Reliable e-nose for air toxicity monitoring by filter diagonalization method

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 2, April 2022, pp. 1286~1298
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i2.pp1286-1298  1286
Journal homepage: http://ijece.iaescore.com
Reliable e-nose for air toxicity monitoring by filter
diagonalization method
Ricardo Macías-Quijas1
, Ramiro Velázquez1
, Roberto de Fazio2
, Paolo Visconti2
,
Nicola Ivan Giannoccaro2
, Aimé Lay-Ekuakille2
1
Facultad de Ingeniería, Universidad Panamericana, Aguascalientes, Mexico
2
Department of Innovation Engineering, University of Salento, Lecce, Italy
Article Info ABSTRACT
Article history:
Received Jul 9, 2021
Revised Sep 14, 2021
Accepted Oct 10, 2021
This paper introduces a compact, affordable electronic nose (e-nose) device
devoted to detect the presence of toxic compounds that could affect human
health, such as carbon monoxide, combustible gas, hydrogen, methane, and
smoke, among others. Such artificial olfaction device consists of an array of
six metal oxide semiconductor (MOS) sensors and a computer-based
information system for signal acquisition, processing, and visualization. This
study further proposes the use of the filter diagonalization method (FDM) to
extract the spectral contents of the signals obtained from the sensors.
Preliminary results show that the prototype is functional and that the FDM
approach is suitable for a later classification stage. Example deployment
scenarios of the proposed e-nose include indoor facilities (buildings and
warehouses), compromised air quality places (mines and sanitary landfills),
public transportation, mobile robots, and wireless sensor networks.
Keywords:
Electronic nose (e-nose)
Filter diagonalization method
MOS gas sensor
Sensor characterization
Toxic compounds
This is an open access article under the CC BY-SA license.
Corresponding Author:
Ramiro Velázquez
Facultad de Ingeniería, Universidad Panamericana
Josemaría Escrivá de Balaguer 101, Aguascalientes, 20290, Mexico
Email: rvelazquez@up.edu.mx
1. INTRODUCTION
An electronic nose (e-nose) is a sensing instrument comprising a set of different gas sensors that
react when exposed to a wide range of chemical particles. According to the outputs obtained from the
sensors, many conclusions about air quality and toxicity can be obtained. They have shown great promise and
utility in three domains: food control, disease diagnosis, and environmental monitoring. In food control, e-
noses have been devoted toward assuring quality and safety for consumers. Some features already addressed
in this domain are food freshness, ageing, contamination during processing, shelf life, and authenticity
confirmation. Viejo et al. [1] implemented an e-nose to assess the aroma profiles in beer and automate the
industrial quality inspection process. Similarly, Radi et al. [2] developed an e-nose for classifying odors from
synthetic flavors such as grapes, strawberry, mango, and orange. To alert about rancidity, Xu et al. [3]
introduced an e-nose monitoring the changes of pecans during storage. Timsorn and Wongchoosuk [4]
explored the use of an e-nose device to identify odors from formalin contamination in seafood. To prevent
products’ adulteration, Świgło and Chmielewski [5] proposed an e-nose to assist in the authenticity testing of
products such as meat, honey, milk, and plant oils. To discourage meat dealers from committing food fraud,
Laga and Sarno [6] presented an e-nose discriminating pork from beef. Wang et al. [7] deployed an e-nose
inside a domestic refrigerator to assess the food freshness level of fruits, vegetables, and meat.
The ability of humans to detect diseases with smell has played a significant role in clinical
diagnosis. E-noses can ease the detection of volatile organic compounds (VOC) exhibiting bacterial

