Evaluating Correlations in IoT Sensors for Smart Buildings
Davide Andrea Guastella
1,3 a
, Nicolas Verstaevel
2 b
, Cesare Valenti
3 c
,
Bilal Arshad
4 d
and Johan Barth
´
elemy
4 e
1
Institut de Recherche en Informatique de Toulouse, Universit
´
e Toulouse III Paul Sabatier, France
2
Institut de Recherche en Informatique de Toulouse, Universit
´
e Toulouse 1 Capitole, France
3
Dipartimento di Matematica e Informatica, Universit
`
a degli Studi di Palermo, Italy
4
SMART Infrastructure Facility, University of Wollongong, Wollongong 2522, NSW, Australia
Keywords:
Smart Building, Smart Cities, IoT Sensors, Evolutionary Approach, Sensors Correlation.
Abstract:
In this paper we introduce a dataset of environmental information obtained via indoor and outdoor sensors
deployed in the SMART Infrastructure Facility of the University of Wollongong (Australia). The acquired
dataset is also made open-sourced along with this paper. We also propose a novel approach based on an evolu-
tionary algorithm to determine pairs of correlated sensors. We compare our approach with three other standard
techniques on the same dataset: on average, the accuracy of the evolutionary method is about 62,92%. We also
evaluate the computational time, assessing the suitability of the proposed pipeline for real-time applications.
1 INTRODUCTION AND
MOTIVATIONS
Smart buildings, like smart cities, are truly complex
systems (Nigon et al., 2017). Indeed, if we set aside
the social and organizational aspects of smart build-
ings and focus only on technology, smart buildings
are equipped with numerous heterogeneous sensors
and networks. Smart buildings are open, meaning
that new sensors and data can appear or disappear at
any time, and they have to face non-linear and unpre-
dictable dynamics. Putting humans in the loop to de-
sign adaptive people-centric control systems will add
another layer of complexity. This involves designing
systems equally complex, as expressed by Ashby’s
law of requisite variety claiming that if a system is
to be stable, the number of states of its control mech-
anism must be greater than or equal to the number of
states in the system being controlled (Ashby, 1991).
This level of complexity, combining a dynamic
and unpredictable nature of malfunction events, the
noisiness of the data, and human activities make those
a
https://orcid.org/0000-0002-6865-1833
b
https://orcid.org/0000-0002-7879-6681
c
https://orcid.org/0000-0002-4961-2054
d
https://orcid.org/0000-0002-6078-0235
e
https://orcid.org/0000-0002-7800-5309
systems difficult to design through traditional meth-
ods, as not all states of the system are known a priori.
In this paper, we propose and evaluate a novel
methodology to dynamically exploit the correlations
between sensors to provide experts with automatic
information about the building state. Studying how
sensors are correlated or uncorrelated allows detect-
ing variations of behavior in the building that might
be indicative of an anomaly such as a malfunction-
ing air conditioning system, or a window that stays
open during the night (Houssin et al., 2020). Because
those anomalies can occur unexpectedly, the correla-
tions must be processed in real-time (i.e. within fixed
time constraints).
The contribution of this article is twofold:
• we introduce an open dataset of one year of
heterogeneous environmental information ac-
quired through internal and external sensors de-
ployed in the SMART Infrastructure Facility at
the University of Wollongong (Australia). Mak-
ing this dataset open sourced will allow the scien-
tific and industrial communities to have a consis-
tent set of environmental information to be used
for simulations, testing applications and compar-
ing methodologies in a smart building context;
• we propose a novel solution based on an evolu-
tionary algorithm to detect in real-time highly
correlated pairs of heterogeneous sensors.
224
Guastella, D., Verstaevel, N., Valenti, C., Arshad, B. and Barthélemy, J.
Evaluating Correlations in IoT Sensors for Smart Buildings.
