Gas Concentration Analysis with Curve Fitting Neural Network for Electronic Nose

Transcription

1 Gas Concentration Analysis with Curve Fitting Neural Network for Electronic Nose Shoffi Izza Sabilla Department of Informatics Institut Teknologi Sepuluh Nopember Surabaya, Indonesia Riyanarto Sarno Department of Informatics Institut Teknologi Sepuluh Nopember Surabaya, Indonesia Abstract It is necessary to provide the accurate calculation method of Ratio, Resistance, and PPM or mg/l for developing the prediction tool of fruit types by using Electronic Nose sensor. Some of curve fitting methods is used to forecasts the result of fruit types, such as exponential and power. But both of them are less accurate and less proper, due to less real data of determining Resistention and PPM result for groundtruth. In this research, This research using Neural network curve fitting methods with 7 rounds/nodes of hidden layer and 15 times of data training. Neural Network is used by using data training for exactly 7 nodes. The data are taken from 10 sensors with 55 kind of gas additives, each has 35 nodes of Rs/Ro and PPM value, so there are 1925 data that will be used. This method is also used for predicting Rs/Ro and PPM for the ripining of mangoes. The result gives an appropriate and accurate value of Rs/Ro and PPM along with the datasheets given in Electronic Nose sensor. Keywords Neural Network; Electronic Nose; Mangoes I. INTRODUCTION Electronic Nose or E-Nose is a tool for simulating human sense of smell [1]. Human nose has capability to smell variety of gases, but E-nose is not only simulating but also identifying gas additives and detecting the harmness of gases. There are plenty of free gas up in the air which provide smell and taste. E-nose is proved in some of case, for example fisheries [2], medicine and pharmacy [3], health [4] which provides the appropriate result. E-nose featured some sensors, that are: 2, 4, 5, 6, 135, and more. Each sensor has different characteristics to identify smell and taste in the gas. Sensor 2 is able to detect H2, Alcohol, LPG, CH4 [5]. Sensor 3: Alcohol, CH4, CO [6]. You can see in 2 and 3, both has the capability for detecting alcohol, but it has different sensitivity level to measure same gas. arious kinds of gas content on the sensor can be seen in the TABLE I. Each sensor has graph datasheets provided from the manufacturer, which has X for PPM value and Y for Rs/Ro. In this development of E-nose system, Rs/Ro has main role of determining the appropiate PPM value based on the datasheets given. However there is no further research for measuring the accuracy of Rs/Ro and PPM value or mg/l. Electronic nose is a portable sensor with a relatively cheap price With the help of Arduino as microcontroller, sensor readings will be easier. Gas 2 TABLE I. LIST OF GAS SENSOR H2 LPG 137 CH4 CO Alkohol Smoke Propane Benzine Hexane NH4 Co2 NH3 O2 C2H6O n- Hexane II. MATERIALS AND METHODS 138 A. Collect Data for Ratio The next step is observing the Rs/Ro movement in both sensor and gas additives itself. Calibration tools can be used to make sure the readability of Rs/Ro and mg/l value in the datasheets, or using manual measurement if necessary. More data provides high level of accuracy, so this research is using the image editor tool, Adobe Illustrator, for ease measurement of millimeter unit. The data are taken from 10 sensors with 55 kind of gas additives, each has 35 nodes of Rs/Ro and PPM value, so there are 1925 data that will be used. The result of this observation can be seen on the Figure 1. Figure 1. Rs/Ro of 2 gas Alcohol

