This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT 1

Monitoring of Large-Area IoT Sensors Using a

LoRa Wireless Mesh Network System:
Design and Evaluation
Huang-Chen Lee , Senior Member, IEEE, and Kai-Hsiang Ke

Abstract— Although many techniques exist to transfer data complex, and data transfer needs must be balanced with
from the widely distributed sensors that make up the Internet operating considerations and infrastructure costs. In contrast to
of Things (IoT) (e.g., using 3G/4G networks or cables), these existing short-range wireless communication technologies cur-
methods are associated with prohibitively high costs, making
them impractical for real-life applications. Recently, several rently used for indoor IoT sensors [4], several techniques have
emerging wireless technologies have been proposed to provide been proposed for providing long-range and outdoor wireless
long-range communication for IoT sensors. Among these, LoRa communication, such as Sigfox [5], LoRa/LoRaWAN [6],
has been examined for long-range performance. Although LoRa NB-IoT, and LTE-M [7]. Among these, LoRaWAN, which is
shows good performance for long-range transmission in the based on the LoRa physical layer (PHY) [8] implementation,
countryside, its radio signals can be attenuated over distance,
and buildings, trees, and other radio signal sources may interfere is a network standard for telecom operators. It allows them
with the signals. Our observations show that in urban areas, to provide network service and enable devices to wirelessly
LoRa requires dense deployment of LoRa gateways (GWs) to transfer data over long distances to remote gateways (GWs).
ensure that indoor LoRa devices can successfully transfer data LoRaWAN uses a star-network topology for communication
back to remote GWs. Wireless mesh networking is a solution for between LoRa GWs and IoT devices; only one hop is allowed
increasing communication range and packet delivery ratio (PDR)
without the need to install additional GWs. This paper presents a between a GW and a LoRa device. Some experiments have
LoRa mesh networking system for large-area monitoring of IoT shown that a LoRa device can transmit data 15 km in an
applications. We deployed 19 LoRa mesh networking devices over open area, which is sufficient for most current long-range
an 800 m × 600 m area on our university campus and installed IoT applications. However, indoor LoRa devices may still be
a GW that collected data at 1-min intervals. The proposed unable to communicate wirelessly with a nearby GW, due to
LoRa mesh networking system achieved an average 88.49% PDR,
whereas the star-network topology used by LoRa achieved only obstacles between sensors, which can attenuate wireless signal
58.7% under the same settings. To the best of our knowledge, strength and result in data losses and communication errors.
this is the first academic study discussing LoRa mesh networking Increasing the spreading factor (SF) of LoRa PHY to 12 can
in detail and evaluating its performance via real experiments. significantly extend the communication range by increasing
Index Terms— Data collection, Internet of Things (IoT), LoRa, the receiver sensitivity [9], but the drawback is that doing
long-range wireless area network (LoRaWAN), multihop, physi- so can lower the data throughput rate and cause more severe
cal layer (PHY), sensor, wireless mesh network. data collision due to the longer times required for transmis-
sion. This issue can be exacerbated when a large number
I. I NTRODUCTION of high-density LoRa devices send wireless data and receive

T HE Internet of Things (IoT) aims to enable conventional

sensing devices to communicate with other devices and
to cooperatively provide intelligent service. For example, IoT
acknowledgments at the same time. Some studies [11] have
shown that these conditions can cause the packet delivery
ratio (PDR) to drop significantly. To solve this issue, telecom
can be used to monitor air quality in a city to provide real-time operators need to deploy more LoRaWAN GWs to increase the
information and warnings to inhabitants as well as to control probability of receiving the wireless data from the LoRaWAN
air pollution [1], [2]. Another example is monitoring of solar devices; this solution increases the cost of the network
panels [3]; smart panels can be installed on the roofs of several infrastructure.
buildings to optimize their collective efficiency. As the number Wireless mesh networks are one feasible solution to increase
of sensors in an IoT system grows, however, the issue of how the communication performance of devices in indoor net-
to transfer data among those devices becomes increasingly works. This approach allows all LoRa devices on the network
Manuscript received October 9, 2017; revised February 22, 2018; accepted to act as routers and relay data from other devices. In this
February 23, 2018. This work was supported by ICP DAS Company Ltd., paper, which is based on [10], we explore the use of wireless
Taiwan. The Associate Editor coordinating the review process was mesh networks based on LoRa PHY with the aim of collecting
Dr. Domenico Grimaldi. (Corresponding author: Huang-Chen Lee.)
The authors are with the Department of Communications Engineering, reliable data on the performance of this technique.
Advanced Institute for Manufacturing with Hightech Innovations, National Fig. 1 shows a snapshot of the network topology of the
ChungCheng University, Chiayi 62102, Taiwan (e-mail: [email protected]). proposed LoRa mesh network on our university campus.
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. In Fig. 1, GW denotes the gateway that manages the network
Digital Object Identifier 10.1109/TIM.2018.2814082 and collects data from the IoT sensors, which are indicated
0018-9456 © 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

Fig. 1. Network topology snapshot of the LoRa mesh network devices distributed around a 600 m × 800 m university campus.