Int J Elec & Comp Eng ISSN: 2088-8708 
Reliable e-nose for air toxicity monitoring by filter diagonalization method (Ricardo Macías-Quijas)
1287
pathogens and have the potential to become a valuable disease diagnosis tool for humans, animals, and
plants. Sanchez et al. [8] monitored exhaled human breath with an e-nose to help diagnose and monitor
certain digestive and respiratory diseases. Siyang et al. [9] analyzed the odor of human urine samples with an
e-nose to detect diseases such as diabetes. In response to the current COVID-19 pandemic, airbus is
deploying e-nose devices in its aircrafts to detect the virus [10]. In veterinary medicine, Jia et al. [11]
explored an e-nose device for detecting three types of wound infections in rats. To detect bovis-infected
cattle in farms, Peled et al. [12] reported the use of an e-nose to analyze VOC in breath samples. In botanics,
Baietto et al. [13] explored the use of an e-nose to sense the bole-rot fungi that affects trees. Wilson [14]
developed and tested an e-nose for the rapid identification of insecticide residues in crops.
Nowadays, e-noses in environmental monitoring have found application in four fields: i) air quality,
ii) water quality, iii) process control, and iv) odor control systems [15]. Some representative work include
Wongchoosuk’s WiFi e-nose sensing and quantifying indoor air contaminants even in very low
concentrations [16]. The e-nose prototype proposed by Mishra devoted to identify poison gases emanated
from waste [17]. The work of Baby reporting on the use of an e-nose to sense contaminating residues and
pesticides in water [18]. The system in [19] using different gas sensors, monitors gas concentration and
temperature in a biogas reactor. The smart system proposed by [20] was capable of sensing ammonium
nitrate that could lead to fire and explosion in storage warehouses. Applications of e-noses outside these three
main domains also involve explosive detection [21] and the space industry [22].
Within the context of air quality assessment, this paper presents the characterization and
implementation of a novel, affordable e-nose device. The prototype consists of a set of six metal oxide
semiconductor (MOS) gas sensors capable of detecting, i) combustible gas, ii) alcohol, iii) methane,
iv) carbon monoxide, v) hydrogen, and vi) smoke, which might represent a threat for the human health. The
implemented device has small dimensions allowing for seamless integration into mobile robots, indoor
facilities, urban transport, risky environments (such as mines and sanitary landfills), or sensor networks
monitoring broad surface areas.
Similar e-nose prototypes exploring MOS sensors can be found in the literature [1], [16], [23]–[25];
sensor arrays range from four to ten, depending on the application. Prototypes rely on MOS technology
because it offers small-size and robust sensors, quite good sensitivity, simple signal processing, commercial
availability, and low cost. The main difference across devices is the method for data processing. Approaches
such as artificial neural networks (ANN), principal component analysis (PCA), Fourier transform, and
wavelets have been explored with satisfactory results for further developing predictive models.
In this paper, the sensors’ signals were spectrally analyzed using the filter diagonalization method
(FDM). To our knowledge, this work is the first one that reports on the use of FDM for e-nose signal
processing. The remainder of this work paper is structured as follows: section 2 introduces the e-nose device,
the MOS sensors used, the main system components, the FDM, and its implementation in the prototype. In
section 3, the experiments that have been carried out are described and the results are shown. Finally, the
concluding remarks and the future work perspectives are given in section 4.
2. RESEARCH METHOD
An operational prototype was implemented with low-cost materials, and it is the first approach for
the research work. The proposed device involves six gas sensors, each one focusing on a specific gas. They
capture the components present in the air for later off-line analysis. The device comprises an electronic
module for data acquisition and a software to visualize the sensors’ behavior. To ensure an adequate gas
concentration around the e-nose, two items have also been considered: an air pump conveying the samples
and a hermetic box enclosing the device.
2.1. Experimental apparatus
A 3D printed plastic base was designed to host the set of sensors. It is small (6.5x5x7 cm) and
displays six perforations around its side-faces for the proper installation of the MOS sensors as shown in
Figure 1(a). All electronic boards and electrical connections are placed inside the plastic base allowing a
clean and safe device handling as shown in Figure 1(b).
The prototype employs six different types of MOS sensors (Hanwei Electronics Co., Ltd., Henan,
China), which change their electric resistance when its sensing material comes in contact with the gas.
Figure 2 illustrates the MOS sensors external and internal structures. The external structure as shown in
Figure 2(a) involves a mesh-like enclosure that protects the sensing material and filters out the suspended
particles on the air, so that only gaseous elements access into the chamber. The clamping ring secures the
mesh and serves as the base for the sensing material and the electrical connections.
Internally, the sensor is composed of a tin dioxide (SnO2) layer, which is the sensing material that
reacts to the input gas. In clean air, no electric current flows through this layer, but when gas is detected,

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 2, April 2022: 1286-1298
1288
electrons are released, allowing current to flow. The current output signal is conveyed via platinum (Pt) wires
as shown in Figure 2(b). Using a simple voltage-divider configuration, the gas concentrations can be derived.
Each of the six MOS sensors comprised in the prototype is sensitive to a specific gas providing valuable data
upon the analysis of their responses. We previously reported a complete experimental characterization of the
MOS sensors [26]. Characteristics such as sensitivity, behavior to temperature, and step response were
examined. Figure 3 presents, as an example, the behavior obtained from the alcohol (ethanol) gas sensor (S2).
Figure 3(a) shows its resistance ratio Rs/Ro; here, Rs represents the sensor resistance to the target gas
(C2H5OH) given a specific concentration, while Ro represents the sensor resistance in clean ambient air. Note
that the resistance ratio Rs/Ro decreases as the concentration of the gas increases. Figure 3(b) shows how the
sensitivity curve of the sensor is affected by temperature. Here, Rso represents the resistance of the sensor in
125 ppm (parts-per-million) alcohol at 20 °C. Note that the resistance ratio Rs/Rso decreases as
environmental temperature increases. Figure 3(c) shows the sensor response to a step input of 125 ppm
alcohol gas. Note that S2 delivers 3.5 V in the presence of the gas and 0.5 V (the baseline) in the absence of gas.
(a) (b)
Figure 1. The plastic base structure: (a) 3D design and (b) actual prototype
(a) (b)
Figure 2. MOS sensors: (a) external and (b) internal structures
Table 1 lists the MOS sensors used, their model number, and their target detection gases. Note that
the sensors are sensitive to a primary gas but also react to secondary gases, resulting in an overlap of their
responses as shown in Figure 3(a) and Table 1. This overlap is advantageous since it provides more
information about the chemical compounds contained in the air sample, thus facilitating the signal
classification stage.
The prototype also comprises an electronic module which encompasses a 32-bit microcontroller
(Texas instruments Stellaris LM4F120) responsible for the sensors’ signal acquisition, universal serial bus
(USB) interfacing with a computer, and the external air pump control. Figure 4(a) shows the block-level
diagram of the electronic module while Figure 4(b) shows the electrical diagram for the air pump control. To
synchronize the start and end of signal collection and transfer the acquired signals for processing and
visualization, a simple communication protocol was built between the electronic module and the computer.
Figure 5 shows the activity diagram of the system operation. The computer initiates the process
which lasts 180 s. During this time interval, the sensor array collects data from its vicinity. At t=20 s, the
pump starts introducing air into the hermetic box. At t=50 s, the pump is set off; the sensors reset (i.e., signals
start going to their baseline). The electronic module captures this behavior as well.

