An Internet of Things Sensor Array For Spatially A

sensors
Article
An Internet of Things Sensor Array for Spatially and
Temporally Resolved Indoor Climate Measurements
Alexander Rusch * and Thomas Rösgen
Institute of Fluid Dynamics, ETH Zurich, 8092 Zurich, Switzerland

* Correspondence: arusch@ethz.ch
Abstract: The COVID-19 pandemic has emphasized the need for infection risk analysis and assess-
ment of ventilation systems in indoor environments based on air quality criteria. In this context,
simulations and direct measurements of CO2 concentrations as a proxy for exhaled air can help to
shed light on potential aerosol pathways. While the former typically lack accurate boundary condi-
tions as well as spatially and temporally resolved validation data, currently existing measurement
systems often probe rooms in non-ideal, single locations. Addressing both of these issues, a large and
flexible wireless array of 50 embedded sensor units is presented that provides indoor climate metrics
with configurable spatial and temporal resolutions at a sensor response time of 20 s. Augmented by
an anchorless self-localization capability, three-dimensional air quality maps are reconstructed up to
a mean 3D Euclidean error of 0.21 m. Driven by resolution, ease of use, and fault tolerance require-
ments, the system has proven itself in day-to-day use at ETH Zurich, where topologically differing
auditoria (at-grade, sloped) were investigated under real occupancy conditions. The corresponding
results indicate significant spatial and temporal variations in the indoor climate rendering large
sensor arrays essential for accurate room assessments. Even in well-ventilated auditoria, cleanout
time constants exceeded 30 min.
Keywords: COVID-19 pandemic; CO2 measurement; airborne transmission; indoor climate sensing;
Citation: Rusch, A.; Rösgen, T. An internet of things; wireless sensor array; ultra-wideband ranging
Internet of Things Sensor Array for
Spatially and Temporally Resolved
Indoor Climate Measurements.
Sensors 2022, 22, 4377. https:// 1. Introduction
doi.org/10.3390/s22124377
The challenges posed by the COVID-19 pandemic have highlighted the need for
Academic Editor: further development of disease-spreading models, in this case, specifically for airborne
Michalis Michaelides virus transmission via aerosol transport.
Received: 10 May 2022
Considering indoor scenarios, in particular, the employment of detailed computa-
Accepted: 6 June 2022
tional fluid dynamics (CFD) simulations [1–3], simpler zonal [4–6], or even multizone
Published: 9 June 2022
simulations [7–11] appears feasible and may help to solve some of the specific ventilation
issues. The capture of turbulent flow features at the full resolution scale, however, remains
Publisher’s Note: MDPI stays neutral
a limiting feature, and some degree of physical approximation is usually required [12].
with regard to jurisdictional claims in
This, in turn, generates a need for code calibration and adaptation to the scenario under
published maps and institutional affil-
investigation. For complex settings including forced ventilation, simulations and theoretical
iations.
models show large discrepancies [4,13] such that experimental measurements present an
attractive and necessary tool for validation.
Local CO2 concentrations are directly linked to human exhalation and serve as a
Copyright: © 2022 by the authors.
marker for the transport of aerosols with a viral load [14,15]. Especially indoors, infection
Licensee MDPI, Basel, Switzerland. risks can be estimated from measurements of CO2 concentrations at inhalation height [16].
This article is an open access article Numerous indoor air quality (IAQ) measurement systems already exist and typically
distributed under the terms and use nondispersive infrared (NDIR) sensors. The presented work also utilizes that very
conditions of the Creative Commons sensor technology due to a trade-off between accuracy on the one hand and availability,
Attribution (CC BY) license (https:// as well as unit price, on the other hand. It must be noted, though, that recent advance-
creativecommons.org/licenses/by/ ments in gas sensor technology promise further improvements in terms of sensitivity and
4.0/). selectivity [17,18].
Sensors 2022, 22, 4377. https://doi.org/10.3390/s22124377 https://www.mdpi.com/journal/sensors

Sensors 2022, 22, 4377 2 of 23
Most indoor climate sensing systems focus on “smart monitoring” [19,20] in the context
of multi-room settings [20–28], e.g., they provide room-specific air quality warnings [29]
via webserver/smartphone or ventilation control systems. These internet of things (IoT)
devices not only enable the direct monitoring and control of IAQ on an individual smart
home basis [30,31] but present a means of generating extensive cloud datasets that are
analyzed to understand and optimize air flow paths and, ultimately, improve living [32]
and working [33] conditions at a generalized scale. Such big data approaches, for example,
utilize cloud processing to extract insights about air quality by applying machine learning
techniques [32,33].
Existing indoor sensor networks are typically not aware of their mesh topology such
that data are only visualized and inspected on a per-node basis, and manual mapping of
sensor locations to the respective sampled datasets is necessary [26–29,34–36]. Furthermore,
rooms are frequently only sampled in a single [21,24] or very few locations [22,35] instead
of resolving entire rooms spatio-temporally with high resolution. Therefore, such systems
are insufficient to capture the spatio-temporal complexity of natural and forced convective
indoor flows. Furthermore, the respective measurement devices are often mounted in
locations not representative of inhalation such as directly to a wall or ceiling. However,
CO2 spreading in large rooms is inhomogeneous, inherently unsteady, and can feature sig-
nificant stratification [16], which renders single-point measurements in non-ideal locations
inadequate to assess infection risks [37]. Furthermore, the effectiveness of countermeasures
and ventilation systems can only be evaluated from measurements that capture both spatial
and temporal trends with high resolution.
To address these issues, in the scope of this work, a large and scalable 50-unit sensor
array has been developed that provides spatially and temporally resolved room climate
measurements in typical indoor environments, for example, seminar rooms and larger
lecture halls with or without active ventilation. The system is designed for temporary
deployment as a high-resolution measurement instrument rather than being permanently
installed as part of smart building infrastructure. The battery-powered wireless measure-
ment units can be flexibly distributed in a target room and placed at inhalation height.
Besides the local CO2 concentration, temperature, relative humidity, and barometric pres-
sure are measured. Despite being developed during the COVID-19 pandemic, the sensor
array addresses the much wider topic of general indoor air quality measurement and
assurance [20–27,29,34,38].
For large rooms and arbitrary sensor unit placement, reconstructing the 3D locations
of all 50 sensor units manually is non-trivial, tedious, and error-prone. To bypass this
problem, the sensor nodes are augmented with an anchorless ultra-wideband (UWB)
ranging capability. Fused with an accurate height profile determination from calibrated
barometric pressure readings, the sensor mesh topology is reconstructed with a mean
Euclidean 3D position error of 0.21 m.
In the following, the design considerations, technical implementation, and measure-
ment results for the 50-node IoT sensor array are described. The system was deployed
during the COVID-19 pandemic to dynamically monitor the room climate and CO2 concen-
trations in various auditoria at ETH Zurich. Therefore, the presented results correspond to
measurements under real occupancy conditions. The data indicate that, in particular, the
time constants of the changes in the room climate have to be considered carefully.
The developed indoor sensor network is the first to use anchorless UWB ranging fused
with barometric sensing for network topology awareness enabling the reconstruction of 3D
indoor climate maps. Existing non-location-aware sensor networks visualize the data of the
sensor nodes in the form of individual graphs. For large networks, this is impractical and
does not allow for the easy detection of multi-scale spatio-temporal trends. In contrast, the
proposed sensor network enables the creation of 3D evolution maps that provide intuitive
insights into large datasets even for rooms with complicated topologies. Furthermore, a
novel indirect approach for calibrating the CO2 sensors of all measurement nodes via a
single reference unit is implemented. This work is further distinguished from most other
Sensors 2022, 22, 4377 3 of 23
indoor climate sensor networks by its size, the use of fully customized commercial sensing
nodes, and the employed IoT approach that does not stream any data to the internet/into
a third-party cloud but gathers and processes the information using only a local network
and a fully (physically and virtually) operator-controlled server backend. To the best
of the authors’ knowledge, no such large, flexible, robust, and location- as well as data
privacy-aware IAQ sensing array has previously been reported.
2. Materials and Methods

The design and implementation of the sensor array were driven by a series of spe-
cific requirements, geared towards repeated application in different, possibly unknown
indoor environments.
2.1. Resolution
The requirements on resolution include the spatial, temporal, and signal domains.
It was considered sufficient to place sensing nodes at approximately 2 m distance for
“high-resolution” measurements, comparable to typical values for social distancing. With a
target size of the array of 50 nodes, this enables a regular coverage of around 200 m2 .
Similarly, a temporal resolution of one sample per two minutes and node was con-
sidered useful. This figure relates to the need for also recording concentration changes
during the room emptying phase where the effectiveness of the ventilation system is of
particular interest.
Finally, the sensitivity of the individual nodes should be sufficient to detect—at least
locally—the CO2 emissions of a single person. Assuming a production per person of
approximately 0.7 g/min and a local “observation volume” of 10 m3 , this translates into
a value of 57 ppm/min. This figure has to be compared to the noise level and temporal
stability of the actual CO2 sensors employed in the sensing nodes.
As will be detailed later, the sensing node functionality was extended to include a spa-
tial ranging capability to automate the process of reconstructing the sensor node placement.
Again, the resolution of the UWB ranging modules and the additional barometric height
sensing was chosen to be better than 1 m for accurate position measurements.
2.2. Remote Operation