by circles with numbers. Nineteen sensors were installed and that the proposed design can mitigate this problem
in different buildings across the 600 m × 800 m university by using mesh networking to increase PDR without
campus. The yellow lines connecting the circles indicate the deploying additional GWs.
topological paths generated by the data being relayed to the 3) The proposed LoRa mesh network module can be
GWs at the moment of the snapshot. integrated with other IoT applications to help collect
The example in Fig. 1 shows that data generated by sensor data from distributed sensors, bypassing complex data
4 were relayed by sensors 1 and 3 before finally arriving at the transmission, and collection issues. Users do not need
GW. Note that the network topology is formed automatically to know how to build a network or route the data,
and may change due to environmental changes. Each sensor because all the communication protocols are contained
decides locally which sensor is the best one to help relay in the module. This significantly simplifies the use of
its data. Using this approach, a sensor that cannot directly long-range wireless communication and facilitates IoT
communicate with the GW may find other sensors to help it applications over large regions.
to transfer data to the GW, thereby raising the packet delivery
performance (PDR) of each sensor. II. R ELATED W ORK AND D ESIGN G OALS
We analyzed the communication performance of the wire- Several long-range wireless communication technologies
less mesh network and the result shows that mesh networks have recently been proposed. LoRa was one of the first such
can deliver better performance without incurring the cost of technologies to become commercially available; therefore,
installing additional GWs. We also analyzed the performance many academic studies have been devoted to discuss and ana-
tradeoffs of different configurations. lyze LoRa’s performance. LoRa (specifically, LoRa PHY) is a
Briefly, the contributions of this paper are as follows. long-range wireless transmission technique, and LoRaWAN is
a networking design that enables telecom operators to provide
1) Whereas other studies [11]–[16] have continued to focus subscription services based on LoRa PHY. LoRa PHY is a
on the analysis of the standard LoRaWAN protocols, proprietary chirp spread spectrum scheme that uses the sub-
discussing how to alleviate the issues caused by sig- 1 GHz wireless frequency band. It features extremely high
nal attenuation and collision in high-density wireless sensitivity, down to −137 dBm, and a maximal link budget
devices, this paper proposes a LoRa mesh network- of up to 157 dB [23]. Several parameters of LoRa PHY can
ing system and evaluates its performance. We thereby be adjusted for different performance goals, including power
identify a new approach for fulfilling the practical level, SF, bandwidth (BW), and coding rate (CR). Changes
need to collect data from IoT sensors scattered over in power level, SF, BW, and CR involve multiple tradeoffs
a campus-sized area without using expensive 3G/4G among transmission time on air (data throughput), power
networks or laying cables for data transmission. consumption, and transmission range, and different tradeoffs
2) This paper is one of a few works to empirically result in varying receiver sensitivities, transmission time on
evaluate the performance of LoRa mesh networking. air, and signal-to-noise ratios (SNRs).
The evaluation illustrates that LoRa signals can be In a study of how SF affects receiver sensitivity in LoRa and
attenuated or blocked by buildings and other obstacles LoRaWAN [13], researchers used a Semtech SX1276 Mbed
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE et al.: MONITORING OF LARGE-AREA IoT SENSORS USING A LoRa WIRELESS MESH NETWORK SYSTEM 3

shield as the end device and a Cisco 910 industrial router as information is provided in its datasheet, including information
the GW. The BW setting of the LoRa device was 125 kHz, about how to form a mesh network, how to relay data, and
the CR was 4/5, and the power output level was set to the the communication performance of the device. The device
minimum, 2 dBm, with a 3-dBi antenna. A packet transmission datasheet [18, p. 8] indicates that it is simply based on repeat-
test showed that the receiver sensitivity (the minimum signal ing the received data in the same network ID. Therefore, this
strength needed to decode the incoming packet correctly) was approach is similar to the Adhoc on-demand distance vector
−123 dBm with an SF of 7. This improved to −135 dBm routing method of Zigbee and the managed flood method
when the SF was increased to 12, that is, larger SF resulted in of Bluetooth 5.0 mesh networks [19]. This approach should
better receiver sensitivity and longer transmission range. The work when network traffic is low, but it may be difficult
researchers also tested the transmission range under different to maintain the communication quality as traffic increases,
SF settings in an outdoor, urban area; the results indicated that, and it may not be able to manage the network topology.
with an SF of 12, the transmission range could reach 3500 m In contrast, the industrial application standard time-slotted
with a 40% PDR; the PDR collapsed to 0% at SF 7 or 9. channel hopping wireless mesh network systems such as
However, the larger SF also resulted in slow data through- WirelessHART [20] and the ISA100.11a [21] standard both
put. The researchers tested the throughput of LoRaWAN at support graph routing as well as source routing; both stan-
BW = 125 kHz, and the results illustrated that the throughput dards are based on the IEEE 802.15.4. In the architecture of
was about 780 bytes/s at an SF of 7 and dropped to only WirelessHART, a special computer network manager manages
28 bytes/s at an SF of 12. This shows that the channel data the network topology and communication schedule. It also
capacity was limited if the SF was large. manages the communication quality and adapts to different
The design and drawbacks of LoRaWAN are similar to those types of traffic and changes in its environment. ISA100.11a
of the ALOHA [17] protocol, as there is no mechanism to implements its routing mechanism in a data link layer for
arbitrate access to a shared wireless frequency. Transmissions data forwarding based on routing graphs, which are created
by more than one transmitter at the same time may result in by the system manager in the ISA100.11a architecture. Both
data collision. LoRaWAN uses this approach to simplify the WirelessHART and ISA100.11a can provide better communi-
design of the media access control layer for battery-powered cation reliability as the network is centrally managed, but they
LoRa device, which is required for saving valuable energy, are not specifically designed for monitoring large areas such
thus prolonging its lifetime. Therefore, a LoRa device can as a university campus. Therefore, it may be inappropriate to
transmit data to the GW at any time, causing significant packet make a direct comparison between these standards and the
collision if many devices transmit signal simultaneously [11]. design proposed here.
To test long-range outdoor transmission, The conclusion based on these previous studies, focusing
Petäjäjärvi et al. [14] installed a LoRa node on the roof rack on evaluating the performance of LoRa/LoRaWAN, is clear:
of a car and another on the radio mast of a boat. They tested to serve more high-density and indoor LoRa devices requires
the communication range of their system using a frequency deployment of more LoRa GWs. However, this approach
of 868 MHz and an output power level of 14 dBm, with an SF defeats the original purpose of LoRa/LoRaWAN, that is,
of 12. They observed that the maximal communication range to provide an inexpensive and efficient IoT communication
in an open area was more than 15 km on land and 30 km service. Another issue is that users may need to pay telecom
over water. However, this good performance is mainly due to operators for subscriptions to a LoRaWAN service, and, even
the lack of obstructions and use of only a few LoRa devices as subscribers, they might not receive good communication
transmitting data on the same frequency simultaneously. quality if the operator’s GW is far away.
When a GW needs to serve a large number of LoRa devices In response to these issues, this paper takes a different
at the same time, the capacity of the uplink channel available approach by building a wireless mesh network system based
to a LoRaWAN node strongly depends on the distance from on LoRa PHY rather than LoRaWAN. The proposed design
the base station [15]. With so many devices, the capacity can collect data from IoT sensors distributed across a large
could drop to a mere 100 bits/s on average, especially for region, such as a tall building or a university campus, using
LoRa channels with high SF and for the most distant nodes. only a single LoRa GW. The design allows all LoRa devices
Additionally, in LoRaWAN, there is no clear channel assess- to act as data routers, thereby enhancing the communication
ment mechanism, which increases the probability of packet performance of indoor devices that may have difficulty in
collision. The PDR can drop to just 25% when node density communicating with the GW directly. The design goals of
is very high [11]. Deploying additional GWs can increase the the proposed system are as follows.
performance in terms of receiving data, but at added cost. 1) The campus-wide wireless mesh network must be able
Directional antennas have also been considered as a means to collect data from IoT sensors, even indoor sensors.
of alleviating interference in LoRa networks [12], but only The network should be able to function in environments
simulations have been evaluated to date. The time-related with physical obstacles such as buildings, and it should
performance of LoRa and LoRaWAN was also evaluated [16] provide automatic routing of data back to the GW. This
to determine the uncertainty of schedule to transmission and is very important if nonprofessional users are to manage
long-term clock stability. the system.
One commercial device from NiceRF [18] can suppos- 2) The deployment procedure for the IoT sensor should not
edly help to form a LoRa mesh network, but very limited require any prior configuration. Users should be able
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