1289
Figure 6 illustrates all six sensors’ responses detailing the different stages. Note that the response of
S2 (alcohol gas sensor) is the slowest one both in its rising and recovery times. This result is due to the low
volatility that alcohol exhibits compared to the other substances explored.
(a) (b) (c)
Figure 3. Sensor characterization example for S2 (alcohol sensor), (a) sensitivity curve, (b) behaviour to
temperature, and (c) step response
Table 1. Specifications of the MOS sensors in the e-nose
Sensor Hanwei Item Primary Gas Secondary Gases
S1 MQ2 Combustible gas H2, LPG, CH4, CO, Alcohol, Propane
S2 MQ3 Alcohol CO, H2
S3 MQ4 Methane Propane, Butane, Alcohol
S4 MQ7 Carbon monoxide H2, LPG, CH4
S5 MQ8 Hydrogen CO
S6 MQ135 Air quality NH3, NOx, Alcohol, Benzene, Smoke, CO2
(a) (b)
Figure 4. E-nose system: (a) block-level diagram and (b) electrical diagram of the air pump control
Figure 5. The activity diagram for the e-nose software

 ISSN: 2088-8708
1290
Figure 6. The sensors’ response during the acquisition process
Once the 180 s acquisition process has concluded, the microcontroller gathers the sensors readings,
converts the continuous data into digital form, and builds a data frame which is sent to the computer via
USB. The computer further filters the samples with a ten-tap moving average filter to eliminate the
undesirable peaks in the sensors’ responses. To obtain reliable plots with comparable maximum and
minimum values, the sensors’ individual baselines were determined and eliminated using (1):
𝑦𝑖 = 𝑦𝑖−𝑦𝑖0 (1)
where yi represents the output of the ith
-sensor and yi0 its baseline.
The final prototype is shown in Figure 7(a). Note that the microcontroller is fitted in the plastic base.
The necessary connections between the sensor array and the electronic module are located inside. Figure 7(b)
shows the implemented system. Inside the hermetic box, the e-nose (plastic base, sensor array, and electronic
module) can be seen. Outside it, the air pump conveying the air samples can be appreciated. The air pump
control circuitry and the power supply are also shown. The final prototype exhibits compact dimensions as
shown in Figure 1(a); low mass (500 g), and low cost (150 USD).
2.2. FDM analysis
The filter diagonalization method (FDM) was initially developed for quantum dynamics calculations
[27] and later used for nuclear magnetic resonance (NMR) signal processing [28], and leak detection in
pipelines [29], [30]. It provides a nonlinear parametric method for time-domain signal analysis using a sum
of damped sinusoids. FDM is traditionally used to solve harmonic inversion problems (HIP), delivering high-
resolution spectra. Compared to Fourier analysis, it is not limited by the incertitude principle, thus providing
high-quality spectra together with a high signal to noise ratio (SNR) without needing a large number of
samples.
(a) (b)
Figure 7. First prototype is (a) control unit and sensors in the plastic base and (b) the prototype with the
external air pump and the hermetic box