DOI: 10.5220/0010210502240231
In Proceedings of the 13th International Conference on Agents and Artificial Intelligence (ICAART 2021) - Volume 1, pages 224-231
ISBN: 978-989-758-484-8
Copyright
c
 2021 by SCITEPRESS – Science and Technology Publications, Lda. All rights reserved

Our proposal addresses the properties of open-
ness, heterogeneity and unpredictability (Guastella
and Valenti, 2018). The novelty of our contribution
is:
• proposed an approach to determine pairs of highly
correlated sensors whose information dynamics
are similar in time;
• it does not make any assumption on the topogra-
phy of the environment where the sensors are de-
ployed. This property makes the solution generic
so that it can be applied regardless of the building
configuration.
• proposed a methodology that uses heterogeneous
information to measure the correlation between
pairs of sensors. Moreover, our approach exploits
sensors whose type of perceived information is
not known a priori.
The rest of this paper is organized as follows: Sec-
tion 2 presents the proposed approach, describing how
it is capable of determining pairs of correlated sen-
sors based on their perceptions. Section 3 presents
the results obtained from our method using the data
acquired from the SMART Facility. In Section 4, we
conclude and point out some future perspectives.
2 METHODOLOGY
In this section, we present a novel approach that re-
minds an evolutionary algorithm for determining cor-
relations among sensors in smart buildings.
Along with this method, we present three other
original pipelines to compare the results obtained
from the evolutionary approach and also to illustrate
techniques that can be used to exploit the introduced
dataset. Fig. 1 shows the main steps of our method-
ology and the three other pipelines. The described
techniques do not require any a priori information on
the topology of the environment. All the input param-
eters we used have been obtained on an experimental
basis; our results refer to their best configuration.
The following section describes the proposed
pipeline. Sections 2.2, 2.3 and 2.4 present the tech-
niques used to compare the results obtained through
the evolutionary method.
2.1 Evolutionary Technique
The proposed approach (diagram ¶ in Fig. 1) consists
of three steps: (i) acquiring information from the sen-
sors; (ii) smoothing the data; (iii) applying the evolu-
tionary approach using the information resulting from
the previous operations.
Our technique takes inspiration from genetic algo-
rithms, even though for this research we considered
only one generation round. Indeed, although in our
particular example the number of sensors is limited,
we propose a general technique (i.e. with any number
of sensors, eventually heterogeneous). To make sure
that computation is performed in a predefined amount
of time, we applied both crossover and mutation op-
erators on the population every time until at least one
sensor acquires new information.
2.1.1 Coding
The population has 30 individuals (possible solu-
tions). Each individual consists of a vector of 5 genes,
where each gene represents the information acquired
by a couple of sensors during the last two hours. In
our experiments, we have made sure that all sensors
acquire data every 2 minute, thus a gene is composed
of two vectors of length 60. These values can be tai-
lored to a specific problem.
2.1.2 Fitness Function
To evaluate the goodness of a solution, it is necessary
to define a numerical function that returns a score for
each individual. Let S
k
= {P
1
,P
2
,...,P
n
} be the k-th
individual with n = 5 genes. A gene P
i
= (T
(1)
i
,T
(2)
i
)
contains the times series T
(1)
i
and T
(2)
i
obtained by a
given pair of sensors.
For performance reasons, we pre-calculate the
smoothed version T
(1)
i
and T
(2)
i
of T
(1)
i
and T
(2)
i
as
described in (Guastella et al., 2019); this technique
was modified to provide an estimate for each avail-
able information.
The correlation ρ
i
between T
(1)
i
and T
(2)
i
is:
ρ
k
i
= |d(T
(1)
i
,T
(1)
i
) − d(T
(2)
i
,T
(2)
i
)|, (1)
with the distance between two time series T
a
and T
b
:
d (T
a
,T
b
) =
∑
j∈[1,γ]
|t
(a)
j
−t
(b)
j
|, (2)
where t
(a)
j
and t
(b)
j
are the j-th information of the data
series T
a
and T
b
respectively. The smaller the differ-
ence d between two time series is, the more similar
these two time series are.
The fitness of the individual S
k
is calculated as the
average correlation among the sensors in P
i
∈ S
k
,∀i:
f (S
k
) =
∑
i
ρ
k
i
n
. (3)
Evaluating Correlations in IoT Sensors for Smart Buildings
225

1
3
2
4
START
END
ACTIVE?
