Available via license: CC BY 4.0
Content may be subject to copyright.
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
Digital Object Identifier
An end to end indoor air monitoring
system based on machine learning and
SENSIPLUS platform
MARIO MOLINARA1,(Member, IEEE),MARCO FERDINANDI1,2,3, (Member, IEEE),GIANNI
CERRO4,(Member, IEEE), LUIGI FERRIGNO1,(Member, IEEE), ETTORE MASSERA5
1Department of Electrical and Information Engineering, University of Cassino and Southern Lazio, Italy, (email: m.ferdinandi, m.molinara, ferrigno @unicas.it)
2Sensichips s.r.l., Via delle Valli, Aprilia, Italy (email: marco.ferdinandi@sensichips.com)
3The author contributed to the paper until October 31, 2019.
4Department of Medicine and Health Sciences, University of Molise, Italy, (email: gianni.cerro@unimol.it)
5ENEA, Portici, Italy (email: massera@enea.it)
Corresponding author: Mario Molinara (e-mail: m.molinara@unicas.it)
ABSTRACT In the framework of indoor air monitoring, this paper proposes an Internet of Things
ready solution to detect and classify contaminants. It is based on a compact and low–power integrated
system including both sensing and processing capabilities. The sensing is composed of a sensor array on
which electrical impedance measurements are performed through a microchip, named SENSIPLUS, while
the processing phase is mainly based on Machine Learning techniques, embedded in a low power and
low resources micro controller unit, for classification purposes. An extensive experimental campaign on
different contaminants has been carried out and raw sensor data have been processed through a lightweight
Multi Layer Perceptron for embedded implementation. More complex and computationally costly Deep
Learning techniques, as Convolutional Neural Network and Long Short Term Memory, have been adopted
as a reference for the validation of Multi Layer Perceptron performance. Results prove good classification
capabilities, obtaining an accuracy greater than 75% in average. The obtained results, jointly with the
reduced computational costs of the solution, highlight that this proposal is a proof of concept for a pervasive
IoT air monitoring system.
INDEX TERMS contaminant detection, air monitoring, sensor networks, neural networks, deep learning,
IoT.
I. INTRODUCTION
Air monitoring is a topic for which last few years have
witnessed a deep increase of interest in many different
fields. Lots of efforts have been addressed particularly for
applications regarding the citizens’ health and safety care,
as cities’ pollution control [1]–[3]. Pollution monitoring for
the detection and identification of contaminants are only
some but surely among the principal applications driving the
development of new ubiquitous and low–cost air monitoring
sensing technologies [4]. Nowadays, bulky and costly de-
vices requiring the employment of professional technicians
are currently adopted for air monitoring related tasks. Indeed,
in the pollution control application, static or mobile (mounted
on vehicles) cumbersome stations are used for air quality
monitoring in urban centers. Their high costs are usually the
main reason for a very sparse or, in same cases (e.g. poorer
countries), non–existent air quality control. For this reason,
a system that is, at the same time, reliable from the sensing
point of view and allowing a dense diffusion with low costs
is the goal of many research activities [5], [6]. The interest
of the scientific community is focused on either the sensing
technology or the data analysis. The first issue is addressed
with the development of miniaturized sensors able to respond
to contaminants with the same performance level than bulky
and costly systems, while the second is pursued through the
rise of novel algorithms and processing techniques able to
retrieve information about contaminants starting from raw
measurement data. In this sense, the authors propose an
integrated system, able to optimize both the sensing and
processing, responding to desirable requirements such as
low cost, integration, portability, light computational bur-
den, good sensitivity and classification capability. In detail,
stemming from the authors’ experience in gas recognition
[7], water analysis [8], [9], an indoor air monitoring system,
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
based on a compact and low cost sensing technology and
Artificial Intelligence (AI) techniques for the detection and
classification of air contaminants, is proposed. The system
development starts from the assumption, well known in sci-
entific literature on this field, that a single sensor is seldom
sensitive to more contaminants and therefore the best solution
is the employment of a sensor array, as implemented in this
work. Furthermore, having more than one sensor causes the
rapid growth of the amount of data to be processed to obtain
classification. Such issue has encouraged the researchers
to focus on ad–hoc computing techniques, often based on
decentralized architectures [10]. According to this trend, the
authors have looked for techniques able to manage and fully
exploit large data sizes, generally available under Machine
Learning (ML) based approaches. Aiming to minimize the
computational burden, a lightweight Multi Layer Perceptron
(MLP) solution, as in [11], is proposed and compared with
more complex Deep Learning (DL) architectures [12] as
Convolutional Neural Networks (CNN) and Long Short Term
Memory (LSTM). Three are the main novelties of the paper:
a) the use of an innovative measuring chip (i.e. the SEN-
SIPLUS) capable of embedding up to 6 air quality sensors,
directly inserted on the chip surface; b) the embedding of
sensing, measuring and classification on reduced resources
and costs platforms according to IoT and Edge Computing
paradigms; c) the comparison of DL and ML classification
performance in terms of accuracy and computational burden.
The paper is organized as follows. Section II gives an
overview about the state of art of current detection and
classification advances in research field; in section III a
comprehensive description of the system is given, both in
terms of sensing and data processing. In this section a de-
tailed description of the sensors peculiarities, data acquisition
and preprocessing, and neural network training is provided.
Obtained results are given in section IV and a discussion
regarding the achieved goals in terms of application require-
ments is provided. Finally, conclusions follow in section V.
II. STATE OF THE ART
A. SENSING ISSUES
The scientific sensing scenario has highlighted various tech-
nological limitations as the low sensitivity and selectivity
or the environmental and atmospheric conditions dependen-
cies [13] of miniaturized, low cost and smart solutions. In
[5], Castel et al. analyze low cost commercial platforms’
performance versus CEN (European Committee for Stan-
dardization) reference instruments, highlighting the fact that
despite of a lower accuracy, stability and selectivity they
provide a useful added value represented by the possibility
to perform data aggregation. Spatial analysis as mapping and
gradient could allow to evaluate the pollutants sources. The
well-known chemical micro-sensors limitations are deeply
faced in the scientific literature through various techniques.