1291
2.2.1. FDM formulation
To explain the method, let us consider a complex unidimensional signal cn=c(nτ) with values along
with equidistant time intervals nτ with n = 0, 1,…, N-1. The FDM aims to represent cn as a sum of damped
sinusoids, as shown in (2), where 𝜔𝑘 = 2𝜋𝑓𝑘 − 𝑗𝛾𝑘 are the complex frequencies of the signal, including the
damping factor, and dk are the corresponding amplitudes. To solve (1), the FDM associates a correlation
function described by the Hamiltonian operator 𝛺
̂, which has complex eigenvalues {𝜔𝑘}, thus cn can be
further transformed into (3):
𝑐𝑛 = ∑ 𝑑𝑘𝑒−𝑗𝑛𝜏𝜔𝑘
𝐾
𝑘=1 (2)
𝑐𝑛 = (𝛷0|𝑒−𝑗𝑛𝜏𝛺
𝛷0) (3)
The problem can be simplified to the diagonalization of the Hamiltonian operator 𝛺
̂ or, as discussed
in [28], to the evolution operator 𝑈
̂ = 𝑒−𝑗𝜏𝛺
. Briefly, a symmetric internal product operator defined by
(a|b)=(b|a) without the conjugate complex is used, where 𝛷0 is the initial state. Assuming that an orthonormal
eigenvector set 𝛾𝑘 is used to perform diagonalization of the evolution operator as shown in (4):
𝑈
̂ = ∑ 𝑢𝑘
𝑘
|𝑌𝑘)(𝑌𝑘| = ∑ 𝑒−𝑗𝜔𝑘𝜏|𝑌𝑘)(𝑌𝑘|
𝑘 (4)
and substituting (4) in (3), yields (5):
𝑑𝑘 = (𝛷0|𝑌𝑘)(𝑌𝑘|𝜓0) = (𝑌𝑘|𝛷0)2
(5)
The resulting eigenvalues determine the position and width of the harmonics while the eigenvectors
define their amplitudes and phases. Assuming a set created from the Krylov vectors, generated by the
evolution operator 𝛷𝑛 = 𝑈͡𝑛
𝛷0 = 𝑒−𝑗𝑛𝜏𝛺
̂
𝛷0 and according to (5), it yields:
(𝛷𝑛|𝑈
̂𝛷𝑚) = (𝛷𝑛|𝛷𝑚+1) = 𝑐𝑚+𝑛+1 (6)
as the set in non-orthonormal, the overlapping matrix can be calculated according to (7):
(𝛷𝑛|𝛷𝑚) = (𝑈
̂𝑛
𝛷0|𝑈͡𝑚
𝛷0) = (𝛷0|𝑈
̂𝑚+𝑛
𝛷0) = 𝑐𝑚+𝑛+1 (7)
In (7) is strictly related to the values of the measured signal. Notation U0
can then be used, being
this the overlapping matrix representation of dimensions M+1M+1. Similarly, U1
can be used for 𝑈
̂. To
reformulate (2), it is then necessary to solve the generalized eigenvalues problem as shown in (8):
𝑈1
𝐵𝑘 = 𝑢𝑘𝑈0
𝐵𝑘 (8)
where 𝑢𝑘 = 𝑒−𝑗𝑛𝜔𝑘𝜏
contains the lines of the spectrum and its corresponding widths. Eigenvectors 𝐵𝑘
contain both amplitudes and phases.
2.2.2. FDM implementation for the e-nose
With the aim of analyzing the sensors’ experimental data, the FDM was implemented. The sensors’
readings were used as inputs cn, and the FDM was used to estimate their spectra. The following steps were
performed by the algorithm:
− Taking into account the Nyquist criterion, the frequency interval [fmin fmax] in which the spectral analysis
of signal cn will be performed is selected; cn is sampled at 𝑓𝑠 =
1
𝜏
− An angular frequency equidistant axis with values 2πfmin<φj<2πfmax, j=0, 1, 2,…, Kwin is created. Value
Kwin is chosen as 𝐾𝑤𝑖𝑛 =
𝑁(𝑓𝑚𝑎𝑥−𝑓𝑚𝑖𝑛)
2𝜏
as suggested in the literature
− Three symmetric complex matrices U(p)
of dimensions Kwin x Kwin, with p=0, 1, 2 are determined. To
calculate the elements that do not belong to the diagonal, (9) can be used, where fp and gp are the Fourier
transforms of the first and second part of signal cn [31], detailed in (10).

 ISSN: 2088-8708
1292
𝑈(𝑝)
(𝜑, 𝜑′) =
𝑒𝑗𝜑
𝑓𝑝(𝜑′) − 𝑒𝑗𝜑′
𝑓𝑝(𝜑) + 𝑒𝑗𝑀𝜑′
𝑔𝑝(𝜑) + 𝑒𝑗𝑀𝜑
𝑔𝑝(𝜑′)
𝑒−𝑗𝜑 − 𝑒−𝑗𝜑′
(9)
𝑓𝑝(φ) = ∑ 𝑒𝑗𝑛φ
𝑐𝑛+𝑝
𝑀
𝑛=0
𝑔𝑝(φ) = ∑ 𝑒𝑗(𝑛−𝑚−1)φ
𝑐𝑛+𝑝
2𝑀
𝑛=𝑀+1 (10)
In (11) is used to calculate the elements located in the diagonal:
𝑈(𝑝)
(φ, φ′) = ∑(𝑀 + 1 − |𝑚 − 𝑛|)𝑒𝑗𝑛φ
2𝑀
𝑛=0 (11)
− Solve the generalized eigenvalues problem with (8), where the eigenvalues and eigenvectors are
calculated using the QZ factorization algorithm [32].
− Select the complex amplitudes dk using (12):
𝑑𝑘
1/2
= ∑ 𝑩𝑗𝑘
𝐾𝑤𝑖𝑛
𝑗=1
∑ 𝑐𝑛𝑒𝑗𝑛φ𝑗
𝑀
𝑛=0 (12)
− Use values ωk and dk to estimate the spectrum with (13). Figure 8 shows a flowchart summarizing the
FDM algorithm implementation. The following section will verify its performance and suitability for e-
nose data processing.
𝐶(𝐹) = − ∑ 𝐼𝑚 {
𝑑𝑘
2𝜋𝐹 − 𝜔𝑘
}
𝑘 (13)
Figure 8. FDM implementation algorithm for the e-nose
3. RESULTS AND DISCUSSION
3.1. E-nose data processing with FMD
A series of experiments were designed using common gases to test the performance of the device,
obtaining some preliminary results that are presented in this Section. The environmental conditions registered
during the experiments were a temperature of 23 ºC  2 ºC and relative humidity (RH) of 30%. For each
experiment, the e-nose was subjected to a gas sample, and two features were obtained: i) the sensors response
in the time domain consisted of a series of 180 samples for each of the six sensors; ii) the frequency spectrum
for each sensor response was calculated using the FDM described in section 2.2. The spectrum of a sensor is
given by (14):
𝑌𝑛(𝑓) = ∑ 𝑎𝑘,𝑛𝛿(𝑓 − 𝑓𝑘,𝑛)
𝑘 (14)
where fk and ak are the frequencies and amplitudes belonging to the kth
-harmonic found in the response of the
nth
-sensor, respectively. The system response to clean air was first verified. As expected, the sensor set