PERCEIVE THE
ENVIRONM ENT
APPLY
EVOLUTIONARY
ALGORITHM
SMOO TH
PERCEPTIONS
START
APPLY SAX METH OD
CALCULATE THE
VORONOI
TESSELLATION
CALCULATE DTW
DETERMINE
CORRELATION
END
START
APPLY GAUSSIAN
SMOOTHING
APPLY K-MEANS
END
ANY DATA
WINDOW
LEFT?
NO
NO
CALCULATE
CORRELATION
START
GET WAVELET
COEFF. AT LEVEL 2
APPLY OTSU’S
THRESHOLDING
END
ANY DATA
WINDOW
LEFT?
NO
YES
CALCULATE
CORRELATION
YES
NO
Figure 1: Main steps of the proposed evolutionary method for determining highly correlated pairs of sensors in smart buildings
(diagram ¶) and the pipelines used for comparing the obtained results (diagrams ·–¹).
2.1.3 Crossover Operator
We have implemented the crossover operator as a sin-
gle cut in a random position. This operator returns
two child individuals and to maintain constant the size
of the population over the entire process, both parents
and children are compared together according to their
fitness and only the best two of them are chosen to
remain in the population. This means that the possi-
bility that only parents remain alive is allowed, com-
pared to other techniques that replace parents with
offsprings, anyhow and independently of their qual-
ity. The two individuals to be recombined are chosen
randomly: this does not exactly follow the evolution-
ary strategy but allows a greater gene variability.
2.1.4 Mutation Operator
To ensure the effective exploration of the solution
space, the mutation operator replaces the sensors
coded by a gene with a couple of sensors taken ran-
domly. This operator is carried out on just one gene
per individual.
2.1.5 Output
This evolutionary approach is carried out on a single
generation and identifies quickly couples of possible
correlated sensors. There is no guarantee of obtain-
ing a correct solution, but this can be used as pre-
processing for the criteria described in 3.2.
2.2 SAX-based Technique
The SAX-based pipeline (diagram · in Fig. 1) uses
moving average filtering, the SAX method (Wang
et al., 2019), Voronoi tessellation (Guastella
and Valenti, 2018) and Dynamic Time Warping
(DTW) (Kenji Iwana and Uchida, 2020). Fig. 2
shows the individual steps of the technique. The first
step involves acquiring raw data from sensors, the
noise is then removed by applying a moving average
filter of size 3 to each time series. The SAX method
is applied to transform the denoised time series into
strings, where the number of alphabets is limited
to 20 symbols; this transformation emphasizes the
differences between the values perceived by the
sensors. Each time series, now converted to a string,
is then compared with the others to determine the one
that minimizes the correlation according to the DTW
distance. The Voronoi tessellation is then applied to
determine the rough relevance area sensors.
Let T S = (ts
1
,ts
2
,...,ts
n
) be the set of time series
obtained each one from the sensors S = (s
1
,s
2
,...,s
n
).
Each time series has been denoised and transformed
to string. A sensor s
i
is correlated with a sensor s
k
∈
S if their Voronoi regions are adjacent and the DTW
distance between their time series is minimized:
argmin
k
(DTW(ts
i
,ts
k
) ∧ ad j(R
s
i
,R
s
k
)) (4)
where DTW(ts
i
,ts
k
) calculates the DTW between the
time series ts
i
and ts
k
respectively, ad j is a boolean
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
226

Indoor and
Outdoor
Raw Sensor
Data
(I1, O1)
Symbolic
Aggregate
Approximation
(SAX) of (I1)
‘qsssrpoiigfedddddcc’
Symbolic
Aggregate
Approximation
(SAX) of (O1)
‘roqtsqqqmhffeeeeddcb’
Correlation is
measured by
using DTW
and Voronoi
tessellation
Outdoor sensor is
correlated with
indoor sensor
Noise
Removal
Moving
Average
Filter
(n = 3)
Figure 2: Block Diagram of SAX pipeline.
operator that returns 1 if the Voronoi regions R
s
i
and
R
s
k
, associated to s
i
and s
k
respectively, are adjacent.