The second group of requirements was derived from the need to quickly place and
operate the array in different environments with changing configurations.
This led to a design of the embedded nodes for decentralized, “headless” operation
with a wireless transmission capability, receiving configuration/control information from
and sending measurements to a remote server. Each node would have to include WiFi trans-
mission capability, and a communication protocol compatible with the existing wireless
network had to be chosen. The details of the resulting IoT constellation will be discussed
further below. Importantly, the client–server architecture safeguards the array’s operation
against the failure of individual measurement nodes.
Furthermore, the nodes’ power management should enable operation over extended
periods of time (e.g., several days continuously) to monitor long-term trends. A large
rechargeable battery was to be included in each node, and the hardware components had to
be selected to facilitate a low-power sleep state between measurements. The sensor nodes
were specifically designed as battery-powered devices since power adapters would have
seriously downgraded remote operability and configuration flexibility, especially for the
large number of measurement nodes involved.
In addition to these requirements that translate directly into hardware selection criteria,
there were further design considerations that had an impact primarily on the software
design. One was the scalability of the array, implying that the number of nodes to be
operated should be easy to increase (or decrease) without any reprogramming required.
Furthermore, provisions had to be included to recalibrate the sensing units remotely so as
to maintain consistency of the readings across all sensors.
Sensors 2022, 22, x FOR PEER REVIEW 4 of 23
Sensors 2022, 22, 4377 4 of 23

be operated should be easy to increase (or decrease) without any reprogramming re-
quired. Furthermore, provisions had to be included to recalibrate the sensing units re-
motely
2.3. so as to
Hardware maintain consistency of the readings across all sensors.
Architecture
The sensor array structure will be discussed considering two points of view: Sec-
2.3. Hardware Architecture
tion 2.3.1 describes the measurement system from a macro perspective including auxiliary
The sensor array structure will be discussed considering two points of view: Section
systems while Section 2.3.2 goes into detail on the composition of the individual sen-
2.3.1 describes the measurement system from a macro perspective including auxiliary sys-
sor units.
tems while Section 2.3.2 goes into detail on the composition of the individual sensor units.
2.3.1. IoT Mesh Structure
2.3.1. IoT Mesh Structure
The measurement system consists of three functional blocks:
The measurement system consists of three functional blocks:
1. A flexibly configurable number of independently and asynchronously operating
1. A flexibly configurable number of independently and asynchronously operating em-
embedded measurement nodes. Currently, 50 units are available.
bedded measurement nodes. Currently, 50 units are available.
2. A reference measurement unit that provides calibration values for the measurement
2. A reference measurement unit that provides calibration values for the measurement
nodes. This is necessary as the installed CO2 sensors require recalibration at the
nodes. This is necessary as the installed CO2 sensors require recalibration at the be-
beginning of each measurement campaign for best accuracy. Section 2.5.1 focuses on
ginning of each measurement campaign for best accuracy. Section 2.5.1 focuses on
the CO2 sensor calibration procedure. The reference unit operates sensor hardware
the CO2 sensor calibration procedure. The reference unit operates sensor hardware
identical
identical to
to the measurementnodes.
the measurement nodes.
3.3. AAserver acting as a backend
server acting as a backend thatthat provides
providesremote
remotemanagement
management functionalities
functionalities for for
the sensor array and logs the measurement data in a database (refer to Section
the sensor array and logs the measurement data in a database (refer to Section 2.4.5). 2.4.5).
Furthermore,
Furthermore, thethe server
serveracts
actsas
asthe
theMessage
MessageQueuing
Queuing Telemetry
Telemetry Transport
Transport (MQTT)
(MQTT)
broker of the IoT mesh (refer to Section 2.4.1).
broker of the IoT mesh (refer to Section 2.4.1).
Figure
Figure11shows
shows the
the entire ensembleof
entire ensemble of5050measurement
measurementnodes nodes asas well
well as as
thethe reference
reference
unit.
unit.Note
Notethat
thatthe
thereference
referenceunit
unit is
is the
the only device with
only device with aa display
displayand
andbuttons.
buttons.AAsimplis-
simplistic
graphical useruser
tic graphical interface (GUI)
interface (GUI)implemented
implemented in in the referenceunit
the reference unitguides
guidesthethe operator
operator
through
throughthethecalibration
calibrationprocess
process(detailed
(detailedin inSection
Section2.5.1).
2.5.1). The
The measurement
measurement nodes
nodes have no
provisions for user
no provisions interaction
for user beyond
interaction beyonda magnetically
a magnetically activated
activatedpower
powerswitch.
switch.
Figure 1. All 50 assembled sensor nodes with the reference unit in front.
Figure 1. All 50 assembled sensor nodes with the reference unit in front.
2.3.2.Measurement
2.3.2. Measurement NodeNode Design
Design
The nodes are based on a modular design, using highly integrated commercial-off-the-
The nodes are based on a modular design, using highly integrated commercial-off-
shelf (COTS) components. A bus system using the Inter-Integrated Circuit (I2C) and Serial Pe-
the-shelf (COTS) components. A bus system using the Inter-Integrated Circuit (I2 C) and
ripheral Interface (SPI) protocols was chosen to connect the individual sensors, enabling fast
Serial Peripheral Interface (SPI) protocols was chosen to connect the individual sensors,
assembly of the components without the need for dedicated interconnect circuitry. The bus
enabling fast assembly of the components without the need for dedicated interconnect
architecture also allows for the later addition of further environmental sensing components
circuitry. The bus architecture also allows for the later addition of further environmental
such as particle size analyzers or volatile organic compound (VOC) detectors [19].
sensing components such as particle size analyzers or volatile organic compound (VOC)
The key components integrated in a single node are, at present:
detectors [19].
The key components integrated in a single node are, at present:
1. A master microcontroller with WiFi transceiver (Espressif Systems ESP32/WROOM32).
2. A CO2 , temperature, and humidity sensor (Sensirion SCD30).
3. A high-precision barometric pressure sensor (Bosch BMP390).
Sensors 2022, 22, 4377 1. 5 of 23

A master microcontroller with WiFi transceiver (Espressif Systems ESP32/WROOM32).
2. A CO2, temperature, and humidity sensor (Sensirion SCD30).
3. A high-precision barometric pressure sensor (Bosch BMP390).
4.
4. A UWB ranging module (Decawave DWM3000).
5.
5. A 6600 mAh Li-ion rechargeable battery pack.
pack.
A block
A block diagram
diagram of
of these
these components
componentsisisshown
shownininFigure
Figure2.2.
IEEE 802.11 WiFi IEEE 802.15.4 UWB
VDC
Li-Ion ESP32 SCD30 BMP390 DWM3000

Battery WROOM32 (CO2, rH, T1) (pabs, T2) (Δxy)
I2C
SPI
Figure 2.
Figure 2. Block
Block diagram
diagram of
of the
the node
node hardware.
hardware.
The
Thespecifications
specificationsofofthe two
the main
two sensor
main components
sensor componentsSCD30 and and
SCD30 BMP390 are detailed
BMP390 are de-
in Table 1. The respective data are extracted from [39–41]. For further specifications, refer
tailed in Table 1. The respective data are extracted from [39–41]. For further specifications,
to these references.
refer to these references.
Table 1. Sensor specifications.
Table 1. Sensor specifications.
Response Time Time Temperature
Response Temperature
Component
Component Measurement
MeasurementRange
Range Accuracy
Accuracy Repeatability
Repeatability (τ63% ) Stability
(τ63%) Stability
1 ◦C
SCD30—CO
SCD30—CO2 2 400–10,000 ppmppm
400–10,000 +/−(30 ppm ppm
+/−(30 + 3% )M.V. 1) +/−10 ppm
+ 3% M.V. +/−10 ppm 20 s 20 s +/− 2.5 ppm/
+/−2.5 ppm/°C
2
SCD30—rH 0–100% +/−3% +/−0.1% 8s N.A.
SCD30—rH 0–100% ◦ +/−3%
◦ +/−0.1% 8s N.A.2
◦C +/−(0.4 C + 0.023 C * (M.V.
SCD30—T
SCD30—T1 1 −40–70
−40–70 °C +/−(0.4 °C1 —25 ◦ C))°C * (M.V. 1—25+/
+ 0.023 −0.1 ◦ C+/−0.1 °C
°C)) 10 s 10 s N.A. 2
N.A. 2
◦C
BMP390—p
BMP390—pabsabs 300–1250
300–1250 hPahPa +/−50+/−50 +/−3 Pa
Pa (abs.);
Pa (abs.); +/−3 Pa (rel.) N.A. 2 N.A. 2
(rel.) N.A. 2 N.A. 2 +/−+/−0.6
0.6 Pa/Pa/°C
BMP390—T −40–85 ◦C ◦C
+/−0.5+/−0.5 N.A. 2 N.A. 2 N.A. 2 N.A. 2
BMP390—T
2 2 −40–85 °C °C N.A. 2 N.A. 2
11 Measured
Measuredvalue. 2 Not
value. 2 Not applicable/not
applicable/not available.
available.
All
All sensors
sensors are,
are, in
in principle,
principle,capable
capableof
of rapid
rapid data
data acquisition.
acquisition. In
In the
the interest
interest of
of power
power
conservation,
conservation, though, the nodes operate currently at a (reconfigurable) sampling rate of
though, the nodes operate currently at a (reconfigurable) sampling rate of one
one
sample every two minutes. The data from each sampling sweep (CO2 concentration, relative
sample every two minutes. The data from each sampling sweep (CO2 concentration, relative
humidity, barometric pressure, temperatures T1 and T2 ) are packaged into a message and
humidity, barometric pressure, temperatures T1 and T2) are packaged into a message and
transmitted to the data server via WiFi using the MQTT protocol (see Section 2.4.1).
transmitted to the data server via WiFi using the MQTT protocol (see Section 2.4.1).
The ranging module is activated and read out only at the beginning of a measure-
The ranging module is activated and read out only at the beginning of a measurement
ment cycle since the nodes’ position is assumed to remain unchanged during the subse-
cycle since the nodes’ position is assumed to remain unchanged during the subsequent meas-
quent measurements. This assumption is not valid during the pressure sensor calibration
urements. This assumption is not valid during the pressure sensor calibration setup, though,
setup, though, where all nodes are temporarily moved to the same height level (refer to
where all nodes are temporarily moved to the same height level (refer to Section 2.5.2).
Section 2.5.2).
2.4. Software
2.4. Software Implementation
Implementation
Themeasurement
The measurementnodes’nodes’ firmware
firmware waswas designed
designed withwith special
special emphasis
emphasis on uni-
on uniformity
formity such that all units run the same software differing only by a node-specific
such that all units run the same software differing only by a node-specific identifier (ID). identi-
At
fier (ID). At the same time, the software enables the independent operation of
the same time, the software enables the independent operation of all nodes. This facilitatesall nodes.
Thisflexible
the facilitates the flexible
addition addition
or removal or removal
of sensor nodesofused
sensor
in anodes used in acampaign
measurement measurement
and
campaign
eases and eases
the future theof
process future
scalingprocess
up theofarray.
scaling up the array.
Figure 3 provides a simplified functional overview of the program structure of the
measurement nodes. Details regarding the different functionalities of the software will be
discussed in the subsequent sections.
Sensors 2022, 22, 4377 Figure 3 provides a simplified functional overview of the program structure 6ofof the
23
measurement nodes. Details regarding the different functionalities of the software will be
discussed in the subsequent sections.
Establish communication (I2C, SPI, serial) & enable watchdog
Yes Initial No
run?
Configure SCD30,
BMP390 & DWM3000
Upon RTC timer timeout, wake via hardware interrupt