A review of methodologies to improve the performance of
chemical sensors in different tasks as classification, regres-
sion and clustering is provided in [14].
B. PROCESSING ISSUES
Different Solutions for classification are available in litera-
ture. Among the others, here we mention: i) the Decision tree
induction, ii) the rule-based methods, iii) the support vector
machine and iv) those based on neural networks. Starting
from some considerations reported in [15], i.e. algorithms
based on neural networks show better classification accuracy
and considering the noisy of the measured data that may
prevent the reliability of algorithm based on thresholds [16],
the authors have paid their attention on the iv) category of
classification algorithm. AI and more in particular ML tech-
niques have been adopted in order to find hidden correlations
among sensors array responses to chemical substances. In
[17], Esposito et al. face the problem of on field calibration
of a low cost technology through a Dynamic Neural Network
approach. In [18] and [19] the authors exploit ML techniques
for gas recognition and concentration estimation, respec-
tively. In [20] the authors describe a carbon monoxide and
methane quantification system based on a sensor array and
an Artificial Neural Network (ANN). The technological rev-
olution experienced in the ML field with the born of DL has
furthered its adoption in an increasing number of different
applications outperforming other classical ML techniques. Its
capability to automatically extract data features and to profit
from large amount of data have been key elements for its
adoption also in sensors based applications.In [21] and [22],
proofs of how DL overcomes other classical ML techniques
in air quality monitoring are provided. In [23] Peng et al.
propose a gas classification system based on CNN comparing
its performance against the ones obtained with MLP and Sup-
port Vector Machine (SVM) architectures. CNNs represent a
breakthrough experienced with the born of DL for different
fields, as image or time series analysis as highlighted in [24]
by Bengio et al. An important DL architecture, well suited for
time series analysis is represented by LSTM neural network
[25]. It is a special kind of Recurrent Neural Networks
(RNN), which are networks with loops capable to maintain
the information. The computed output at each step is inputted
for the next step, providing the possibility to find out and
correlate time dependencies. The greater complexity of the
LSTM internal units allows to counteract the long short term
dependency, which represents one of the main RNN limits
[26]. In [27] and [28], deep RNN and LSTM are applied to
different scenarios with important data time dependencies.
III. PROPOSED INTEGRATED SYSTEM
A. OVERVIEW
The proposed integrated system is shown in Figure 1 and
mainly composed of:
•SENSIPLUS Chip (henceforth SPC): it is a micro-
electronic measurement device endowed with on-chip
sensing capabilities developed by Sensichips s.r.l. [29]
and the Department of Information Engineering of the
University of Pisa. Endowed with a versatile analog
front end and different internal and external ports, it
allows to perform electrical impedance measurements
2VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
FIGURE 1. The proposed integrated system. SDM stands for SENSIPLUS Deep Machine.
both on internal and external sensors. It has been already
adopted in other works, as in [7], [30], [31].
•SENSIPLUS Deep Machine (SDM): a hardware/software
module for data acquisition, processing and analysis.
The block diagram, depicted in Figure 2, shows the
logical operation flow and highlights the software and
hardware components exploited for each task. Data
Acquisition task is carried out through the SPC API,
which is a software library, developed either in Java
or C programming language, runnable on both Micro
Controller Unit (MCU) and Linux/Windows/Android
hosts. Acquired data are then preprocessed through an
ad–hoc developed module, which is described in detail
in section III-B4. As for the API, depending on the
application requirements, it can be executed on a MCU
and/or a host. Finally, classification can be performed
through one of the adopted ML techniques (MLP, CNN
or LSTM). The ML technique can be run on a MCU
or on a more computationally endowed device as a PC,
depending on which one is chosen.
B. DATA ACQUISITION AND PREPROCESSING
The experimental campaign has been carried out in order to
acquire a set of raw measurements used as the basis for ML
training phases (further described in III-C).
1) Measurement setup
The adopted measurement setup, shown in Figure 3, is com-
posed of:
•a Personal Computer (PC) running a proprietary (ad hoc
developed) JAVA software for measurement storage and
displaying;
•a MCU running SPC API and transferring acquired val-
ues to the PC through USB. The MCU is connected to
SPC with the proprietary one wire serial communication
protocol, namely SENSIBUS;
•a SPC plugged on the cable disposed on the internal
surface of a transparent glass box and endowed with 3
sensors;
•a little bowl containing the contaminant in the liquid
state used to submit the gas to sensors through the
evaporation process.
The adopted experimental setup is aimed to emulate the real
scenario which correspond to a common indoor environment
where low-concentrations of contaminants can be found.
Tens of ppm is the concentrations expected in reference
application and, for this reason, a small scale emulation of the
same conditions has been pursued. Furthermore, as a proof of
concept research activity, the efforts have been focused on the
system sensitivity and classification capability.
As for the sensing technology, 3 different sensors, whose
operating principle is provided in III-B2, have been used:
•The internal sensor based on aluminum oxide interdigi-
tated electrodes, namely ONCHIP_ALUMINUM_OXIDE;
•An external commercial capacitive humidity sensor,
namely OFFCHIP_HUMIDITY;
DATA ACQUISITION
Software: SENSIPLUS API
Hardware: MCU and/or Host
PREPROCESSING
Software: EMA filtering and
normalization
Hardware: MCU and/or Host
CLASSIFICATION
Software: MLP, CNN or LSTM
Hardware: MCU or Host
FIGURE 2. The SDM block diagram
VOLUME 4, 2016 3
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
•An external sensor based on gold interdigitated elec-
trodes functionalized with graphene as sensing material,
namely OFFCHIP_GRAPHENE.
The electrical impedance measuring capability of the SPC
has been exploited for the measurement acquisition phase.
A preliminary sensitivity analysis has allowed to optimize
the measuring settings (e.g. sinusoidal stimulus frequency
and amplitude) and to select a specific electrical quantity
(e.g. resistance, capacitance, conductance etc.), both aiming
to maximize the sensor sensitivity. Furthermore, the afore-
mentioned analysis phase has brought to witness, as high-
lighted by the scientific literature, the sensors’ dependence
on the environmental condition, in terms of response’s slope
and amplitude. Moreover, in order to follow the continual
variation of the environment (temperature and humidity), a
reference quantity has been computed through Exponential
Moving Average (EMA) (III-B4) for each sensor.
