sensors

Article
A New NILM System Based on the SFRA Technique and
Machine Learning
Simone Mari * , Giovanni Bucci, Fabrizio Ciancetta , Edoardo Fiorucci and Andrea Fioravanti

Dipartimento di Ingegneria Industriale e dell’Informazione e di Economia, Università dell’Aquila,

67100 L’Aquila, Italy; [email protected] (G.B.); [email protected] (F.C.);
[email protected] (E.F.); [email protected] (A.F.)
* Correspondence: [email protected]

Abstract: In traditional nonintrusive load monitoring (NILM) systems, the measurement device is
installed upstream of an electrical system to acquire the total aggregate absorbed power and derive
the powers absorbed by the individual electrical loads. Knowing the energy consumption related to
each load makes the user aware and capable of identifying malfunctioning or less-efficient loads in
order to reduce consumption through appropriate corrective actions. To meet the feedback needs of
modern home, energy, and assisted environment management systems, the nonintrusive monitoring
of the power status (ON or OFF) of a load is often required, regardless of the information associated
with its consumption. This parameter is not easy to obtain from common NILM systems. This article
proposes an inexpensive and easy-to-install monitoring system capable of providing information on
the status of the various loads powered by an electrical system. The proposed technique involves the
processing of the traces obtained by a measurement system based on Sweep Frequency Response
Analysis (SFRA) through a Support Vector Machine (SVM) algorithm. The overall accuracy of the
system in its final configuration is between 94% and 99%, depending on the amount of data used
for training. Numerous tests have been conducted on many loads with different characteristics. The
positive results obtained are illustrated and commented on.

Keywords: machine learning (ML); nonintrusive load monitoring (NILM); smart home; support
vector machine (SVM); sweep frequency response analysis (SFRA)

Citation: Mari, S.; Bucci, G.;

Ciancetta, F.; Fiorucci, E.; Fioravanti,
A. A New NILM System Based on the
1. Introduction
SFRA Technique and Machine
Learning. Sensors 2023, 23, 5226.
The goal of energy saving within modern smart homes and energy management
https://doi.org/10.3390/s23115226
systems is pursued by monitoring and controlling household parameters, such as lighting
and home temperature [1]. This need has led to a significant increase in attention to
Academic Editor: Daniele Cenni
nonintrusive load-monitoring (NILM) systems.
Received: 3 April 2023 Among the energy-monitoring systems, those based on the NILM technique represent
Revised: 26 May 2023 one of the most relevant solutions. The total energy consumption of users is monitored and
Accepted: 29 May 2023 the consumption of each individual load is identified. For this purpose, the measurements
Published: 31 May 2023 of current and voltage are carried out, or often of the current alone; the data collected
are then processed with a so-called “disaggregation” algorithm. The main advantages of
the nonintrusiveness are the simplicity and cost-effectiveness of installation. Therefore,
systems of this type are useful for both consumers and utility companies when analyzing
Copyright: © 2023 by the authors. the use and costs of electricity.
Licensee MDPI, Basel, Switzerland. In the early 1990s, the first NILM system was proposed [2]. Since then, more advanced
This article is an open access article
algorithms have enabled a significant improvement in energy-unbundling systems.
distributed under the terms and
This is especially true over the past decade, which has seen a significant increase in
conditions of the Creative Commons
interest in this topic.
Attribution (CC BY) license (https://
The first NILM systems detected events and classified the various loads using tradi-
creativecommons.org/licenses/by/
tional algorithms [2,3]. The most modern, however, use artificial intelligence algorithms, in
4.0/).

Sensors 2023, 23, 5226. https://doi.org/10.3390/s23115226 https://www.mdpi.com/journal/sensors

Sensors 2023, 23, 5226 2 of 25

particular, through machine-learning (ML) techniques. For example, in some studies [4,5],
the energy disaggregation problem has been reformulated as an adaptive filtering problem;
refs. [6,7] propose model-driven NILM systems and the works proposed in other stud-
ies [8–10] are based on hidden Markov chains, while others [11–15] use artificial neural
networks. The latter types of algorithms learn from the data provided and can perform
certain tasks. Therefore, ML algorithms continue to improve over time by learning from
data with minimal human intervention [16].
With regard to the NILM problem, these systems process the active power—and
sometimes also the reactive power—absorbed by the monitored system [16]. However,
NILM systems have been developed based on transient rather than stationary characteristics
or the analysis of other quantities, which differ due to belonging to different domains (time
or frequency).
In this sense, the sampling frequency is a fundamental parameter used to define the
extractable information. Low-frequency time series were processed to evaluate steady-state
characteristics. On the other hand, time series at other frequencies were processed to
obtain information about the startup and shutdown transients to be able to discriminate
the loads through their dynamic parameters (overshoot, rise time, etc.) or by characterizing
appliances based on the pulses produced on the power line. Other attempts have been
made by processing the trajectories drawn on the V-I plane.
Today, the division of NILM systems into event-driven and non-event-driven sys-
tems is the most widely used division and is the best for defining the state of the art of
these systems.
The former involves the detection of an event (understood as an appliance turning
on, turning off, or switching to a different consumption state) and then classifying it based
on the features associated with the appliance that caused it. This type of approach can
therefore be divided into three basic steps: event detection, feature extraction, and load
identification. In particular, the last step is performed in most cases using ML algorithms
that work well as classification systems. Numerous supervised ML algorithms have been
proposed in the literature, including K Nearest Neighbor (KNN) [17], naïve Bayes [18],
Decision Tree (DT) [19], Support Vector Machine (SVM) [20], Principal Component Analysis
(PCA) [21], and Artificial Neural Network (ANN) [22,23]. Finally, unsupervised [24]
and semi-supervised learning algorithms [25], as well as those related to graph signal
processing [26], have also been proposed.
On the other hand, non-event-based systems are NILM systems that do not have
an event-detection phase. In these cases, the concept of the “signature”/“features” of an
appliance is also lost, as the only feature used by the models is the aggregate power profile.
They use a window of samples of the aggregate signal (therefore time series data) as input;
the samples are processed continuously without waiting for the occurrence of events. For
this reason, this type of system is particularly suitable for low-frequency signals. Indeed, it
was developed precisely to allow the processing of signals acquired with reduced frequen-
cies, for which the detection of events is more difficult. In some cases, the disaggregation
problem is formulated as a blind source separation (BSS) problem—that is, the problem of
recovering a signal from a set of mixed signals. Numerous approaches have been proposed
in the last decade, the most significant being those based on Combinatorial Optimiza-
tion [27], Discriminative Sparse Coding [28], Hidden Markov Model Approaches [9,29–35],
and Deep Learning (DL) [36–43].
NILM systems are used in a wide range of applications. Among these, very promising
are the applications in Ambient Assisted Living, i.e., systems that make it possible to meet
the needs of elderly or disabled users, allowing them to live independently [44]. In fact,
knowing the changes in the status of the various appliances in a relatively short time, it is
possible to infer the Activities of Daily Living (ADL) of the occupants.
This paper presents a measurement system for nonintrusive monitoring (it does not
require modifications to the electrical system) based on the injection of a variable frequency
sinusoidal signal and the characterization of the system based on the response to it. This
Sensors 2023, 23, 5226 3 of 25

technique is called Sweep Frequency Response Analysis (SFRA) and is widely used in
diagnostics and fault-finding in transformers and electric motors.
The proposed solution is very different from the other solutions proposed in the
literature, which provide for the analysis of time-varying electrical signals through different
approaches. Following these approaches, the focus is generally on transients of the absorbed
current, which indicate a change in the connection state. The measurement of the current in
static conditions does not allow the identification of active devices, except in very special
simple cases.
The approach proposed in this paper makes it possible to identify which appliances
are inserted through a measurement performed in static conditions (not in the connec-
tion/disconnection transient). It allows for the detection of a sort of signature that is unique
and independent of the absorbed current. This approach, as illustrated below, allows us to
overcome the typical problems of NILM systems in identifying multi-state or continuous
variable load household appliances (or, in general, electrical loads).
All the SFRA apparatuses available on the market can only work on single devices that
are switched off and disconnected from the grid. The SFRA system proposed in this article
can operate online [45], thus allowing it to extend its operating range to systems for contin-
uous diagnostics on devices while supplied by the mains; no functioning interruptions or
disconnection operations are needed for the standard SFRA apparatuses.
The proposed system is based on a machine-learning algorithm, the Support Vector
Machine (SVM), which is capable of determining the status of individual household appli-
ances starting from the measurement obtained by the SFRA system. It was installed on a
home test system and acquired and processed the data locally.
Extensive measurements were made in order to verify the operational characteristics.
The results obtained from field applications are also included and discussed.

