2018 International Conference on Information and Communications Technology (ICOIACT)
Detection of Organic Solvent Compounds Using Optical Fiber Interferometer Array and Neural Network Pattern Recognition Dwi Sasmita Aji Pambudi, Muhammad Rivai, Achmad Arifin Department of Electrical Engineering Institut Teknologi Sepuluh Nopember Surabaya, Indonesia
[email protected],
[email protected],
[email protected] the identification of the gas is carried out on the gas of the human breath and the gas urine to diagnose a non-invasive disease [3,4]. Human breath gas consists of nitrogen, oxygen, water vapor (H2O) and carbon dioxide (CO2). While most of the other consists of volatile organic compounds.
Abstract—Organic solvent compounds are widely used as production raw materials in the field of chemical industry. Organic compounds are easily changed from liquid to gas conditions at room temperature. Organic solvent compounds are commonly found as gases or vapors, which are flammable, toxic, and explosive. The identification of the gas sensor is required in identifying and classifying some gases of volatile organic compounds, especially to monitor the condition of the organic solvent vapor pollutants in the environment. The latest development of gas sensor was designed based on the optical field by using Fabry-Perot interferometer which is applied to optical fiber to increase the sensitivity of gas sensor. The gas sensor was designed by coating chemical membranes on the tip of the optical fiber to increase the sensor selectivity. Three different types of chemical membranes are coated on the same three optical fibers placed in the sensor chamber. In this study, sensor output data are interpreted into digital form through analog-to-digital converter, while data processing and identification are performed by computer. The identification process of organic solvent is done by using artificial neural network algorithm. The results show that the sensor array could produce a different pattern for each of the gas vapor samples. The Neural network pattern recognition system can identify the type of vapor with 100% accuracy rate. Identification of organic solvent compound types, may be used to detect low-vapor gas vapor exposure applied in monitoring activities and analysis of organic solvent vapor.
The industrial environment is an area susceptible to exposure to vapor and gas exposure generated by organic compounds continuously. Acetone (C3H6O), benzene (C6H6), and n-hexane (C6H14) belong to toxic combustible and flammable organic compounds. These organic compounds are widely used as additives in various products such as nail polish, cigarette, gasoline, dyes, perfumes, explosives, paints and liners, adhesives, and others. Exposure to or contact with the compound for a long time or excessively may cause adverse health effects. Some of the health effects that can be caused are upper respiratory tract irritation, nervous system disorders and cause headaches. When the vapor of the compound into the lungs, it can cause injury and stimulate cancer cells. In recent years, several studies have used semiconductors [5], quartz micro balance [6,7], chemocapacitor [8], and polymer [5, 9, 10] materials in the detection and identification of gas vapors using sensor array. Currently, many types of gas sensors, using metal oxide semiconductors, which commercially have a low sensitivity level, which can only perform measurements of gas concentration in the range 1010000 ppm. While humans only have the ability to smell gas at a concentration of at least 10 ppm.
Keywords—Fabry-Perot interferometer, neural network, optical fiber, organic solvent
I. INTRODUCTION
As an optical sensor development, optical interferometry method is used to improve sensor sensitivity in measuring gas concentration because it has several advantages: high sensitivity, fast response, low power consumption, stable at high temperatures, and robust to interference, both of electric field and magnetic field [10]. The optical interferometer consists of several types, commonly used are Mach-Zehnder interferometer, Michelson interferometer [11], Fabry-Perot interferometer. But in this study used Faber Perot interferometer as a sensor to measure the gas concentration. Fabry-Perot interferometer was chosen because of its relatively simple configuration and is more easily applied to optical fibers as sensors, lower cost, small size and high resolution [12].