2) Notes about adopted sensors
As listed in III-B1, the adopted sensors in the experimental
campaign are characterized by different peculiarities.
The ONCHIP_ALUMINUM_OXIDE is the SPC built-in
sensor based on aluminum oxide interdigitated electrodes.
It is a generic sensor capable to fastly react to volatile
compounds. Water vapour and contaminants molecules de-
positing among the interdigitated tracks and affecting the
aluminum oxide cause a variation of the electrical proper-
ties, allowing their detection. The OFFCHIP_HUMIDITY
is a simple capacitive sensor manufactured by IST ( [32]).
The operating principle is based on the dielectric constant
variation of the polymer used as sensing material. The typical
capacitance value (23 ◦Cand 30% RH) is 140 ±40 pF,
measured in the frequency range 1 kHz – 100 kHz.
Finally, the adoption of OFFCHIP_GRAPHENE has been
enabled by the research activity carried out at ENEA Portici
research center [33]. Here we report a summary of the prop-
erties and its fabrication for sensing purposes. Graphite flakes
were obtained [34] from NGS Naturgraphit GmbHWinner
Company (Leinburg-Germany). Iso-Propyl Alcohol (IPA)
was purchased from Carlo Erba. All aqueous solutions were
prepared with ultrapure water from Type1 Ultrapure Milli-Q
system (Millipore). Pristine graphene was synthesized from
natural graphite powder by a Liquid Phase Exfoliation (LPE)
method. The process is a sonication–assisted exfoliation of
graphite flakes in a hydro–alcoholic solution. Specifically, 80
mg graphite flakes were dispersed into 80 ml of a water/IPA
mixture (7:1 v/v). The dispersion was sonicated in a low–
power bath (around 30 W) for 48 h. Afterwards, unexfoliated
graphitic crystals were separated from the dispersion by cen-
trifugation at 500 rpm for 45 min obtaining a black, homo-
geneous suspension of few–layer graphene at concentration
of 0.1 mg/ml. A flake of graphene, based on a few layers
of carbon sheets, is an almost two-dimensional material and
the chemical change on its surface can drastically change
the electric transport on the entire flake. Graphene flakes
are sensitive to oxidizing and reducing gases. A particular
MCU
Contaminant
SENSIPLUS PC
display
storage
FIGURE 3. Measurement setup
chemical and electrical affinity for nitrogen dioxide is ex-
perienced. It can significantly modify the resistivity of a
film of graphene flakes already to a few tens of ppb in the
humid air. A graphene flake film can be easily dispensed
onto an interdigitated electrode by dropcasting a few tens of
microliters of the graphene suspension.
3) Measurement Procedure
hlThe experimental campaign has been carried out according
to a systematic procedure collecting data with a substance
at a time. All sensors’ responses are acquired with an ac-
quisition rate, empirically chosen according to the observed
phenomenon velocity, of 0.5 S/s performing a measurement
from the three sensors at each acquisition time–step. The
measurements have been achieved pursuing the following
phases:
•Air exposure for 120 seconds;
•Chemical substance introduction inside the glass box for
600 seconds;
•Air exposure for further 120 seconds.
As regards the initial air exposure, it has been empirically
chosen after a preliminary analysis to let the sensors reach the
steady state. The fixed time interval of 600 seconds, used for
the chemicals evaporation phase, has been chosen according
to an experimental evaluation of the sensors response time
and to meet the fastly responsive application requirement.
Finally, the last 120 seconds of clean air exposure have been
used to analyze the sensors recovery capability. With the
selected sampling rate, 420 (= 60 + 300 + 60) samples
are collected for each experiment. 10 repetitions have been
conducted for each substance to analyze the measurement
repeatability and enhance the ML system generalization ca-
pability among different environmental conditions. The fol-
lowing dangerous substances, commonly found in indoor en-
vironments, have been adopted: acetone, alcohol, ammonia,
bleach. Furthermore, to estimate the sensing system measure-
ment background, further data acquisitions have been carried
out submitting water vapour and clean air to the sensors. In
this way, whatever classification technique is involved, it has
been trained to distinguish a chemical substance from air and
water vapour to avoid classification errors.
4) Preprocessing
The huge amount of raw measurements acquired through
the setup described in the previous section has required the
4VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
Raw
measurements Processing &
Labelling Structured ML
suitable dataset
Label Feature Vector
Contaminant (t0=n-1)
Contaminant (t1 = n)
… …
Contaminant (tm = n-1+m)
FIGURE 4. ML suitable Dataset preparation schema. Parameter n and m correspond to the time-window size of each sample and the dataset size, respectively.
development of an ad–hoc preprocessing software module,
whose operation principle is shown in Figure 4. It deals with a
series of preliminary operations on the sensors responses and
the generation of a labelled dataset suited for the successive
ML training phase. The developed software allows to adapt
the raw measurements to any of the adopted ML techniques’
requirements.
In this module, the main operations are: the evaluation
of an environmental reference, normalization and labelling.
The first one is accomplished by employing an Exponential
Moving Average (EMA) filtering stage to acquired samples.
Such operation is addressed to face the problem of the
sensors’ baseline continuous variation, mainly caused by the
environmental conditions fluctuation and sensors’ drift. The
EMA filtering has been computed according to equation 1,
where stis the 3–dimensional measurement vector at time
instant t,α= 10−4the degree of weighting decrease and
etthe environmental reference vector computed at the same
time instant.
et=αst+ (1 −α)et−1(1)
The αcoefficient has been chosen to have a high relevance
to initial samples, during the clean air sensors exposition.
