See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/268685010

Calibration of a cluster of low-cost sensors for the measurement of air

pollution in ambient air

Conference Paper · November 2014

DOI: 10.1109/ICSENS.2014.6984922

CITATIONS READS

12 1,703

5 authors, including:

Michel Gerboles Maria Gabriella Villani

European Commission ENEA
71 PUBLICATIONS   1,403 CITATIONS    38 PUBLICATIONS   927 CITATIONS   

SEE PROFILE SEE PROFILE

Manuel Aleixandre
Spanish National Research Council
85 PUBLICATIONS   1,634 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Smart Ring in l'Aquila (ITA), City 2.0 View project

Special Issue Atmosphere "Recent Advances of Air Pollution Studies in Italy View project

All content following this page was uploaded by Maria Gabriella Villani on 24 November 2014.

The user has requested enhancement of the downloaded file.

 Calibration of a cluster of low-cost sensors for the
measurement of air pollution in ambient air
Laurent Spinelle, Michel Gerboles
Air and Climate Unit, IES
JRC, European Commission
Ispra (VA), Italy
laurent.spinelle@jrc.ec.europa.eu
Manuel Aleixandre
Institute for applied Physics, CSIC
Madrid, Spain
manuel.aleixandre@gmail.com
Abstract—Using a clustered system of sensors able to measure
O3, NO/NO2, CO and CO2, we want to compare the performance
of several calibration methods. Based on the outcome of the
single sensor’s evaluation, the suitability of this cluster is
investigated using deterministic or statistical methods such as
multivariate regression and neural networks. The objective of the
study is to assess if the measurement uncertainty of sensors can
reach the Data Quality Objective of the European Air Quality
Directive for indicative methods.
Keywords—gas sensors, ozone, nitrogen dioxide, air quality,
measurement uncertainty, multivariate linear regression, neural
network
I. INTRODUCTION
Compared to the reference methods defined in the Air
Quality Directive [1], the use of low-cost gas sensors would
reduce air pollution monitoring costs and allow larger spatial
coverage even in remote areas. However, calibrating these
sensors for monitoring air quality keeps on being a challenge.
The selectivity and stability of sensors are generally found
problematic [2,3,4]. Consequently, more sophisticated
algorithms for quantifying air pollution are being developed.
Among the studied methods, the temperature cycle operation
was shown to limit cross sensitivities and aging of sensors [5]
under laboratory conditions. This method is also relevant for
the identification of organic compounds [6]. Other authors
reported the use of several metal oxides (MOx) sensors
operated at different heating temperature [7,8]. Theses multi-
sensors were either calibrated against standard gas mixtures or
using artificial neural network under field conditions. The latter
method resulted mixed results either satisfactory for short
periods or generally weak for longer data series. Neural
network calibration has mainly been implemented for the
identification of organic compounds and smell [9,10] or for
monitoring compounds such as CO or CH4 at high levels [11].
Few attempts have been made to use neural network for the
calibration of sensor for monitoring in the low ppb range
[12,13]. One of these studies looked at neural network
calibration for benzene at ppb levels [14]. However, the
Maria Gabriella Villani
ENEA, Agenzia Nazionale per le nuove tecnologie,
l'energia e lo sviluppo economico sostenibile
Ispra (VA), Italy
mariagabriella.villani@enea.it
Fausto Bonavitacola
Phoenix Sistemi & Automazione s.a.g.l.
Muralto (TI), Switzerland
fausto.bonavitacola@ingpec.eu
majority of the studies cited before used the sole MOx-type
sensors which are known to suffer from a lack of stability and
long response times [15]. A recent study describes a new real-
time field calibration by comparing mobile sensor responses
with reference measurements of existing reference monitoring
stations [16].
Recently, within the EURAMET MACPoll project [17], the
performance of single commercial sensors has been evaluated
[18,19,20,21] according to a precise protocol [22]. This study
produced large datasets of measurements for several
compounds under laboratory conditions and field campaigns.
Such data sets were not previously available in literature,
especially considering the number of controlled parameters