React to management
Startup/Initialization
commands (if any)
Receive reference values

& calibrate SCD30
Perform UWB ranging
Publish ranging data
Sample CO2, rH, T1, T2 & pabs
Publish measurement data

Sampling
React to management
commands (if any)
Disable watchdog
Enter deep sleep
Figure3.3.Program
Figure Programsequence
sequenceofofthe
themeasurement
measurementnode
nodesoftware.
software.
2.4.1.
2.4.1.MQTT
MQTTStructure
Structure
MQTT
MQTTwas waschosen
chosenasasthe thenetwork
networkprotocol
protocoltototransmit
transmitdata datafrom
fromthethemeasurement
measurement
units
units to the server backend and to implement management functionalities for sensor
to the server backend and to implement management functionalities for the ar-
the sensor
ray. MQTT
array. MQTT adheres to a to
adheres strict client–server,
a strict client–server,publish–subscribe
publish–subscribe paradigm
paradigmwherewhere
all network
all net-
traffic
work passes through
traffic passes the server,
through whichwhich
the server, is called the broker
is called in MQTT
the broker terms.
in MQTT Messages
terms. are
Messages
published underunder
are published so-called topics,topics,
so-called to which MQTTMQTT
to which clients can subscribe.
clients The broker
can subscribe. The routes
broker
incoming information
routes incoming to all clients
information to allthat have
clients subscribed
that to the corresponding
have subscribed topic under
to the corresponding topic
which a message is published. However, the broker itself remains
under which a message is published. However, the broker itself remains passive and can- passive and cannot
publish any messages
not publish by itself.
any messages by In the specific
itself. setup atsetup
In the specific hand,atthe measurement
hand, units, as units,
the measurement well
asasthe reference unit, act as MQTT clients, while the server backend acts
well as the reference unit, act as MQTT clients, while the server backend acts both as both as the MQTT
broker and client
the MQTT brokeratand
the same
client time.
at theAs such,
same the As
time. server
such,runs
thean MQTT
server broker
runs software
an MQTT as
broker
well as a concurrent
software as well as adaemon
concurrentthat daemon
connectsthat to the local host
connects as an
to the MQTT
local host asclient, receives
an MQTT all
client,
nodes’ measurement data, and logs them in a database. Upon startup
receives all nodes’ measurement data, and logs them in a database. Upon startup of the of the sensor array,
the broker
sensor relays
array, thea broker
retainedrelays
MQTT message MQTT
a retained containing the reference
message containing unit’s
thecalibration data
reference unit’s
tocalibration
all sensordata
nodes as they come online. A multitude of MQTT
to all sensor nodes as they come online. A multitude of MQTT messagemessage topics allows
clients
topics(including the server
allows clients backend)
(including to selectively
the server backend)gather and process:
to selectively gather and process:
1. Reference unit measurement data (“calibration values”).
2. Sensor unit ranging data.
3. Sensor unit measurement data.
Sensors 2022, 22, 4377 7 of 23
4. User-triggered remote management commands.

5. Automatic client management commands.
6. Operational status messages.
2.4.2. Startup/Initialization
As indicated in Figure 3, the sensor nodes operate in two distinct modes: Startup/initialization
mode and sampling mode. During the startup of the sensor array, each node performs three tasks:
1. Configuring the SCD30, BMP390, and DWM3000 sensor modules.
2. Receiving the calibration values from the reference unit and calibrating the CO2 sensor
accordingly (refer to Section 2.5.1).
3. Performing UWB-based ranging with all other nodes that can be reached and sending
the ranging information to the database backend.
In addition, during initialization, the sensor nodes may receive management com-
mands and act upon them. For instance, the operator can remotely instruct nodes in the
IoT array to restart.
2.4.3. Failsafe Operation

When operating dozens of sensor units, scale-up-specific problems start to appear that
are not encountered with a prototype-scale lower number of units. The erroneous behavior
of single nodes must be accepted as inevitable and factored in. For example, single units
may not be able to successfully range with their peers. Moreover, they may fail to return
messages when trying to connect to the WiFi or the MQTT network. Furthermore, nodes
may become stalled during operation due to irregular sensor behavior or hardware failures.
With a measurement interval of two minutes per node, at full array capacity, the IoT mesh
yields 50 × 60/2 = 1500 measurement cycles per hour—each cycle consists of multiple
sensor readouts and network connection attempts. Therefore, it is crucial to ensure the
failsafe operation of the single measurement nodes and—even more importantly—of the
sensing array as a whole. Fault tolerance was implemented considering three aspects.
Due to the decentral and symmetric software design, a faulty sensor node does not
influence the rest of the array. Except for the UWB ranging stage, during the startup of the
IoT mesh, the sensor nodes operate asynchronously and independently from each other.
No node synchronization is required once the sampling stage is entered.
A deep sleep mode (refer to Section 2.4.4) powers down the host microcontroller and
most peripherals between measurements. The wake-up from sleep mode effectively resets
possibly erroneous internal states of the microcontroller. That is, even if a node is faulty, a
possibility is created that it returns to a valid state after the next wake-up from a deep sleep.
Finally, a timer-based hardware watchdog is present in each node that resets the sensor
units in case they become stuck during operation (compare Figure 3). While awake, the
sensor units must periodically reset their watchdog so as to indicate that the program
sequence is actively progressing. Failure to do so causes the watchdog to time out and
signals a potentially hung sensor unit. Upon timeout, the host microcontroller is reset,
forcing the node into a clean startup.
2.4.4. Sleep Mode and Power Consumption

The sensor nodes are battery-powered to ensure maximum configurational flexibility
and ease of operation. This poses efficiency constraints on the IoT mesh to maximize the
runtime of the nodes. For this purpose, the measurement units enter a low-power deep
sleep mode of fixed duration between measurements. The sleep interval effectively defines
the maximum frequency at which the nodes send data to the server and is currently set
to two minutes. Once the measurement data are transmitted to the server backend, the
host microcontroller powers down all components including its own subsystems that are
not needed to recover the node from its deep sleep state. During sleep, the sensor node
is unresponsive and cannot receive MQTT messages. Table 2 lists the measurement node
components and subsystems along with their respective states during deep sleep.
Sensors 2022, 22, 4377 8 of 23
Table 2. State of sensor node components during sleep mode.
Component Subsystem State During Deep Sleep

CPU Down
Memory (ROM, SRAM, flash) Down
ULP coprocessor Up
ESP32 WiFi, Bluetooth & radio transmitter Down
RTC controller & peripherals Up
RTC fast memory Down
RTC slow memory Up
Microcontroller peripherals Down
DWM3000 Down
BMP390 Down
SCD30 Up
Before entering deep sleep at the end of a measurement cycle, a timeout of the real-time
clock (RTC) is registered as a wakeup source. This causes the RTC controller and the ultra-
low-power (ULP) coprocessor to remain up and running as the ESP32 enters sleep mode.
In this state, the ESP32 draws a current of only 0.2 mA. Upon timeout of the RTC timer, the
ESP32 is woken via hardware interrupt from the ULP coprocessor and starts up normally
just as it would during initial powering—with one crucial exception: Before going into deep
sleep, a binary flag is set in the RTC slow memory indicating whether the sensor node has
already passed its initialization stage including UWB ranging and CO2 sensor calibration.
The ESP32 is then configured to keep the RTC slow memory powered even during sleep.
At each wake-up, the initialization flag is checked so that ranging and calibration are only
performed once. This avoids re-running these time- and power-consuming stages of the
sensor array operation even though the microcontroller is newly initialized every time it
is woken.
For additional power consumption optimization, the sensor node software is designed
in such a way that, during the sampling stage, the uptime of the units in between sleeps is as
short as possible. Typically, a sensor unit is awake for less than one second per measurement
cycle. However, this is strongly dependent on the WiFi conditions, as establishing the
network connection takes up the largest part of the uptime.
The DWM3000 module is only used during the UWB ranging in the initialization
stage of the sensor unit (compare Figure 3). Afterward, it is permanently powered off and
also remains inactive during deep sleep. The BMP390 sensor is triggered by the ESP32
to perform a single measurement “on demand”. Therefore, it is inactive by default and
only wakes once per measurement cycle to acquire a single sample. The SCD30, on the
other hand, needs multiple samples for its CO2 readings to converge. As such, the response
time constant is inversely proportional to the SCD30’s sampling frequency. Due to this
accumulating measurement principle, the SCD30 cannot be triggered to perform a single
instantaneous measurement. Instead, the sensor samples continuously with the configured
frequency and needs to be polled by the host for a new, converged reading. This has several
implications for the measurement nodes:
1. The SCD30 sensor cannot be powered down during deep sleep. It must remain up
and running so that it can sample the necessary data points to provide a converged
CO2 reading at the next measurement cycle of the node.
2. To avoid having the host stay awake for longer than necessary due to unsuccessful
attempts at polling the SCD30, the sampling frequency of the sensor needs to be
chosen fast enough such that a converged CO2 value is available to be read by the
next time the ESP32 wakes up.
3. At the same time, the sampling frequency of the SCD30 should be set as low as possible
such that the power-consuming sampling process of the SCD30 is not performed
unnecessarily often.
Sensors 2022, 22, 4377 9 of 23
The SCD30 CO2 sensor is the most power-consuming component of the sensor nodes
next to the ESP32 processor since its operation includes driving an infrared light-emitting
diode (LED) as an optical sampling beam. The manufacturer specifies the maximum
current drawn by the sensor to be 75 mA [40]. As elaborated above, the sensor must sample
periodically and cannot be powered down in sleep mode. For an optimum tradeoff between
measurement accuracy and power consumption, the sampling frequency of the SCD30 and
the sleep interval of the sensor units need to be adjusted. A reference sheet for the CO2
sensor relates sampling frequencies and response times [42]. This technical specification
also prescribes a lower and upper bound on the sampling interval of [5 s; 60 s]. Ignoring this
operating limit would result in inaccurate CO2 measurements. Given a desired temporal
resolution of the sensor array of two minutes, an optimized SCD30 sampling interval of
20 s is selected.
Another power-intensive task is the UWB-based ranging. However, it is only per-
formed once at the startup of the sensor array. Thus, its influence on the sensor unit’s
total runtime remains limited—it consumes a small, fixed amount of the battery capacity
available at the sampling stage.
While the effective capacity of the used rechargeable battery packs varies, the sensor
units achieve an average continuous runtime of more than seven days with the chosen
power management. This operating time easily covers typical time scales observed in
indoor climate investigations. If necessary, the runtime of the sensor mesh can be prolonged
by extending the sleep interval or shortened to increase the effective sampling rate.
2.4.5. Database Structure and Management