Although this approach suffers long term dependency on the
response to contaminants exposition, the proposed system
is not deeply influenced since it is designed to provide
early detection and recognition. As for the normalization, the
ratio between each sensor response and its etparameter is
computed, as in equation (2).
fi
t=si
t/ei
t,i=1,2,3 with ft= (f1
t, f 2
t, f 3
t)(2)
Finally, the labelling operation is performed on the nor-
malization output, assigning the substance identifier to each
sample selected through a specific time window and step. A
time window of parametric size (n) can be generated for each
sample, providing in such a way a bi-dimensional feature
vector (Ft=ft−n+1, .., ft−1, ft).
C. ADOPTED MACHINE LEARNING ARCHITECTURES
As stated in the introduction, the goal is to provide the user
with a classification system able to perform most operations
on the MCU. Therefore, we implemented and performed
classification by means of a lightweight MLP. Nevertheless,
to compare our results with heavier, more complex and
powerful classifiers, we also tested our dataset with two
other ML and particularly DL approaches, namely the CNN
and LSTM. In the following, a more detailed description
of the designed architectures is presented. A preliminary
tuning phase of the networks’ hyper-parameters has been
executed for each architecture. Setting parameters, defining
the networks’ structure (e.g. number of layers and internal
neurons) and the ones determining the way they are trained
(e.g. learning rate, drop–out and batch normalization), have
been selected in this stage.
1) Multi Layer Perceptron
The designed MLP, as depicted in Figure 5, is characterized
by 3 input neurons, 64 neurons in the hidden layer and 6
output neurons. The Rectified Linear Unit (ReLU) activation
function has been applied on the output of the hidden layer
and the softmax function has been chosen for the output layer.
The 3–dimensional feature vector, computed according to
equation 2, is used as input for this network. In such a way,
no time latency and memory usage are required since a clas-
sification output is generated for each acquisition time–step.
Since, as described in section III-C2, for the DL architectures
a 2–dimensional feature vector (Ft) has been used as input,
a further implementation of the MLP model has been carried
out for a fairer comparison. Here, the same Ft, flattened in
a mono–dimensional feature vector, has been used as input.
For this reason, a wide input layer with 120 neurons has been
adopted, while the number of neurons in the hidden layer
FIGURE 5. Implemented MLP3architecture
VOLUME 4, 2016 5
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
has been preserved. In section IV-B, results corresponding
to both implementations, from now on referred as MLP and
MLP120, are presented.
2) Convolutional Neural Network
The developed CNN, as shown in Figure 6, is characterized
by the following stages:
•Input: a bi–dimensional feature vector (Ft=
ft−n+1, .., ft−1, ft) is used as input, where nis equal
to 40.
•Automatic features’ extraction: 2 consecutive Convolu-
tional Layers (CLs) are used in this stage for automatic
features’ extraction. Both the CLs execute the same op-
erations: zero padding of size 1, convolution, batch nor-
malization and ReLU activation function. Convolutional
kernels sizes are (16 ∗3∗3) and (14 ∗16 ∗3∗3) for CL1
and CL2, respectively. In such a way, the output of this
stage is a 3–dimensional matrix with size (14 ∗3∗40).
•Classification: the computed feature maps in the pre-
vious stage are flattened in a mono-dimensional vec-
tor and used as input for the Fully Connected Layers
(FCLs). The latter is composed of 128 neurons (hidden
layer) that are fully connected to the 6neurons of the
output layer.
f1
t1t2t3.. .. t40
Input stage
p(Analyte #1)
p(Analyte #2)
p(Analyte #6)
Probability
vector:
Fully
Connected
Layer Output
Layer
16
Feature
maps
3
40
f2
f3
Automatic Features’
extraction stage Classification stage
128
neurons
2D buffer
3x40x14
neurons
Flattening
6
neurons
Convolutions
14
Feature
maps
3
40
Convolutions
FIGURE 6. Developed Convolutional Neural Network (CNN) architecture.
3) Long Short Term Memory (LSTM) neural network
A multivariate 8–layer LSTM neural network has been
adopted and the same input as for the CNN has been used for
this model. Figure 7 shows an unrolled view of the proposed
architecture, where each LSTM cell is composed of 240
internal units that contains the trainable parameters.
FIGURE 7. Unrolled LSTM architecture.
4) Networks training
As reported in section B, for each substance and for each
of the experiments about 420 measurements have been col-
lected. A subset of 300 values has been considered for the
network training (the samples obtained in step two of the
procedure, after the chemical substance introduction). In this
way, a dataset of about (6 ∗10 ∗300) = 18000 samples
has been collected. A 10–fold cross–validation has been
performed in order to have a statistical analysis of the results
over the whole dataset. 1800 samples have been used as test
set for each fold (an entire experiment for each substance
(6∗1∗300)) while the remaining samples have been exploited
for training and validation sets (6 experiments for training
and 3 for validation with 10800 and 5400 samples respec-
tively). The network training has been carried out using the
Cross Entropy as loss function and Adam Optimizer as op-
timizing algorithm. Furthermore, an early stopping strategy,
based on a maximum number of epochs (patience) without
improvements on the validation set, has been used to avoid
the network overfitting. Training experiments have been car-
ried out through a Linux server characterized by an Intel(R)
Xeon(R) E5-2609 v4 (8 cores) CPU and 256 GB RAM. GPU
acceleration has been exploited for all the experiments using
a NVIDIA®TITAN X Pascal.
IV. RESULT
A. RAW AND PRE–PROCESSED MEASUREMENTS
Ten different experiments for each substance have been per-
formed according to the previously explained measurement
phases (III-B3). Once the whole raw dataset has been ac-
quired, it has been submitted to the preprocessing module
described in section III-B4. For instance, the EMA filtering
output computed on the ONCHIP_ALUMINUM_OXIDE
sensor to ammonia is depicted in Figure 8. The blue line
represents the raw sensor response and the orange one is the
filtered version.
As highlighted in the Figure 8, a 13.47% Percentage
Variation (PV) has been obtained with the raw measurement
while a 1.96% PV with the filtered version. The reported PVs
prove the negligible effect of the EMA long term dependency
FIGURE 8. Raw sensor response to ammonia. ONCHIP_ALUMINUM_OXIDE
sensor response is depicted here as example
6VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
FIGURE 9. Normalized sensors responses to all tested substances.
problem characterized by a short observation period.