(NOx, O3, CO, SO2, CO2 temperature, relative humidity, wind
and pressure).
In this study, an analysis of the performance of different
calibration models over a great number of sensors tested in the
same conditions was performed. The study considered the
calibration of single sensor and of a cluster of sensors able to
measure a broad range of pollutants. The performance of the
calibration methods were compared taking as indicator the
measurement uncertainty of each method. An example is given
in this paper for the estimation of hourly O3 values. It was then
evaluated if the uncertainty could meet the Data Quality
Objective (DQO) of the European Air Quality Directive [1].
II. EXPERIMENTS
Experiments were carried out in collaboration with the
European Reference Laboratory for Air Pollution (ERLAP) at
the EMEP station of the Joint Research Centre (45°48.881’N,
8°38.165’E). The station is located in a semi-rural area at the
NW edge of the Po valley (Italy) and it is equipped with
meteorological sensors (temperature, relative humidity, wind
and pressure) and reference gas analyzers for NOx, O3, CO,
CO2 and SO2. These reference measurements were used for
data validation, comparison and data treatment of sensor
responses.
 TABLE I. LIST OF CLUSTERED SENSORS TABLE II. MLR MODELS OF SINGLE SENSORS
2
R of linear Sensor’s model Multivariate linear model R2
Manufacturer Sensor’s model Pollutant
regression O3-B4 0.49
αSense O3-B4 O3 0.07
Citytech O3_3E1F O3 0.87 O3_3E1F 0.91
CairPol CairclipO3/NO2 O3/NO2 Unknown
αSense NO2-B4 NO2 0.06 CairclipO3/NO2 Unknown
NO2_3E50 NO2 0.01
Citytech
NO_3E100 NO 0.05 NO2-B4 0.56
2710 sensor NO2 0.31
e2V
4514 sensor NO2 0.34 NO2_3E50 0.63
CairPol CairClip NO2 NO2 0.37
Figaro 5042 sensor CO 0.17 2710 sensor 0.36
e2V 4514 sensor CO 0.56
Edinburgh Sensors Gascard NG CO2 0.14 4514 sensor 0.42
ELT Sensors S-100H CO2 0.58
CairClip NO2 0.74
Based on the evaluation and validation of low-costs sensors 0.23
5042 sensor
carried out within the EURAMET MACPoll project [16-19],
several sensors were chosen to be grouped in a clustered 4514 sensor 0.58
system able to detect O3, NO/NO2 and CO. These pollutants are Gascard NG 0.47
regulated in the air quality Directive 2008/50/EEC [1]. It was
decided to keep the most efficient sensors with shortest S-100H 0.62
response time, highest sensitivity, and smallest repeatability. C. Artificial neural network (ANN)
To extend the range of species, CO2 sensors were added. The
cluster consisted of 5 NO2 sensors and 2 CO sensors, both An artificial neural network was designed based on a
electrochemical and metal oxide type, 1 NO and 3 O3 multilayer perceptron (MLP) with feedback, input bias and one
electrochemical sensors and 2 infrared CO2 sensors (see hidden layer of variable number of nodes [23]. The input
Table 1). For NO2, MOx and electrochemical sensors were consisted of sensor responses and meteo data. The output of the
used in order to benefit from the different inherent cross- ANN consisted in the estimate of O3 reference values.
sensitivities of both types of sensors. Initially, all sensors which were correlated with O3 and
found independent between each other were selected. Using
III. CALIBRATION METHODS AND CHOICE OF VARIABLES sensitivity analysis, variables that were not found significant
Three calibration methods were tested: simple linear were discarded. Sensitivity analysis was performed by the
models (LR), multivariate linear estimation (MLR) established estimation of the effect of one parameter (also called predictor)
within MACPoll [16-19] and artificial neural networks (ANN) on the model residuals when this one is significantly modified.
with raw, standardized and calibrated sensor responses. For the analysis, we computed the sums of squares residuals
for the model when the parameter was eliminated from the
A. Linear regression (LR)
neural network and we estimated the ratio of the reduced model
Calibration functions were established by ordinary linear against the full one. In this method called Custom Neural
regression, using the minimization of residuals of the sensor Networks (CNN), we defined individual network architectures
responses vs. reference measurement (UV photometry analyzer and training algorithms. Several networks were designed to
for O3, chemiluminescence for NOx, non-dispersive infrared solve the model and then we selected those networks that best