Fig. 3. System architecture.

Fig. 2. Architecture of the proposed LoRa wireless mesh network module.

to simply install their IoT sensors without considering

communication issues. Unlike LoRaWAN, there should
be no need to add GWs to increase performance; only
one GW should be needed.
3) The wireless mesh network must be manageable. Infor-
mation about the network routing path, data delivery
ratio, received signal strength (RSS), and SNR of each
path should be collected and displayed on the GW.
This information should help the manager understand
the status of the system and correct communication
problems in the case of weak data links.
4) The design should be empirically verified in an outdoor
Fig. 4. Basic flowchart of a GW.
environment to ensure that the system works in practical
scenarios.
These goals guided the design of our system, which is
network automatically. The GW of this system will sequen-
described in Section III.
tially collect the data accumulated in the IoT sensors. The GW
outputs the data to the management user interface, displaying
III. S YSTEM A RCHITECTURE AND D ESIGN the status of the wireless mesh network and showing the data
In this section, we describe the design of the system. in the IoT sensors. The user’s application can read the data
The basic design of the LoRa wireless mesh network from the user interface, which is useful in applications such as
module is shown in Fig. 2. It is based on a Nuvoton electricity/water–meter monitoring or solar panel management.
Nano100LE3BN [22], an ARM M0 microprocessor, and a We introduce a network formation scenario, schematically
Semtech SX1278 [23] LoRa RF 430 MHz transceiver with shown in Fig. 4. Child_list is a data structure that is initially
a 1.9-dBi gain helical antenna. The current design focuses on empty. It is held by the GW and will store the list of LoRa
wireless communication performance rather than, for instance, nodes joined to its network. In the initial stage, the GW will
low-power requirements; therefore, all nodes and the GWs are broadcast beacons periodically (i.e., every 60 s) inviting other
wall powered via 5-V USB adapters. LoRa nodes (hereafter called nodes) to join its network, as
The microprocessor communicates with the LoRa trans- shown in the flowchart. A LoRa node hearing the beacon from
ceiver via a serial peripheral interface (SPI) to send and receive the nearby GW may join this network by sending a JOIN
data wirelessly. The microprocessor is also integrated with a request and setting the GW as its parent.
USB-UART converter Silicon Lab CP2102 to convert UART Fig. 4 shows how the GW adds the joined node to its
to USB for communicating with the application processor. The child_list and then begins to gather data by calling the subfunc-
nodes and the GWs use the same hardware platform. All mesh tion node_query, as illustrated in Fig. 5. This function sends
networking protocols are run on the ARM M0 microprocessor. a QUERY request to the target node to gather data from the
This module can act as a radio module, similar to XBee [12], node in the child_list. If the QUERY request times out, then
to accept AT-commands from an application processor with the function adds one to the counter timeout_count, marking
USB or UART. An example of an AT-command list can be the target node as missing if timeout_count > 5. The GW
found in [24]; therefore, the application processor can use this will remove the missing node from its child_list and will not
module as a network interface, so users do not need to worry attempt to gather data from it in the next node query. Once the
about how the wireless network is formed. GW has the first child, and child_list is not empty, the GW
Fig. 3 illustrates the architecture of the proposed LoRa wire- will stop broadcasting the periodical beacon. Instead, the GW
less mesh network system. The integration of the IoT sensors will use the data packets sent to its children as beacons for
with the LoRa wireless mesh network module forms the mesh notifying other nodes to join it if necessary, as shown in Fig. 4.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE et al.: MONITORING OF LARGE-AREA IoT SENSORS USING A LoRa WIRELESS MESH NETWORK SYSTEM 5