A summarized overview of the normalized sensors re-
sponses for each experiment is provided in Figure 9.
The vertical dotted orange lines divide the sensors
responses to the various substances, alphabetically or-
dered in these graphs. All the air exposure phases
(initial and final) are removed to focus the attention
on the responses to chemicals. Similar behavior char-
acterizes the ONCHIP_ALUMINUM_OXIDE and OF-
FCHIP_HUMIDITY sensors’ responses, confirming their
analogous sensing nature, a great complementarity is pro-
vided by the OFFCHIP_GRAPHENE sensor. Having an
evident sensitivity to only 2 (ammonia and bleach) of the
6 total substances and, more in particular, an opposite trend
for them, the OFFCHIP_GRAPHENE sensor is expected to
improve the system classification capability.
B. CLASSIFICATION RESULTS
Results obtained with the adopted ML architectures are sum-
marized in Table 1 in terms of global accuracy mean and
standard deviation values (evaluated on the 10 folds), while
detailed metrics for each class are provided in Tables 2, 3,
4. As for the latters, Accuracy (eqn. 3), Precision (eqn. 4),
Recall (eqn. 5) and F1–score (eqn. 6) metrics are shown.
Accuracy =Correctly Classified Samples
Total Samples (3)
Precision =Positive Correctly Classified
Total Positive Classified Samples (4)
Recall =Positive Correctly Classified
Total Positive Samples (5)
F1–score = 2 ∗Precision ∗Recall
Precision +Recall (6)
As shown, best performance results have been obtained
with the CNN but, considering the concept of measure-
ment compatibility, no considerable improvements have been
reached with the two DL techniques. A more detailed
overview of the obtained results, except for the MLP120
which has been used as benchmark, is provided in the form
of Confusion Matrices in Figures 10,11 and 12.
The µand σvalues shown in a generic (i,j) matrix cell
represent the mean and the standard deviation (computed on
the 10-fold cross validation testing results) of the samples
that belong to class iand have been classified as j. High
FIGURE 10. MLP Global Confusion Matrix
VOLUME 4, 2016 7
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
FIGURE 11. CNN Global Confusion Matrix
FIGURE 12. LSTM Global Confusion Matrix
accuracies with relative small standard deviations have been
obtained for ammonia, bleach and water. A performance
decrease is shown by the CNN for air, with respect to MLP
and LSTM which have obtained 90% and 100% accuracy
values, respectively.
Finally, most of confusion regards acetone and alcohol
classes: for these contaminants all 3 architectures have shown
worst results. For such phenomenon, further considerations
are provided in section IV-C.
For the sake of completeness has been evaluated the Re-
ceiver Operating Characteristics of the classifiers (see Fig-
ures 13–16). As it is possible to see from the figures, the
trend of the three curves is equivalent to the ranking of the
three classifiers obtained in terms of accuracy, precision, etc.
Indeed the three mean curves have an Area Under the Curve
(AUC) equal to 0.92 for CNN, 0.91 for MLP and 0.87 for
LSTM.
C. DISCUSSION
1) Further investigations on classification results
As introduced in section IV-B, a common trend has been
obtained with all ML architectures. Two key observations can
be further highlighted: for air, ammonia, bleach and water
classes a high accuracy is obtained, meaning that the inte-
grated system has a good sensitivity and recognition capabil-
ity for these substances. As regards acetone and alcohol, they
are both polar substances that can attack oxygen atoms on the
film sensitive surface. Furthermore, they both have low boil-
ing points; therefore, they are difficult to be distinguished by
the adopted sensor array. For this reason, in Figures IV-C1–
IV-C1 the classifiers’ outputs, obtained with the aggregation
of data coming from both acetone and alcohol, are reported.
In detail, we have considered samples belonging to these
classes labeling them as one only substance, namely acet–
alc. With this operations, the element (1,1) of the confusion
Architecture Mean (µ) Standard Deviation (σ)
MLP 71.1 % 7.9 %
MLP120 62.4 % 10.8 %
CNN 75.1 %5.6 %
LSTM 69.9 % 12.1 %
TABLE 1. Mean and accuracy values evaluated for each ML architecture
FIGURE 13. The ROC curves for MLP
8VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
Substance Precision Recall F1-score Accuracy
Acetone µ0.50 0.51 0.48 0.51
σ0.32 0.35 0.28 0.32
Air µ0.53 0.88 0.66 0.90
σ0.09 0.19 0.13 0.18
Alcohol µ0.52 0.44 0.44 0.44
σ0.32 0.35 0.32 0.34
Ammonia µ0.96 0.85 0.89 0.84
σ0.13 0.09 0.06 0.08
Bleach µ0.93 0.84 0.88 0.84
σ0.09 0.06 0.06 0.06
Water µ0.81 0.76 0.77 0.82
σ0.13 0.16 0.13 0.12
TABLE 2. Synthetic performance metrics for MLP architecture: mean value
and standard deviation
Substance Precision Recall F1-score Accuracy
Acetone µ0.49 0.47 0.43 0.51
σ0.32 0.38 0.32 0.36
Air µ0.58 0.74 0.62 0.74
σ0.19 0.32 0.25 0.31
Alcohol µ0.53 0.40 0.40 0.44
σ0.38 0.32 0.28 0.30
Ammonia µ0.99 0.95 0.97 0.96
σ0.03 0.06 0.03 0.06
Bleach µ0.90 0.98 0.94 0.98
σ0.10 0.03 0.03 0.03
Water µ0.86 0.81 0.82 0.92
σ0.16 0.16 0.13 0.07
TABLE 3. Synthetic performance metrics for CNN architecture: mean value
and standard deviation
Substance Precision Recall F1-score Accuracy
Acetone µ0.51 0.41 0.42 0.40
σ0.38 0.38 0.38 0.39
Air µ0.62 0.89 0.72 1.00
σ0.25 0.25 0.25 0.00
Alcohol µ0.58 0.51 0.48 0.22
σ0.32 0.35 0.32 0.23
Ammonia µ0.95 0.87 0.90 0.85
σ0.16 0.09 0.09 0.15
Bleach µ0.67 0.64 0.64 0.67
σ0.35 0.38 0.35 0.26
Water µ0.66 0.71 0.67 0.77
σ0.28 0.28 0.25 0.12
TABLE 4. Synthetic performance metrics for LSTM architecture: mean value
and standard deviation
matrices shows a higher recognition rate, comparable with
those belonging to the remaining classes.