2. Frequency Response Analysis of Household Appliances

SFRA has been successfully used to perform diagnostics on the windings of electric
machines during the production process [46,47]. An electric machine can be considered
a complex electrical network of capacitances, inductances, and resistors. As shown in
Figure 1, the SFRA instrument injects a sinusoidal excitation voltage (the typical amplitude
is 10 Vpp) with a continuously increasing frequency into one end of the transformer
winding and measures the signal returning from the other end. This test is conducted
with the machine disconnected from the power line. More details are reported in another
Sensors 2023, 23, x FOR PEER REVIEW 4 of 26
study [46].

Figure1.1.SFRA
Figure SFRA applied
applied to
to aa star-connected
star-connected electric
electric machine.
machine.

The comparison of input and output signals generates a frequency response, which
can be compared with reference data. Degradation of the insulating materials or a change
in the shape of the windings will result in a change in the RLC components of the network
and, consequently, in the frequency response curve. Faults can therefore be detected by
Sensors 2023, 23, 5226 4 of 25

The comparison of input and output signals generates a frequency response, which
can be compared with reference data. Degradation of the insulating materials or a change
in the shape of the windings will result in a change in the RLC components of the network
and, consequently, in the frequency response curve. Faults can therefore be detected by
processing correlation indices between different curves.
In the proposed application, shown in Figure 2, the SFRA technique is applied to
the electrical system supplied by the mains in order to obtain a signature that allows for
discriminating different power supply conditions of a domestic system. The applied signal
and the output signal, between the terminal of the neutral conductor and the ground, are
Sensors 2023, 23, x FOR PEER REVIEW 5 of 26
acquired and processed by the system. The proposed measurement system can therefore
be conveniently installed on a standard domestic socket.

Figure2.2.SFRA
Figure SFRAsystem.
system.

AAlow-voltage
Hanning window (±5 Vpp) withsinusoidal
a width equalsignalto the
withacquisition time (corresponding
variable frequency (from 2 kHz to to
64
cycles of the generated frequency) was used to process the FFT.
1.5 MHz) is superimposed on the supply voltage (240 Vrms and 50 Hz) and applied between Downstream of the FFT
processing,
the power phase the system
conductor calculated
terminal the
and /𝑉 ratio. For example, the 1 kHz response is
𝑉 ground.
achieved by injecting
The signal generator a 1 kHz sinusoidal
is coupled signal
to the generated
network at a frequency
by means of 100 filter
of a band-pass MS/s.that The
appliedonly
allows signaltheand the output
passage of thesignal were sampled
test signal. The twoatinput a sampling
channels rateofofthe
25 measurement
× 1000 = 25,000
Hz. A time
circuit window
are also of (1/1000)
decoupled from× the
64 =power
0.064 s supply
was considered for theband-pass
by two other processingfilters.
of the FFTs,
The
filters block bothtothe
corresponding 1600fundamental
samples. This frequency
process(50was Hz) and thefor
repeated harmonic componentsof(up
all the frequencies to
inter-
2est.
kHz)
The [48].
block diagram of the LabVIEW code is shown in Figure S1.
As
In the
order first
to part of thethe
evaluate work, the system’s
validity response
of the signature forwas evaluated
different over a ranges,
frequency fairly widefour
frequency
sub-bandsrange were and by acquiring a sufficiently high number of points.
defined:
(1) The
2–10frequency
kHz; response was obtained by injecting a signal generated at 100 MS/s. In
order
(2) 10–100 kHz; memory, the sampling frequencies to acquire both applied and output
to optimize the
signals
(3) 100 were
kHz–1 adapted
MHz;according to the frequency to be analyzed. In detail, the sampling
frequency was chosen as being equal to 25 times the analyzed frequency. To obtain a
(4) 1–1.5 MHz.
better resolution, the FFT was performed by fixing a frequency bin at the frequency of the
For each
generated sub-band,
sinusoid. 200 points
The FFT was also were initially on
performed acquired. These
the output sub-divisions
signal and the samplewere ob- at
tained considering
the same bin was considered. the possible response to this type of excitation signal. Figure 3 sche-
matically
A Hanning showswindowthe installation
with a widthof the SFRA
equal system
to the in the time
acquisition test (corresponding
system. From the to
knowledge in the literature about SFRA [48], the low-frequency
64 cycles of the generated frequency) was used to process the FFT. Downstream response (2–10 kHz)of theis
characterized
FFT processing, bythean ohmic-inductive
system calculated behavior
the Voutin/V which the characteristics of the grid up-
in ratio. For example, the 1 kHz re-
stream of the system are predominant; therefore,
sponse is achieved by injecting a 1 kHz sinusoidal signal generated the contribution of the
at loads is usually
a frequency of
notMS/s.
100 significant. The medium-frequency
The applied signal and the outputresponse (10 were
signal kHz–1sampled
MHz) isatcharacterized
a sampling rate by res-
of
onance
25 × 1000 phenomena.
= 25,000 Hz. AsAthis
timeband is generally
window the most
of (1/1000) × 64interesting
= 0.064 s was in terms of the effect
considered for the of
loads on the response, it has been split into two sub-bands to
processing of the FFTs, corresponding to 1600 samples. This process was repeated for all increase resolution. The
high-frequency
the frequencies ofresponse
interest.(1–1.5 MHz)
The block is characterized
diagram by capacitive
of the LabVIEW code is effects
shown duein both
Figure to S1.
the
network and the user loads and the connection of the measuring instrument
In order to evaluate the validity of the signature for different frequency ranges, four itself, which
generally determine
sub-bands were defined: a poor reproducibility of the measurement.
(1) 2–10 kHz;
(2) 10–100 kHz;
(3) 100 kHz–1 MHz;
(4) 1–1.5 MHz.
Sensors 2023, 23, 5226 5 of 25

For each sub-band, 200 points were initially acquired. These sub-divisions were
obtained considering the possible response to this type of excitation signal. Figure 3
schematically shows the installation of the SFRA system in the test system. From the
knowledge in the literature about SFRA [48], the low-frequency response (2–10 kHz) is
characterized by an ohmic-inductive behavior in which the characteristics of the grid
upstream of the system are predominant; therefore, the contribution of the loads is usually
not significant. The medium-frequency response (10 kHz–1 MHz) is characterized by
resonance phenomena. As this band is generally the most interesting in terms of the effect
of loads on the response, it has been split into two sub-bands to increase resolution. The
high-frequency response (1–1.5 MHz) is characterized by capacitive effects due both to the
network and the user loads and the connection of the measuring instrument6 itself,
Sensors 2023, 23, x FOR PEER REVIEW of 26 which
generally determine a poor reproducibility of the measurement.

Figure 3. Installation of the SFRA in the test system.

Figure 3. Installation of the SFRA in the test system.