In addition to the MQTT broker software and the daemon for receiving and logging
measurement data, the backend server also runs a MariaDB database. It is structured into
three separate tables storing information regarding:
1. Calibration/reference values.
2. UWB-based ranging.
3. Measurement data.
The calibration/reference table may contain as little as one entry if the calibration
procedure via the reference unit is performed only once and in case it yields meaningful
values right away. Typically, the reference measurement is repeated several times to ensure
consistent values.
Each entry in the ranging table records the point-to-point distance between two mea-
surement nodes as initiated by one of the two units. It is complemented by an entry of the
same node-pairing but with reversed initiator. This table is populated upon startup of the
sensor array and remains unchanged throughout the actual climate measurements. At most,
this table contains n × (n − 1) entries, where n is the number of sensor nodes deployed.
Whenever a measurement unit sends sampled data to the backend via MQTT, the
logger daemon receives it and creates a new entry in the measurement data table. Each
entry in this table contains the unique ID of the measurement unit and the sampling
timestamp. Furthermore, the sampled entities, i.e., CO2 concentration, temperature mea-
surement of the SCD30, relative humidity, temperature measurement of the BMP390, and
barometric pressure, are logged. For a representative measurement campaign of 24 h with
all 50 sensor nodes and the default two-minute sampling interval, this table contains up to
24 × 60/2 × 50 = 36000 entries.
2.5. Sensor Calibration

Even though both the SCD30 and the BMP390 are factory pre-calibrated, recalibration
for each measurement campaign is necessary for best accuracy. Section 2.5.1 describes how
the CO2 sensor’s internal reference value is reset in each measurement node without having
to manually expose each unit to a reference atmosphere. The alignment of the barometric
pressure readings of the BMP390 sensors is similarly described in Section 2.5.2.
Sensors 2022, 22, 4377 10 of 23
2.5.1. CO2 Calibration

The Sensirion SCD30 CO2 sensors must be calibrated before each measurement cam-
paign. For that purpose, a reference measurement with a known CO2 concentration is
needed. To avoid the large operational overhead of calibrating every individual sensor
node in an environment of a known CO2 concentration, an indirect, two-stage calibration
procedure is employed that makes use of a single reference unit. In doing so, a common
sampling baseline is established that guarantees consistent sensor readings across the entire
sensor network at the beginning of a measurement campaign.
First, the reference unit is calibrated—it contains the same SCD30 sensor as the mea-
surement nodes. For this “self-calibration”, the reference unit is placed outdoors in fresh
air and a two-minute CO2 measurement is triggered, sampling a known CO2 concentra-
tion of approximately 400 ppm [43]. At the end of the self-calibration measurement, the
reference unit resets its own SCD30-internal reference such that the current sensor output
corresponds to a CO2 concentration of 400 ppm. Next, the now-calibrated reference unit
is used to perform a two-minute measurement at the target measurement site where the
IoT mesh is to be deployed (“reference measurement”). Ideally, this is performed while the
target location is still unoccupied such that the CO2 concentration in the entire room can be
assumed to be homogeneous. In doing so, the reference/starting CO2 concentration at the
measurement site is determined. Both the self-calibration and the reference measurement
duration are chosen to conform with the sampling interval of the sensor array. Further-
more, the SCD30 sampling frequency of the reference unit matches the measurement units’
configuration. At the end of the reference measurement, the reference values are sent to the
MQTT broker as a “retained message”, which all measurement units receive upon initial
startup (refer to Section 2.4.2). Once the measurement nodes receive the reference values
from the MQTT broker, they reset their own SCD30-internal CO2 reading to the reference
value. This concludes the relay-style two-stage calibration, and the measurement nodes are
ready to sample.
2.5.2. Pressure Calibration

The BMP390 senses absolute barometric pressure, which is used to determine the
vertical position of each measurement node with respect to other units. As such, differences
in the barometric pressure readings are of interest to determine relative height profiles.
Even though the BMP390 sensors are factory calibrated, they exhibit a spread in absolute
barometric pressure readings in the order of +/−0.5 hPa, corresponding to height differ-
ences of approximately +/−4 m [41]. To enable precision height measurement comparisons
between the sensor units, a common reference needs to be established. For this, all sensor
units are positioned at the same elevation level, and a few barometric readings are per-
formed. Typically, the floor of the investigated room is chosen as the calibration level since
it is readily referenced during post-processing. Any one of the units’ barometric pressure
values at the calibration level can be chosen as a reference, and the offset of each BMP390
reading to this reference is then used to compensate for the sensor-specific variations. In a
typical measurement campaign, the barometric pressure calibration is performed at the end
of the actual measurement before the deactivation of the array. Once calibrated, the BMP390
readings are highly consistent even over long measurement periods. Figure 4 shows the
BMP390 sensor readings of all sensor units of an exemplary measurement run (inclined
auditorium case, see Section 3.2) in an uncompensated form, whereas Figure 5 depicts them
after calibration. Each graph and color represent a single measurement node. In this spe-
cific setup, the sensor units were located at nine discrete heights during the measurement,
which is reflected in the compensated barometric pressure readings as nine time-varying
ensembles of barometric pressure values. The calibration procedure is evident from the
sudden change in the barometric readings at the end of the measurement. In this case, the
actual spread of the uncompensated barometric pressure readings is approximately half
as large as stated in the BMP390 datasheet (+/−25 Pa corresponding to +/−2 m height
difference). Calibration practically eliminates any relevant spread, and the relative accuracy
heights during the measurement, which is reflected in the compensated barometric pres-
sure readings as nine time-varying ensembles of barometric pressure values. The calibra-
sure readings as nine time-varying ensembles of barometric pressure values. The calibra-
tion procedure is evident from the sudden change in the barometric readings at the end
tion procedure is evident from the sudden change in the barometric readings at the end
of the measurement. In this case, the actual spread of the uncompensated barometric pres-
of the measurement. In this case, the actual spread of the uncompensated barometric pres-
sure readings is approximately half as large as stated in the BMP390 datasheet (+/−25 Pa
Sensors 2022, 22, 4377 sure readings is approximately half as large as stated in the BMP390 datasheet of (+/−25 Pa
corresponding to +/−2 m height difference). Calibration practically eliminates11any 23
relevant
corresponding to +/−2 m height difference). Calibration practically eliminates any relevant
spread, and the relative accuracy of the sensors of +/−3 Pa (+/−0.25 m) is exploited. The
spread, and the relative accuracy of the sensors of +/−3 Pa (+/−0.25 m) is exploited. The
changes in the barometric pressure over the course of the measurement are solely due to
changes in the
of the sensors ofbarometric
+/−3 Pa (+/pressure
−0.25ism) over
is the course
exploited. of the measurement are solely due to
atmospheric fluctuations. This evident from The
the changes in the barometric
highly correlated pressure
evolution of the dif-
atmospheric
over the coursefluctuations. This is evident
of the measurement are solelyfromduethe highly correlated
to atmospheric evolution
fluctuations. Thisofis the dif-
ferent BMP390 readings at every time step. Figure 5 indicates that the offset between
evidentBMP390
ferent from thereadings
highly correlated
at everyevolution
time step. of the different
Figure BMP390 that
5 indicates readings at every
the offset between
grouped
time step.barometric
Figure 5 pressure
indicates thatplots
the is almost
offset between perfectly
grouped constant over
barometric time. Thus,
pressure plots the rela-
is
grouped barometric pressure plots is almost perfectly constant over time. Thus, the rela-
tive vertical
almost positions
perfectly of the time.
constant sensor units are reliably determined.
tive vertical positions ofover
the sensor Thus, theare
units relative vertical
reliably positions of the sensor units
determined.
are reliably determined.
Figure 4. Uncompensated BMP390 values (prior to calibration) in an inclined auditorium.