2) Computational analysis and memory footprint of MLP and
CNN
An evaluation of the FLoating point OPerationS (FLOPS)
and Memory Footprint (MF) in bytes both for MLP and CNN
is provided in this section, since they are the lightest and the
one providing best accuracy, respectively.
The CNN, as described in section III-C2, is composed of
two convolutional layers (CL1, CL2) as feature extraction
stage and a final set of 2 FC layers used for classification,
FIGURE 14. The ROC curves for CNN
Symbol Description Value Value
(MLP) (CNN)
w timesteps 1 40
s sensors 3 3
k11st layer kernels N/A 16
k22nd layer kernels N/A 14
k kernel size N/A 3
n0input neurons 3 w·s·k2
n1hidden neurons 64 128
n2output neurons 6 6
TABLE 5. Adopted notation symbols for computational analysis and memory
footprint.
FIGURE 15. The ROC curves for LSTM
VOLUME 4, 2016 9
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
FIGURE 16. The mean ROC curve of the three classifiers with the AUC for
CNN, MLP and LSTM respectively of 0.92, 0.91, 0.87
while the MLP contains only the FC layers. A general evalu-
ation of the needed FLOPS for the two convolutional layers
and for the FC is provided in equation 7, whose parameters
are defined in Table 5. Please note that the MLP is completely
characterized by the FC layers.
F LOP S =
wsk1(2k2+ 1) CL1
wsk2(2k2k1+ 1) CL2
2n1(n0+n2)FC
(7)
FIGURE 17. MLP Global Confusion Matrix
FIGURE 18. CNN Global Confusion Matrix
The overall memory footprint (X) is the sum of two
contributions, the first related to the model parameters (Xp),
the second related to the run-time (Xr). By considering that
on the ESP32 the float type is represented with 4 bytes, in the
equations (8) appear 4 as a multiplicative coefficient.
FIGURE 19. LSTM Global Confusion Matrix
10 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
MF =
4k1k2CL1p
4wsk1CL1r
4k2k1k2CL2p
4wsk2CL2r
4[(n0+ 1)n1+ (n1+ 1)n2]FCp
4(n0+n1+n2)FCr
(8)
The obtained results in terms of FLOPS and MF for each
layer of the two ML techniques are presented in Table 6
and 7. As expected, MLP is clearly much lighter than CNN
because of the lack of the features’ extraction layers and the
reduced neurons number used as FC input (s vs wsk2) and
hidden layers (64 vs 128).
In the next section detail about the adopted MCU is given.
Here can be observed that from Table 7 is evident that
the number of bytes requested by CNN (21656 + 872408)
is greater than the pSRAM dimension. The problem has
been resolved by using the flash memory for parameters and
pSRAM for runtime. Please note that generally speaking,
data can be read from flash memory as many times as needed,
while most devices are designed and tested for about 100,000
to 1,000,000 write operations.
3) Full system Time and Power analysis
Both MLP and CNN have been trained on a powerful server,
therefore the implementation on the MCU regarded only
the classification phase. This means that, once the network
weights have been correctly set, they are used for all classi-
fication trials. A characterization of execution time for mea-
surement acquisition and classification through MLP or CNN
has been performed through an embedded implementation of
the SDM on the ESP32 MCU. The adopted MCU has a 32-bit
CPU Xtensa dual-core LX6 microprocessor, operating at 160
or 240 MHz and performing at up to 600 DMIPS; in the MCU
is also available an Ultra low power (ULP) co-processor.
The MCU is equipped with 512kB of pSRAM (pseudo
Static Random Access Memory) and 4MB of flash memory.
The execution times have been measured through suitable
Technique CL1 CL2 FC Tot.
MLP N/A N/A 1152 1152
CNN 36480 485520 462848 984848
TABLE 6. Evaluated number of floating point operations for MLP and CNN.
Technique CL1 CL2 FC Tot.
MLPpN/A N/A 2584 2584
MLPrN/A N/A 292 292
CNNp576 8064 863768 872408
CNNr7680 6720 7256 21656
TABLE 7. Evaluated MF in bytes for MLP and CNN both for parameters (p)
and runtime (r) variables).
MLP CNN
Phases Power
[mW]
Elapsed
[s]
Energy
[µWh]
Elapsed
[s]
Energy
[µWh]
DAcq 201.0 0.372 20.77 0.372 20.77
Infe 218.0 0.001 0.06 0.854 51.71
Sleep 2.6 1.627 1.18 0.774 0.56
Duty Cycle / 2.000 22.01 2.000 73.04
TABLE 8. Elapsed time and energy Consumption by MCU during three
different phases: DAcq = Data Acquisition; Infe = inference phase; Sleep =
sleep phase between different acquisition. Values are reported for both MLP
and CNN
software routines placed inside the classification code and
they are not derived from the number of operations, time
complexity or analytical analyses that are anyway present
for the techniques’ analysis in Subsection IV-C2. Regarding
the acquisition phase, the minimum required time to perform
a single measurement on the 3 sensors has resulted in 372
ms. Whatever classification technique is adopted, the data
acquisition time interval is the same (after that the CNN
input buffer has been filled). Once the buffer has reached
the steady state, a classification output is provided for each
time–step as for the MLP. For this reason, an initial latency
is needed for the CNN usage to acquire the wtime–steps.