1293
showed constant low magnitude output values. The FDM-based spectrum revealed low amplitude peaks and
frequencies with inexistent harmonics. It can be therefore concluded that the sensors’ response to clean air is
negligible. Figure 9 shows the e-nose response to acetone (C3H6O). Note in Figure 9(a) how the sensors
respond to the stimulus represented by the dotted line, especially sensors S4 and S2 followed by S6, this last
indicating the air quality. In addition, a spectrum containing harmonics of relevant magnitude in a narrow
frequency range can be observed as shown in Figure 9(b). Table 2 summarizes the frequency (Sn_F) and
amplitude (Sn_A) values found by the FDM algorithm.
(a)
(b)
Figure 9. The e-nose response to acetone on (a) time response and (b) FDM-based spectrum
Figures 10(a) and (b) show the time response and FDM-based spectrum to ethanol at 71.5%,
respectively. As shown in Figure 10(a), sensors S2 and S4 exhibit the most significant responses; while S2
shows the highest amplitude value, S4 shows the fastest rise time. Table 3 lists the harmonic values found by
the FDM algorithm. Note the correspondence with Figure 10(b), values obtained from S2 and S4 outstand
from the rest of the sensors.

 ISSN: 2088-8708
1294
Table 2. Acetone: amplitude (Sn_A) and frequency (Sn_F) values got for each sensor reading using FDM
S1_A S1_F S2_A S2_F S3_A S3_F S4_A S4_F S5_A S5_F S6_A S6_F
0.039 0.464 27.466 0.001 0.063 0.176 0.012 0.433 2.859 0.003 4.580 0.014
7.571 0.003 28.083 0.006 0.311 0.005 0.016 0.410 1.024 0.015 1.061 0.030
2.132 0.016 5.256 0.028 0.154 0.017 10.011 0 0.178 0.031 0.542 0.051
0.545 0.037 1.195 0.045 0.025 0.030 4.978 0.012 0.166 0.048 0.584 0.063
0.539 0.051 0.322 0.060 0.011 0.051 0.295 0.029 0.015 0.065 0.373 0.079
0.041 0.072 0.110 0.082 0.761 0.042 0.027 0.082 0.206 0.090
0.031 0.082 0.071 0.105 0.058 0.073 0.019 0.116 0.060 0.132
0.011 0.121 0.041 0.122 0.020 0.087 0.016 0.151 0.201 0.153
0.022 0.229 0.015 0.168
0.015 0.197
0.034 0.134
0.014 0.172
0.026 0.167
(a)
(b)
Figure 10. The e-nose response to ethanol on (a) time response and (b) FDM-based spectrum

1295
Table 3. Ethanol: amplitude (Sn_A) and frequency (Sn_F) values got for each sensor reading using FDM
0.012 0.423 0.468 0.002 0.048 0.002 3.146 0.003 0.189 0.007 0.428 0.004
1.852 0.005 0.393 0.011 0.025 0.012 1.487 0.017 0.109 0.023 0.248 0.017
1.135 0.019 0.187 0.023 0.531 0.032 0.216 0.031 0.083 0.031
0.421 0.032 0.071 0.037 0.489 0.047 0.029 0.042 0.024 0.048
0.203 0.049 0.021 0.050 0.178 0.063 0.010 0.059 0.012 0.066
0.071 0.067 0.071 0.079
0.017 0.085 0.018 0.118
0.108 0.182 0.032 0.129
0.023 0.142 0.013 0.151
0.015 0.161
Finally, Figures 11(a) and (b) show the time response and FDM-based spectrum to gas butane
(C4H10), respectively. Note in Figure 11(a) that S4 exhibits the highest amplitude value. Table 4 shows the
harmonics found by the FDM analysis. Note the correspondence with Figure 11(b), values obtained from S4
outstand from the rest of the sensors.
(a)
(b)
Figure 11. The e-nose response to gas butane on (a) time response and (b) FDM-based spectrum