gas-filter correlation spectroscopy for CO and differential non represented the relationship between the sensor input and the
dispersive infrared for CO2). Thus, the calibration functions target reference measurements.
were of the type Rs = a.X +b where Rs represents sensor
responses and X is the corresponding reference measurements. Three studies have been performed by changing the input
Finally, the measuring function, the converse equation data: raw, standardized and calibrated by MLR sensors data.
X = f(Rs) was applied to each sensor. The coefficient of To realize the study, the dataset was divided in three periods: a
determination (R²) of each calibration is given in Table 1. training period (1 week), a test period (1 week) and a validation
period (3 weeks). The output of ANN consisted in the average
B. Multivariate linear regression (MLR) of an ensemble of the 5 best networks within 100 tested
The calibration was carried out using the least square networks with different MLP architectures.
method taking into consideration more than one explanatory
variables. Models were established during the MACPoll studies IV. RESULTS
(see Table 2). As for the LR, the calibration functions consisted A. Linear regression
of equations of the type Rs = f(X,Yi), where f(X,Yi) is a
The linear regression was performed using data of the first
function of multiple reference measurements. The resulting
week in order to calibrate. As we focused on ozone, we
measuring function, X = f(Rs,Yi), was applied to each sensor.
selected the sensor that showed the best correlation factor: the
The coefficient of determination (R²) of each calibration is
O3_3E1F sensor with R2 = 0.87 (see Table 1). This sensor will
given in Table 2.
be used throughout the whole document.
 The measuring equation was applied to the extrapolation set TABLE III. PARAMETERS OF THE REGRESSION OF THE 3ANN MODELS
of data, the remaining 4 weeks. Fig. 1 gives the scatterplot of Input Intercept Slope R2 RMSE
the estimated sensor values vs. the reference measurements. It Raw data -2.9 1.04 0.95 64.0
shows that although the strength of association remain similar Standardized data -3.9 1.07 0.96 64.8
to calibration (R² = 0.86 compared to 0.87), the sensor Calibrated data (MLR) 1.5 0.99 0.95 60.7
responses drifted both in slope (1.13) and intercept (-7).
(O3_3E1F) and 2 NO2 MOx (2710 and 4514). The
B. Multivariate linear estimation combination of electrochemical/metal oxide and O3/NO2
During the MACPoll evaluations the O3_3E1F sensor was sensors appears to be the most efficient.
not affected by meteorological parameters such as temperature
or humidity but it showed cross-sensitivity to NO2. As needed Table 3 gives the regression parameters for the three types
in the MLR model (see Table 2), the other input (gaseous or of input data for the validation period. The R2 of the three
meteorological) have been selected within the reference models are equivalent. The ANN based on calibrated data
measurements to maximize the benefits of the calibration. (MLR) gave the best slope and intercept (nearer to 1 and 0,
respectively) and the lower root mean square error (RMSE).
The constants a, b and c of the sensor model were fitted However, the improvement of the ANN, even on raw data, is
during the first week of measurements. Subsequently, this significant as shown by the increase of R2 from 0.87 to 0.95.
function was applied to the extrapolation set of data. Fig. 2 Fig. 3 gives the sensor values vs. reference UV-photometry
shows the scatterplot of the calibrated sensor data using the values data using the ANN on raw data.
MLR method against the reference measurement. The use of
NO2 reference values led to an improvement of R2 from 0.87 to D. Comparing the different models
0.91. To assess the performance of each calibration method, we
calculated the measurement uncertainty using orthogonal
C. Artificial neural network regression of the estimated outputs against reference data. The
For the estimation of O3 with ANN, the significant input uncertainty was compared with the DQO for indicative method
parameters were found by sensitivity analysis and consisted in that corresponds to a relative expanded uncertainty of 30%.
3 sensors without meteo data: one O3 electrochemical Fig. 4 shows the relative expanded uncertainty of the five