Regarding the classification phase, 854 ms and 1 ms have
been the elapsed times for CNN and MLP, respectively. In
such a way, the CNN classification time results have an
impact of 69% on the total elapsed time (acquisition +
classification) while a impact lower than 1% is obtained for
MLP. As regards the system response time, in terms of correct
classification output, the system takes 2 minutes in average
after the substance submission. Considering the distribution
of correct classification during the whole submission, it has
resulted that most of errors are concentrated in the initial
phase. Finally, an evaluation of the energy consumption in
a duty cycle has been carried out. The same rate used for
dataset acquisition (one output every 2 seconds) has been
maintained for the complete sequence. In such a way, after
the acquisition and classification phases, the ESP32 turns
in light–sleep mode, with a significant decrease of needed
power. In Table 8 a summary of the required times and energy
for each phase is reported. As result, an estimation of the
energy consumption for each acquisition/classification cycle
has been performed, obtaining 22.01 µW h and 73.04 µW h
for MLP and CNN, respectively. Such values are particularly
meaningful in order to obtain an estimation of the proposed
system average lifetime in real battery operated scenarios.
V. CONCLUSIONS
A low–cost and low–power integrated system for pervasive
indoor air monitoring is developed to detect the presence of
contaminants. To this aim, both sensing, performed with a
sensor array, and processing tasks are addressed in this work.
In terms of processing, three different machine–learning
based classification techniques have been tested. Due to ob-
tained results and expected computational burden, MLP and
VOLUME 4, 2016 11
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
CNN have been further investigated in terms of: i) computa-
tional complexity analysis, ii) ESP32 MCU implementation;
iii) execution time and power consumption analyses. Six
classes have been considered for the data acquisition and
ML technique training. Four of them are reliably recognized,
while acetone and alcohol detection is generally confused by
the system. We explained the motivation of such confusion
by noting chemical composition similarities between them.
A possible solution to solve the problems is to employ
other sensor typologies that are able to discriminate between
them. The proposed solution is now under improvement and
final deployment, to be released in the next future as rapid,
flexible, distributed and reliable system to be used in indoor
environments, especially in public or industrial buildings.
VI. ACKNOWLEDGMENT
The research leading to these results has received funding
from the European Union’s Horizon 2020 research and in-
novation programme under grant agreement SYSTEM No.
787128. The authors are solely responsible for it and it
does not represent the opinion of the Community and the
Community is not responsible for any use that might be made
of information contained therein.
The authors gratefully acknowledge Sensichips s.r.l. for the
support during the experimental phases and NVIDIA Corpo-
ration for the donation of the Titan Xp GPUs. This work was
also supported by MIUR (Minister for Education, University
and Research, Law 232/216, Department of Excellence).
REFERENCES
[1] M. Bentayeb, V. Wagner, M. Stempfelet, M. Zins, M. Goldberg,
M. Pascal, S. Larrieu, P. Beaudeau, S. Cassadou, D. Eilstein,
L. Filleul, A. L. Tertre, S. Medina, L. Pascal, H. Prouvost, P. Quénel,
A. Zeghnoun, and A. Lefranc, “Association between long-term exposure
to air pollution and mortality in france: A 25-year follow-up study,”
Environment International, vol. 85, pp. 5 – 14, 2015. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0160412015300349
[2] M. Kampa and E. Castanas, “Human health effects of air pollution,”
Environmental Pollution, vol. 151, no. 2, pp. 362 – 367, 2008, proceedings
of the 4th International Workshop on Biomonitoring of Atmospheric
Pollution (With Emphasis on Trace Elements). [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0269749107002849
[3] C. Santos, J. A. Jiménez, and F. Espinosa, “Effect of event-based sensing
on iot node power efficiency. case study: Air quality monitoring in smart
cities,” IEEE Access, vol. 7, pp. 132577–132 586, 2019.
[4] L. Morawska, P. K. Thai, X. Liu, A. Asumadu-Sakyi, G. Ayoko,
A. Bartonova, A. Bedini, F. Chai, B. Christensen, M. Dunbabin, J. Gao,
G. S. Hagler, R. Jayaratne, P. Kumar, A. K. Lau, P. K. Louie,
M. Mazaheri, Z. Ning, N. Motta, B. Mullins, M. M. Rahman, Z. Ristovski,
M. Shafiei, D. Tjondronegoro, D. Westerdahl, and R. Williams,
“Applications of low-cost sensing technologies for air quality monitoring
and exposure assessment: How far have they gone?” Environment
International, vol. 116, pp. 286 – 299, 2018. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0160412018302460
[5] N. Castell, F. R. Dauge, P. Schneider, M. Vogt, U. Lerner, B. Fishbain,
D. Broday, and A. Bartonova, “Can commercial low-cost sensor platforms
contribute to air quality monitoring and exposure estimates?” Environment
International, vol. 99, pp. 293 – 302, 2017. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0160412016309989
[6] Y. Li, W. Pang, C. Sun, Q. Zhou, Z. Lin, Y. Chang, Q. Li, M. Zhang, and
X. Duan, “Smartphone–Enabled Aerosol Particle Analysis Device,” IEEE
Access, vol. 7, pp. 101 117–101 124, 2019.
[7] P. Bruschi et al., “A novel integrated smart system for indoor air monitor-
ing and gas recognition,” in 2018 IEEE Int. Conf. on Smart Computing
(SMARTCOMP), June 2018, pp. 470–475.
[8] M. Ferdinandi et al., “A novel smart system for contaminants detection and
recognition in water,” in 2019 IEEE International Conference on Smart
Computing (SMARTCOMP), June 2019, pp. 186–191.
[9] G. B. et al., “Contaminants detection and classification through a cus-
tomized iot-based platform: A case study,” IEEE Instrumentation Mea-
surement Magazine, vol. 22, no. 6, pp. 35–44, Dec 2019.
[10] Y. Deng, Z. Chen, X. Yao, S. Hassan, and J. Wu, “Task Scheduling for
Smart City Applications Based on Multi–Server Mobile Edge Comput-
ing,” IEEE Access, vol. 7, pp. 14410–14 421, 2019.
[11] X. Zhai, A. A. S. Ali, A. Amira, and F. Bensaali, “MLP Neural Network
Based Gas Classification System on Zynq SoC,” IEEE Access, vol. 4, pp.
8138–8146, 2016.