The sinusoidaltest
The sinusoidal test signal
signal introduces
introduces no problems
no problems to thetosystem.
the system.
This is This is essentially
essentially
due
due to the reduced amplitude of the test signal with respect to the line voltage (1.54%), (1.54%),
to the reduced amplitude of the test signal with respect to the line voltage
which
which is fullywithin
is fully withinthe thelimits
limits imposed
imposed by standard
by the the standard
[49]. [49].
During
During thethetests,
tests,it it
waswas verified
verified thatthat the signal
the signal does does not create
not create problemsproblems in intelligent
in intelligent
automation systemsoperating
automation systems operatingwith withconveyed
conveyed waves
waves [50].
[50]. ThisThis is also
is also because
because thesethese
sys- systems
tems adopt
adopt sophisticated
sophisticated signal-modulationalgorithms
signal-modulation algorithms that
that encode
encodethe thedata
datatransmitted
transmitted with
with different
different sub-carriers
sub-carriers or that
or that widen
widen thethe transmissionband
transmission band (Spread
(Spread Spectrum),
Spectrum),ob- obtaining
a better resistance to interference and noise. Other systems adopt OrthogonalFre-
taining a better resistance to interference and noise. Other systems adopt Orthogonal Frequency
quency Division
Division Multiplexing
Multiplexing (OFDM) (OFDM) modulation
modulation techniques,
techniques, which
which areare even
even moreeffective.
more ef-
fective.
Several tests were performed at a residential test facility. A wide variety of loads
Several tests were performed at a residential test facility. A wide variety of loads were
were taken into consideration, powering them individually or simultaneously and under
taken into consideration, powering them individually or simultaneously and under dif-
different working
ferent working conditions:
conditions:
(1)
(1)Hairdryer;
Hairdryer;
(2)
(2)Microwaveoven;
Microwave oven;
(3)
(3)Lamp;
Lamp;
(4)
(4)Laptop;
Laptop;
(5)
(5) Induction
Inductionhob;
hob;
(6) Heater;
(7) Drill;
(8) TV.
Figure 4 shows the frequency response of these appliances when powered individu-
ally. The measurements were conducted in 24 different power supply scenarios, as sum-
Sensors 2023, 23, 5226 6 of 25

(6) Heater;
(7) Drill;
(8) TV.
Figure 4 shows the frequency response of these appliances when powered individually.
The measurements were conducted in 24 different power supply scenarios, as summarized
in Table 1. It is important to note that Scenario 1 represents the case in which none of the
appliances was powered (condition indicated with “Open Circuit” in Figure 1). Scenarios 2
Sensors 2023, 23, x FOR PEER REVIEW 7 of 26
to 9 represent the single power supply conditions of household appliances. Scenarios 10 to
24 represent the simultaneous power conditions.

Figure 4. Frequency response

Figure of individually
4. Frequency poweredpowered
response of individually household appliances.
household appliances.

Table 1.scenarios.
Table 1. Power supply Power supply scenarios.
Hairdryer Microwave Oven Lamp Laptop Induction hob Heater Drill TV
Hairdryer Microwave Oven 1 Lamp Laptop Induction hob Heater Drill TV
1 2 x
2 x 3 x
4 x
3 x 5 x
4 6 x x
5 7 x x
6 8 x x
9 x
7 10 x x
x
8 11 x x x
9 12 x x x
13 x x x
10 x x
14 x x x
11 x 15 x x
x x
12 x 16 x xx x
13 17 xx x x x x
18 x x x
14 x x x
19 x x x x
15 x 20 x x xx x x
16 21 x x x xx x
17 x 22 x x x xx x
23 x x
18 x x xx x
24 x x x x
19 x x x x
20 x x
To support x
an objective x 5 shows the lower and upper envelopes
evaluation, Figure
21 of the traces obtained inxthe presence and absence of each of the eight considered
x x
appli-
22 x
ances, obtained x the measurements performed for the different scenarios.
following x x
Meas-
23 x urements were performed for each of the 24 scenarios reported x in Table 1,xthus obtaining
x
24 x 24 SFRA traces. For each envelope (related toxeach appliance), the traces were
x divided into
x
Sensors 2023, 23, 5226 7 of 25

To support an objective evaluation, Figure 5 shows the lower and upper envelopes of
the
Sensors 2023, 23, x FOR PEER REVIEW traces obtained in the presence and absence of each of the eight considered appliances,8 of 26
obtained following the measurements performed for the different scenarios. Measurements
were performed for each of the 24 scenarios reported in Table 1, thus obtaining 24 SFRA
traces. For each envelope (related to each appliance), the traces were divided into two
two groups according to the presence or absence of the appliance in the power supply
groups according to the presence or absence of the appliance in the power supply scenario.
scenario. The envelopes were then obtained by considering the maximum and minimum
The envelopes were then obtained by considering the maximum and minimum values of
values of each of the two groups for each frequency bin. From these envelopes, it is im-
each of the two groups for each frequency bin. From these envelopes, it is immediately evi-
mediately
dent evident
that the that theofcontribution
contribution of the low-frequency
the low-frequency measurement
measurement (2–10 (2–10
kHz) is not kHz) is
influenced
by the different load configurations; therefore, in the rest of the work, we will only referwe
not influenced by the different load configurations; therefore, in the rest of the work, to
willother
the only three
refer to the other three sub-bands.
sub-bands.

(a) (b)

(c) (d)

(e) (f)

Figure 5. Cont.
Sensors
Sensors 2023,
2023, 23,
23, x5226
FOR PEER REVIEW 9 8ofof 26
25

(g) (h)
Figure
Figure 5.
5. Envelopes
Envelopes of
of the
the traces
traces obtained
obtained in
in the
the presence
presence and
andabsence
absenceof
ofthe:
the:(a)
(a)hairdryer,
hairdryer,(b)
(b)mi-
mi-
crowave oven, (c) lamp, (d) laptop, (e) induction hob, (f) heater, (g) drill, and (h) TV.
crowave oven, (c) lamp, (d) laptop, (e) induction hob, (f) heater, (g) drill, and (h) TV.

These traces were

These traces wereusedusedasasinputs
inputsto atomachine-learning-based
a machine-learning-based classification
classification algo-
algorithm,
rithm, the Support Vector Machine (SVM), to determine the
the Support Vector Machine (SVM), to determine the correct combination of poweredcorrect combination of pow-
ered appliances.
appliances. A NILM
A NILM system
system basedbasedon on this
this type
type ofof inputisiseasy
input easytotoinstall,
install, as
as it
it can
can be
be
connected
connected to to aa standard
standard domestic
domestic socket,
socket, such
such as any household
as any household appliance.
appliance. Traditional
Traditional
NILM systems, on
NILM systems, onthetheother
otherhand,
hand,measure
measure thethe aggregate
aggregate power
power upstream
upstream of plant
of the the plant
and
and therefore
therefore require
require a more a more difficult
difficult installation.
installation.
The
The measurement
measurement obtainedobtained represents
represents thethe transfer
transfer function
function of the equivalent
equivalent RLC
circuit
circuit [23]. Therefore,
Therefore, the result
result mainly influenced by the physical characteristics of the
is mainly influenced by the physical characteristics the
appliances rather than
appliances rather thanby bytheir
theirpower
powerabsorption.
absorption.ThisThis represents
represents a great
a great advantage
advantage for for
the
the discrimination
discrimination of multi-state
of multi-state or continuously
or continuously variable
variable load appliances
load appliances (such as(such as whose
drills) drills)
whose identification
identification is ofteniscritical
often critical for systems
for systems based onbased
the on the analysis
analysis of power of power consump-
consumption.
tion. The transfer function is minimally influenced by the choice of the socket in which to
The
install thetransfer
measuringfunction is minimally
system. Tests were influenced
carried outbyin the
allchoice of the shown
the sockets socket inin which
Figureto 3;
install thepossible
all of the measuring system.
positions ofTests were carried
the instrument on out
the in all thesockets
various socketsallow
shown theinmaximum
Figure 3;
reproducibility
all of the possible of positions
the measurement. Regardless,
of the instrument the various
on the instrument is meant
sockets allowtothebemaximum
used on a
single socket. The
reproducibility proposed
of the algorithm
measurement. is described
Regardless, theininstrument
Section 3. is meant to be used on a
single socket. The proposed algorithm is described in Section 3.
3. Machine-Learning Systems
Machine learningSystems
3. Machine-Learning is the field of study that allows computers to learn without being
explicitly
Machine learning is theUnlike
programmed [51]. field oftraditional
study that programming, whichtoprovides
allows computers a list ofbeing
learn without more
or less complex
explicitly rules defined
programmed [51]. by the programmer
Unlike to obtain certain
traditional programming, outputs,
which machine
provides learn-
a list of
ing automatically learns patterns and correlations to solve extremely complex
more or less complex rules defined by the programmer to obtain certain outputs, machine problems. In
problems where existing solutions require a lot of manual adjustments or
learning automatically learns patterns and correlations to solve extremely complex prob-long lists of rules,
a machine-learning
lems. In problems wherealgorithm can often
existing simplify
solutions the code
require andmanual
a lot of achieveadjustments
better performance.
or long
Sometimes they allow us to find solutions to problems that otherwise
lists of rules, a machine-learning algorithm can often simplify the code and achievewould not be solved
better
through traditional approaches. These algorithms are used to process large amounts of
performance. Sometimes they allow us to find solutions to problems that otherwise would
data in order to discover patterns that are not immediately apparent. They are also used in
not be solved through traditional approaches. These algorithms are used to process large
situations where the algorithm needs to dynamically adapt to new patterns in the data or
amounts of data in order to discover patterns that are not immediately apparent. They are
when the data itself is generated as a function of time, such as stock price prediction; in this
also used in situations where the algorithm needs to dynamically adapt to new patterns
case, we speak of online learning.
in the data or when the data itself is generated as a function of time, such as stock price
Machine-learning algorithms can be classified into supervised learning, unsupervised
prediction; in this case, we speak of online learning.
learning, semi-supervised learning, and reinforcement learning. This classification is made
Machine-learning algorithms can be classified into supervised learning, unsuper-
in relation to the quantity of data available during the training phase and the type of
vised learning, semi-supervised learning, and reinforcement learning. This classification
supervision during the training.
is made in relation to the quantity of data available during the training phase and the type
Specifically, in supervised learning, the training data provided to the algorithm in-
of supervision during the training.
clude desired solutions called labels. Supervised learning solves two types of problems:
classification and regression.
Sensors 2023, 23, x FOR PEER REVIEW 10 of 26