Figure
Figure 4. Uncompensated
UncompensatedBMP390
BMP390 values
values (prior
(prior to calibration)
to calibration) in an inclined
in an inclined auditorium.
auditorium.
Figure 5. Compensated BMP390 values (after calibration) in an inclined auditorium.
Due to the high precision of the pressure readings, once calibrated, these values
are more accurate than the UWB lateral ranging values and thus stabilize the spatial
reconstruction algorithm and reduce the number of positional ambiguities.
3. Results and Discussion

As described above, the measurement of local CO2 concentrations is the key purpose
of the sensor array as the production and exhalation of the gas by humans is directly linked
to the aerosolized virus load produced. Differences in the diffusivity between CO2 and
small aerosol droplets are sufficiently small for flows dominated by natural convection.
Sensors 2022, 22, 4377 12 of 23
The following sections discuss the results from a measurement campaign in a large
at-grade seminar room (Section 3.1) and a campaign in an inclined lecture hall (Section 3.2).
Additionally, a discussion on the ventilation time constants is provided in Section 3.3.
Furthermore, the accuracy of the proposed self-localization strategy via combined UWB
ranging and barometric sensing is assessed in Section 3.4.
Both in-field measurement campaigns (Sections 3.1 and 3.2) were conducted over
several hours under real occupancy conditions and include distinct events such as people
entering and leaving the rooms. Evolutionary 3D maps of the complete datasets of both cam-
paigns are available as Videos S1 and S2 in the Supplementary Materials. Accompanying
annotations embedded in the videos highlight significant events. It is highly recommended
to inspect these animations in addition to the static plots provided in this section.
3.1. Measurements in an At-Grade Seminar Room

The first measurement campaign was conducted on 27 August 2021 from 2:30 p.m. to
6:45 p.m. during an exam at ETH Zurich. The investigated room has a size of 1400 m2 , in
which 226 students were seated at individual and separate desks. This particular room is
normally used for technical drawing classes for architectural and civil engineering students
and also serves as an exam room due to its large size and the possibility to seat students at
individual desks. The measurement started while the room was still unoccupied such that
the transients from students entering could be captured and compared against the baseline.
The room features two doors in the front and one door in the back, which were kept open
during the entire duration of the exam/measurements. The sensor units were evenly
distributed in the room and placed directly on the desks. Figure 6 shows the schematic
topology of the examination room with the location of the 50 sensor units annotated by
their respective IDs. Note that unoccupied desks are marked by diagonal lines.
Since the entire room is at-grade, so were the measurement units. As a result, baro-
metric pressure readings must only vary due to atmospheric pressure fluctuations and
coincide after calibration, as can be seen in Figure 7. Note that the barometric pressure
sensor calibration was performed prior to the measurement. The slightly diverging plots
towards the end of the measurement indicate that some BMP390 sensors drifted over the
course of the measurement campaign. However, the drift is minor with the largest absolute
discrepancy of 10 Pa. Furthermore, in this case, the inconsistency is irrelevant for the room
topology reconstruction as all sensor units were placed at the same height.
When students started entering the room through the back door, the CO2 readings
increased rapidly with a distinct gradient from back to front as visualized in Figure 8. This
conforms with the applied seating strategy: Students were seated in the front first and rows
in the back were occupied later. Note that this highly unsteady behavior is best viewed in
the Video S1 supplied in the Supplementary Materials.
A similar, yet spatially opposite effect was observed when the students left after the
exam. In contrast to the orderly seating procedure prior to the exam, all students were
allowed to exit the room at the same time through the back door after the exam. Naturally,
the single back door presented a bottleneck such that the occupation of the room was
shifted towards the back as students gathered in front of the exit. Accordingly, the zones of
highest CO2 concentration first shifted towards the back of the room (see Figure 9) with the
gradients pointing from front to the rear middle before the levels slowly dropped towards
the baseline values when the room was completely empty again.
Again, this effect is best seen in the Video S1, Supplementary Materials. Note that the
same effects during entering and emptying of the room are visible in the relative humidity
readings. A zone in the front of the room featured high CO2 concentrations even after
students had already cleared this area. This was caused by the supervisors gathering in that
particular region after the exam. Figure 9 clearly depicts this in the lower right part of the
plot corresponding to the measurement units 11, 21, and 31 (see Figure 6 for measurement
unit mapping).
Sensors 2022, 22, 4377 13 of 23
Figure 6. Layout of at-grade seminar room with measurement unit placements annotated. Unoccu-
Figure 6. Layout of at-grade seminar room with measurement unit placements annotated. Unoccu-
pied desks are crossed out.
pied desks are crossed out.
Since the entire room is at-grade, so were the measurement units. As a result, baro-
metric pressure readings must only vary due to atmospheric pressure fluctuations and
coincide after calibration, as can be seen in Figure 7. Note that the barometric pressure
sensor calibration was performed prior to the measurement. The slightly diverging plots
towards the end of the measurement indicate that some BMP390 sensors drifted over the
room
Sensors 2022, 22, x FOR PEER REVIEW topology reconstruction as all sensor units were placed at the same height. 14 of 23
course of the measurement campaign. However, the drift is minor with the largest abso-
Sensors 2022, 22, 4377 14 of 23
lute discrepancy of 10 Pa. Furthermore, in this case, the inconsistency is irrelevant for the
room topology reconstruction as all sensor units were placed at the same height.
Figure 7. Calibrated barometric pressure readings in at-grade exam room. Sensor unit IDs are color-
coded.
conforms with the applied seating strategy: Students were seated in the front first and
Figure
Figure
rows in7.
7. Calibrated
the back barometric
Calibrated pressure
barometric
were readings
pressure
occupied later. Noteinthat
readings at-grade
this exam
in at-grade room.
exam
highly SensorSensor
room.
unsteady unit IDs
unitareIDs
behavior iscolor-
are
best
coded.
color-coded.
viewed in the Video S1 supplied in the Supplementary Materials.
conforms with the applied seating strategy: Students were seated in the front first and
rows in the back were occupied later. Note that this highly unsteady behavior is best
viewed in the Video S1 supplied in the Supplementary Materials.
Sensors 2022, 22, x FOR PEER REVIEW 15 of in

23
Figure 8. CO
Figure 8. concentration gradient
CO22 concentration gradient from
from back
back to
to front
front during
during seating
seating of
of students
students prior
prior to
to exam
exam in
at-grade
at-grade room.
room.
Figure 8. COback
the single 2 concentration gradient from back to front during seating of students prior to exam in
door presented a bottleneck such that the occupation of the room was
at-grade room.
shifted towards the back as students gathered in front of the exit. Accordingly, the zones
of highest CO2 concentration first shifted towards the back of the room (see Figure 9) with
the gradients pointing from front to the rear middle before the levels slowly dropped to-
wards the baseline values when the room was completely empty again.
the single back door presented a bottleneck such that the occupation of the room was
shifted towards the back as students gathered in front of the exit. Accordingly, the zones
of highest CO2 concentration first shifted towards the back of the room (see Figure 9) with
the gradients pointing from front to the rear middle before the levels slowly dropped to-
wards the baseline values when the room was completely empty again.
Figure CO
9. CO
Figure 9. 2 concentration
2 concentration gradient
gradient fromfrom
frontfront to as
to back back as students
students left the left the at-grade
at-grade room
room after after
the exam.
the exam.
Again, this effect is best seen in the Video S1, Supplementary Materials. Note that the
same effects during entering and emptying of the room are visible in the relative humidity
readings. A zone in the front of the room featured high CO2 concentrations even after
students had already cleared this area. This was caused by the supervisors gathering in
that particular region after the exam. Figure 9 clearly depicts this in the lower right part
of the plot corresponding to the measurement units 11, 21, and 31 (see Figure 6 for meas-
Sensors 2022, 22, 4377 15 of 23
The local room temperature is an obvious parameter in the description of a room

climate environment. Any observed non-uniformities have a direct impact on the natural
convective and forced ventilation flows established.
Figure 10 shows the temporal evolution of the temperature in the at-grade seminar
room for a location close to a sun-lit window (node 42), in a shaded area (node 8), and the
room average computed over all sensor nodes. Note that the average of the sensor readings
is obtained at desk height and is not the same as, say, the average ceiling temperature. These
visible differences highlight the need to include realistic thermal boundary conditions in
any “high fidelity” numerical flow simulation. A constant increase in room temperature is
visible in Figure 10 on average, as well as for the exemplified locations of nodes 8 and 42
right from the beginning of the measurement due to the room heating up from sunlight.
The increased slope in the average profile and in the profile of node 8 around 3:45 p.m.
marks the filling of the room. Note that node 42 already features a kink and a much steeper
increase in temperature before students enter. This is due to direct sunlight impinging on
the node and is evident for the entire ensemble of nodes 42–50 as they were all located
at the sun-lit window front. This temperature inhomogeneity is best visualized in the
Supplementary Video S1. Except for the temperature, at that point in time, the room climate
was still relatively homogeneous. The corresponding local decrease in relative humidity
is also apparent for nodes 42–50 (see Supplementary Video S1). Note that Figure 10 even
allows one to infer shaded and sunny phases of the day during the exam depicted by the
fluctuating temperature evolution of node 42 between 4:50 p.m. and 5:50 p.m. At around
5:50 p.m., students started leaving the examination room. Accordingly, the average room
temperature, as well as the temperatures at nodes 8 and 42, decrease.
Figure 10.
Figure 10. Temperature
Temperature evolution
evolution in
in the
the large
largeplanar
planarseminar
seminarroom.
room.
3.2.
3.2. Measurements
Measurements in in an
an Inclined
Inclined Auditorium
Auditorium
The
The second room investigated with
second room investigated with the
the sensor
sensor network
network isis aa large,
large, inclined
inclined auditorium
auditorium
at 2
at ETH
ETHZurich
Zurichwith
witha abase
basearea of of
area 460460
mm and 600600
2 and seats. An An
seats. image of the
image ofroom as seen
the room as from
seen
afrom
fronta corner is shown
front corner in Figure
is shown 11. The
in Figure 11.top seats
The top in the in
seats back
theare located
back approximately
are located approxi-
7mately
m above7 mthe frontthe
above rows.
front rows.
Figure 10. Temperature evolution in the large planar seminar room.
3.2. Measurements in an Inclined Auditorium