Fig. 5. Flowchart of subfunction node_query of a GW.

Fig. 7. Flowchart of the subfunction process_packet of a LoRa node.

the list based on its own subjective decision, i.e., the node with
the strongest RSSI and with the smallest hop-count distance
to the GW in the present moment. The reason behind this is
that a node will try to find a parent that is the closest to the
GW but that also has a good communication quality.
Therefore, considering buildings, obstacles, interferences in
the field, and the location of the node, that is, whether it is
indoors or outdoors or on the first floor or the top floor, the
geographically nearest node based on the 2-D top-down view
is not always the best parent candidate. The various methods
for deciding on a suitable parent candidate are not discussed
deeply in this paper. In the practical implementation, the GW
has a global-view topological map of the mesh network after
the network is formed, and it can force assignment of a new
parent to a node according to its comprehensive information.
Fig. 6. Basic flowchart of a LoRa node. This approach is reasonable, as it is very difficult for a node
to find the best parent with its own local-view and limited
knowledge.
If a node can hear the beacon from the GW, it must be able to After the node m decides on a suitable parent, it sends a
overhear the data packet from the same GW. So this approach JOIN request and waits for confirmation before joining the
saves time by eliminating unnecessary beacon broadcasting by network. If the node m receives a confirmation, then it will join
the GW. the network. Afterward, if node m receives a data or command
As shown in Fig. 6, a power-ON node m will listen for packet, it will process this packet and respond as appropriate
a beacon from a nearby GW or will overhear a data packet by calling the subfunction process_packet.
(functioning as a beacon for m) from a nearby GW or node. As shown in Fig. 7, process_packet determines what type
If m has not yet joined the network, it will choose a suitable of packet has been received and processes the packet as
network parent (either the GW or a node). The choice of a appropriate. If a node n receives a JOIN request from another
suitable network parent is based on multiple factors. In this node, then n may accept or reject this request based on its
design, the node stores the RSS indicator (RSSI) of recently current status (for example, n may reject the JOIN request
received beacons/packets from different GWs or nodes in a if n already has too many children) and respond accordingly.
candidate list. The node’s hop-count distance to the GW is also If the incoming packet is a data or command packet, and if the
recorded in the list. The node m selects the best parent from packet destination is n itself, then n will process the packet
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

Fig. 8. Procedure for a node joining the network and querying the GW.

Fig. 9. PDR at different distances under different SF settings.

and respond accordingly; for example, n will respond to a data

request or command from the GW. If the destination is not n
This experiment tested the PDR of two nodes under different
itself but the packet source is n’s parent or child, then n will
SF settings, including SF = 6, 9, and 12. According to the
help to relay this packet to the destination; otherwise, it will
design of the LoRa PHY layer, larger SF paired with higher
drop the packet.
receiver sensitivity should result in higher PDR. Fig. 9 shows
We present an example scenario, illustrated in Fig. 8, in
that SF = 12 yielded the best PDR among all the tested
which node 1 joins the network and sets the GW as its parent.
distances. With SF = 6, the PDR was 56% at 300 m and
The GW adds node 1 to its child_list. Then, the GW queries
dropped to 0% at 500 m. With SF = 12, the PDR was 98%
nodes in its child_list for data. In this design, the GW initiates
at 500 m, observably better than SF = 6. Even when the
the query procedure. As shown in Fig. 8, the GW sends query
distance was extended to 1000 m, the PDR remained at 54%.
M1 to node 1, and node 1 reports its data M2 back to the
This experiment confirmed that large SF can increase the
GW. Meanwhile, the nearby node 3 overhears M2. Note that
transmission range and data delivery performance.
M2 for node 3 is a beacon. Thus, node 3 joins the network and
sets node 1 as its parent. Node 1 notifies the GW of the newly
joined node 3. The mesh network is formed and all nodes are B. In-Building Range Test
connected to the GW based on the procedure. Therefore, in the In order to further investigate the communication per-
next query round, the GW will query node 3 to gather its data formance of the designed LoRa module, we conducted an
using the message sequence M3, M4, M5, and M6. Based experiment in a five-story classroom building on our university
on this sequence, all the nodes can join this multihop wireless campus to test how the LoRa signal traveled vertically across
mesh network, and the GW can query all the nodes. the building’s floors. The receiver node was installed on the
In addition, because the GW has a global-view topologi- fifth floor (5F), the top floor of this building, and transmitter
cal map of the mesh network and has all the routing path nodes were installed on 1F, 2F, 3F, 4F, and 5F. For two
information, in this design, the node can ask the GW for nodes, we used the same test procedure as in the previous test.
the routing path to another node; therefore, the routing is Different LoRa PHY settings were tested, including various
implemented. However, in the real-world application, in most SF and BW values. The RSSI, PDR, and time on air were
cases, the GW will initiate the query procedure for collecting collected for evaluation.
data from the nodes; direct communication between nodes is Fig. 10 shows the PDR at BW = 250 kHz. When the
used less frequently. transmitter was placed on 1F, the LoRa radio signal had to
travel through four concrete stories to 5F. Therefore, all the
IV. E VALUATION AND D ISCUSSION PDRs for 1F are low compared with those of the other floors.
We evaluated the proposed system to verify the applicability For the LoRa transceiver, the larger BW can increase the PHY
of the design. In this section, we describe and discuss our rate and reduce the time on air, but it also lowers the PDR
evaluation. as the output power is distributed across a larger BW. Setting
S F = 6 further lowers the receiver sensitivity and causes even
lower PDR. As a result, the PDRs at SF = 6 are the lowest
A. Preliminary Experiments among the SF settings in this experiment.
First, we tested the proposed LoRa module at different Fig. 11 shows the PDRs at BW = 125 kHz. All
transmission distances by setting the output power level to PDRs at BW = 125 kHz were higher than the PDRs at
20 dBm and the BW to 125 kHz. Two nodes were placed BW = 250 kHz, particularly at SF = 6 on 1F. However,
1 m above the ground. The transmitter sends 200 packets in both tests, the PDR of SF = 6 was still low compared
to the receiver, which is placed in the transmitter’s line of with PDR at other SF settings. Although the PDRs were all
sight on a street in an urban area, with moving and stationary almost 100% at SF settings from 7 to 12 (Fig. 11), the PDRs
obstacles (i.e., pedestrians, cars, and buildings) between them. with SF = 6 ranged from 80% to 93%. We conclude that,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE et al.: MONITORING OF LARGE-AREA IoT SENSORS USING A LoRa WIRELESS MESH NETWORK SYSTEM 7