Fig. 1. Scatterplot of the calibrated sensor data using the linear regression Fig. 2. Scatterplot of the calibrated sensor data using the MLR against reference
against reference measurements. measurements

Fig. 3. Scatterplot of the calibrated sensor data using the ANN based on raw
sensor’s data against reference measurements. Fig. 4. Relative expanded uncertainty of the estimated values versus reference
data as a function of the level of O3 for the five calibration models.
models as a function of the O3 level. This uncertainty was
calculated using the following equation [24].
( ) (√ ( ) [ ( ) ] )⁄
( )
The LR model met the DQO with an uncertainty of 20% at
the O3 limit value of 60 nmol/mol [1]. Even if at low levels the
MLR is similar to the LR, it slightly raised up to a value of
25% at high concentration .This could be explained by a drift
of the sensor response over time. The major drawback of this
model was that it needed meteo and reference data for the
calibration of most sensors. The artificial neural network gave
the lowest uncertainty especially for the ANN based on MLR
calibrated measurements. Though, as it was based on MLR
data, it required the use of reference values. ANN on raw data
gave the most interesting results with an uncertainty of 10% at
60 nmol/mol, with the combination of only 1 electrochemical
O3 sensor and 2 MOx NO2 sensors.
The ANN method increased the strength of association
between estimated and reference data (higher R² and lower
RMSE). Moreover, it also allowed the decrease of the bias to
reference data, with the slope and intercept of orthogonal
regression being nearest to 1 and 0 respectively.
V. CONCLUSION
Based on the measurement uncertainty estimated by
orthogonal regressions between the sensor outputs versus
reference data, the most suitable calibration method appeared
to be ANN using raw sensors data. Simple LR and MLR were
shown to produce the highest measurement uncertainty. While
MLR needed meteo data for calibration of most sensors, ANN
using only 3 sensors of different types (1 O3 chemical and 2
NO2 resistive sensors) was able to solve the main interferences
on the O3 sensor without the need for reference parameters.
It is likely that by combining electrochemical O3 sensors
and NO2 MOx sensors, the ANN can solve the O3/NO2
interference of chemical sensors. Actually, O3 electrochemical
sensors are generally slightly affected by NO2. The
humidity/temperature dependence was also corrected thanks to
the difference of influence of these parameters on chemical and
MOx sensors. Finally, using a cluster of sensors for calibration
purpose, the data quality objectives of the European Directive
for indicative methods could be met for O3 (uncertainty, , of
30 %).
ACKNOWLEDGMENT
The authors wish to acknowledge the collaboration of
F. Lagler, C. Gruening, I. Goded, N. R. Jensen and
A. Dell’Acqua our JRC colleagues, for carrying out air
pollution measurements. This work has been partly funded by
EMRP MACPoll. The EMRP is jointly funded by the EMRP
participating countries within EURAMET and the European
Union.
REFERENCES
[1] Directive 2008/50/EC of the European Parliament and the Council of 21
May 2008 on ambient air quality and cleaner air for Europe
[2] H. Nakagawa, S. Okazaki, S. Asakura, K. Fukuda, H. Akimoto, S. Taka-
hashi, S. Shigemori, “An automated car ventilation system”, Proceeding
of the 7th International Meeting on Chemical Sensors, Beijing, 1988, pp.
187–189.
View publication stats
[3] M.C. Carotta, G. Martinelli, L. Crema, M. Gallana, Merli F., G. Ghiotti,
E.Traversa, “Array of thick film sensors for atmospheric pollutant
monitoring”, Sens. Actuators B: Chem. 68 (2000) 1–8.
[4] C. Pijolat, C. Pupier, M. Sauvan, G. Tournier, R. Lalauze, “Gas
detection for automotive pollution control”, Sens. Actuators B: Chem.
59 (1999) 195–202.
[5] M. Lösch, M. Baumbach, and A. Schütze, “Ozone detection in the ppb-
range with improved stability and reduced cross sensitivity”, Sensors