Specifically, in supervised learning, the training data provided to the algorithm in-
Sensors 2023, 23, 5226 9 of 25
clude desired solutions called labels. Supervised learning solves two types of problems:
classification and regression.
Classification is the problem of cataloging data into two or more classes; so, by
Classification
providing input to isthethe problem of cataloging
machine-learning system, itdata
mustinto twoits
return orclass
moreofclasses; so, by
belonging.
providing input to the machine-learning system, it must return its class of belonging.
On the other hand, regression interpolates data to associate two or more features with
eachOn the other
other. hand, regression
By providing interpolates
the algorithm with andata to associate
input feature,two
the or more features
regressor returnswith
the
each
other feature. A system of estimating the price of houses starting from features, suchthe
other. By providing the algorithm with an input feature, the regressor returns as
other feature. of
size, number A rooms,
systemandof estimating the price of
area, is a regression houses starting from features, such as
system.
size, number
The most of popular
rooms, and area, is a regression
supervised-learning system. are k-Nearest Neighbors, linear
algorithms
The most
regression, popular
logistic supervised-learning
regression, Support Vector algorithms
Machine are k-Nearest
(SVM), DecisionNeighbors, linear
Trees, Random
regression, logistic regression, Support Vector Machine (SVM), Decision Trees, Random
Forests, and Neural Networks.
Forests, and Neural Networks.
The NILM problem can be set up either as a regression problem—for example, when
The NILM problem can be set up either as a regression problem—for example, when
the algorithm is called to estimate the power absorbed by the single appliance starting
the algorithm is called to estimate the power absorbed by the single appliance starting from
from the aggregate power measurement [52]—or as a classification problem [53], as in the
the aggregate power measurement [52]—or as a classification problem [53], as in the case
case in which starting from the aggregate power measurement is necessary to determine
in which starting from the aggregate power measurement is necessary to determine which
which appliances are powered and which are not.
appliances are powered and which are not.
The system proposed in this manuscript solves a multi-label classification problem
The system proposed in this manuscript solves a multi-label classification problem
since, starting from an SFRA trace, it is possible to identify several powered appliances
since, starting from an SFRA trace, it is possible to identify several powered appliances
simultaneously. The algorithm used is the SVM; the system configuration and its opera-
simultaneously. The algorithm used is the SVM; the system configuration and its operation
tion are illustrated in the following paragraphs.
are illustrated in the following paragraphs.
3.1. Support
3.1. Support Vector
Vector Machine
Machine
A SVM
A SVM is is one
one ofof the
the most
most popular
popular models
models in in machine
machine learning,
learning, as
as itit is
is very
very powerful
powerful
and versatile [51]. SVMs are best suited for classifying complex
and versatile [51]. SVMs are best suited for classifying complex but small- to medium- but small- to medium-
sized datasets.
sized datasets. While
While classic
classic classification
classification algorithms
algorithms discriminate
discriminate based
based on on characteristics
characteristics
common to
common to each
each class,
class, the
the SVM
SVM algorithms
algorithms buildbuild the
the model
model based
based onon thethe mostmost difficult
difficult
samplesto
samples todiscriminate,
discriminate,i.e., i.e.,the
the most
most similar
similar samples
samples belonging
belonging to different
to different classes.
classes. In
In this
this sense,
sense, the only
the only samples
samples usedused inconstruction
in the the constructionof theof model
the model are called
are called supportsupport vec-
vectors.
tors.other
The The other
samples samples are therefore
are therefore useless.useless.
Based on the support
Based on the support vectors, vectors, the algorithm finds the optimal hyperplane that sepa-
rates them,
rates them, which
which can can then
then bebe used
used to to discriminate
discriminate new new samples.
samples. In In other
other words,
words, adding
adding
more formation
more formation samples
samples far far from
from thethe hyperplane
hyperplane (therefore
(therefore notnot particularly
particularly complex complex to to
classify) will
classify) willnot
notaffect
affectthethedecision
decision boundary,
boundary, which
which willwill be completely
be completely determined
determined by
by the
samples
the sampleslocated at the
located at edge of the
the edge of hyperplane.
the hyperplane.
Consider
Consideraa case case in
in which
which the samples to be classified are defined by only two features. features.
This
This case can be represented on a two-dimensional plane, as shown in Figure 6.6. A
case can be represented on a two-dimensional plane, as shown in Figure A
SVM
SVM algorithm
algorithm looks looks for
for the
the line
line capable
capable of of maximizing
maximizing the the margin
margin between
between the the most
most
similar
similar samples
samples belonging
belonging to to different
differentclasses,
classes,i.e.,
i.e.,the
thesupport
supportvectors.
vectors.

Figure 6. Representation of a linear classification problem in which the samples are defined by only
two features.
Sensors 2023, 23, 5226 10 of 25

Consider a linear classification problem in which n-dimensional inputs X are divided

into two classes y ∈ {−1, 1}. The classifier can be formulated as follows:

f (x) = wT φ(x) + b, (1)

where w is the vector of weights, b is the bias, and φ(x) is the feature space of the inputs.
The sign of f(x) will be the output yi of the classification.
Since the inputs are linearly separable, it will be possible to choose several linear
decision boundaries, each of which will not produce classification errors in the training data.
Training a SVM model positions the boundary to maximize the margin—that is, the
distance from the hyperplane to the nearest data point in either class. More specifically, we
want to optimize the following objective function:
 
maxmin dist(xi , w, b) ∀i yi wT φ(xi ) + b ≥ 0, (2)
w,b i

where dist(x, w, b) is the Euclidean distance from the feature point φ(x) to the hyperplane
defined by w and b. With this objective function, the distance from the decision boundary
wT φ(x) + b = 0 to the nearest point i is maximized. The constraints force finding a decision
boundary that correctly classifies all the training data. In other words, for the classifier,
a correct training point yi and wT φ(xi ) + b must have the same sign, in which case their
product must be positive.
It is known from Euclidean geometry that the distance between the point φ(xi ) and the
|w T φ(x ) +b
hyperplane wT φ(x) + b = 0 can be defined as i
||w||
. Since yi is the sign of f (xi ) , it can
be written as follows:

yi (wT φ(xi ) + b)
maxmin | ∀i yi (wT φ(xi ) + b) ≥ 0, (3)
w,b i ||w||

We can observe that, due to the normalization of ||w|| in (3), the scale of w is arbitrary
in this objective function. That is, if w and b are multiplied by a real scalar α, the factors of
α in the numerator and denominator will cancel each other out. Now, suppose we choose
the scale so that the point closest to the hyperplane, xi satisfies yi (wT φ(xi ) + b) = 1. With
this assumption, the min in Equation (3) becomes redundant and can be removed. The
i
objective function and constraint can be rewritten as:

1  
max | ∀i yi wT φ(xi ) + b ≥ 0, (4)
w,b ||w||

Finally, we convert the problem into a quadratic program (QP). In this way, the objec-
tive function is quadratic in the unknowns and all constraints are linear in the unknowns. A
QP has a single global minimum, which can be found efficiently with current optimization
packages [54].
1  
max ||w||2 | ∀i yi wT φ(xi ) + b ≥ 0, (5)
w,b 2

However, not all classification problems are linear; in fact, in some cases, it is not
possible to separate the classes with a straight line; therefore, we speak of non-linear classifi-
cation. The kernel trick [55] solves non-linear classification problems with SVM algorithms.
In more detail, a polynomial kernel was used to determine the presence, or absence, of
an appliance starting from the SFRA traces. Using a polynomial kernel means determining
similarity, not only by processing the features of the input samples but also by their
combinations, as shown in Figure 7.
Moreover, in real scenarios, data belonging to different classes overlap. As a result, it
will not be possible to satisfy all the constraints in (5). One way to deal with this problem
and still train useful classifiers is to relax some constraints by introducing so-called slack
Sensors 2023, 23, 5226 11 of 25

variables [56]. Normally, a Lagrangian transformation addresses the optimization problem,

Sensors 2023, 23, x FOR PEER REVIEW
which 12 of 26
allows the constrained optimization problem expressed in (5) to be reformulated
into a non-constrained optimization problem.

7. Representation
Figure 7. Representationof
ofaanon-linear
non-linearclassification
classificationproblem
probleminin which
which thethe examples
examples areare defined
defined by
by only
only twotwo features.
features.

The Lagrangian
Moreover, forscenarios,
in real the SVM objective functionto
data belonging indifferent
(5), with Lagrange multipliers
classes overlap. i ≥ 0, is:
As aaresult, it
will not be possible to satisfy all the constraints 1 in (5). One way to deal with this problem
L(a1:N )is=to∑
2∑ ∑
and still train useful classifiers ai −some
relax ai a j yi y j k (xby
constraints i ,xj )introducing
, (6)
so-called slack
i j
variables [56]. Normally, a Lagrangian transformation addresses the optimization prob-
where k (xi ,xj )allows
lem, which is called
theaconstrained
kernel function. For example,
optimization problem if weexpressed
used the in basic
(5) linear features,
to be reformu-
lated into a non-constrained T
optimization problem.
i.e., φ(x) = x, then k (xi ,xj ) = xi xj . Instead, because a polynomial kernel has been chosen in
The Lagrangian
the implemented SVMfor the SVM
classifier, objective
it will function
be defined as: in (5), with Lagrange multipliers
𝑎 ≥ 0, is:  b
k (xi ,xj ) = a + xi T xj , (7)
𝐿( : ) = ∑ 𝑎 − ∑ ∑ 𝑎 𝑎 𝑦 𝑦 𝑘(𝒙𝒊,𝒙𝒋 ) , (6)
3.2.
where The𝑘Proposed Structure
(𝒙𝒊 ,𝒙𝒋 ) is called a kernel function. For example, if we used the basic linear features,
𝑻 input is the trace obtained from the SFRA system; thus,
i.e., 𝜙(𝒙) = 𝒙, then 𝑘(𝒙system,
In the proposed 𝒊 ,𝒙𝒋 )
= 𝒙𝒊the 𝒙𝒋 . Instead, because a polynomial kernel has been chosen
each
in thepoint of the trace
implemented represents
SVM a feature
classifier, it will of be the SVM.as:
defined The algorithm must have a number
of input functions equal to the number of bins of the measured frequency response.
The problem is also attributable𝑘to (𝒙𝒊 ,𝒙 = (a + 𝒙𝒊 𝑻classification
a𝒋 )multi-label 𝒙𝒋 ) , problem, where a single (7)
sample can belong to multiple defined classes, unlike in multi-class classification, where
each sample
3.2. The Proposed can uniquely
Structure belong to only one class.
In fact, the purpose of the system is to determine the status (ON or OFF) of the
In the proposed
appliances. This means system,
that the the input is
number ofthe traceisobtained
classes fromofthe
equal to that theSFRA system;
appliances andthus,
the
each point of the trace represents a feature of the SVM. The algorithm
belonging of an SFRA trace to a certain class will indicate the ON state of that appliance. must have a number
of single
A input functions
SFRA trace equal
must to the numberbeofable
therefore binstoof be
theassociated
measured withfrequency response.
multiple classes (or
The problem is also attributable to a multi-label classification
labels), as the system must be able to recognize the loads even under simultaneous problem, where apower
single
sample can belong to multiple defined classes, unlike in multi-class
supply conditions. SVMs are not natively capable of performing multi-class or multi-label classification, where
each sample can
classifications uniquely
since, belongabove,
as explained to onlya oneSVMclass.
defines a hyperplane that separates classes
In fact, the purpose of the
equidistantly in order to guarantee the maximum system is to determine the status
margin. When(ONthe ornumber
OFF) of oftheclasses
appli-
ances.toThis
rises threemeans or more, that
thusthepassing
numberfrom of classes
a binary is classification
equal to thattoofmulti-class,
the appliances and the
it is possible
belonging of an SFRA trace to a certain class will indicate
to guarantee equidistance only between two of the classes, discarding this property with the ON state of that appliance.
A single
all the otherSFRA trace must therefore be able to be associated with multiple classes (or la-
classes.
bels),Toassolve
the system must be able
this classification to recognize
problem, which the loads even
involves under multiple
assigning simultaneous
labelspower
to an
supply conditions.
instance, we converted SVMs it toaremultiple
not nativelybinary capable of performing
classification multi-class
problems. A SVM was or multi-label
therefore
classifications
associated with since,
eachashousehold
explainedappliance,
above, a SVM defines aahyperplane
performing that separates
binary classification classes
in order to
equidistantly
determine its ONin order
or OFF to status,
guarantee the maximum
starting from the SFRAmargin. When
trace. Thethe number
proposed of classes
structure is
rises to in
shown three
Figure or more,
8. thus passing from a binary classification to multi-class, it is possible
to guarantee equidistance only between two of the classes, discarding this property with
all the other classes.
To solve this classification problem, which involves assigning multiple labels to an
instance, we converted it to multiple binary classification problems. A SVM was therefore
associated with each household appliance, performing a binary classification in order to
Sensors 2023, 23, x FOR PEER REVIEW 13 of 26

Sensors 2023, 23, 5226 12 of 25

determine its ON or OFF status, starting from the SFRA trace. The proposed structure is
shown in Figure 8.

Figure
Figure 8.
8. The
The proposed
proposed structure.

4. Experimental
4. Experimental Results
Results
As part of the development
As part of the phase, the
development phase, the proposed
proposed algorithm was implemented
algorithm was implemented and
and
tested to
tested to evaluate
evaluate its
its performance
performance with
with real
real data.
data.

4.1. The Proposed System Setup

4.1. The Proposed System Setup
As explained in Section 2, the SFRA technique was performed by plugging the in-
As explained in Section 2, the SFRA technique was performed by plugging the in-
strument into a standard household socket. As previously discussed, the input signal
strument into a standard household socket. As previously discussed, the input signal is a
is a variable frequency sinusoidal signal applied between the phase conductor terminal
variable frequency sinusoidal signal applied between the phase conductor terminal and
and ground, while the output signal is the measured signal between the neutral conduc-
ground, while the output signal is the measured signal between the neutral conductor
tor terminal and ground. Both signals are acquired and processed. Figure 9 shows the
terminal and ground. Both signals are acquired and processed. Figure 9 shows the meas-
measurement system used.
urement system used.
The measurement system must be connected to the test system by means of cables
with suitable bandwidth and the same characteristic impedance of the generator to avoid
reflection and signal mismatch and to improve the sensitivity, repeatability, and reliability
of the measurement.
The input signal and related acquisition for the SFRA were performed using the
Digilent Analog Discovery 2 NI Edition card with a BNC adapter.
The control system was developed using LabVIEW and run on a PC; this software
automatically programs the Discovery FPGA at startup, with a configuration file designed
to implement the measurement application. Once programmed, the integrated FPGA
communicates with the PC via a USB 2.0 connection. The PC enables the creation of the
user interface to access the data and process them in the experimental phase. A final NILM
system can bypass the PC by integrating post-processing directly into the system.
The Discovery FPGA has a ±25 V input range, a 14-bit resolution, a 100 MS/s sampling
frequency, and a 30 MHz bandwidth. It is equipped with an arbitrary function generator
with an output range of ±5 V, a bandwidth of 20 MHz, and a sampling rate of 100 MS/s.
For appropriate interfacing with the network, the instrument is equipped with a
coupling circuit for each of the three channels (one for generation and two for acquisition),
as shown in Figure 9. The coupling circuit includes a third-order Butterworth filter with a
flat passband and high attenuation outside the desired frequency range. The generation
section and acquisition section coupling circuits both involve a 50 Ω resistor in series and
parallel, respectively, to allow impedance adaptation. In addition, all coupling circuits are
Sensors 2023, 23, 5226 13 of 25