2 B. Curve Fitting There are some methods that can be used for measuring the curve fitting accuracy in the Rs/Ro and mg/l. In previous research, to do the curve fitting accuracy of rs/ro using adjust power method [7] can be found on formula (5) and power method [8] can be found on formula (4). Each method has its own advantages and disadvantages, so observations made to determine the appropriate method for developing the E-nose sensor system. This research use 4 methods which has low error possibility, that are: exponential curve fitting method on formula (1), normalization exponential curve fitting method on formula (2) or (3), power curve fitting method, and adjust power curve fitting method. Exponential Method Curve fitting Where Yi is the given data, Yi2 is a prediction value, and n is the calculation data. RSE measures the deviation of value that are predicted and observed, while R^2 measures the value around the data. TTS means Total Sum of squares. Residual Standart Error can be found on formula (7).! 2 = %&& -!&& %&& = 1!&& %&& (7)! = # $%+' (1) Expand the formula will be or, Power curve fitting method Adjust Power Method ln # = ln (' ()*+ ) (2) ln # = %& + ( (3)! = $ %& %'! = 10 &'( ( )* )' -, -, $* + %, (4)., 0, 1, 2 (5) C. Mathematical Models Comparison There are two ways for determining the chosen method, the Residual Standard Error (RSE) and R^2, and also there is Root Mean Squared Error (RMSE). RSE provided from the means of error, if it has least RSE value so it gives more accurate result.!"# =!"# = 1 '-2 1 '-2!"" / 012 1(+,-+,2). (6) Yi is a data which is observed, fi is predicted value, y is mean from observed data. If the error value closes to 1, the curve fitting methods is correct, otherwise it is not correct if the error value closes to 0. Besides the two steps before, this formula squares the error value and divided by the all the tested data, then giving the roots as result. On the other hand, root mean squared error can be shown in formula (8) or (9).!"#$ = This is for simplified formula!"#$ = 2 ( ) (+( - -() / (8) 1 (!## (9) Where Yi is the actual data, Fi is the prediction value, and n is the total data. A is the Exponential curve fitting method B is the Normalization Exponential curve fitting method C is the Exponential Orde 2 curve fitting method D is the Power curve fitting method E is the Adjust Power curve fitting method F is the Neural Network (JST) curve fitting method Sens or 2 3 TABLE II. COMPARISON OF ERROR Gas A B C D E F CO 3,,4202 0,3501 Nan 0,1359 0,1316 0,1096 Alkohol Smoke 4,0028 0,2120 0,0734 0,2577 0,0944 0,0356 Alkohol 2,1389 0,4736 0,125 0,1318 0,1157 0,0791 Benzine 16,5369 0,4640 0,1538 0,2057 0, Hexane Nan LPG Nan

3 CO O C2H6O NH Benzen e n- Hexane CH CO Nan Alcohol Propane Based on the research result can be seen in TABLE II, those methods has smaller error level rather than the other methods. Curve fitting methods that is implemented for calculating the accuracy of Rs/Ro and PPM value use Neural Network curve fitting methods. Figure 2 is the result of comparison to five formulas with data from datasheet or groundtruth. Input layer... Hidden layer Output layer ppm Figure 3. Neural Network Principal In the Arduino Uno system, there is a calibration step of free air (RO AIR CLEANER FACTOR) right before sensor is read, to provide more accurate data response while detecting the gas concentration. Arduino scans the response data from the sensor periodically in real-time. The sensor uses the digital (ADC) value to get Resistance (Rs), can be found on formula (10) or (11). $% - $!(!" = )!( $!( (10) Figure 2. Comparison to five curve fitting methods From the figure we can see the blue line is the data from the datasheet 3. The red line is Rs/Ro using the Neural Network curve fitting method. The green line is the Power method. The black line is the power adjust method. The magenta colored line is the exponential normalization method. And the cyan colored line is the exponential method D. A Implementation of new formula ANN known as a way of stimulating human brain work. This method consist of 3 layers, that are: input layer, output layer, and hidden layer [9]. The data is processed from input layer to the output layer one by one, where hidden layer placed between those two. But only a few ANN doesn't have hidden layer, just input layer and output layer. Figure 3 is the visualization understanding of the neural network. III. RESULT AND DISCUSSION This experiment uses Arduino as a microcontroller. Based on the previous research, more alcohol concentration means if the fruit is ripe [10]. The sensor is placed on the Arduino microcontroller board that matched with the PIN input, then it scans the analog signal that converted to digital. The digital PIN will only recognize 0 as low value and 5 olts as a high value.!"# = %&' -!* 1023 (11) c is the voltage from microcontroller board, RL is the voltage generated from the sensor in tested area. RL is the resistance value from the sensor datasheet with kilo Ohm unit. Digital value is the result from the analog to digital conversion (ADC). In Arduino, microcontroller has 10 bits resolution from 0 to Next step is to calculate Rs/Ro after knowing the resistance (Rs) value. A. Sensors All sensor will be tested to determine the best use of detecting alcohol concentration by using Slope-Deflection and ariant Coefficient. The more Slope-deflection value, the sensor will provide good result of gas concentration detection [11].!"#$% = ( )- ( +, ) -, + (12) Y2 is the end node from Y line and Y1 is the initial node from Y line. X2 is the end node from line X and X1 is the initial node from line X. The formula of slope can be shown on (12). The ratio of the slope value on the alcohol gas content in the sensor 2, 3, 138 can be seen in the TABLE III. Most small value from comparison is sensor 3. ariant coefficient is used to find the stable point of the sensor. Rs/Ro value is observed when the sensor moves close to mango fruit. Rs/Ro value is used to determine which sensor that most appropriate for calculating the gas concentration with variant coefficient formula.