2.3 Gaussian Smoothing based
Technique
This technique (diagram ¸ in Fig. 1) is based on the
Gaussian smoothing technique (Pociask et al., 2018)
and k-means clustering (Shyr-Shen et al., 2018): this
technique calculates the difference between time se-
ries and smoothed versions, then measures the cor-
relation and apply a k-means clustering algorithm to
separate the sensors into two groups (i.e. indoor, out-
door).
This pipeline iterates all the possible pairs of sen-
sors. Furthermore, for each pair the algorithm iterates
on blocks of 60 information.
Let D
i
and D
k
be the windows of 60 information
perceived from the sensors s
i
and s
k
respectively, rel-
ative to the time interval T = [t − 60,t] with |T | = 60.
The method calculates two sets DS
i
and DS
k
where
ds
`
i
∈ DS
i
and ds
`
k
∈ DS
k
are computed by smooth-
ing ` times the data windows D
i
and D
k
through the
Gaussian algorithm. The pipeline evaluates a set C of
correlations, where c
`
∈ C is the average of the dif-
ferences between D
i
and D
k
and the respective data
windows obtained by applying ` times the Gaussian
smoothing:
c
`
=
∑
(|ds
`
i
− D
i
| − |ds
`
k
− D
k
|)
w
(5)
where ds
`
i
∈ DS
i
and ds
`
k
∈ DS
k
are the smoothed time
series, D
i
and D
k
the input time series, w = |D
i
| =
|D
k
| = 60 is the number of samples for each time se-
ries. The standard deviation of the values in C is the
output of the method for a pair of sensors s
i
and s
k
.
Compared to the previous pipeline, this approach
process data from 60 samples, consecutive in time, at
each iteration. We examine all the possible pairs of
sensors for each sample window: this requires a sig-
nificant amount of computational time. Although the
algorithm has determined 89% of correctly correlated
pairs, the computational time (∼ 82 seconds) makes
the application unsuitable in real-time contexts.
2.4 Pearson based Technique
The third technique (diagram ¹ in Fig. 1) uses the
Pearson correlation coefficient (Zhou et al., 2016) and
wavelet decomposition (Sciortino et al., 2017). This
pipeline iterates over windows of 60 samples, con-
secutive in time, for all the pairs of sensors in the
dataset. The first step of the pipeline consists of cal-
culating the wavelet decomposition of both the time
series of the current sensors pair. The wavelet decom-
positions are calculated for the entire time series us-
ing the Daubechies wavelet of order 4 (Vonesch et al.,
2007). Then, the technique iterates all the informa-
tion in blocks of 60 samples for all pairs of sensors.
At each iteration, the Pearson coefficient is calculated
from the wavelet coefficient of both current sensors,
for the current time instant. This process results in a
set P = {ρ
1
,ρ
2
,...,ρ
n
} of Pearson coefficients. The
correlation between the two sensors considered in the
current iteration is calculated as the average of the
Pearson coefficients in P. The last step consists of de-
termining which pairs of sensors have been correctly
matched. For this purpose, we applied Otsu’s thresh-
olding technique (Otsu, 1979) to calculate a thresh-
old value used to separate the set of correlation coef-
ficients into two separate and disjointed classes. Once
the two classes have been obtained, we identified the
correctly matched pairs of sensors.
The computation time of this technique is lower
as compared to the previous ones (it takes just about
1 second to process the entire dataset), however, the
accuracy is considerably lower, returning only 22% of
correctly matched pairs.
3 EXPERIMENTAL RESULTS
This section introduces the real dataset acquired by
the sensors installed at SMART Facility building and
the results obtained through the techniques described
in the previous section. We also carry out a descrip-
tion of the evaluation methods used to compute the
results obtained by the presented pipelines along with
their computational times.
3.1 Experimental Context
The SMART Infrastructure Facility at the Univer-
sity of Wollongong (NSW, Australia) was created in
2011 and provides specialist laboratories for 150 staff
and 200 postgraduate research students. In 2018, the
Evaluating Correlations in IoT Sensors for Smart Buildings
227

Table 1: Temperature and Humidity comparison for both
Pycom and Droplet modules.