The second room investigated with the sensor network is a large, inclined auditorium
at ETH Zurich with a base area of 460 m2 and 600 seats. An image of the room as seen
Sensors 2022, 22, 4377 16 of 23
from a front corner is shown in Figure 11. The top seats in the back are located approxi-
mately 7 m above the front rows.
Figure 11. Inclined auditorium at ETH Zurich. Image taken from a front corner. Image used with
Figure 11. Inclined auditorium at ETH Zurich. Image taken from a front corner. Image used with
permission from ETH Zurich.
permission from ETH Zurich.
Themeasurement
The measurementwas wasperformed
performedon on2020September
September20212021between
between9:25
9:25a.m.
a.m. and
and 2:50
p.m.p.m.
2:50 During
Duringthatthat
time, twotwo
time, subsequent
subsequent introductory
introductorystudent
studentevents
eventswere
werecarried
carried out,
out, both
of which
both usedused
of which the auditorium
the auditorium at approximately
at approximately 50%50%
capacity. NoNo
capacity. detailed
detailedseating plan as
seating
provided
plan in Section
as provided 3.1 is3.1
in Section available since
is available students
since were
students allowed
were toto
allowed join
jointhe
theevents
eventslater,
later,
leaveleave early,
early, and and
move move around/change
around/change seats.
seats. Again,
Again, the measurement
the measurement waswas started
started prior to
prior to the
the first first students
students entering
entering to establish
to establish some some reference
reference datadata points
points ininundisturbed
undisturbedroom
room conditions.
conditions. WithWith an even
an even distribution
distribution of of
thethe sensorunits,
sensor units,every
everysecond
second seating
seating row
row and
and approximately every fifth to sixth seat was covered by the measurement system. In
this campaign, two features are of specific interest: First, the sloped topology of the room
in combination with a ventilation system injecting air at the floor level and extracting at the
ceiling should establish forced and natural convective flows directed towards the upper
rear seating rows. Second, the capture of two subsequent events in the same room should
provide insights into the cleanout and ramp-up time constants. The full 3D evolution map
of all sensor readings is visualized in Video S2 available in the Supplementary Materials.
Refer to this animation for the most intuitive presentation of the collected data.
Figure 12 shows the CO2 concentration distribution at 11 a.m. in a bird’s eye view.
Each rectangle represents a seat in the lecture hall and the colored spots mark the locations
of the measurement units. That particular time instance is chosen as it represents an
extreme situation with close to maximum room average and top-level CO2 concentrations,
while the sensor readings were not yet influenced by people leaving the room. One clearly
recognizes the significant concentration gradient from front to back, with critical levels
above 1200 ppm being approached in the rear. As expected, exhaled air is convected up
the topological ramp of the inclined auditorium towards the rear. Clearly, seats in the
front of the lecture hall are much less exposed to potentially virus-loaded air than seats
further in the back. As was the case in the previous measurement campaign, relative
humidity readings closely follow the CO2 concentration evolution. Refer to Video S2 in the
Supplementary Materials for a side-to-side animation of both readings.
above 1200 ppm being approached in the rear. As expected, exhaled air is convected up
the topological ramp of the inclined auditorium towards the rear. Clearly, seats in the
front of the lecture hall are much less exposed to potentially virus-loaded air than seats
further in the back. As was the case in the previous measurement campaign, relative hu-
midity
Sensors 2022, 22, 4377 readings closely follow the CO2 concentration evolution. Refer to Video S2 in the 17 of 23
Supplementary Materials for a side-to-side animation of both readings.
Figure 12. CO2 distribution

Figure 12.inCOthe inclined auditorium.
2 distribution A auditorium.
in the inclined strong concentration gradient from
A strong concentration front
gradient from front to
to back is visible. back is visible.
The significantofcontribution
The significant contribution of natural convection
natural convection to theupwards
to the overall overall upwards
directeddirected flow
profile becomes apparent when inspecting the temperature distribution
flow profile becomes apparent when inspecting the temperature distribution in the audi- in the auditorium.
Figure 13 depicts the temperature profile in the auditorium at the end of the first student
torium. Figure 13 depicts the temperature profile in the auditorium at the end of the
2, x FOR PEER REVIEW
first
18 of 23
event. Clearly, the rear middle region of the lecture hall exhibits the largest temperatures
student event. Clearly, the rear middle region of the lecture hall exhibits the largest tem-
indicating that hot air is predominantly flowing and accumulating towards the back of
peratures indicating that hot air is predominantly flowing and accumulating towards the
the auditorium.
back of the auditorium.
13. Temperature
Figuredistribution
Figure 13. Temperature distribution
in the in the inclined
inclined auditorium atauditorium
the end of at
thethe endstudent
first of the first student event.
event.
Highest temperatures are observed in the
Highest temperatures are observed in the rear of the room. rear of the room.
When studying the ventilation in such an environment, a clear distinction has to be

made between the average CO2 concentration (or temperature) and its local values. This
is visualized in Figure 14 where the temporal evolution of the average value (across all
Sensors 2022, 22, 4377 Figure 13. Temperature distribution in the inclined auditorium at the end of the first student18
event.
of 23
Highest temperatures are observed in the rear of the room.

sensor nodes) is compared with readings in a “hot spot” location in the rear (node 50) and
sensor nodes) is compared with readings in a “hot spot” location in the rear (node 50) and
a “clean” patch (node 3) at the front. See Figure 12 for the locations of nodes 3 and 50.
a “clean” patch (node 3) at the front. See Figure 12 for the locations of nodes 3 and 50.
Figure 14.
Figure 14. Local
Local and
and average
average CO
CO22 concentrations
concentrations in
in the
the inclined
inclined auditorium.
auditorium.
In this highly inhomogeneous environment, resolved measurements can provide the

necessary values for tuning corresponding numerical simulations, which would normally
rely on empirical
empirical turbulence
turbulence and
and mixing
mixing models
models [12].
[12].
Figure
Figure 1414intuitively
intuitivelydepicts
depictsthethe
timeline of events
timeline during
of events that that
during measurement campaign:
measurement cam-
Between 9:40 a.m. and 10 a.m. students entered the room. Consequently, in that
paign: Between 9:40 a.m. and 10 a.m. students entered the room. Consequently, in that time frame,
the
timeaverage
frame, and local COand
the average 2 concentrations increased steeply.
local CO2 concentrations Around
increased 10 a.m.,
steeply. the slope
Around of
10 a.m.,
the curves decreases, indicating the start of the first student event in the auditorium. The
average CO2 concentration keeps increasing up until around 11 a.m., which was the time at
which the first student event ended. Within the next 10 min, people exited the lecture hall.
During that time, the average CO2 concentration remains constant, while it even keeps
increasing at the location of node 50. At 11:10 a.m., all students had left the room marking
the beginning of the clean-out phase. Students started entering the lecture hall again at
11:50 a.m. at a residual average CO2 concentration of 500 ppm for the second event. Even
though fresh air conditions could not be fully recovered, due to the asymptotic behavior of
the curve, a clean-out time of at least one hour can be estimated. The analysis of the second
introductory student event starting at 12:15 p.m. follows the same reasoning except for
the fact that both the CO2 concentration at node 50 as well as the room average show a
distinct maximum around 12:40 p.m. that is not visible in the trendline of the first event.
Furthermore, the maximum CO2 levels are lower than during the preceding event. This
is because the room ventilation was ramped up during the second event at 12:40 p.m. to
increase the air exchange in the lecture hall. Consequently, the clean-out phase after the
second student event starting at 1:15 p.m. is also shorter. This time, the room returned to
fresh air levels, which took approximately 70 min after the lecture hall had been emptied.
At this point, a remark has to be made on the array’s sensitivity. The majority of
optical CO2 sensors are specified with sensitivities/noise floors of approximately 50 ppm
CO2 . While this appears to be sufficient to detect the emissions of a single person, such
a measurement depends critically on the environmental conditions. In a large room with
forced ventilation, the gas exchange rates of several thousand m3 /h can easily be so large
that the contributions of a single person fall well below that threshold. This remains a
CO2. While this appears to be sufficient to detect the emissions of a single person, suc
measurement depends critically on the environmental conditions. In a large room w
forced ventilation, the gas exchange rates of several thousand m3/h can easily be so la
that the contributions of a single person fall well below that threshold. This remain
Sensors 2022, 22, 4377 challenge if the array were to be used for measurements of realistic “point 19 ofspread
23 fu
tions”, i.e., with local CO2 production rates on the order of grams per minute.
challenge if the array were to be used for measurements of realistic “point spread functions”,
3.3. Ventilation Time Constants
i.e., with local CO2 production rates on the order of grams per minute.
A figure of specific interest is the decay rate of the CO2 concentration once a ro
3.3. Ventilation Time Constants
has been emptied. This has obvious consequences for the overall occupation policy a
turns outAtofigure
also ofbespecific
one ofinterest
the moreis the decay rate of the CO2 concentration once a room
difficult/decisive parameters to reproduce in any hi
has been emptied. This has obvious consequences for the overall occupation policy and
fidelity
turnssimulation. Here
out to also be the
one of thereadings of average CO
more difficult/decisive 2 concentrations in the two roo
parameters to reproduce in any
presented before (H : 460 m 2, inclined; H2: 1400 m2, at-grade) are examined, both at
high-fidelity simulation. Here the readings of average CO2 concentrations in the two rooms
1
startpresented
of occupancy
before and
(H1 : 460 m2 , inclined;
during the cleanout m2 , at-grade)
phase
H2 : 1400 after theareaudience
examined,had
bothleft (Figure 1
at the
start of occupancy and during the cleanout phase after the audience had left (Figure 15).
O ccu p atio n S tart O ccu p atio n E n d
850 900
H (1) H (1)
1 1
800 H1 (2) 850 H1 (2)
H2 H2
800
750
CO2 concentration [ppm]
CO2 concentration [ppm]

750
700
700
650
650
600
600
550
550
500 500
450 450
0 10 20 30 40 0
Time [min.] Time [min.]
Figure 15. Rise

Figure and
15. Rise decay
and decayof
of CO concentration
CO22 concentration during
during roomroom population
population and emptying
and emptying phases. phases.
One recognizes rather significant time constants in the order of several tens of minutes
One recognizes rather significant time constants in the order of several tens
for both transient phases. It is noteworthy that even in the same room (H1 ), the time
minutes for both
constants differtransient phases.
significantly. It isofnoteworthy
In the case that even
H1 (2), as elaborated in feedback
above, the same room (H1),
control
of the ventilation system was activated, in contrast to cases H1 (1) and H2 , where constant
ventilation was maintained. The ramping up of the ventilation in the former case led to a
reduction of approximately a factor of two in the ventilation turnover times.
3.4. Sensor Node Position Reconstruction

In order to demonstrate the functionality and accuracy of the ranging approach, a
subset of ten sensing nodes was placed in a controlled environment with placement heights
and separations determined by independent laser ranging. This allows one to compare
the ranging results based on the UWB and barometric pressure data with the “ground
truth”. Processing of the ranging data is performed with multidimensional scaling (MDS),
specifically the SMACOF algorithm (Scaling by Majorizing a Complicated Function, [44]).
Additionally, the algorithm is augmented to include the barometric node heights as con-
straints. These constraints prove essential for the MDS reconstruction in the absence of
anchor nodes with known spatial reference positions.
Figure 16 displays the position errors with respect to the known ground truth positions.
The error magnitudes are roughly the same in all three directions, and the mean Euclidean
3D error is 0.21 m. This value must be compared to the size of the sensing volume covered
by the nodes, which was 4 × 5 × 2 m3 .
absence of anchor nodes with known spatial reference positions.
Figure 16 displays the position errors with respect to the known ground truth p
tions. The error magnitudes are roughly the same in all three directions, and the m
Euclidean 3D error is 0.21 m. This value must be compared to the size of the sensing
Sensors 2022, 22, 4377 20 of 23
ume covered by the nodes, which was 4 × 5 × 2 m3.
Figure 16. 16.