Fig. 10. PDR on different floors under different SF settings

(BW = 250 kHz). Fig. 13. Another network topology snapshot of the LoRa mesh network
devices.

SPI between the microprocessor and the LoRa transceiver is

SPI, and the time measurement is not very accurate, some
uncertainties may be introduced, varying estimations of the
time on air. However, this might be acceptable as this result
is for understanding how SF and BW affect the time on
air. We can see that larger SF and smaller BW resulted in
longer time on air. The time on air was only 16 ms for
SF = 6 and BW = 250 kHz, but it increased to 2881 ms
for SF = 12 and BW = 62.5 kHz. Therefore, although
larger SF and smaller BW settings increase the PDR, these
settings also significantly prolong the time on air and may
cause serious packet collision if multiple transmitters send
data simultaneously. We can compare this drawback to single-
Fig. 11. PDR on different floors under different SF settings
(BW = 125 kHz).
hop LoRaWAN systems, which suffer from low PDR with a
high density of LoRa devices, as mentioned in [13]. In the
proposed system, the GW initiates all communications and
queries devices. Therefore, the LoRa node cannot actively
send out data, thus avoiding packet collision at the cost of
long transmission delays. Although this delay is a drawback,
in the context of the IoT data collecting, this approach prevents
packet collision without introducing complications into the
simplified design.

C. Campus-Scale Experiment: 19 LoRa Nodes on a

University Campus
In the campus-scale experiment, we placed 19 LoRa nodes
(originally, we installed 20 nodes, but node 6 was lost during
Fig. 12. Time on air (in millisecond) under different BW and SF settings. the experiment), distributed in an 800 m × 600 m area of
our university campus. We allowed the GW to collect data at
1-min intervals for eight days. In this setting, for each node,
if we want to increase the PDR, we should set the SF to a the maximum latency is 60 s, depending on when the node
larger number and the BW to a smaller number. is queried by the GW. Each node is queried by the GW one
Fig. 12 illustrates how time on air (the amount of time time every minute, so it will be queried 60 (min) × 24 (h) ×
needed to transmit data wirelessly) was affected by different 8 (days) = 11 520 times (packets) if a node is connected to
BW and SF settings for a packet with a 20-byte payload. The the network during all eight days. The LoRa PHY parameters
time on air is measured using the GPIO interrupt of the node’s were set to SF = 12 and BW = 250 kHz. In this experiment,
microprocessor; the GPIO interrupt was triggered when the we focused only on communication performance; the content
LoRa transceiver to send data and receiving the notification of data packets was not considered. The nodes form a wireless
of data sent out from the transceiver, and to measure the time mesh network automatically and the GW collects data from
interval between two interrupt triggering with timer. As the these nodes periodically based on a multihop network.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

Fig. 14. PDR under different hop-count limitations.