and Actuators B: Chemical, vol. 130, no. 1, pp. 367–373, 2008.
[6] A. Gramm and A. Schütze, “High performance solvent vapor
identification with a two sensor array using temperature cycling and
pattern classification”, Sensors and Actuators B: Chemical, vol. 95, no.
1–3, pp. 58–65, Oct. 2003.
[7] M. Kamionka, P. Breuil, and C. Pijolat, “Calibration of a multivariate
gas sensing device for atmospheric pollution measurement”, Sensors and
Actuators B: Chemical, vol. 118, no. 1–2, pp. 323–327, Oct. 2006.
[8] M. Kamionka, P. Breuil, and C. Pijolat, “Atmospheric pollution
measurement with a multi-materials sensing device”, Materials Science
and Engineering: C, vol. 26, no. 2–3, pp. 290–296, Mar. 2006.
[9] D. B. Reza Nadafi, S. N. Nejad, M. Kabganian, and F. Barazandeh,
“Neural network calibration of a semiconductor metal oxide micro smell
sensor”, in 2010 Symposium on Design Test Integration and Packaging
of MEMS/MOEMS (DTIP), 2010, pp. 154–157.
[10] Baumbach, M., Sossong, A., Delprat, H., Soulantica, K., Schuetze, A.,
Borrel, H., Maisonnat, A., Chaudret, B.: New Micro Machined Gas
Sensors Combined with Intelligent Signal Processing Allowing Fast Gas
Identification after Power-Up. Proceedings Sensor 2005. Vol. 2.91–96.
[11] G. Huyberechts, P. Szekowka, J.R.N. Licznerski, Simultaneous
quantification of carbon monoxide and methane in humid air using a
sensor array and an artificial neural network, Sens. Actuators B: Chem.
45 (1997) 123–130.
[12] S. De Vito, G. Di Francia and L. Martinotto, “Neural calibration of
portable multisensor device for urban atmospheric pollution
measurement”, Sensors and Microsystems, pp. 283–290, 2008.
[13] K. Brudzewski and S. Osowski, “Gas analysis system composed of a
solid-state sensor array and hybrid neural network structure”, Sensors
and Actuators B: Chemical, vol. 55, no. 1, pp. 38–46, 1999.
[14] S. De Vito, E. Massera, M. Piga, L. Martinotto, and G. Di Francia, “On
field calibration of an electronic nose for benzene estimation in an urban
pollution monitoring scenario”, Sensors and Actuators B: Chemical, vol.
129, no. 2, pp. 750–757, Feb. 2008.
[15] L. Spinelle, M. Aleixandre and M. Gerboles, Comparison between the
performances of commercially availbale low cost sensors for the
monitoring of O3 and No2 in ambient air. Unpublished.
[16] D. Hasenfratz, O. Saukh, and L. Thiele, “On-the-Fly Calibration of
Low-Cost Gas Sensors”, in Wireless Sensor Networks, G. P. Picco and
W. Heinzelman, Eds. Springer Berlin Heidelberg, 2012, pp. 228–244.
[17] http://macpoll.eu
[18] L. Spinelle, M. Gerboles and M. Aleixandre, “Report of the laboratory
and in-situ validation of micro-sensors for monitoring ambient air
pollution - O12: CairClipO3/NO2 of CAIRPOL (F)”, Publications
Office of the European Union, EUR 26373, 2013.
[19] L. Spinelle, M. Gerboles and M. Aleixandre, “Report of laboratory and
in-situ validation of micro-sensor for monitoring ambient air pollution -
NO9: CairclipNO2 sensor of CairPol (F)”, Publications Office of the
European Union, EUR 26394, 2013.
[20] L. Spinelle, M. Gerboles and M. Aleixandre, “Report of laboratory and
in-situ validation of micro-sensor for monitoring ambient air - Ozone
micro-sensors, AlphaSense, model B4-O3 sensors (UK)” Publications
Office of the European Union, EUR 26681, 2013.
[21] L. Spinelle, M. Gerboles and M. Aleixandre, “Report of laboratory and
in-situ validation of micro-sensor for monitoring ambient air pollution –
NO2: NO2-B4 sensor of AlphaSense (UK)”, unpublished.
[22] L. Spinelle, M. Aleixandre and M. Gerboles, “Protocol of evaluation and
calibration of low-cost gas sensors for the monitoring of air pollution”,
Publications Office of the European Union, EUR 26112, 2013.
[23] StatSoft, Inc. (2011). STATISTICA (data analysis software system),
version 10. www.statsoft.com
[24] Guide to the demonstration of equivalence of ambient air monitoring
methods, Report by an EC Working, Group on Guidance