The gas identification is needed to know the content of the gas vapor compounds contained therein. Sometimes, the gas is measured in low concentrations, so it will be very difficult to detect by the human sense of smell. The importance of gas measurements in low concentrations, led to various studies related to the development and improvement of gas sensitivity sensors. Measurements in low gas concentrations can be applied to several purposes. In the military field, gas identification is carried out on explosive and flammable gases [1]. While in the industrial field, gas identification is conducted to control the air quality level with the growth rate of pollutants produced by industrial processes [2]. In the biomedical field,
978-1-5386-0954-5/18/$31.00 ©2018 IEEE
477
2018 International Conference on Information and Communications Technology (ICOIACT)
Relay Module
Computer
DAQ Sensor Chamber SoC Arduino Nano Inlet
Outlet
Serial interface
EoC
Optical Fiber Tip 16 bit
Valve Cleaning
Valve In
Analog to Digital ADS1115
Valve Out Y Coupler
Silica Gel
Gas Sample
Light Source (White led)
Photodiode Low Pass Filter 1
Non-Inverting Amplifier 1
Low Pass Filter 2
Non-Inverting Amplifier 2
Low Pass Filter 3
Non-Inverting Amplifier 3
Air Pump Singlemode Fiber Optic Air Signal Conditioning Circuit
Fig. 1. Schematic diagram of experiment setup for organic solvent vapor detection and measurement system using optical fiber interferometer sensor
interferometer is a function of the phase difference between the reflected light beam sequentially.
This research was designed to identify gas to several types of gases from volatile organic compounds. Gas vapor sensors was designed to have high sensitivity by applying Fabry-Perot interferometer on optical fiber coated with a particular chemical membrane at optical fiber tip. Optical fiber interferometer was used to increase the ability and sensitivity of sensors to very small changes. While the use of chemical compounds was to increase the sensitivity and selectivity to certain types of test gases. Some gas molecules absorbed by chemical membranes also increases sensor sensitivity in low gas concentrations. The chemical membranes will be in swelling process or expanding effect when sensor was exposed to gas [4].
The construction interference occurs when the transmitted light is in phase. If the transmission light has a half phase difference, there is destructive interference. The reflected light repeats in phase and depends on the wavelength (O), the incident angle to etalon (T), the thickness of etalon (L), and the material refractive index between the reflected (n) surfaces. Thus the phase difference (G) between the reflections occurring can be calculated according to Eq. (1),
G
The chemical membranes will interact with a sample gas based on the similarity of the polarity level between sample gas and chemical membranes. Differences in the output value of the gas concentration produced from each type of chemical membranes coated on the optical interferometer are used to perform the process of classification of certain types of gases. The classifying process of gas vapor is implemented by using Backpropagation Neural Network algorithm. Neural Network algorithm has the ability to learn adaptively and the ability to generalize from a certain pattern so that it is resistant to the existence of data errors due to noise [7].
2S
O
2 nL cos T
The working principle of peripheral fabric interferometer systems on optical fiber sensors can be shown in Fig. 2. The optical fiber sensor consists of two mirrors formed between the ends of optical fibers with chemical membranes and chemical membranes layers with air. Differences in the refractive index between two different materials, resulting in a continuous and reflected light wave. The two reflective waves of light will merge into one and interfere with each other. Interference does not occur when light is passed in low intensity [13]. The change in the thickness of the chemical membranes will result in the value of etalon (d). The light source is passed to the same optical fiber line as the reflected light path detected by the photodiode to detect the change in the resulting light intensity value. Rated power of the light intensity can be expressed in Eq. (2), § 4Sd · I I1 I 2 2 I1I 2 cos¨ ¸ © O ¹
A. Design of Optical Fiber Interferometer Interferometer is an instrument used to generate a coherent wave so that the interference occurs. Two waves are coherent if the phase difference remains. The phase difference is caused by the difference in path length and also because of the phase reversal as the wave is reflected by a denser medium. FabryPerot interferometer is an optical device to produce a certain frequency or wavelength (monochromatic). This interferometer utilizes multiple reflection effects as a wave intensity division mechanism. The intensity of the reflected and transmitted by
where I1 and I2 are the reflected light intensity values of reflected beams, d is the length of etalon, λ is the wavelength of the light source [13].
478
2018 International Conference on Information and Communications Technology (ICOIACT)
Chemical membrane SMF
Air VCC d (cavity)
U1A LM358AN
8
Fig. 2. Chemical membrane coated on optical fiber sensor
2 Sensor 1
The optical fiber interferometer is designed with three fiber-optic cables coated with different polarity types of chemical membranes which consist of polar, mid-polar and non-polar types, on the surface of optical fibers tip. Three types of chemical membranes coated on optical fibers tip are polyethylene glycol (PEG-1540), and phenylmethyldimethylpolysiloxane (OV-17), dimethylpolysiloxane (OV101). The vapor absorbed into the chemical membranes will produce a swelling process on chemical membranes and change the electron density of the chemical membranes chain.