still achieved at most an average PDR of 69.74%, as shown

in Fig. 14.
Fig. 15 shows how the GW and the nodes were deployed in
this experiment. The GW uses a 3-dBi omniantenna installed
on the roof of the building. All nodes were installed at
locations that were not in the line of sight of the GW.
Fig. 14 shows the end-to-end PDR of this campus-scale
experiment. The PDR was counted on the GW side, based
on the data log collected in this experiment. For ease of
evaluation, data were analyzed by applying different limited-
hop counts. In Fig. 14, 1Hop indicates that all nodes were
restricted to sending data to the GW with only one hop, that is,
with no relay from other nodes. If a data packet was delivered
back to the GW with more than two hops, this packet was not
counted in 1Hop PDR. Therefore, the performance of 1Hop
includes only packets delivered with one hop, so it can be
seen as a LoRaWAN simulation. We can see that 1Hop PDRs
are relatively low. Some nodes did not successfully deliver
any data to the GW, resulting in PDRs of 0% (i.e., nodes
Fig. 15. Pictures of the GW and nodes deployed in this experiment. 1, 4, 8, 12, 17, 18, and 19 in Fig. 14). This result indicates
that the one-hop LoRa network does not perform well for
indoor nodes, which again confirmed the analysis of [11]–[13].
Referring back to the snapshot of the network topology The PDR could fall even further if nodes are allowed to send
shown in Fig. 1, the color of the circles with node numbers data actively (such as in the ALOHA protocol in LoRaWAN
indicates the floor of the building on which each node was approach). The proposed design, by contrast, uses a polling-
installed (e.g., 1F or RF). Nodes at 1F, like node 12, need the based approach, whereby all the nodes listen for requests from
help of other nodes to relay data to the GW (GW in Fig. 1), the GW, then report data only when queried, avoiding the
because they cannot communicate with the GW directly. problem of intranetwork packet collision.
In this case, node 14 at 5F and node 3 at 8F-assisted For these low-PDR nodes, with the exception of node 4,
node 12 in relaying its data to the GW (the path is denoted the PDRs increased dramatically if the hop-count constraint
by 12-14-3-GW). was relaxed to 3Hop, allowing a node to use, at most, three
Another snapshot of the network topology was taken and hops to deliver data to the GW. For example, the PDR of node
is shown in Fig. 13. In this snapshot, node 2 has changed its 12 was 0% at 1Hop, but it increased to 94.94% at 3Hop. For
parent to node 3 and has skipped node 14, communicating all nodes, the average 1Hop PDR was 58.7% and the average
with the GW using the path 12-3-GW. These results confirm 3Hop PDR was 88.49%. These results support our assertion
that our LoRa mesh network module can dynamically form a that the proposed LoRa mesh networking solution can increase
wireless mesh network and adapt to environmental changes. PDR without deploying an additional GW, avoiding additional
Note that, in both snapshots, node 18 was not connected to infrastructure costs associated with the collecting IoT sensor
the network at the moment the snapshot was taken. Further data.
examination revealed that it was located indoors and was not With regard to the relationship between transmission
in a good position to contact a potential parent. Node 18 distance and PDR, multiple factors affect the PDR,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE et al.: MONITORING OF LARGE-AREA IoT SENSORS USING A LoRa WIRELESS MESH NETWORK SYSTEM 9

Fig. 16. Parent selections of node 10 in this experiment.

Fig. 17. Relationship between PDR and RSSI.

Fig. 18. Relationship between PDR and SNR.

including buildings, obstacles, interferences in the field, and current parent for a period of time (i.e., 5 min pass without
the location of the node, such as whether it is indoors or out- receiving any command or data request from its parent), then
doors, or whether it is on a first floor or top floor. Thus, the node will try to find a new parent by sending a JOIN
the geographical distance based on the top-down view does request. Fig. 16 shows that if node 10 cannot join the GW,
not always correlate with the PDR. Transmission distance is then it has four parent candidates: nodes 3, 9, 14, and 15.
not a good factor for estimating PDR, which is very difficult Referring to the pie chart in Fig. 16 (left), we can see that
to model without considering these factors. node 9 was node 10’s parent 74% of time and transferred
Fig. 16 shows the parent candidates for node 10 in the 7085 packets. The bar chart in Fig. 16 (right) indicates that
experiment. Note that this does not include the case in which node 9 had the highest average PDR, almost 100%, and the
node 10 sends data directly to the GW. In this design, a node strongest RSSI among all the parent candidates.
chooses its parent according to its current local information. Fig. 17 shows the relationship between RSSI and PDR.
If the environment changes or the node loses contact with its These signal quality statistics are based on the delivery of
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