[12] Y. Sun, F. P. . Lo, and B. Lo, “A deep learning approach on gender and age
recognition using a single inertial sensor,” in 2019 IEEE 16th International
Conference on Wearable and Implantable Body Sensor Networks (BSN),
May 2019, pp. 1–4.
[13] F. Rock, N. Barsan, and U. Weimar, “Electronic nose: Current status and
future trends,” Chemical Reviews, vol. 108, no. 2, pp. 705–725, 2008,
pMID: 18205411.
[14] S. Marco and A. Gutierrez-Galvez, “Signal and data processing for ma-
chine olfaction and chemical sensing: A review,” IEEE Sensors Journal,
vol. 12, no. 11, pp. 3189–3214, Nov 2012.
[15] G. Kesavaraj and S. Sukumaran, “A study on classification techniques
in data mining,” in 2013 Fourth International Conference on Computing,
Communications and Networking Technologies (ICCCNT), July 2013, pp.
1–7.
[16] G. Betta, C. Liguori, and A. Pietrosanto, “Uncertainty evaluation in algo-
rithms with conditional statement,” IEEE Transactions on Instrumentation
and Measurement, vol. 53, no. 4, pp. 969–976, Aug 2004.
[17] E. Esposito, S. D. Vito, M. Salvato, V. Bright, R. Jones, and O. Popoola,
“Dynamic neural network architectures for on field stochastic calibration
of indicative low cost air quality sensing systems,” Sensors and Actuators
B: Chemical, vol. 231, pp. 701 – 713, 2016. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S092540051630332X
[18] B. Mondal, M. Meetei, J. Das, C. R. Chaudhuri, and H. Saha,
“Quantitative recognition of flammable and toxic gases with artificial
neural network using metal oxide gas sensors in embedded
platform,” Engineering Science and Technology, an International
Journal, vol. 18, no. 2, pp. 229 – 234, 2015. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S2215098615000129
[19] S. D. Vito, A. Castaldo, F. Loffredo, E. Massera, T. Polichetti,
I. Nasti, P. Vacca, L. Quercia, and G. D. Francia, “Gas concentration
estimation in ternary mixtures with room temperature operating sensor
array using tapped delay architectures,” Sensors and Actuators B:
Chemical, vol. 124, no. 2, pp. 309 – 316, 2007. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0925400506008562
[20] G. Huyberechts, P. Szecówka, J. Roggen, and B. Licznerski, “Simultane-
ous quantification of carbon monoxide and methane in humid air using
a sensor array and an artificial neural network,” Sensors and Actuators
B: Chemical, vol. 45, no. 2, pp. 123 – 130, 1997. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0925400597002839
[21] X. Li, J. Shao, and T. Chi, “Deep learning architecture for air quality
predictions,” Environmental Science and Pollution Research, vol. 23, 10
2016.
[22] H. Karimian, Q. Li, C. Wu, Y. Qi, Y. Mo, G. Chen, S. Sachdeva, and
X. Zhang, “Evaluation of different machine learning approaches in fore-
casting pm2.5 mass concentrations,” Aerosol and Air Quality Research,
vol. 19, 01 2019.
[23] P. Peng, X. Zhao, X. Pan, and W. Ye, “Gas classification using deep
convolutional neural networks,” Sensors, vol. 18, no. 1, 2018. [Online].
Available: http://www.mdpi.com/1424-8220/18/1/157
[24] Y. LeCun and Y. Bengio, “Convolutional networks for images, speech,
and time series,” in The Handbook of Brain Theory and Neural Networks,
M. A. Arbib, Ed. Cambridge, MA, USA: MIT Press, 1998, pp. 255–258.
[Online]. Available: http://dl.acm.org/citation.cfm?id=303568.303704
[25] S. Hochreiter and J. Schmidhuber, “Long short–term memory,” Neural
Computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[26] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies
with gradient descent is difficult,” IEEE Transactions on Neural Networks,
vol. 5, no. 2, pp. 157–166, March 1994.
12 VOLUME 4, 2016
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2987756, IEEE Access
[27] A. Graves, A. Mohamed, and G. Hinton, “Speech recognition with deep
recurrent neural networks,” in 2013 IEEE International Conference on
Acoustics, Speech and Signal Processing, May 2013, pp. 6645–6649.
[28] X. Ma, Z. Tao, Y. Wang, H. Yu, and Y. Wang, “Long short-term
memory neural network for traffic speed prediction using remote
microwave sensor data,” Transportation Research Part C: Emerging
Technologies, vol. 54, pp. 187 – 197, 2015. [Online]. Available:
http://www.sciencedirect.com/science/article/pii/S0968090X15000935
[29] “Sensichips.” [Online]. Available: www.sensichips.com
[30] G. Cerro et al., “Preliminary realization of a monitoring system of activated
carbon filter RLI based on the SENSIPLUS® microsensor platform,”
in 2017 IEEE International Workshop on Measurement and Networking
(M&N), Sep. 2017, pp. 1–5.
[31] G. Cerro et al., “Metrological characterization of a novel microsensor
platform for activated carbon filters monitoring,” IEEE Transactions on
Instrumentation and Measurement, vol. 67, no. 10, pp. 2504–2515, Oct
2018.
[32] “Humidity Modules & Sensors.” [Online]. Available: https://www.ist-
ag.com/en/products-services/humidity-modules-sensors
[33] F. Ricciardella, S. Vollebregt, T. Polichetti, M. Miscuglio, B. Alfano,
M. L. Miglietta, E. Massera, G. Di Francia, and P. M. Sarro,
“Effects of graphene defects on gas sensing properties towards no2
detection,” Nanoscale, vol. 9, pp. 6085–6093, 2017. [Online]. Available:
http://dx.doi.org/10.1039/C7NR01120B
[34] B. Alfano, T. Polichetti, M. Mauriello, M. Miglietta, F. Ricciardella,
E. Massera, and G. D. Francia, “Modulating the sensing properties of
graphene through an eco-friendly metal-decoration process,” Sensors and
Actuators, B: Chemical, vol. 222, pp. 1032–1042, January 2016.
VOLUME 4, 2016 13