Pycom Droplet
Temperature
variation
-40 to +85 ±0.5
◦
0 to +65 ±1
◦
Humidity
variation
0 to 100%, ±0.4% RH 0 to 100%, ±0.008% RH
SMART building was equipped with a Droplet envi-
ronmental sensors from Nube iO. A fleet of 140 sen-
sors monitor the temperature, humidity, atmospheric
pressure and movements inside every room inside the
Facility. These sensors are battery-powered and rely
on a long-range, low-power LoRaWAN network for
transmitting data every 30 second. The LoRaWAN
gateway then transmits the incoming data to a local
database. The physical location of the sensors and the
open database containing all the records and time se-
ries data for 2019 are publicly available and detailed
in (Barth
´
elemy et al., 2020b).
In addition to the existing Droplet sensors, mobile
sensors based on a Pycom platform have been devel-
oped to capture the temperature and humidity at a dif-
ferent location inside and outside the building, every 2
minutes, thus enabling the rapid deployment and test-
ing of different scenarios. Fig. 3 show the two types
of sensors used for the experimentation.
Figure 3: Pycom (left) and Nube iO (right) modules.
The comparison between the Droplet and Pycom
modules for both temperature and humidity is shown
in Table 1, as these are the two data parameters used
to perform correlation in this study.
The scenario investigated in this work is illus-
trated in Fig. 4 and the data used in the remainder
of this Section is also open and available (Barth
´
elemy
et al., 2020a).
3.2 Evaluation Method
The techniques proposed in the previous section have
been validated in four versions of the presented
dataset, containing different sensors deployed in the
SMART Facility building:
• Test #1: all sensors, inside and outside the build-
1st floor
2nd floor
Figure 4: Location of the Droplet sensors (blue squares) and
Pycom sensors (red dots). B35, B2 and B3 are the buildings
next to the SMART Facility Building.
ing;
• Test #2: only the sensors inside the building;
• Test #3: external sensors and internal sensors on
the 2nd floor;
• Test #4: external sensors and internal sensors on
1st floor.
The techniques have been evaluated on the same set
of sensors but using different types of information: (i)
temperature, (ii) humidity and (iii) temperature and
humidity combined. This last combination was tested
only on the evolutionary technique because it is the
only one capable of integrating heterogeneous infor-
mation.
We used two criteria to assess the quality of the
solution obtained by the proposed techniques:
1. a pair of sensors are considered to be correctly
matched if their information is correlated and both
its sensors are internal or external to the building;
2. a pair of sensors are considered to be correctly
matched if their information is correlated and the
relevant areas of sensors are close one to each
other.
The second criterion introduces a spatial proximity
constraint: rather than evaluating if two sensors are
both inside or outside the building, we calculate the
Voronoi tessellation by using the position of sensors,
which results in a set of adjacent regions that repre-
sent the rough relevance area of sensors. Two sen-
sors are paired if their Voronoi regions overlap. In the
case of the anomalous behavior of a sensor, this infor-
mation can be useful for the domain expert to choose
the sensor characterized by both spatial proximity and
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
228

Table 2: Percentage of sensor pairs correctly matched by
the evolutionary technique.
Temperature Humidity
Temperature&
Humidity
Test #1 56,67% 63,33% 70,00%
Test #3 46,67% 50,00% 36,67%
Test #4 46,67% 40,00% 56,67%
high informative correlation. For each technique, the
following steps are carried out to determine whether
two sensors are correlated according to this criterion:
1. compute the Voronoi regions separately for the
sensors of each floor of the building;
2. determine the pairs of sensors using one of the de-
scribed techniques;
3. stack the Voronoi tessellations calculated for each
floor;
4. check if the two regions overlap: if so, sensors are
correlated.
We applied this criterion to the techniques that use the
evolutionary approach, the Pearson correlation and
Gaussian smoothing. The SAX-based method already
uses the Voronoi tessellation to determine whether
sensors are correlated or not.