all packets to the GW. We can see that the average PDR a command to do so from the GW. If nodes are allowed to
was above 96.1% when RSSI was larger than −110 dBm, send data actively, without waiting for a data request from the
and the average PDR dropped to 5.8% when the RSSI GW, as in the ALOHA LoRaWAN approach, packet collision
was smaller than −130 dBm. In the range from −110 to may decrease PDR in the event that many nodes send data
−130 dBm, the RSSIs decreased as the standard deviation simultaneously. Striking a balance between transmission delay
of PDR increased, indicating unreliable communication. The and packet collision is a complex task.
maximum standard deviation of PDR occurred when RSSI was Regarding the timing performance degradation, as the pro-
in the range from −120 to −130 dBm. In this range, packet posed method uses a mesh network, the number of nodes that
delivery was unreliable, so nodes should avoid low RSSI a network can serve is smaller than for conventional star-
parents if a stable communication path is to be established. topology networks. Referring to Fig. 8, for a one-hop node
Fig. 18 illustrates the relationship between PDR and SNR. (i.e., node 1 in Fig. 8), the message transmission time is 2t,
We can see that larger SNR accompanies higher PDR and where t is the time to transfer a message M. For a two-
smaller standard deviation of SNR. For all SNR values larger hop node (i.e., node 3 in the figure), the time is 4t, and
than −5, the PDRs are greater than 94.2%. A significant cliff for a three-hop node it is 6t. Considering a given period
appears in the PDR trend at SNR below −20, dropping the time p, the number of nodes from which a GW can collect
average PDR to only 5.4%; by contrast, the PDR was still data in a star-topology network is p ÷ 2t, but this drops
65% when the SNR was in the range from −15 to −20. to only one-third of that ( p ÷ 6t) when all the nodes are
The above-mentioned information in combination with three-hop nodes in the mesh network. The mesh network is a
the results suggests that a stable LoRa PHY link with solution for trading off long range with multiple hops, and
a PDR > 90% can be achieved with RSSI > −110 dBm it can reduce the number of nodes that can be served in
and SNR > −5. By contrast, this would be impractical for given period of time. Therefore, this degradation of timing
LoRaWAN systems, which require the deployment of more performance should be considered in designing the IoT data
GWs near the nodes to achieve the same RSSI and SNR. This collecting system. Meanwhile, in this design, in order to
further supports the proposed design, which can increase PDR avoid multiple nodes simultaneously uploading their data over
while circumventing the requirement for additional GWs. the wireless network, causing data collisions, the GW is
responsible for polling data from the nodes and the nodes
are not allowed to actively upload data. The drawback of this
V. D ISCUSSION AND C ONCLUSION
approach is that the latency between generating data in a node
This paper presents the design of a LoRa wireless mesh to uploading that data to the GW is high. However, the polling
network system for collecting data from IoT sensors distrib- approach has greater flexibility by allowing the GW to decide
uted across a large geographical area. Our preliminary and when to query which node, in contrast to the approach in
in-building experiments confirmed the findings of previous which communication is initiated by nodes or centralized
studies regarding the effect of various LoRa PHY parameters scheduling.
(e.g., SF and BW) on communication performance, includ- In addition, questions remain about the security of such a
ing PDR, and time on air. In our campus-scale experiment, network. As this paper focuses on data transmission, security
19 LoRa nodes were distributed over an 800 m × 600 m issues have not been discussed in this paper. However, we can
area. Data analysis comparing the LoRa PHY one-hop wireless directly apply LoRaWAN’s security solution, which uses appli-
network (i.e., simulated LoRaWAN) with our proposed LoRa cation server to provide application session keys for end-to-end
mesh network shows that the mesh network can significantly security and network server to provide network session keys
increase PDR without installing additional GWs. In addition, for network session security in a multihop network.
parent selections, PDR versus RSSI and PDR versus SNR, Finally, the cost in terms of power consumption may also
were comprehensively analyzed. The results suggest thresholds be an issue if this system is to be implemented with, for
for RSSI and SNR for maintaining communication quality. instance, battery-powered nodes. Regarding the issue of low
These thresholds may be impractical for LoRaWAN architec- power at nodes, the router node needs to help relay data
ture, where achieving the ideal RSSI and SNR would require from other nodes, so its wireless transceiver must be on
deployment of many additional GWs, thus raising the cost of at all times; thus, high power consumption is unavoidable.
the infrastructure. Therefore, in the current design, all the router nodes are wall
As far as we know, this is the first academic study extending powered. If some nodes have limited energy, a straightforward
LoRa PHY to mesh networking. It is also the first to evaluate approach can be used to disable the data relaying function
such a mesh network in practical experiments over a large of these nodes, similar to LoRa Class A end devices [25].
geographical area. This paper has explored the potential of Thus, energy used in wireless communication can be saved
IoT sensor deployment and monitoring and collecting of data as a simple node, that is, one with its data relaying function
from those IoT sensors in an area requiring long-range trans- disabled, only enables its wireless transceiver while it needs to
missions. Further study on this topic will focus on adjusting send or receive data. However, this requires further modifying
the SF/BW parameters dynamically to raise PDR. Further the protocol and introducing friend nodes as defined in the
study will also be required to understand how to minimize Bluetooth 5.0 mesh network [19] to assist the simple nodes
delays in data transmission from the nodes. In the current in buffering and relaying data. As the current design has been
design, the node can send out data only if it has received demonstrated the correctness and performance of the system,
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