4 TABLE III. RESULT OF SLOPE Slope- Sensor Deflection The ratio of the variant coefficient value on the alcohol can be seen in the TABLE I. M1 is raw mango and M2 is ripe mango. Most small value from comparison is sensor 3. The less value of variant coefficient, the more stable of the sensor to calculate the concentration of alcohol before and after the presence of mango [12]. TABLE I. RESULT COMPARISON OF ARIANT COEFFICIENT Mean Sensor Mango (M1, M2) ariant ariant Coefficient Coefficient M M M M M Figure 3. Response of Alcohol gas content on sensor 3 Based on the test, the result shows that sensor 3 has good accuracy for detecting the alcohol concentration of the mango. It also detects the difference between the ripe fruit, raw fruit, or rotten fruit concentration which shows in the graph. B. Proposed Mathematical Models PPM or mg/l value is correct compares with the data given from the 3 sensor datasheet. This is proved by the small value of RMSE, that is Prediction of mg/l value using ANN methods very close to data from datasheet or groundtruth. TABLE. RESULT OF CONERSION TO Mg/L mg/l Tessting Rs/Ro mg/l ANN Datasheet M M M2 0, M M Raw Mango Ripe Mango Figure 3 is the response of the sensor 3 where the Rs/ Ro value of the resulting mango for 400 seconds. Blue line is clean air, red line is mature green line is ripe mango, and line cyan is rotten mango. Can be seen differences in response and alcohol gas content in clean air, raw mangoes, ripe mango, and rotten mango. Rotten Mango RO =

5 Figure 4. Prediction mg/l value from Rs/Ro raw mango Figure 6. Prediction mg/l value from Rs/Ro rotten mango Figure 5. Prediction mg/l value from Rs/Ro ripe mango Experiments were performed on 8 mangoes, 3 raw mangoes, 2 ripe mangoes, and 3 rotten mangoes. The experiment is carried out for 400 seconds in a closed place so it is very unlikely denoising to occur. Seen from the Figure 4, the blue line shows the Rs/Ro and mg/l data of the 3 datasheet and the green line is the prediction of the Rs/Ro and mg/l values for raw mangoes. This also shows the prediction of the value of Rs/Ro and mg/l in ripe mango and rotten mango. Figure 5 is plotting for ripe mangoes with red stripes. The results of the data also matches the 3 datasheet and the higher the alcohol content. The ANN method also succeeds in predicting the value of Rs/Ro and mg/l in the rotten mango and the results correspond to the 3 datasheet. Prediction rotten mango can be seen on Figure 6. Figure 7 is the result of prediction mg/l value for raw, ripe, and rotten mango experiments made into one graph. Looks different value of mg/l for mango maturity level. Figure 7. Result of all prediction ripening mango I. CONCLUSION After doing the research on several sensors in electronic nose, each sensor has its own characteristics of calculating the gas concentration. There are some ways to determine the accuracy of ppm or mg / L using a curve fitting method, but the ANN method is suitable to be implemented in electronic nose sensors with the RMSE of These method gives a good result as the 3 datasheet for estimating the value of mg/l in the alcohol concentration of mango. Sensor 3 is an appropriate sensor to detect the alcohol gas content in the maturity of mango fruit. In addition, these sensors can identify raw mango, ripe mango, and rotten mango. A. Future Works Further research will be done classification of mango fruit maturity level of Rs/Ro value by SM, k-nn method, decision tree, neural network (JST), fuzzy and Extreme Learning. ACKNOWLEDMENT Authors would like to thank Institut Teknologi Sepuluh Nopember for supporting the research.