 ISSN: 2088-8708
1296
Table 4. Gas butane: amplitude (Sn_A) and frequency (Sn_F) values got for each sensor reading using FDM
0.060 0.011 0.131 0.003 1.760 0.006 3.603 0.002 2.782 0.004 2.388 0.004
0.467 0.008 0.098 0.015 1.136 0.017 1.191 0.013 2.049 0.019 1.322 0.017
0.191 0.021 0.033 0.029 0.632 0.021 0.333 0.034 1.230 0.033 0.581 0.032
0.017 0.046 0.012 0.197 0.175 0.045 0.056 0.057 0.677 0.046 0.496 0.046
0.036 0.069 0.014 0.074 0.095 0.070 0.140 0.060
0.015 0.115 0.016 0.240
0.010 0.247 0.014 0.134
3.2. Comparison with Fourier analysis
The discrete Fourier transform (DFT) is the most commonly used method to solve the harmonic
inversion problem (HIP) using the reliable and efficient fast Fourier transform (FFT) algorithm. However, the
FFT is sensitive to time-frequency uncertainties which limit the resolution of the resulting spectrum. As the
FFT spectrum resolution depends on the number of processed samples, the use of excessively large data sets
is often necessary. In contrast, the FDM is a parametric method that, upon the use of linear algebra, extracts
the parameters relevant for the construction of the signal spectrum. The FDM outperforms the FFT as it
requires a lower amount of data to build the spectrum and does not restrain the spectrum resolution with
uncertainties.
Figure 12 compares the FDM and FFT spectra for the acetone sample. The comparison is limited to
the signals of sensors S2 and S4, which show the clearest response to acetone. Note that the FDM produces
more accurate high-resolution spectra with well-defined peaks. Harmonics are well separated among them,
clearly showing the signal frequencies in the spectrum. In contrast, the FFT produces wide peaks that tend to
merge between them; this low-resolution effect might lose or miss information on the spectrum harmonics.
Note that the FDM-based spectrum exhibits a higher number of harmonics represented by well-defined
peaks. It can be therefore concluded that the FDM-based spectrum is a clearer and more accurate
representation for the e-nose signal processing.
Current work focuses on using these spectra to perform gas classification. Approaches such as
random forest [33] and convolutional neural networks (CNN) [34] are currently being considered. Future
work will evaluate the possibility of optimizing the number of sensors while keeping satisfactory results and
migrating the electronic unit to an field-programmable gate array (FPGA) system on chip (SoC) architecture
[35]. Applications in wireless sensor networks (WSN) for domotics/inmotics [36] and automated farming
[37] are foreseen.
Figure 12. Performance comparison between FDM and FFT (180 samples)
4. CONCLUSION
This research work has described a self-developed e-nose device implemented with low-cost
components. It is devoted to detect potentially harmful volatile compounds in the environment. The prototype
involves an array of six metal oxide semiconductor (MOS) sensors, an electronic module for data acquisition,
and a computer-based information system for signal analysis and visualization. MOS sensors offer small size,
low power consumption, fast response, and recovery times. Their response can be divided in three time-
regions: the reference value (baseline), the rising time, and the resetting time.
The hardware design and testing stages are now complete ensuring that the requirements have been
fulfilled and that the prototype is functional. The filter diagonalization method (FDM) was implemented to
calculate the harmonics involved in the acquired signals. The algorithm showed high precision using a low