Figure Position
Positionerrors
errors following MDS/SMACOF
following MDS/SMACOF reconstruction.
reconstruction.
Tests in larger rooms and locations have shown that ranging up to distances around
Tests in larger rooms and locations have shown that ranging up to distances around
20 m does not pose a problem for the current implementation with the DWM3000 UWB
m does not pose
modules. a problemit for
Nevertheless, the be
cannot current implementation
excluded with thesettings
that more complicated DWM3000 UWB modu
(e.g., array
Nevertheless, it cannot
patterns without be excluded
line-of-sight that
visibility more
of the complicated
nodes and clutteredsettings (e.g., in
environments array patterns w
general)
may pose some
out line-of-sight as of yet of
visibility unknown
the nodeschallenges (loss of signal,
and cluttered multipath problems,
environments in general)etc.).
may pose so
as of4.yet unknown challenges (loss of signal, multipath problems, etc.).
Conclusions
The paper presents a large IoT sensor array to monitor the variations in the indoor cli-
4. Conclusions
mate of large auditoria with high spatial and temporal resolution. While the development
ofThe
the sensor
paperarray was driven
presents by the
a large IoTneed for rapid
sensor arrayverification
to monitor of numerical predictions
the variations in the ind
in the management of such rooms under epidemic restrictions, the demonstrated func-
climate of large auditoria with high spatial and temporal resolution. While the deve
tionalities point to a broader field of applications in the general area of indoor air quality
ment of the sensor
assessment array Highly
and control. was driven
unsteadybyevents,
the need forasrapid
as well verification
significant of numerical p
spatial gradients
dictions
in COin the management
2 concentration, of such
temperature, androoms under
relative epidemic
humidity, restrictions,
are successfully the and
captured demonstra
resolved, as well as visualized readily in 3D evolution maps. Based on scalable
functionalities point to a broader field of applications in the general area of indoor IoT and
MQTT
quality technology,and
assessment the sensing array
control. is extensible
Highly unsteadyto even largeras
events, sizes.
wellThrough the use ofspatial
as significant
highly integrated COTS components, the cost per node is kept low such that the hardware
investment does not appear to be a limiting factor in this regard.
To overcome operational and management issues imposed by the size of the sensor
network, a simple relay-style calibration method was developed that reduces the operator’s
manual overhead to a minimum. Furthermore, a novel 3D sensor node positional awareness
mechanism is presented that eliminates the need for spatial anchors by combining UWB
ranging and barometric pressure sensing. Finally, special care was taken at the sensor array
conceptualization stage to design the entire system and its software for a maximum level
of robustness and data privacy-awareness.
The insights gained from the proposed sensor network enable identifying seating
positions with high infection risks in lecture halls and optimizing temporal and spatial
room occupation schedules such that they conform with cleanout times and local flow
patterns. Furthermore, the large and high-resolution datasets can serve both as input
for numerical simulations and their subsequent validation at different scales and levels
of complexity.
Sensors 2022, 22, 4377 21 of 23
Future work will focus on the incorporation of additional sensing components such as
particle counters, as well as other gas (VOC) detectors to yield an even more complete air
quality assessment. Furthermore, the extension of the array to a larger number of sensor
nodes is anticipated. Moreover, as more and more datasets are collected, machine learning
techniques may be used to infer IAQ metrics in rooms with, as of yet, untested topologies.
Supplementary Materials: The following supporting information can be downloaded at: https://
www.mdpi.com/article/10.3390/s22124377/s1, Video S1. SuppInfo_evolution_map_atgrade_auditor
ium.mp4: Evolution of indoor climate in at-grade auditorium during exam, Video S2. SuppInfo_evolu
tion_map_inclined_auditorium.mp4: Evolution of indoor climate in inclined auditorium during
two subsequent student events
Author Contributions: Conceptualization, A.R. and T.R.; methodology, A.R. and T.R.; node software,
A.R. and T.R.; server software: A.R.; validation, A.R. and T.R.; formal analysis, A.R. and T.R.;
investigation, A.R. and T.R.; resources, A.R. and T.R.; data curation, A.R.; writing—original draft
preparation, A.R. and T.R.; writing—review and editing, A.R. and T.R.; visualization, A.R. and T.R.;
supervision, T.R.; project administration, A.R. and T.R.; funding acquisition, T.R. All authors have
read and agreed to the published version of the manuscript.
Funding: This research received no external funding.
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Data Availability Statement: The data presented in this study are available on request from the cor-
responding author. The data are not publicly available due to privacy reasons since the measurements
were conducted in the presence of students and staff. The full measurement data are visualized in the
Supplementary Videos.
Conflicts of Interest: The authors declare no conflict of interest.
References
1. Hassan, A.M.; Megahed, N.A. COVID-19 and urban spaces: A new integrated CFD approach for public health opportunities.
Build. Environ. 2021, 204, 108131. [CrossRef] [PubMed]
2. Mariam Magar, A.; Joshi, M.; Rajagopal, P.S.; Khan, A.; Rao, M.M.; Sapra, B.K. CFD simulation of the airborne transmission
of COVID-19 vectors emitted during respiratory mechanisms: Revisiting the concept of safe distance. ACS Omega 2021, 6,
16876–16889. [CrossRef] [PubMed]
3. Mohamadi, F.; Fazeli, A. A Review on Applications of CFD Modeling in COVID-19 Pandemic. Arch. Comput. Methods Eng. 2022.
[CrossRef] [PubMed]
4. Teshome, E.; Haghighat, F.F. Zonal Models for Indoor Air Flow—A Critical Review. Int. J. Vent. 2004, 3, 119–129. [CrossRef]
5. Megri, A.C.; Haghighat, F. Zonal Modeling for Simulating Indoor Environment of Buildings: Review, Recent Developments, and
Applications. HVAC&R Res. 2007, 13, 887–905. [CrossRef]
6. Song, F.; Zhao, B.; Yang, X.; Jiang, Y.; Gopal, V.; Dobbs, G.; Sahm, M. A new approach on zonal modeling of indoor environment
with mechanical ventilation. Build. Environ. 2008, 43, 278–286. [CrossRef]
7. Emmerich, S.J. Validation of multizone IAQ modeling of residential-scale buildings: A review. Trans. Am. Soc. Heat. Refrig. Air
Cond. Eng. 2001, 107, 619–628.
8. Li, Y.; Duan, S.; Yu, I.T.S.; Wong, T.W. Multi-zone modeling of probable SARS virus transmission by airflow between flats in Block
E, Amoy Gardens. Indoor Air 2005, 15, 96–111. [CrossRef]
9. Chen, Y.; Gu, L.; Zhang, J. EnergyPlus and CHAMPS-Multizone co-simulation for energy and indoor air quality analysis. Build.
Simul. 2015, 8, 371–380. [CrossRef]
10. Szczepanik-Scislo, N.; Scislo, L. Comparison of CFD and Multizone Modeling from Contaminant Migration from a Household
Gas Furnace. Atmosphere 2021, 12, 79. [CrossRef]
11. Ma, L.; Zheng, X.; Guo, Y.; Zhong, W.; Zhang, Y.; Qian, H. A multi-zone spatial flow impact factor model for evaluating and
layout optimization of infection risk in a Fangcang shelter hospital. Build. Environ. 2022, 214, 108931. [CrossRef]
12. Mirzaie, M.; Lakzian, E.; Khan, A.; Warkiani, M.E.; Mahian, O.; Ahmadi, G. COVID-19 spread in a classroom equipped with
partition—A CFD approach. J. Hazard. Mater. 2021, 420, 126587. [CrossRef] [PubMed]
13. Foster, A.; Kinzel, M. Estimating COVID-19 exposure in a classroom setting: A comparison between mathematical and numerical
models. Phys. Fluids 2021, 33, 021904. [CrossRef] [PubMed]
14. Peng, Z.; Jimenez, J.L. Exhaled CO2 as a COVID-19 Infection Risk Proxy for Different Indoor Environments and Activities.
Environ. Sci. Technol. Lett. 2021, 8, 392–397. [CrossRef]
Sensors 2022, 22, 4377 22 of 23
15. Schade, W.; Reimer, V.; Seipenbusch, M.; Willer, U. Experimental Investigation of Aerosol and CO2 Dispersion for Evaluation of
COVID-19 Infection Risk in a Concert Hall. Int. J. Environ. Res. Public Health 2021, 18, 3037. [CrossRef]
16. Bhagat, R.K.; Wykes, M.D.; Dalziel, S.B.; Linden, P.F. Effects of ventilation on the indoor spread of COVID-19. J. Fluid Mech. 2020,
903, F1. [CrossRef]
17. Liu, X.; Ma, T.; Pinna, N.; Zhang, J. Two-Dimensional Nanostructured Materials for Gas Sensing. Adv. Funct. Mater. 2017,
27, 1702168. [CrossRef]
18. Joshi, N.J.; Braunger, M.L.; Shimizu, F.M.; Riul, A., Jr.; de Oliveira, O.N., Jr. Insights into nano-heterostructured materials for gas
sensing: A review. Multifunct. Mater. 2021, 4, 032002. [CrossRef]
19. Blocken, B.; van Druenen, T.; Ricci, A.; Kang, L.; van Hooff, T.; Qin, P.; Xia, L.; Alanis Ruiz, C.; Arts, J.; Diepens, J.; et al.
Ventilation and air cleaning to limit aerosol particle concentrations in a gym during the COVID-19 pandemic. Build. Environ.
2021, 193, 107659. [CrossRef]
20. Scislo, L.; Szczepanik-Scislo, N. Air Quality Sensor Data Collection and Analytics with IoT for an Apartment with Mechanical
Ventilation. In Proceedings of the 2021 11th IEEE International Conference on Intelligent Data Acquisition and Advanced
Computing Systems: Technology and Applications (IDAACS), Cracow, Poland, 22–25 September 2021; pp. 932–936. [CrossRef]
21. Kim, J.J.; Jung, S.K.; Kim, J.T. Wireless monitoring of indoor air quality by a sensor network. Indoor Built Environ. 2010, 19,
145–150. [CrossRef]
22. Kim, J.-Y.; Chu, C.-H.; Shin, S.-M. ISSAQ: An Integrated Sensing Systems for Real-Time Indoor Air Quality Monitoring. IEEE
Sens. J. 2014, 14, 4230–4244. [CrossRef]
23. Abraham, S.; Li, X. A Cost-effective Wireless Sensor Network System for Indoor Air Quality Monitoring Applications. Procedia
Comput. Sci. 2014, 34, 165–171. [CrossRef]
24. Marques, G.; Pitarma, R. An Indoor Monitoring System for Ambient Assisted Living Based on Internet of Things Architecture.
Int. J. Environ. Res. Public Health 2016, 13, 1152. [CrossRef] [PubMed]
25. Tran, T.V.; Dang, N.T.; Chung, W.-Y. Battery-free smart-sensor system for real-time indoor air quality monitoring. Sensors Actuators
B Chem. 2017, 248, 930–939. [CrossRef]
26. Nelson, J.; Andoh, C.; Comia, A.; Echeveria, L.; Hopkins, J.; Maniti, M.; Pierce, T. Wireless Sensor Network with Mesh Topology
for Carbon Dioxide Monitoring in a Winery. In Proceedings of the 2021 IEEE Topical Conference on Wireless Sensors and Sensor
Networks (WiSNeT), San Diego, CA, USA, 17–20 January 2021; pp. 30–33. [CrossRef]
27. Saad, S.M.; Saad, A.R.M.; Kamarudin, A.M.Y.; Zakaria, A.; Shakaff, A.Y.M. Indoor air quality monitoring system using wireless
sensor network (WSN) with web interface. In Proceedings of the 2013 International Conference on Electrical, Electronics and
System Engineering (ICEESE), Kuala Lumpur, Malaysia, 4–5 December 2013; pp. 60–64. [CrossRef]
28. Brunelli, D.; Minakov, I.; Passerone, R.; Rossi, M. POVOMON: An Ad-hoc Wireless Sensor Network for indoor environmen-
tal monitoring. In Proceedings of the 2014 IEEE Workshop on Environmental, Energy, and Structural Monitoring Systems
Proceedings, Naples, Italy, 17–18 September 2014; pp. 1–6. [CrossRef]
29. Bhattacharya, S.; Sridevi, S.; Pitchiah, R. Indoor air quality monitoring using wireless sensor network. In Proceedings of the 2012
Sixth International Conference on Sensing Technology (ICST), Kolkata, India, 18–21 December 2012; pp. 422–427. [CrossRef]
30. Schieweck, A.; Uhde, E.; Salthammer, T.; Salthammer, L.C.; Morawska, L.; Mazaheri, M.; Kumar, P. Smart homes and the control
of indoor air quality. Renew. Sustain. Energy Rev. 2018, 94, 705–718. [CrossRef]
31. Omidvarborna, H.; Kumar, P.; Hayward, J.; Gupta, M.; Nascimento, E. Low-Cost Air Quality Sensing towards Smart Homes.
Atmosphere 2021, 12, 453. [CrossRef]
32. Arroyo, P.; Herrero, J.L.; Suárez, J.I.; Lozano, J. Wireless Sensor Network Combined with Cloud Computing for Air Quality
Monitoring. Sensors 2019, 19, 691. [CrossRef]
33. Molka-Danielsen, J.; Engelseth, P.; Olesnanikova, V.; Sarafin, P.; Zalman, R. Big Data Analytics for Air Quality Monitoring at a
Logistics Shipping Base via Autonomous Wireless Sensor Network Technologies. In Proceedings of the 2017 5th International
Conference on Enterprise Systems (ES), Beijing, China, 22–24 September 2017; pp. 38–45. [CrossRef]
34. Lozano, J.; Suárez, J.I.; Arroyo, P.; Ordiales, J.M.; Alvarez, F. Wireless sensor network for indoor air quality monitoring. Chem.
Eng. Trans. 2012, 30, 319–324. [CrossRef]
35. Bamodu, O.; Xia, L.; Tang, L. An indoor environment monitoring system using low-cost sensor network. Energy Procedia 2017,
141, 660–666. [CrossRef]
36. Jelicic, V.; Magno, M.; Brunelli, D.; Paci, G.; Benini, L. Context-Adaptive Multimodal Wireless Sensor Network for Energy-Efficient
Gas Monitoring. IEEE Sens. J. 2012, 13, 328–338. [CrossRef]
37. Pastor-Fernández, A.; Cerezo-Narváez, A.; Montero-Gutiérrez, P.; Ballesteros-Pérez, P.; Otero-Mateo, M. Use of Low-Cost Devices
for the Control and Monitoring of CO2 Concentration in Existing Buildings after the COVID Era. Appl. Sci. 2022, 12, 3927.
[CrossRef]
38. Stamatescu, G.; Sgârciu, V. Evaluation of wireless sensor network monitoring for indoor spaces. In Proceedings of the 2012
International Symposium on Instrumentation & Measurement, Sensor Network and Automation (IMSNA), Sanya, China, 25–28
August 2012; pp. 107–111. [CrossRef]
39. Datasheet Sensirion SCD30 Sensor Module. Available online: https://sensirion.com/media/documents/4EAF6AF8/61652C3C/
Sensirion_CO2_Sensors_SCD30_Datasheet.pdf (accessed on 27 May 2022).
40. SCD30 CO2 Sensor Accuracy. Available online: https://sensirion.com/de/produkte/katalog/SCD30 (accessed on 26 April 2022).
Sensors 2022, 22, 4377 23 of 23
41. BMP390 Digital Pressure Sensor. Available online: https://www.bosch-sensortec.com/media/boschsensortec/downloads/