provided with a high-voltage ac blocking capacitor, connected in series with a 1:1 pulse
Sensors 2023, 23, x FOR PEER REVIEW 14 of 26
transformer. The features of the filters developed for the SFRA apparatus are shown in
Figures 10 and 11.

Sensors 2023, 23, x FOR PEER REVIEW 15 of 26

Sensors 2023, 23, x FOR PEER REVIEW 15 of 26

transformer.
transformer. The
The features
features of
of the
the filters
filters developed
developed for
for the
the SFRA
SFRA apparatus
apparatus are
are shown
shown in
in
Figures 10 and 11.
Figures 10 and 11.
Figure 9.
Figure 9. The
The SFRA
SFRA measurement
measurement system.
system.

The measurement system must be connected to the test system by means of cables
with suitable bandwidth and the same characteristic impedance of the generator to avoid
reflection and signal mismatch and to improve the sensitivity, repeatability, and reliability
of the measurement.
The input signal and related acquisition for the SFRA were performed using the Digi-
lent Analog Discovery 2 NI Edition card with a BNC adapter.
Figure 10.
FigureThe10. Coupling
control circuit
circuit for
Couplingsystem was
for the signal
signal generation
thedeveloped using section.
generation
generation LabVIEW
section.
section. and run on a PC; this software
automatically programs the Discovery FPGA at startup, with a configuration file designed
to implement the measurement application. Once programmed, the integrated FPGA
communicates with the PC via a USB 2.0 connection. The PC enables the creation of the
user interface to access the data and process them in the experimental phase. A final NILM
system can bypass the PC by integrating post-processing directly into the system.
The Discovery FPGA has a ±25 V input range, a 14-bit resolution, a 100 MS/s sampling
frequency, and a 30 MHz bandwidth. It is equipped with an arbitrary function generator
with an
Figure
Figure 11.
Figure 11.output
11. Coupling
range
Coupling circuit
Coupling of ±5
circuit for
circuit V,
for the
for asignal
bandwidth
the signal
the of 20section.
signal acquisition
MHz, and a sampling rate of 100 MS/s.
acquisition section.
For appropriate interfacing with the network, the instrument is equipped with a cou-
Incircuit
plingIn ordertotoavoid
avoid the
unwanted over-voltages due to resonance phenomena at high
In order
orderfor avoidofunwanted
to each three over-voltages
unwanted channels (one due
over-voltages for to
to resonance
duegeneration andphenomena
resonance two at
at high
high fre-
for acquisition),
phenomena as
fre-
frequencies,
quencies,
shown in thethe
Figure amplitude
amplitude
9. The of of
the
couplingthe applied
applied
circuit signal
signal must
must
includes a not
not exceed
exceed
third-order a a few
few volts
volts
Butterworth (5
(5 Vpp
Vpp
filter in the
in
withthea
quencies, the amplitude of the applied signal must not exceed a few volts (5 Vpp in the
present case).
case). The
flat passband
present and accuracy
The of
of the
high attenuation
accuracy the adopted
outsidemeasurement
adopted system,
system, as
the desired frequency
measurement discussed
asrange. The in
discussed in aa previ-
generation
previ-
ous paper
section and [57], has been
acquisition evaluated
section using
coupling a reference
circuits both parallel
involve LCR
a 50 circuit.
ous paper [57], has been evaluated using a reference parallel LCR circuit. This circuitΩ This
resistor incircuit
seriescon-
and
con-
sists of
parallel, a 50 Ω resistive
respectively, adapter,
to allow a fixed
impedance inductance,
adaptation.and Ina variable
addition, capacitance.
all coupling
sists of a 50 Ω resistive adapter, a fixed inductance, and a variable capacitance. The refer- The refer-
circuits are
enced
provided
enced values of
with
values the
ofathe circuit
high-voltageimpedance were
ac blocking
circuit impedance measured
werecapacitor, with
measuredconnected a Keysight E4980AL
in series
with a Keysight precision
with a precision
E4980AL 1:1 pulse
Sensors 2023, 23, 5226 14 of 25

present case). The accuracy of the adopted measurement system, as discussed in a previous
paper [57], has been evaluated using a reference parallel LCR circuit. This circuit consists
of a 50 Ω resistive adapter, a fixed inductance, and a variable capacitance. The referenced
values of the circuit impedance were measured with a Keysight E4980AL precision LCR
meter. The estimated accuracy of the Vout/Vin ratio was better than ±0.2 dB in the interval
from +5 to −25 dB and in the frequency range of 5 kHz to 1.5 MHz.
The SVM was implemented on a desktop computer (based on the Windows 10 × 64-bit
operating system) using the open-source Python 3.7 from Anaconda [58]; the machine-
learning algorithm was developed using the Scikit-learn library. Python is the programming
language mostly used in artificial intelligence (AI) applications due to the availability of
numerous libraries for continuous data acquisition and processing.

4.2. The Achieved Results

The proposed measurement technique is innovative and does not appear to have
been tested by other authors. Due to the specificity of the acquired data (frequency re-
sponse), there are no public datasets used by other authors against which to compare the
performance of the proposed algorithm [59].
The measurement system was installed on a test facility, which was designed to gener-
ate electrical loads created by domestic users as part of the “non-intrusive infrastructure for
monitoring loads in residential users” research project. The facility, located in the Electrical
Engineering Laboratory of the University of L’Aquila (I), allows for the generation of
electrical loads in a single or simultaneous way.
During the test phase, various parameters were evaluated in order to define the most
significant sub-bands, the number of measurement points to be acquired, and the number
of training examples needed to obtain a satisfactory performance. To this end, the precision,
recall, and F1-Score during classification were evaluated [60]. These parameters were
obtained using the numbers of true positive (TP), false positive (FP), true negative (TN),
and false negative (FN) as follows:

TP
Precision = , (8)
TP + FP

TP
Recall = , (9)
TP + FN

precision × recall
F1 − score = 2 × , (10)
precision + recall
The concept of positive has been attributed to the ON state of household appliances
and that of negative to the OFF state. Precision indicates all of the times the system has
provided an indication of the ON state of an appliance and how many times the prediction
has been correct. Precision does not take FNs into account. On the other hand, Recall
indicates how many times the system has provided a correct indication about the ON state
of the appliance compared to all of the samples in which the appliance was actually in the
ON state. Recall does not take FPs into account. To have a metric capable of taking into
account both FPs and FNs, the F1-Score is used, which is a harmonic mean of Precision
and Recall.
Since, as already explained above, each appliance is associated with a SVM algorithm
that reveals its presence, or not, the performance of each SVM was evaluated individually.
We started by acquiring 20 samples for each of the 24 scenarios, for a total of 480
training samples. Each sample consisted of an SFRA trace in which 200 points were
acquired for each of the 3 sub-bands. Performance was evaluated on a test set consisting of
50 samples for each scenario, for a total of 1200 test samples. The obtained results, shown in
Table 2, are already excellent, as 480 training samples is a relatively low number considering
that acquiring a single sample takes about 40 s. The system does not make mistakes for five
Sensors 2023, 23, 5226 15 of 25

of the eight appliances analyzed and also shows high performance regarding the other three
appliances. To define which of the three sub-bands made the most significant contribution
to the identification of household appliances, the system’s performance was evaluated by
providing the three sub-bands separately as input to the machine-learning system. The
results are reported in Table 3 and a graphical comparison is provided in Figure 12.