LEE et al.: MONITORING OF LARGE-AREA IoT SENSORS USING A LoRa WIRELESS MESH NETWORK SYSTEM 11

such modifications of specific requirements can be pursued in [17] N. Abramson, “THE ALOHA SYSTEM: Another alternative for com-
the future work. puter communications,” in Proc. Fall Joint Comput. Conf. (AFIPS Fall),
New York, NY, USA, 1970, pp. 281–285.
[18] (2017). NiceRF LoRa-MESH Series Mesh Network Modules. [Online].
ACKNOWLEDGMENT Available: http://nicerf.com/product_149_203.html
[19] (2017). Bluetooth 5.0 Mesh Networking Specifications.
The authors would like to thank ICP DAS Company Ltd., Accessed: Dec. 30, 2017. [Online]. Available: https://www.bluetooth.
Taiwan, for its technology support. They would also like to com/specifications/mesh-specifications
[20] Industrial Communication Networks—Wireless Communica-
thank P.-J. Lee and P.-C. Kuo for their excellent assistance. tion Network and Communication Profiles—WirelessHART,
Standard IEC 62591:2010, 2010.
R EFERENCES [21] (2017). The ISA100 Standards Overview & Status. [Online]. Available:
http://isa100wci.org/Documents/PDF/ISA100_Standards-Overview_
[1] Y. Cheng et al., “AirCloud: A cloud-based air-quality monitoring system WCI_Webinar_V3_12Apr10.aspx
for everyone,” in Proc. SenSys, Nov. 2014, pp. 251–265. [22] Nuvoton. (2017). NUC100 Advanced Series. Accessed: Sep. 19, 2017.
[2] J. Shah and B. Mishra, “IoT enabled environmental monitoring system [Online]. Available: http://www.nuvoton.com/hq/products/
for smart cities,” in Proc. Int. Conf. Internet Things Appl. (IOTA), microcontrollers/arm-cortex-m0-mcus/nuc100-200-advanced-series/
Jan. 2016, pp. 383–388. [23] Semtech. (2017). SX1278 137 MHz to 525 MHz Low Power Long
[3] B. Ando, S. Baglio, A. Pistorio, G. M. Tina, and C. Ventura, “Sentinella: Range Transceiver. Accessed: Sep. 19, 2017. [Online]. Available:
Smart monitoring of photovoltaic systems at panel level,” IEEE Trans. http://www.semtech.com/wireless-rf/rf-transceivers/sx1278/
Instrum. Meas., vol. 64, no. 8, pp. 2188–2199, Aug. 2015. [24] (2017). XBee/XBee-PRO S1 802.15.4 (Legacy) RF Modules.
[4] G. Mois, S. Folea, and T. Sanislav, “Analysis of three IoT-based wireless Accessed: Dec. 30, 2017. [Online]. Available: https://www.digi.com/
sensors for environmental monitoring,” IEEE Trans. Instrum. Meas., resources/documentation/Digidocs/90000982/Default.htm
vol. 66, no. 8, pp. 2056–2064, Aug. 2017. [25] (2017). LoRaWAN Classes. Accessed: Dec. 30, 2017. [Online]. Avail-
[5] Sigfox. (2017). Sigfox—The Global Communications Service Provider able: https://www.lora-alliance.org/technology
for the Internet of Things (IoT). Accessed: Sep. 2, 2017. [Online].
Available: https://www.sigfox.com/en
[6] LoRa-Alliance. Accessed: Sep. 2, 2017. [Online]. Available:
https://www.lora-alliance.org/ Huang-Cheng Lee (SM’12) received the
[7] K. Flynn. (2017). Standardization of NB-IOT Completed. Ph.D. degree from National Tsing-Hua University,
Accessed: Sep. 2, 2017. [Online]. Available: http://www.3gpp.org/news Hsin-Chu, Taiwan, in 2010.
-events/3gpp-news/1785-nb_iot_complete Since 2000, he has been in the industry and
[8] (2017). LoRa Modulation Basics. Accessed: Sep. 3, 2017. [Online]. has a wide breadth of experience designing
Available: http://www.semtech.com/images/datasheet/an1200.22.pdf personal digital assistants, cellular phones, and
[9] M. Centenaro, L. Vangelista, A. Zanella, and M. Zorzi, “Long-range low-power embedded systems. In 2011, he joined
communications in unlicensed bands: The rising stars in the IoT the Department of Communications Engineering
and smart city scenarios,” IEEE Wireless Commun., vol. 23, no. 5, and Electrical Engineering, National Chung-Cheng
pp. 60–67, Oct. 2016. University, Chiayi, Taiwan, where he has been an
[10] K.-H. Ke, Q.-W. Liang, G.-J. Zeng, J.-H. Lin, and H.-C. Lee, “Demo Associate Professor since 2015. He has a strong
abstract: A LoRa wireless mesh networking module for campus-scale track record of collaborating with industry partners to transfer technologies
monitoring,” in Proc. 16th Int. Conf. Inf. Process. Sensor Netw. (IPSN), from academic research into practice. His current research interests include
Pittsburgh, PA, USA, Apr. 2017, pp. 259–260. wireless sensors, mesh networking, Internet of Things, and low-power
[11] N. Varsier and J. Schwoerer, “Capacity limits of LoRaWAN tech- embedded systems, especially for natural and industrial environment
nology for smart metering applications,” in Proc. IEEE Int. Conf. monitoring.
Commun. (ICC), Paris, France, May 2017, pp. 1–6. Dr. Lee has been an Associate Editor of the IEEE T RANSACTIONS ON
[12] T. Voigt, M. C. Bor, U. Roedig, and J. M. Alonso, “Mitigating inter- I NSTRUMENTATION AND M EASUREMENT since 2015 and an Associate
network interference in LoRa networks,” in Proc. Int. Conf. Embedded Editor of the IEEE S ENSORS J OURNAL since 2017.
Wireless Syst. Netw. (EWSN), Uppsala, Sweden, Feb. 2017, pp. 323–328.
[13] A. Augustin, J. Yi, T. Clausen, and W. M. Townsley, “A study of LoRa:
Long range & low power networks for the Internet of Things,” Sensors,
vol. 16, no. 9, p. 1466, 2016. Kai-Hsiang Ke received the B.S. and M.S. degrees
[14] J. Petäejäejäervi, K. Mikhaylov, A. Roivainen, T. Hänninen, and M. Pet- from the Department of Communications Engi-
tissalo, “On the coverage of LPWANs: Range evaluation and channel neering, National Chung-Cheng University, Chiayi,
attenuation model for LoRa technology,” in Proc. 14th Int. Conf. ITS Taiwan, in 2015 and 2017, respectively.
Telecommun., Dec. 2015, pp. 55–59. He is currently an Engineer with Delta Electronics,
[15] K. Mikhaylov, J. Petäejäejäervi, and T. Haenninen, “Analysis of capacity Tainan, Taiwan.
and scalability of the LoRa low power wide area network technology,” in
Proc. 22th Eur.Wireless Conf., Eur. Wireless (VDE), May 2016, pp. 1–6.
[16] M. Rizzi, P. Ferrari, A. Flammini, and E. Sisinni, “Evaluation of the
IoT LoRaWAN solution for distributed measurement applications,” IEEE
Trans. Instrum. Meas., vol. 66, no. 12, pp. 3340–3349, Dec. 2017.