3.3 Experimental Results
Table 2 shows the percentage of correct pairs obtained
by the proposed evolutionary method using the first
criterion. The average of the percentages of correct
pairs for each information is as follows: 62, 50% for
temperature, 63,33% for humidity and 65.83% for
combined temperature and humidity. The results ob-
tained by combining temperature and humidity are on
average better (about 3%) than the tests conducted us-
ing only information of the same type. Moreover, we
report a better accuracy in test #1 using heterogeneous
information: the 70% of correlated pairs have been
correctly matched by the evolutionary method.
The evolutionary approach carries out a single
evolution by using the data acquired from the last 2
hours, as described in Section 2.1. This behavior is
contradictory to the normal functioning of the genetic
algorithms, where usually a variable number of evo-
lutions are carried out on a single set of information.
By using the SAX-based technique, 85% of pairs
of sensors are considered as correct. However, this
pipeline requires about 15 seconds to compute the re-
sult. Concerning other methods, the SAX-based tech-
nique determines pairs of sensors that have a high cor-
relation and spatial proximity thanks to the Voronoi
tessellation. The technique based on Pearson correla-
tion provided an average of 48% of correlated sensor
pairs using temperature and humidity separately. The
Table 3: Percentage of sensor pairs correctly matched by
the described techniques.
Temperature Humidity
SAX 87,80% 78,05%
Pearson 17,32% 6,95%Test #1
Gaussian F. 28,42% 23,66%
SAX 91,67% 83,33%
Pearson 21,74% 9,42%Test #3
Gaussian F. 42,75% 48,19%
SAX 82,35% 82,35%
Pearson 12,50% 2,94%Test #4
Gaussian F. 24,27% 21,32%
Table 4: Computational times, in seconds, required by the
described techniques. The evolutionary is omitted,as it re-
quires about 50ms on average to compute a solution.
Temperature Humidity
SAX 24,25 26,05
Pearson 2,1803 2,1775Test #1
Gaussian F. 146,3742 156,1693
SAX 20,24 20,22
Pearson 1,3 1,4Test #2
Gaussian F. 107,0053 100,1627
SAX 10,87 11,45
Pearson 0,7 0,8Test #3
Gaussian F. 50,7633 51,0154
SAX 7,1 5,33
Pearson 0,3 0,3Test #4
Gaussian F. 25,3639 23,8448
technique based on the Gaussian smoothing provided
an average of 37,89% of correctly correlated sensor
pairs using temperature, 54,09% using humidity.
Table 3 shows the percentage of correct pairs ob-
tained by these techniques. These results refer to the
pairs obtained by applying the first criterion, based
on the choice of sensors both indoor or outdoor. This
means that the 85% of pairs of sensors returned by
this pipeline contain either indoor sensors or outdoor
sensors.
Table 4 reports the computational time required by
the described techniques to determine the pairs of cor-
related sensors. The table does not list the computa-
tional time for the evolutionary method: this requires
about 50ms to evaluate a solution for each test and
each combination of data.
Among the techniques listed in Table 4, the
one based on the Pearson correlation requires the
least computational time. However, this depends on
the number of sensors available in the environment:
test #1 requires about 2 seconds whereas test #4,
which considers a smaller set of sensors, requires 0,3
seconds. Therefore, these techniques are not suitable
for real-time applications.
Table 5 shows the percentage of pairs obtained by
applying the Voronoi proximity criterion on the so-
lutions obtained by all the described techniques. For
each pair of sensors, we verify whether the Voronoi
Evaluating Correlations in IoT Sensors for Smart Buildings
229

regions of the two sensors, which describe their rel-
evance area, are in proximity; if so, the sensors are
considered as correlated. The remaining pairs of sen-
sors are omitted.
This criterion filtered a small number of pairs in
the case of test #4 for Pearson and Gaussian smooth-
ing based techniques. This is because the number of
pairs obtained by the two techniques is small for this
test, and many of these resulting correlated pairs are
also spatially correlated. The test #4 using the Pear-
son based method on humidity data returns 100% of
pairs using the Voronoi criterion: only four pairs of
sensors were determined by the technique, all charac-
terized by spatial proximity.
On average, the application of the proximity crite-
rion has filtered a greater number of pairs obtained by
the evolutionary technique, about 14%, against 24%
of the technique based on Gaussian smoothing and
30% of the technique using the Pearson correlation.