Table 2. The results obtained with 480 training samples and 200 points for each sub-band.

Total Errors FP FN Precision Recall F1-Score

Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 27 0 27 1.00 0.92 0.96
Laptop 0 0 0 1.00 1.00 1.00
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 5 5 0 0.98 1.00 0.99
Drill 29 29 0 0.93 1.00 0.97
TV 0 0 0 1.00 1.00 1.00

Table 3. Performance evaluation for each sub-band.

10–100 kHz
Total Errors FP FN Precision Recall F1-Score
Hairdryer 98 98 0 0.75 1.00 0.86
Microwave Oven 50 0 50 1.00 0.83 0.91
Lamp 110 96 14 0.78 0.96 0.86
Laptop 51 51 0 0.89 1.00 0.94
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 61 61 0 0.83 1.00 0.91
Drill 48 48 0 0.89 1.00 0.94
TV 0 0 0 1.00 1.00 1.00
100 kHz–1 MHz
Total Errors FP FN Precision Recall F1-Score
Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 141 30 111 0.89 0.68 0.77
Laptop 9 0 9 1.00 0.98 0.99
Induction Hob 59 9 50 0.97 0.83 0.89
Heater 5 5 0 0.98 1.00 0.99
Drill 116 106 10 0.79 0.98 0.87
TV 0 0 0 1.00 1.00 1.00
1–1.5 MHz
Total Errors FP FN Precision Recall F1-Score
Hairdryer 2 2 0 0.99 1.00 0.99
Microwave Oven 93 85 8 0.77 0.97 0.86
Lamp 115 0 115 1.00 0.67 0.80
Laptop 71 6 65 0.98 0.84 0.90
Induction Hob 79 74 5 0.80 0.98 0.88
Heater 29 29 0 0.91 1.00 0.95
Drill 90 76 14 0.84 0.97 0.90
TV 48 39 9 0.91 0.98 0.94
 Lamp 115 0 115 1.00 0.67 0.80
Laptop 71 6 65 0.98 0.84 0.90
Induction Hob 79 74 5 0.80 0.98 0.88
Heater 29 29 0 0.91 1.00 0.95
Sensors 2023, 23, 5226 Drill 90 76 14 0.84 0.97 16 of 25
0.90
TV 48 39 9 0.91 0.98 0.94

Figure 12. F1-Scores obtained for each considered sub-band.

In light
In light ofof these
these results,
results, it
it was
was decided
decided that
that we
we would
would consider
consider only
only the
the sub-bands
sub-bands ofof
10–100 kHz and 100 kHz–1 MHz in order to reduce the time required for
10–100 kHz and 100 kHz-1 MHz in order to reduce the time required for the measurement. the measurement.
In fact,
In fact, it
it is
is evident from Figure
evident from Figure 12 12 that
that the
the 1–1.5
1–1.5 MHz
MHz band
band never
never allows
allows forfor appliance
appliance
discrimination that
discrimination that outperforms
outperformsthe theprevious
previousbands.
bands.This
This reduces
reduces thethe
timetime it takes
it takes to
to ac-
acquire a single trace to 22.56 s. Table 4 reports the performance evaluation
quire a single trace to 22.56 s. Table 4 reports the performance evaluation using only the using only the
first two
first two sub-bands
sub-bands as as input.
input.

4. The
Table 4.
Table The results
resultsobtained
obtainedwith
with480 training
480 samples
training andand
samples 200200
points for each
points sub-band,
for each usingusing
sub-band, only
the first two sub-bands.
only the first two sub-bands.

Total
Total Errors
Errors FPFP FN FN Precision
Precision Recall
Recall F1-Score
F1-Score
Hairdryer
Hairdryer 00 00 0 0 1.001.00 1.00
1.00 1.00
1.00
Microwave
Microwave Oven
Oven 00 00 0 0 1.001.00 1.00
1.00 1.00
1.00
Lamp
Lamp 2929 00 29 29 1.001.00 0.92
0.92 0.96
0.96
Laptop 4 4 0 0.99 1.00 0.99
Laptop
Induction Hob 0
4 0
4 0
0 1.00
0.99 1.00
1.00
0.99
1.00
Induction
Heater Hob 30 30 0 0 0.991.00 1.00
1.00 1.00
0.99
Drill 7 7 0 0.98 1.00 0.99
TV 0 0 0 1.00 1.00 1.00

Comparing the results with those of Table 2, it can be seen that the system’s perfor-
mance has remained roughly unchanged. However, there is a significant improvement
in the detection of the drill, highlighting that the 1–1.5 MHz sub-band introduced use-
less randomness for identification purposes. In this way, 400 points are acquired in the
10 kHz–1 MHz frequency band.
The possibility of decreasing the number of acquired points has been evaluated.
Therefore, in Table 5, the performances obtained for 200, 134, and 100 points are reported.
Furthermore, Figure 13 shows a graphical comparison of the impact of the number of
acquired points on the F1-Score.
Sensors 2023, 23, 5226 17 of 25

Table 5. Performance evaluation as the points acquired decrease.

10 kHz–1 MHz (200 Points)

Total Errors FP FN Precision Recall F1-Score
Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 39 0 39 1.00 0.89 0.94
Laptop 0 0 0 1.00 1.00 1.00
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 2 2 0 0.99 1.00 1.00
Drill 5 5 0 0.99 1.00 0.99
TV 0 0 0 1.00 1.00 1.00
10 kHz–1 MHz (134 Points)
Total Errors FP FN Precision Recall F1-Score
Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 26 0 26 1.00 0.93 0.96
Laptop 0 0 0 1.00 1.00 1.00
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 5 5 0 0.98 1.00 0.99
Drill 5 5 0 0.99 1.00 0.99
TV 0 0 0 1.00 1.00 1.00
10 kHz–1 MHz (100 Points)
Total Errors FP FN Precision Recall F1-Score
Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 42 0 42 1.00 0.88 0.94
Laptop 0 0 0 1.00 1.00 1.00
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 6 6 0 0.98 1.00 0.99
Sensors 2023, 23, x FOR PEER REVIEW Drill 2 2 0 0.99 1.00 19 of 26
0.99
TV 0 0 0 1.00 1.00 1.00

Figure 13.
Figure 13. Graphical
Graphical comparison
comparison of
of the
the impact
impact of
of the
the number
number of
of acquired
acquired points
points on
on the
the F1-Score.
F1-Score.

The performance
The performance proved
proved to to be
be very
very good,
good, even
even when
when only
only using
using 100
100 measurement
measurement
points as a system input.
points as system input. In these conditions, in fact, the system made errors only for three
of the eight
eight appliances
appliancesanalyzed
analyzedwhilewhilemaintaining
maintaininga aminimum
minimum F1-Score
F1-Scoreof of
0.94. This
0.94. re-
This
duction allowed
reduction alloweda adecrease
decreaseininthe
theexecution
executiontime
timeofofthe
themeasurement
measurement system
system from 22.56 s
to 6.09 s. The performances shown so far always foresaw 480 training samples (20 for each
of the 24 scenarios). As a final analysis, the impact of the number of training samples on
performance was evaluated as shown in Figure 14. Table 6 reports the results obtained
using an SFRA trace consisting of 100 points acquired in the 10 kHz–1 MHz frequency
band, reducing the number of samples used in the training phase.
Sensors 2023, 23, 5226 18 of 25

to 6.09 s. The performances shown so far always foresaw 480 training samples (20 for each
of the 24 scenarios). As a final analysis, the impact of the number of training samples on
performance was evaluated as shown in Figure 14. Table 6 reports the results obtained
using an SFRA trace consisting of 100 points acquired in the 10 kHz–1 MHz frequency
band, reducing the number of samples used in the training phase.

Table 6. Performance evaluation as training samples decrease.