R4
1
A
3
Out 1 R5 100k
1k R6
4
C2 10uF
R7 1k
100k
GND
Low Pass Filter
Non-inverting Amplifier
Fig. 4. Signal conditioning circuit consist of low pass filter and non-inverting amplifier
B. Design of Hardware System and Sensor Principle In this study, the hardware designed consists of the sensor data signal processing and the gas flow system. In Fig. 1, the air is flowed into the sensor chamber by using 3.3Vdc air pump. Dry air for the cleaning process of the sensor chamber is obtained by flowing air into the silica gel in order to reduce the air humidity level at room temperature. The process of cleaning and flowing the gas samples into the sensor chamber is controlled remotely from the computer through the microcontroller by opening and closing the valve using a solenoid valve activated by relay module. When cleaning process, valve 1 (cleaning) is opened, valve 2 (sample in) and valve 3 (sample out) is closed. Otherwise, when the sample gas in flowing process, valve 2 (sample in) and valve 3 (sample out) is opened and valve 1 (cleaning) is closed.
Digital data from A/D converter are transferred to microcontroller Arduino Nano through I2C serial interface with SCL and SDA pin. Data from each sensor received by microcontroller are sent to the computer through serial communication with baud rate 9600 bps to be processed on computer. The data signal received by the computer are processed and displayed the magnitude of changes in the value obtained from each sensor when flowed by gas samples. C. Identification System Using Neural Network The identification process of gas vapor gas applied to some gas samples using artificial neural network with error backpropagation algorithm. The neural network method used is trained to recognize the pattern by adjusting the weights, where a pattern is required to meet an output target [7]. Backpropagation algorithm consists of two passes in different layers on each network, i.e. forward pass and backward pass. During the forward pass process, the synaptic weights of the network have been determined. In the backward pass process, the synaptic weights are adjusted according to the errorcorrection rule with Mean Square Error (MSE) calculation by averaging the total squares of the target response difference with the actual response to obtain a mean error value. The value of this error is then transmitted back through the network, as opposed to the direction of the synaptic connection [14]. The synaptic weights will then be adjusted or changed to get an actual response value that is closer to the desired target response value. Backpropagation algorithm calculation process can be described in accordance with Eq. (3) to Eq. (9).
The sensor block was arranged by the optical fibers tip, 3dB couplers with coupling ratio 50:50, white light sources, and photodiodes to measure the intensity of light received from the optical fiber sensor, as shown in Fig. 3. The voltage signal received by the photodiode is then passed through the Low Pass Filter and amplified 100 times with a non-inverting amplifier, as shown in Fig. 4. The amplifier output signal is then converted into digital data via analog-to-digital (A/D) converter ADS1115 with 16 bits of resolution to measure voltage changes in mV. The ADS1115 has four selectable slave addresses. The ADS1115 can perform data conversion at rates up to 860 samples per second. PGA onboard available on the ADS1115 is able to provide a range of input from the supply to as low as ± 256mV.
The error value Eq is calculated by using formula in Eq. (3), Eq
1 2
3 ¦ d qh xout ,h n3
2
h 1
where d qh is desired output value and xout ,h is output of activation function in output neuron. If the value of E q ! E t arg et (not convergent), then the weights value are
Fig. 3. Sensor Principle
updated. The backpropagation process is stopped when
479
2018 International Conference on Information and Communications Technology (ICOIACT)
convergence is achieved. The updating of weight value w jis k 1 in k -iteration are calculated by formula in Eq. (4),
s 1 w jis k 1 w jis k P s G js x xout ,i
k
1
#1
P0
P
T>@
PEG-1540
W
X1
where P is learning rate value, defined in Eq. (5), and G j is portion of error correction weight adjustment.