1297
number of samples. To our knowledge, FDM has not been previously explored for e-nose data processing
and has the potential to become a valuable tool for this application.
REFERENCES
[1] C. G. Viejo, S. Fuentes, A. Godbole, B. Widdicombe, and R. R. Unnithan, “Development of a low-cost e-nose to assess aroma
profiles: An artificial intelligence application to assess beer quality,” Sensors and Actuators, B: Chemical, vol. 308, Apr. 2020,
Art. no. 127688, doi: 10.1016/j.snb.2020.127688.
[2] B. Radi, D. N. Rohmah, E. Wahyudi, M. D. Adhityamurti, and J. P. L. Yuroto Putro, “Implementation of an electronic nose for
classification of synthetic flavors,” Bulletin of Electrical Engineering and Informatics (BEEI), vol. 10, no. 3, pp. 1283–1290, Jun.
2021, doi: 10.11591/eei.v10i3.3018.
[3] K. Xu, J. Wang, Z. Wei, F. Deng, Y. Wang, and S. Cheng, “An optimization of the MOS electronic nose sensor array for the
detection of Chinese pecan quality,” Journal of Food Engineering, vol. 203, pp. 25–31, Jun. 2017, doi:
10.1016/j.jfoodeng.2017.01.023.
[4] K. Timsorn and C. Wongchoosuk, “Inkjet printing of room-temperature gas sensors for identification of formalin contamination
in squids,” Journal of Materials Science: Materials in Electronics, vol. 30, no. 5, pp. 4782–4791, Jan. 2019, doi: 10.1007/s10854-
019-00772-9.
[5] A. Gliszczyńska-Świgło and J. Chmielewski, “Electronic nose as a tool for monitoring the authenticity of food. a review,” Food
Analytical Methods, vol. 10, no. 6, pp. 1800–1816, Dec. 2017, doi: 10.1007/s12161-016-0739-4.
[6] S. A. Laga and R. Sarno, “Temperature effect of electronic nose sampling for classifying mixture of beef and pork,” Indonesian
Journal of Electrical Engineering and Computer Science (IJEECS), vol. 19, no. 3, pp. 1626–1634, Sep. 2020, doi:
10.11591/ijeecs.v19.i3.pp1626-1634.
[7] M. Wang et al., “Real-time assessment of food freshness in refrigerators based on a miniaturized electronic nose,” Analytical
Methods, vol. 10, no. 39, pp. 4741–4749, 2018, doi: 10.1039/c8ay01242c.
[8] C. Sánchez, J. P. Santos, and J. Lozano, “Use of electronic noses for diagnosis of digestive and respiratory diseases through the
breath,” Biosensors, vol. 9, no. 1, p. 35, Feb. 2019, doi: 10.3390/bios9010035.
[9] S. Siyang, C. Wongchoosuk, and T. Kerdcharoen, “Diabetes diagnosis by direct measurement from urine odor using electronic
nose,” in The 5th 2012 Biomedical Engineering International Conference, Dec. 2012, pp. 1–4, doi:
10.1109/BMEiCon.2012.6465441.
[10] M. Taylor, “Airbus wants its bomb-detecting ‘E-nose’ to sense COVID-19.” laboratoryequipment.com.
https://www.laboratoryequipment.com/563979-Airbus-Wants-its-Bomb-detecting-E-nose-to-Sense-Viral-Diseases-Too/ (accessed
Nov. 16, 2021).
[11] P. Jia, F. Tian, Q. He, S. Fan, J. Liu, and S. X. Yang, “Feature extraction of wound infection data for electronic nose based on a
novel weighted KPCA,” Sensors and Actuators, B: Chemical, vol. 201, pp. 555–566, Oct. 2014, doi: 10.1016/j.snb.2014.05.025.
[12] N. Peled et al., “Detection of volatile organic compounds in cattle naturally infected with Mycobacterium bovis,” Sensors and
Actuators, B: Chemical, vol. 171–172, pp. 588–594, Aug. 2012, doi: 10.1016/j.snb.2012.05.038.
[13] M. Baietto, A. D. Wilson, D. Bassi, and F. Ferrini, “Evaluation of three electronic noses for detecting incipient wood decay,”
Sensors, vol. 10, no. 2, pp. 1062–1092, Jan. 2010, doi: 10.3390/s100201062.
[14] A. D. Wilson, “Identification of insecticide residues with a conducting-polymer electronic nose,” Chemical Sensors, vol. 4, no. 3,
pp. 1–10, 2014.
[15] L. Capelli, S. Sironi, and R. Del Rosso, “Electronic noses for environmental monitoring applications,” Sensors (Switzerland),
vol. 14, no. 11, pp. 19979–20007, Oct. 2014, doi: 10.3390/s141119979.
[16] C. Wongchoosuk, C. Khunarak, M. Lutz, and T. Kerdcharoen, “WiFi electronic nose for indoor air monitoring,” in 2012 9th
International Conference on Electrical Engineering/Electronics, Computer, Telecommunications and Information Technology,
May 2012, pp. 1–4, doi: 10.1109/ECTICon.2012.6254166.
[17] S. O. Mishra and S. H. Saeed, “Optimization of electronic sensors for detecting pollution due to organic gases using PARAFAC,”
International Journal of Electrical and Computer Engineering (IJECE), vol. 9, no. 5, pp. 3441–3449, Oct. 2019, doi:
10.11591/ijece.v9i5.pp3441-3449.
[18] R. E. Baby, M. Cabezas, and E. N. W. De Reca, “Electronic nose: A useful tool for monitoring environmental contamination,”
Sensors and Actuators, B: Chemical, vol. 69, no. 3, pp. 214–218, Oct. 2000, doi: 10.1016/S0925-4005(00)00491-3.
[19] I. Iswanto, A. Ma’arif, B. Kebenaran, and P. Megantoro, “Design of gas concentration measurement and monitoring system for
biogas power plant,” Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 22, no. 2, pp. 726–732,
May 2021, doi: 10.11591/ijeecs.v22.i2.pp726-732.
[20] M. K. Ibrahim, N. M. Hussien, and S. N. Alsaad, “Smart system for monitoring ammonium nitrate storage warehouse,”
Indonesian Journal of Electrical Engineering and Computer Science (IJEECS), vol. 23, no. 1, pp. 583–589, Jul. 2021, doi:
10.11591/ijeecs.v23.i1.pp583-589.
[21] L. Wang, Y. H. Gui, S. P. Zhang, and C. S. Xie, “Research on explosives detection by electronic nose,” Chinese Journal of
Sensors and Actuators, vol. 20, no. 1, pp. 42–45, 2007.
[22] R. C. Young, W. J. Buttner, B. R. Linnell, and R. Ramesham, “Electronic nose for space program applications,” Sensors and
Actuators, B: Chemical, vol. 93, no. 1–3, pp. 7–16, Aug. 2003, doi: 10.1016/S0925-4005(03)00338-1.
[23] B. Gunawan, S. Alfarisi, G. Satrio, A. Sudarmaji, M. Malvin, and K. Krisyarangga, “MOS gas sensor of meat freshness analysis
on E-nose,” TELKOMNIKA (Telecommunication Computing Electronics and Control), vol. 17, no. 2, pp. 771–780, Apr. 2019,
doi: 10.12928/telkomnika.v17i2.11787.
[24] D. R. Wijaya, R. Sarno, and E. Zulaika, “Gas concentration analysis of resistive gas sensor array,” in 2016 International
Symposium on Electronics and Smart Devices, ISESD 2016, Nov. 2017, pp. 337–342, doi: 10.1109/ISESD.2016.7886744.
[25] M. Gancarz et al., “Application of electronic nose with MOS sensors to prediction of rapeseed quality,” Measurement: Journal of
the International Measurement Confederation, vol. 103, pp. 227–234, Jun. 2017, doi: 10.1016/j.measurement.2017.02.042.
[26] R. Macias-Quijas, R. Velazquez, N. I. Giannoccaro, and A. Lay-Ekuakille, “A novel electronic nose instrument for the detection
of volatile hazardous compounds: preliminary results,” in 2021 7th International Conference on Control, Instrumentation and
Automation (ICCIA), Feb. 2021, pp. 1–5, doi: 10.1109/ICCIA52082.2021.9403595.
[27] M. R. Wall and D. Neuhauser, “Extraction, through filter-diagonalization, of general quantum eigenvalues or classical normal
mode frequencies from a small number of residues or a short-time segment of a signal. I. Theory and application to a quantum-
dynamics model,” The Journal of Chemical Physics, vol. 102, no. 20, pp. 8011–8022, May 1995, doi: 10.1063/1.468999.