datasheets/bst-bmp390-ds002.pdf (accessed on 26 April 2022).
42. Low Power Mode for SCD30. Available online: https://sensirion.com/media/documents/0FEA2450/61652EF9/Sensirion_CO2
_Sensors_SCD30_Low_Power_Mode.pdf (accessed on 26 April 2022).
43. Field Calibration for SCD30. Available online: https://sensirion.com/media/documents/33C09C07/620638B8/Sensirion_SCD3
0_Field_Calibration.pdf (accessed on 26 April 2022).
44. de Leeuw, J.; Mair, P. Multidimensional Scaling Using Majorization: SMACOF in R. J. Stat. Softw. 2009, 31, 1–30. [CrossRef]

An Internet of Things Sensor Array For Spatially A

Uploaded by

Copyright:

Available Formats

An Internet of Things Sensor Array For Spatially A

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Internet of Things Sensor Array For Spatially A

Uploaded by

Copyright:

Available Formats

sensors

Institute of Fluid Dynamics, ETH Zurich, 8092 Zurich, Switzerland

Sensors 2022, 22, 4377. https://doi.org/10.3390/s22124377 https://www.mdpi.com/journal/sensors

2. Materials and Methods

2.2. Remote Operation

Sensors 2022, 22, 4377 4 of 23

Sensors 2022, 22, 4377 1. 5 of 23

IEEE 802.11 WiFi IEEE 802.15.4 UWB

Li-Ion ESP32 SCD30 BMP390 DWM3000

Establish communication (I2C, SPI, serial) & enable watchdog

Upon RTC timer timeout, wake via hardware interrupt

commands (if any)

Receive reference values

Perform UWB ranging

Publish ranging data

Sample CO2, rH, T1, T2 & pabs

Publish measurement data

Enter deep sleep

4. User-triggered remote management commands.

2.4.3. Failsafe Operation

2.4.4. Sleep Mode and Power Consumption

Table 2. State of sensor node components during sleep mode.

Component Subsystem State During Deep Sleep

2.4.5. Database Structure and Management

2.5. Sensor Calibration

2.5.1. CO2 Calibration

2.5.2. Pressure Calibration

Figure 4. Uncompensated BMP390 values (prior to calibration) in an inclined auditorium.

Figure 5. Compensated BMP390 values (after calibration) in an inclined auditorium.

3. Results and Discussion

3.1. Measurements in an At-Grade Seminar Room

Sensors 2022, 22, x FOR PEER REVIEW 15 of in

The local room temperature is an obvious parameter in the description of a room

3.2. Measurements in an Inclined Auditorium

Figure 12. CO2 distribution

When studying the ventilation in such an environment, a clear distinction has to be

When studying the ventilation in such an environment, a clear distinction has to be

In this highly inhomogeneous environment, resolved measurements can provide the

CO2 concentration [ppm]

Figure 15. Rise

3.4. Sensor Node Position Reconstruction

Figure 16. 16.

41. BMP390 Digital Pressure Sensor. Available online: https://www.bosch-sensortec.com/media/boschsensortec/downloads/

You might also like