IEEE Circuits and Systems Magazine - Q3 2020 - 13

- ell-known standards that can reach high data-rates.
w
Other options based on cellular communications such as
4 G/5G may be also considered. However, all these RATs
require great amount of energy to work. For that reason,
LPWAN-based solutions such as LoRaWAN, Sigfox, or
Narrow Band-Internet of Things (NB-IoT) [19] are currently being adopted by the IoT market as they provide
very long transmission distances with highly reduced energy consumption, although with the limitation of severely
limiting the transmission rate. Given the heterogeneous
characteristics of the available RATs, it is crucial to precisely select the most adequate technology to guarantee both
the energy efficiency of the device as well as the QoS of
the transmitted packets. At this point, TinyML comes into
play with the aim of smartly supporting this decision. As
discussed in next Section, a number of ML techniques may
be employed to this end, thanks to the algorithm adaptations provided by the afore-reviewed TinyML frameworks.
This multi-R AT smart board can be integrated
within a plethora of devices and gadgets, e.g., vehicles,
wearables, sensors, robots, etc., which will be enriched
with intelligence and communication capabilities. These
elements are connected to the cloud or other devices
through the available communication interfaces in order to gain access to a range of next-generation services,
which are finally accessed by me a n s of e nd - u se r's
personal devices such as smartphones or tablets.

ing the device status, several aspects are considered. The
percentage of remaining battery (B) is one of the key factors
to evaluate as the power consumption of both interfaces
clearly differs. The coverage level (C) of each RAT {c1, c2},
is also taken into account in order to avoid transmissions
over a lossy link. Besides, LoRaWAN makes use of licensefree bands, e.g., Industrial, Scientific and Medical (ISM), so
the access to this wireless medium is normally limited by
international regulations. Thus, we model the availability
of the LoRaWAN interface with the boolean feature L. We
consider that NB-IoT does not present this restriction as it
makes use of licensed spectrum. We also assume that there
are not dedicated queues for each interface; instead, we
consider a One in-One out message-forwarding system.
This assumption is made given the memory constraints of
MCUs, which prevent them of storing great amount of data
in RAM. Two message features are also evaluated in the decision process. The data size (S), in bytes, is considered as
LoRaWAN has a maximum payload length of 240 B. Finally,
the urgency level of the message (U) is also contemplated
with the aim of supporting different classes of services.
With the features identified above, the current
system status is defined by the following vector
(B, " c 1, c 2 ,, L, S, U ), which is employed for determining
the most adequate action A ! " a 0, a 1, a 3 , . We tackle this
issue as a supervised multi-class classification problem.
To this end, a dataset has been generated to train a number of algorithms that have been embedded in a real constrained device as explained in the following.

B. Problem Statement
We consider the problem of intelligently selecting the most
adequate communication interface in a multi-RAT set up C. Dataset
when certain data is generated to be transmitted in a non- A synthetic training dataset of 10,000 samples, i.e., feaconnection-oriented fashion. Concretely, given the architec- tures vectors, has been generated and each of the
ture presented above, we consider an end-device producing -features composing the input vectors has received a
sporadic messages and that is equipped with two commu- value randomly calculated. The values for the message
nication interfaces, namely, LoRaWAN and NB-IoT. As both urgency (U) and the coverage of each RAT (c1, c2) have
of them are LPWAN-based solutions, they provide long been generated by assigning them uniformly distributrange communications with reduced power consumption, ed values in the range [0%, 100%]. The values for the
which are highly valued characteristics for moving devices. device's remaining battery (B) and the packet size (S),
However, given these common general attributes, as shown have been randomly calculated in the ranges [1%, 100%]
in Table II, both present unique characteristics in terms of and [50 bytes, 500 bytes], respectively, following a unibandwidth, uplink/downlink latency, duty-cycle, etc. There- form distribution as well. Finally, the LoRaWAN interface
fore, by considering these RAT features as well as the device availability (L), which was modeled as a boolean input,
status and the message characteristics, it should be decid- has received true/false values uniformly distributed.
Each of the 10,000 samples has been labelled with the
ed whether to transmit a message or not, and if so, the most
most adequate action considering its input features. To
adequate communication interface should be selected.
The set of all possible actions (A) is {a0, a1, a2}, where
Table II.
a0 represents the action of
Main characteristics of RATs under consideration.
dropping the message and {a1,
RAT
Throughput
Latency
RX Sensitivity
Power Consumption
a2} indicates sending the data
LoRaWAN
Low
High
High
Low
using LoRaWAN or NB-IoT inNB-IoT
High
Low
Medium
Medium
terfaces, respectively. RegardTHIRD QUARTER 2020

IEEE CIRCUITS AND SYSTEMS MAGAZINE

IEEE Circuits and Systems Magazine - Q3 2020

Table of Contents for the Digital Edition of IEEE Circuits and Systems Magazine - Q3 2020