 ISSN: 2088-8708
1298
[28] V. A. Mandelshtam, “FDM: The filter diagonalization method for data processing in NMR experiments,” Progress in Nuclear
Magnetic Resonance Spectroscopy, vol. 38, no. 2, pp. 159–196, Mar. 2001, doi: 10.1016/S0079-6565(00)00032-7.
[29] A. Lay-Ekuakille, C. Pariset, and A. Trotta, “Leak detection of complex pipelines based on the filter diagonalization method:
Robust technique for eigenvalue assessment,” Measurement Science and Technology, vol. 21, no. 11, Sep. 2010, Art. no. 115403,
doi: 10.1088/0957-0233/21/11/115403.
[30] A. Lay-Ekuakille, G. Vendramin, and A. Trotta, “Robust spectral leak detection of complex pipelines using filter diagonalization
method,” IEEE Sensors Journal, vol. 9, no. 11, pp. 1605–1614, Nov. 2009, doi: 10.1109/JSEN.2009.2027410.
[31] V. A. Mandelshtam and H. S. Taylor, “Harmonic inversion of time signals and its applications,” Journal of Chemical Physics,
vol. 107, no. 17, pp. 6756–6769, Nov. 1997, doi: 10.1063/1.475324.
[32] D. J. Evans and W. S. Yousif, “The QZ algorithm for the calculation of the eigenvalues of a real matrix,” Parallel Algorithms and
Applications, vol. 4, no. 3–4, pp. 183–192, Jan. 1994, doi: 10.1080/10637199408915463.
[33] V. M. Álvarez-Pato, C. N. Sánchez, J. Domínguez-Soberanes, D. E. Méndoza-Pérez, and R. Velázquez, “A multisensor data
fusion approach for predicting consumer acceptance of food products,” Foods, vol. 9, no. 6, Jun. 2020, Art. no. 774, doi:
10.3390/foods9060774.
[34] B. Calabrese, R. Velázquez, C. Del-Valle-Soto, R. de Fazio, N. I. Giannoccaro, and P. Visconti, “Solar-powered deep learning-
based recognition system of daily used objects and human faces for assistance of the visually impaired,” Energies, vol. 13, no. 22,
Nov. 2020, Art. no. 6104,doi: 10.3390/en13226104.
[35] P. Visconti, R. Velazquez, S. Capoccia, and R. De Fazio, “High-performance AES-128 algorithm implementation by FPGA-based
SoC for 5G communications,” International Journal of Electrical and Computer Engineering (IJECE), vol. 11, no. 5, pp. 4221–
4232, Oct. 2021, doi: 10.11591/ijece.v11i5.pp4221-4232.
[36] C. Del-Valle-Soto, L. J. Valdivia, R. Velázquez, L. Rizo-Dominguez, and J. C. López-Pimentel, “Smart campus: An experimental
performance comparison of collaborative and cooperative schemes for wireless sensor network,” Energies, vol. 12, no. 16, Aug.
2019, Art. no. 3135, doi: 10.3390/en12163135.
[37] P. Visconti, N. I. Giannoccaro, R. de Fazio, S. Strazzella, and D. Cafagna, “IoT-oriented software platform applied to sensors-
based farming facility with smartphone farmer app,” Bulletin of Electrical Engineering and Informatics (BEEI), vol. 9, no. 3,
pp. 1095–1105, Jun. 2020, doi: 10.11591/eei.v9i3.2177.

Reliable e-nose for air toxicity monitoring by filter diagonalization method

More Related Content

Similar to Reliable e-nose for air toxicity monitoring by filter diagonalization method (20)

More from IJECEIAES (20)

Recently uploaded (20)

Reliable e-nose for air toxicity monitoring by filter diagonalization method