OV-101
X2
For the output layer, portion of error correction weight adjustment G j can be defined in Eq. (6). Furthermore, value of
d
G s
§ s 1 s 1 s 1 · s ¨ ¦ G h whj ¸ g v j ¨ ¸ ©h 1 ¹
j
qh
s g v s xout j ,j
OV-17
1 s j
#4
O3
T>@
T>@
X3
Ethyl Acetate
O4
Acetone
Benzene
T>@
Methane
N-Hexane
#100
Input Layer
Hidden Layer
Output Layer
Input training data used are the result of the normalization of a set value voltage change gas samples at each time of testing. Neural network architecture consists of an input layer with three neurons, a hidden layer with 100 neurons, and an output layer with six neurons. So, there are 300 weights that connect the input layer with the hidden layer and 600 weights connecting the hidden layer with the output layer. Besides, the neural network architecture uses 100 biases in hidden neurons and 6 biases in output neuron.
Dv sj s j
II. RESULTS AND DISCUSSION A. Characteristics of chemical membranes coated on tip optical fiber sensor Initially, dry air is flowed into the gas chamber with speed of 0.1 LPM (liter per minute) until a stable value is achieved without any significant change in the value of the voltage. Furthermore, the sample gas is flowed into the gas chamber until a stable sensor value changes. First experiments were performed with two sensors, one sensor with chemical membranes, and other sensors without chemical membranes.
The error in the output neuron determines the error count in the previous hidden layer output used in the weight connection setting between input layer and hidden layer. The arrangement of two sets of weights from the two pairs of layers is recalculated in a looping process until it reaches the desired error limit [14]. The parameter of the learning speed (μ) scales the adjustment of the weights. The process of repeating to reach this desired error limit uses the MSE method of error calculation. Mean Square Error (MSE) values are evaluated after all sets of training data (n samples of sets data) are processed by formula in Eq. (11),
1 n ¦ Eq nq 1
O2 T>@
O6
where D is a stochastic variable sampled from a uniform distribution at training time.
MSE
#3 T>@
T>@
wv js
D >1 f v @>1 f v @
g v js
T>@
Fig. 5. Neural Network Architecture
wf v js
1 e
T>@
Ammonia
O5
Because it uses sigmoid bipolar activation function [15], define in Eq. (9), then the derivative of activation function can be defined in Eq. (10),
f v js
O1
T>@
where G j is derivative of activation function in neuron, defined in Eq. (8).
#2
g v js
T>@
#5
G j for the hidden layer can be defined as formula in Eq. (7),
G js
T>@
There are three gas vapor sample, i.e. acetone, n-hexane, ammonia, used in gas flow system to investigate the effect of chemical membranes coated on optical fiber tip. The output value of each sensor is normalized to determine the effect on each gas sample. The normalized difference between the sensor with chemical membrane and without chemical membrane showed that the sensor with chemical membrane indicates a value greater than the sensors that are not coated with chemical membranes.
The input value from the neural network system is obtained by normalizing the value of the voltage change obtained from each optical fiber sensor for each gas vapor sample.
480
2018 International Conference on Information and Communications Technology (ICOIACT)
Gas Flow System
Voltage Change, 'V (V)
0.020 0.018 0.015 0.013
Cleaning process
0.010 0.008
Gas sample flowed
Cleaning process
0.005 0.003 0.000 0
20
40
60
80
100
120
140
160
180
200
Time (s)
Signal Processing
OV-101
Fig. 6. Realization of experimental system
OV-17
PEG-1540
Fig. 7. Sequence of cleaning and flowing gas samples when the voltage changes on acetone vapor
The chemical membranes coated on the fiber optic sensor can increase the sensitivity of the sensor. The process of swelling in the chemical membrane can result in increasing value of voltage change. Acetone and ammonia vapor have a high response to the polymer PEG-1540, cause acetone and ammonia vapor are classified as polar compounds. B. The pattern of the optical fiber sensor output The next experiment was carried out using three sensors each coated with different chemical membranes. The measurement was performed in time 200 seconds. In Fig. 7, cleaning the gas chamber with dry air is carried out at 1-50 seconds, while at 51-80 seconds the sample vapor is flowed into the gas chamber, then re-cleaning the gas chamber at a later time.
Fig. 8. The pattern of voltage change to the vapors
In Fig. 8, it can be seen that the polymer PEG-1540 results a voltage change response greater than chemical membranes OV-101 and OV-17. Acetone, ammonia, and ethyl acetate vapor are polar and tend to bind to polar chemical membrane of PEG-1540. While benzene, methane, and n-hexane are nonpolar and also tends to bind to non-polar chemical membrane of OV-101. It can be seen that the chemical membranes have a good selectivity capability when exposed by organic solvent vapor.