6 REFERENCES [1] Syahbanisyah, Fun dengan Fungsi Linier, 7 September [Online]. Available: [Diakses 14 Agustus 2017]. [2] P. Sidik, Rumus Koefisien ariasi, 21 Januari [Online]. Available: [Diakses 14 Agustus 2017]. [3] TECHNICAL DATA -3 GAS SENSOR, [Online]. Available: [Diakses 15 Agustus 2017]. [4] Wijaya, D.R; Sarno, R,; Zulaika, E, Sensor array optimization for mobile electronic nose: Wavelet transform and filter based feature selection approach, Int. Rev. Comput. Softw., vol. 11, pp , [5] Linder R.; Perriello G.; Di Daniele N.; Poppl S.J.; De Lorenzo A. Mohamed E.I., Predicting Type 2 Diabetes using an Electronic Nose-based Artificial Neural Network Analysis, Diabetes Nutr Metab, vol. 15, pp , [6] O'Connell, M; aldora, G; Peltzer, G, Practical Approach for Fish ness Determinations Using A Portable Electric Nose, Sens Actuator, vol. 80, pp , [7] Szczurek, P.M.; Licznerski, B.W. Szecówka, On reliability of neural network sensitivity analysis applied for sensor array optimization, Sensors and Actuators B: Chemical, pp , [8] -2 Semiconductor Sensor for Combustible Gas, [Online Available: [Diakses 15 Agustus 2017]. [9] Nouri, Farhad Gholizadeh; Chen, Zhiyuan; Maqbool, Mehdi;, Monitoring Mango Fruit Ripening after Harvest using Electronic Nose (znose) Technique, IACSIT Press, vol. 65, [10] Wijaya, D. R; Sarno, R.; Zulaika, E., Gas concentration analysis of resistive gas sensor array, Int. Symp. Electron. Smart Devices, pp , [11] Zhu, L; Seburg, R.A; Puech, S; Mifsud, J.C, Flavor Analysis in A Pharmaceutical Oral Solution Formulation Using an Electric Nose, Journal of Pharmaceutical and Biomedical Analysis, vol. 34, pp , [12] Baietto, Manuela; Wilson, Alphus D., Electronic-Nose Applications for Fruit Identification, Ripeness and Quality Grading, Sensors, vol. 15, pp , [13] Sarno, Riyanarto; Hariyanto, Dedy R. W;, Detection of Diabetes from Gas Analysis of Human Breath using E- Nose, [14] Siswantoro, J; Hilman, M. Y; Widiasri, M, Computer ision System for Egg olume Prediction Using Backpropagation Neural Network, [15] Garner, C.E; Smith, S; Bardhan, P.K; Ratcliffe, N.M; Probert, C.S, A Pilot Study of Fecal olatile Organic Compounds in Feces from Cholera Patients in Bangladesh to Determine their Utility in Disease Diagnosis, Transactions of the Royal Society of Tropical Medicine and Hygiene, vol. 103, pp , 2009.