This is because the evolutionary solution determines
a limited number of sensor pairs, while the other tech-
niques evaluate the correlation for each possible pair
of sensors, resulting in large sets of pairs. Despite
this, having a limited set of sensor pairs that have not
only a high informative correlation but also physical
proximity enables the domain expert to easily choose
which sensor to use in case of unexpected anomalous
situations, where the information of a malfunctioning
sensor can be compensated by the one of the corre-
lated and also near sensor.
For illustrative purposes, Fig. 5 shows the two
floors of SMART Facility building and the Voronoi
regions resulting from the spatial proximity criterion
on the test #1 using the evolutionary technique. We
observe that the method determines pairs of sensors
situated outside and inside the building and that are in
immediate proximity.
Figure 5: Voronoi tessellation resulting from the proximity
criterion on test #1 using the evolutionary technique. Over-
lapping regions, depicted as the colored polygons, deter-
mine relevance area of correlated sensors.
Table 5: Pairs of sensors determined through the Voronoi
criterion on the results obtained by the described tech-
niques.
Temperature Humidity
Temperature&
Humidity
Pearson 21,83% 14,04%
Gaussian F. 17,60% 19,59%
X
Test #1
Evolutionary 16,67% 6,67% 13,33%
Pearson 26,19% 12,77%
Gaussian F. 19,01% 23,04%
X
Test #2
Evolutionary 6,67% 20,00% 26,67%
Pearson 23,33% 7,69%
Gaussian F. 16,95% 19,70%
X
Test #3
Evolutionary 6,67% 10,00% 3,33%
Pearson 35,29% 100,00%
Gaussian F. 42,42% 37,93%
X
Test #4
Evolutionary 16,67% 26,67% 20,00%
The experiments were carried out on a machine
equipped with i7-7820HQ, 32GB RAM and Win-
dows 10. The computation takes less than 50ms to
generate a new solution for a set of information per-
ceived at a given time instant. The techniques were
coded in MatLab language without particular opti-
mizations; the evolutionary method was coded in Java
language.
3.4 Discussion
The techniques described show promising results on
the data acquired from the SMART Facility build-
ing. Among these, the evolutionary one allows ad-
dressing the need for on-line computation. Moreover,
this evolutionary approach allows integrating hetero-
geneous, which is not possible by the other presented
techniques. The criteria to determine the correlation
between sensors allow the domain expert to extract
useful knowledge from the perceived information: by
using the criterion based on the correlation between
internal or external sensors it is possible to deter-
mine the topography of the environment, discriminat-
ing between internal and external sensors. The crite-
rion based on Voronoi tessellation allows determining
pairs of sensors that have a high informative correla-
tion and are also in the immediate vicinity. In both
cases, the proposed techniques produced a significant
amount of pairs of sensors.
The evolutionary approach determines pairs of
sensors in real-time by using blocks of consecutive
information in time. Nevertheless, the algorithm car-
ries out only one evolution for each data window; this
does not allow the achievement of an optimum. Using
the evolutionary technique determines pairs of sen-
sors whose correlation persists in time: at time instant
t, the pairs of sensors determined by the technique are
those that have survived the selection, therefore their
correlation is high over time. The advantage of the
evolutionary technique over the others also concerns
ICAART 2021 - 13th International Conference on Agents and Artificial Intelligence
230

the computation time: this requires on average 50ms
to compute a solution, while the SAX method requires
15 seconds.
4 CONCLUSIONS
The contribution of this article is twofold: firstly, we
present a novel and open dataset of environmental
information acquired from indoor and outdoor sen-
sors deployed at SMART Infrastructure Facility at the
University of Wollongong, Australia; secondly, we
present a novel evolutionary approach to determine
in real-time the correlation between pairs of different
sensors.
This correlation is computed by using the infor-
mation coming from the sensors and two criteria: the
former based on the presence of both sensors inside
or outside a building, the latter on the spatial distance
among the sensors themselves. We experimentally
verify that this allows determining accurately corre-
lated pairs of sensors.