The classification process uses artificial neural network algorithm with parameter of learning rate equal to 1.0 and sum square error of 1x10-4. In the learning phase, convergence is achieved when the iteration process reaches 171,126 epochs. Profile of sum square error at each iteration can be seen in Fig. 9. The parameter values used in Neural Network shown in Table I.
Each sensor shows different response patterns for each gas sample. The normalized results of each pattern of changes in sensor voltage values. In Fig. 8, the voltage change pattern can be generated from each of the different organic solvent vapors. Differences in the pattern of voltage change can be used to do the classification of organic solvent vapor type. The normalization result of characteristic pattern response from three different sensors is used as input on artificial neural network algorithm.
Testing is performed to determine the characteristics of the neural network. The test was performed by the number of neurons in a different hidden layer i.e. 10 neurons, 50 neurons, 100 neurons and with different error limits of 1x10-3 and 1x10-4, as shown in Table I. Test results show that the greater the number of neurons in the hidden layer then the number of iterations produced tend to be less. The smaller MSE will result in a more accurate identification process and an increasing number of iterations.
MSE
The learning phase of the neural network uses three input neurons representing three optical fiber sensors coated by PEG1540, OV-101, and OV-17 chemicals. 100 neurons were used in the hidden layer to improve the accuracy of learning on neural networks. While the output neurons used 6 neurons that represent the output of the vapor gas sample, i.e. acetone (C3H6O), ammonia (NH3), ethyl acetate (C4H8O2), benzene (C6H6), methane (CH4), n-hexane (C6H14), as shown in Fig. 5. The process of classifying the organic solvent vapor type with the backpropagation neural network algorithm, is done by training data in the form of normalization of the voltage changes of the three sensors at three times of measurement.
Iteration (epoch) Fig. 9. The error profile of neural network in training phase
481
2018 International Conference on Information and Communications Technology (ICOIACT)
TABLE I. MSE
PARAMETER OF NEURAL NETWORK
ACKNOWLEDGMENT This Research was carried out with financial aid support from the Ministry of Research, Technology and Higher Education of the Republic of Indonesia (Kemenristekdikti RI).
Parameter Neuron of Hidden Layer
Learning rate (μ)
Derivative of Activation Function (α)
Iteration (epoch)
10
1.0
1.0
86490
50
1.0
1.0
63342
100
1.0
1.0
63306
10
1.0
1.0
165636
50
1.0
1.0
150804
100
1.0
1.0
171126
0.001
0.0001
TABLE II.
REFERENCES [1]
[2]
[3]
TESTING OF NEURAL NETWORK WITH DIFFERENT ERROR MSE
Gas vapor sample Benzene Methane N-Hexane Amonia Ethyl Acetate Acetone
[4]
0.0001
0.001 Benzene
Benzene
Benzene
Benzene
Methane
Methane
Methane
Methane
N-Hexane
N-Hexane
N-Hexane
N-Hexane
Amonia
Amonia
Amonia
Amonia
Ethyl Acetate
Ethyl Acetate
Ethyl Acetate
Ethyl Acetate
Acetone
Acetone
Acetone
Acetone
[5]
[6]
[7]
[8]
[9]
The identification system test is done by taking each gas vapor sample with two samples, so there are totally 12 test input data. The result of the classification process using different error limit values, shown in Table II. Both with error limit value of 0.001 and 0.0001, it can be shown that the identification system can identify the gas vapor with 100% accuracy rate.
[10]
[11]
III. CONCLUSION
[12]
From the experiment, it can be shown that the optical fiber sensor is proven to increase the sensitivity when it is exposed to the sample vapor. Chemical membranes coated on optical fiber tip can be used to detect organic solvent vapor in which PEG-1540 has a higher sensitivity to polar organic solvent vapor, whereas OV-101 has a higher sensitivity to non-polar organic solvents. At the training phase of artificial neural network algorithm, it can be shown that the iteration process requires 171,126 epochs with the sum square error of 1x10-4. Vapor samples of acetone, ammonia, ethyl acetate, benzene, methane, n-hexane can be identified with 100% accuracy. Artificial neural network algorithms that have been designed, proven to make the process of classification with good results. Future work is focused on the identification of organic solvent vapors with different temperatures and humidities, in order to be able to recognize under different conditions.
