Galan-Vidal, J. Munoz, C. Dominguez, S. Alegret, Glucose biosensor strip in a three electrode configuration based on composite and biocomposite materials applied by planar thick film technology, Sensors and Actuators B, Vol., 1998, pp.7-6. [] Food Act, 198 (Act 81) and Food Regulation, 198, Int. Law Book Services, 199.
[] Alison Mitchell, Signal transduction : A taste of things to come, Nature Vol.98, 1999, pp.. [6] Robert H. Cagan, Morley R. Kare, Biochemistry of Taste and Olfaction, Academic Press 1981,. pp.. [7] F.Winquist, S.Holmin, C.Krantz-Rulcker, P.Wide, I.Lundstrom, A hybrid electronic tongue, Analytical Chimica Acta, Vol.6,, pp.17-17. [8] Simon Haykin, Neural network A Comprehensive Foundation nd edition, Prentice Hall International, Inc. 1999 [9] M. Baret, D.L Massart, P.Fabry, F.Conesa, C. Eichner, C. Menardo, Applications of neural network calibrations to an halide ISE array, Talanta, Vol.1,, pp.86-877.