These techniques can be applied for different
smart building applications such as environmen-
tal monitoring to optimize energy consumption or
anomalies detection. In a real environment, sensors
can be subject to unpredictable anomalies that cause
missing information. A domain expert could be able
to understand if a given sensor is malfunctioning or
otherwise there is an emergency by using the pro-
posed system. For example, a high temperature could
indicate a malfunction, rather than the presence of
fire. In this case, the system should assist the domain
expert to determine whether a sensor is malfunction-
ing or there is an emergency.
We plan to extend our research by integrating
more information (e.g. luminosity or noise) and by
determining the state of devices (e.g. determining if a
door is open or closed).
REFERENCES
Ashby, W. R. (1991). Requisite variety and its implications
for the control of complex systems. In Facets of Sys-
tems Science, pages 405–417. Springer.
Barth
´
elemy, J., Arshad, B., Verstaevel, N., Guastella, D.,
and Perez, P. (2020a). Smart building additional data.
Barth
´
elemy, J., Arshad, B., Verstaevel, N., and Perez, P.
(2020b). Smart infrastructure facility building data.
Guastella, D. A., Camps, V., and Gleizes, M.-P. (2019). Es-
timating missing environmental information by con-
textual data cooperation. In Baldoni, M., Dastani,
M., Liao, B., Sakurai, Y., and Zalila Wenkstern, R.,
editors, PRIMA 2019: Principles and Practice of
Multi-Agent Systems, pages 523–531. Springer. doi:
10.1007/978-3-030-33792-6 37.
Guastella, D. A. and Valenti, C. (2018). Estimating miss-
ing information by cluster analysis and normalized
convolution. In 2018 IEEE 4th International Forum
on Research and Technology for Society and Industry
(RTSI), pages 1–6. doi: 10.1109/RTSI.2018.8548454.
Houssin, M., Combettes, S., Gleizes, M.-P., and Lartigue,
B. (2020). SANDMAN: a Self-Adapted System for
Anomaly Detection in Smart Buildings Data Streams.
In (to appear in) Proceedings of the 18th Adaptive
Computing (and Agents) for Enhanced Collaboration
(ACEC) at WETICE 2020.
Kenji Iwana, B. and Uchida, S. (2020). Time series
classification using local distance-based features in
multi-modal fusion networks. Pattern Recognition,
97:107024.
Nigon, J., Verstaevel, N., Boes, J., Migeon, F., and Gleizes,
M.-P. (2017). Smart is a matter of context. In Interna-
tional and Interdisciplinary Conference on Modeling
and Using Context, pages 189–202. Springer.
Otsu, N. (1979). A threshold selection method from gray-
level histograms. IEEE Transactions on Systems,
Man, and Cybernetics, 9(1):62–66.
Pociask, E., Malinowski, K. P.,
´
Sle¸zak, M., Jaworek-
Korjakowska, J., Wojakowski, W., and Roleder, T.
(2018). Fully automated lumen segmentation method
for intracoronary optical coherence tomography. Jour-
nal of Healthcare Engineering, 2018:1–13.
Sciortino, G., Tegolo, D., and Valenti, C. (2017). Auto-
matic detection and measurement of nuchal translu-
cency. Computers in Biology and Medicine, 82:12–
20.
Shyr-Shen, Y., Shao-Wei, C., Chuin-Mu, W., Yung-Kuan,
C., and Ting-Cheng, C. (2018). Two improved k-
means algorithms. Applied Soft Computing, 68:747
– 755.
Vonesch, C., Blu, T., and Unser, M. (2007). Generalized
Daubechies wavelet families. IEEE Transactions on
Signal Processing, 55(9):4415–4429.
Wang, L., Lu, F., Cui, M., and Bao, Y. (2019). Survey of
methods for time series symbolic aggregate approxi-
mation. In Cheng, X., Jing, W., Song, X., and Lu, Z.,
editors, Data Science, pages 645–657. Springer.
Zhou, H., Deng, Z., Xia, Y., and Fu, M. (2016). A new
sampling method in particle filter based on pearson
correlation coefficient. Neurocomputing, 216:208 –
215.
Evaluating Correlations in IoT Sensors for Smart Buildings
231