[13]
[14]
[15]
482
M. Bae, J. Lim, S. Kim, Y. Song, “Ultra Highly Sensitive Optical Gas Sensors based on Chemomechanical Polymer-Incorporated Fiber Interferometer”, OPTIC EXPRESS, Vol. 21, No. 2, pp. 2018-2023, January 2013. C.L. Yuana, C.P. Chang, Y. Song, “Hazardous Industrial Gases Identified Using a Novel Polymer/MWNT Composite Resistance Sensor Array”, Materials Science and Engineering B, Vol. 176, Issue 11, pp. 821–829, June 2011. P. Kaushal, “Advances in Electronic Nose Technology for Clinical Applications”, International Journal of Engineering and Innovative Technology (IJEIT), Vol. 3, No. 8, hal. 130-135, February 2014. R.Y. Shah, Y.K. Agrawal, “Introduction to Fiber Optics: Sensors for Biomedical Applications”, Indian Journal of Pharmaceutical Sciences, Vol. 73, No. 1, pp. 17-22, January 2011. M. Rivai, T. Mujiono, H. Juwono, “Identification of Organic Solvent Vapors Using Polymer Coated SiO2 Crystal Array”, Conference Proceedings: The Fourth Saudi Technical Conference and Exhibition, Vol.III, pp. 244-249, December 2006. H. A Sujono, M. Rivai, “Vapor identification system using quartz resonator sensor array and support vector machine”, ARPN Journal of Engineering and Applied Sciences, Vol. 9, No. 12, pp. 2426-2430, December 2014. M. Rivai, T. Mujiono, Tasripan, “Chemically Coated Quartz Crystal Sensors for Fragrance Recognition”, International Journal of Scientific & Engineering Research, Vol. 3, Issue 6, pp. 637-641, June 2012. A. R. Indrapraja, M. Rivai, A. Arifin, D. Purwanto, “The detection of organic solvent vapor by using polymer coated chemocapacitor sensor”, International Conference on Physical Instrumentation and Advanced Materials, Vol. 853, pp. 241-245, May 2017. R.S. Gelais, G. Mackey, J. Saunders, J. Zhou, A.L. Hotte, A. Poulin, J.A. Barnes, H.P. Loock, R.S. Brown, Y.A. Peter, “Gas Sensing Using Polymer-Functionalized Deformable Fabry-Perot Interferometers”, Sensors and Actuators B: Chemical, Vol. 182, pp. 45-52, June 2013. Y. Shang, X. Wang, E. Xu, C. Tong, J. Wu, “Optical Ammonia Gas Sensor Based on a Porous Silicon Rugate Filter Coated with PolymerSupported Dye”, Analytica Chimica Acta, Vol. 685, Issue 1, pp. 58–64, January 2011. B.H. Lee, Y.H Kim., K.S. Park, J.B. Eom, M.J. Kim, B.S. Rho, H.Y. Choi, “Interferometric Fiber Optic Sensors”, Sensors, Vol. 12, No.3, pp. 2467-2486, February 2012. Marzuarman, M. Rivai, T. A. Sardjono, D. Purwanto, “Investigation of michelson interferometer for volatile organic compound sensor”, International Conference on Physical Instrumentation and Advanced Materials, Vol. 853, pp. 120-123, May 2017. X. Zhang, W. Peng, Z. Liu and Z. Gong, "Fiber Optic Liquid Level Sensor Based on Integration of Lever Principle and Optical Interferometry," IEEE Photonics Journal, Vol. 6, No. 2, pp. 1-7, April 2014. M. Rivai, E.L. Talakua, “The Implementation of Preconcentrator in Electronic Nose System to Identify Low Concentration of Vapors Using Neural Netowrk Method”, International Conference on Information, Communication Technology and System (ICTS), pp. 31-36, September 2014. H. Widyantara, M. Rivai, D. Purwanto, “Neural Network for Electronic Nose using Field Programmable Analog Arrays”, Iternational Journal of Electrical and Computer Engineering (IJECE), Vol. 2, No. 6, pp. 739747, December 2012.