10 kHz–1 MHz (100 Points. 15 Samples for Each Scenario)

Total Errors FP FN Precision Recall F1-Score
Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 42 0 42 1.00 0.88 0.94
Laptop 0 0 0 1.00 1.00 1.00
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 6 6 0 0.98 1.00 0.99
Drill 3 3 0 0.99 1.00 0.99
TV 0 0 0 1.00 1.00 1.00
10 kHz–1 MHz (100 Points. 10 Samples for Each Scenario)
Total Errors FP FN Precision Recall F1-Score
Hairdryer 0 0 0 1.00 1.00 1.00
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 59 0 59 1.00 0.83 0.91
Laptop 2 0 2 1.00 0.99 0.99
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 5 5 0 0.98 1.00 0.99
Drill 23 21 2 0.95 1.00 0.97
TV 0 0 0 1.00 1.00 1.00
10 kHz–1 MHz (100 Points. 5 Samples for Each Scenario)
Total Errors FP FN Precision Recall F1-Score
Hairdryer 5 5 0 0.98 1.00 0.99
Microwave Oven 0 0 0 1.00 1.00 1.00
Lamp 71 0 71 1.00 0.80 0.89
Laptop 26 0 26 1.00 0.94 0.97
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 16 16 0 0.95 1.00 0.97
Drill 31 29 2 0.93 0.99 0.96
TV 0 0 0 1.00 1.00 1.00
10 kHz–1 MHz (100 Points. 1 Sample for Each Scenario)
Total Errors FP FN Precision Recall F1-Score
Hairdryer 5 5 0 0.98 1.00 0.99
Microwave Oven 1 1 0 0.99 1.00 0.99
Lamp 125 0 125 1.00 0.64 0.78
Laptop 95 2 93 0.99 0.77 0.87
Induction Hob 0 0 0 1.00 1.00 1.00
Heater 17 17 0 0.95 1.00 0.97
Drill 53 53 0 0.88 1.00 0.94
TV 8 8 0 0.98 1.00 0.94
 of the eight appliances analyzed while maintaining a minimum F1-Score of 0.94. This re-
duction allowed a decrease in the execution time of the measurement system from 22.56 s
to 6.09 s. The performances shown so far always foresaw 480 training samples (20 for each
of the 24 scenarios). As a final analysis, the impact of the number of training samples on
performance was evaluated as shown in Figure 14. Table 6 reports the results obtained
Sensors 2023, 23, 5226 19 of 25
using an SFRA trace consisting of 100 points acquired in the 10 kHz–1 MHz frequency
band, reducing the number of samples used in the training phase.

Figure
Figure 14.
14. Graphical
Graphical comparison
comparison of
of the
the impact
impact of
of the number of
the number of training
training samples
samples on
on the
the F1-Score.
F1-Score.

The system maintains interesting performances even when trained with only one
training sample for each scenario (therefore with 24 total training samples). This is mainly
because the SVM natively suffers more from the quality of the training samples rather than
the quantity, which is precisely because it builds a model based only on the most difficult
samples to discriminate.
Lower performance was found in the detection of the Lamp, Laptop, and Drill. In the
case of the Lamp, this is due to the insignificance of its related load compared to the overall
network, while in the case of the Laptop and Drill, it is due to the extreme variability of their
working conditions. However, F1-Score values of 0.78, 0.87, and 0.94, respectively, can be
considered largely satisfactory for a trained system with such a small number of samples.
In order to provide an overall assessment of the system’s performance, metrics widely
used for multi-label classification systems were used, including micro-average and macro-
average. As reported in (11)–(13), in the micro-average, all TPs, TNs, FPs, and FNs are
summed for all of the labels and subsequently averaged:

∑nN=1 TPn
Precisionmicro− averaging = , (11)
∑nN=1 TPn + FPn

∑nN=1 TPn
Recall micro−averaging = , (12)
∑nN=1 TPn + FN n
2 × Precisionmicro− averaging × Recall micro−averaging
F1 − scoremicro− averaging = , (13)
Precisionmicro− averaging + Recall micro− averaging
On the other hand, the macro-average, as reported in (14)–(16), is simply the average
of the Precision and Recall for each label:

∑nN=1 Precisionn
Precisionmacro− averaging = , (14)
N

∑nN=1 Recall n
Recall macro− averaging = , (15)
N
Sensors 2023, 23, 5226 20 of 25

2 × Precisionmacro− averaging × Recall macro− averaging

F1 − scoremacro−averaging = , (16)
Precisionmacro− averaging + Recall macro− averaging
The difference between the two lies is the fact that the micro-average reflects any
imbalances in the dataset. Unbalance means there are test samples in a greater number
of one or more classes than the others. In other words, having more samples for a given
scenario, the macro-average, by creating a simple average of Precision, Recall, and F1-Score,
does not consider this imbalance. On the contrary, the micro-average takes these situations
into account.
In the case in question, the dataset is balanced; therefore, both averages are functional
and adequate for verifying the performance of this system. Table 7 reports the micro-
averages and macro-averages calculated based on the values reported in Table 6.

Table 7. Impact of the size of the training set on multi-label classification.

Micro-Average Macro-Average
Training
Samples for Precision Recall F1-Score Precision Recall F1-Score
Each Scenario
20 0.99 0.98 0.99 0.99 0.98 0.99
15 0.99 0.98 0.99 0.99 0.98 0.99
10 0.99 0.97 0.98 0.99 0.97 0.98
5 0.98 0.96 0.97 0.98 0.96 0.97
1 0.96 0.92 0.94 0.97 0.92 0.94

An additional consideration needs to be made to integrate the proposed system into

an electrical system. As explained above, there is no interference with the normal operation
of the devices during system operation. Furthermore, the system poses no problems to the
EMI filters, which are the input stage of the monitored devices, as the powers involved—
which can be associated with the test signal—are extremely low.
To analyze the operating conditions of the measurement system in detail, it was
simulated in a SPICE environment.
Specifically, the simulation was oriented to analyze the effects produced by the test
signal on commercial EMI filters that could be connected (to other devices) in proximity
to the system being tested. The analysis was extended to the entire range of frequencies
involved; as a reference, a commercial EMI filter family was considered [61] for standard
use in commercial and residential apparatuses for AC currents up to 16 Arms in single-
phase systems.
The analysis was extended to the entire range of frequencies involved. Figure 13 summa-
rizes the scheme considered for the simulation. The resistance RLoad equal to 50 Ω was chosen
in order to simulate the load of a generic household appliance (230 Vrms /50 Ω = 4.6 Arms ).
The system’s response was evaluated by varying the frequency in the range in which
the proposed system operates in the final configuration (10 kHz–1 MHz). The frequency
response of the current entering the EMI filter was evaluated. Several simulations were
carried out by varying the RLC parameters of the EMI filter. The current was found to be
harmless across the entire spectrum. As an example, Figure 15 shows the input current
response obtained with the RLC parameters reported in Figure 16. The spectrum shows
two resonance peaks and a maximum current draw of 4.64 mA.
 Arms).The system’s response was evaluated by varying the frequency in the range in which
The system’s
the proposed response
system was in
operates evaluated
the finalbyconfiguration
varying the frequency
(10 kHz–1inMHz).
the range
Theinfrequency
which
the proposed system operates in the final configuration (10 kHz–1 MHz). The frequency
response of the current entering the EMI filter was evaluated. Several simulations were
response of the current entering the EMI filter was evaluated. Several simulations were
carried out by varying the RLC parameters of the EMI filter. The current was found to be
carried out by varying the RLC parameters of the EMI filter. The current was found to be
harmless across the entire spectrum. As an example, Figure 15 shows the input current
Sensors 2023, 23, 5226 harmless across the entire spectrum. As an example, Figure 15 shows the input current 21 of 25
response obtained with the RLC parameters reported in Figure 16. The spectrum shows
response obtained with the RLC parameters reported in Figure 16. The spectrum shows
two resonance peaks and a maximum current draw of 4.64 mA.
two resonance peaks and a maximum current draw of 4.64 mA.

Figure
Figure 15.
15.The
Figure15. Thefrequency
The frequencyresponse
frequency responseof
response ofthe
of theinput
the inputcurrent
input currenttoto
current the
to EMI
the
the filter.
EMI
EMI filter.
filter.

Figure 16. The scheme used for SPICE simulation.

Figure16.
Figure 16. The
The scheme
scheme used
used for
for SPICE
SPICE simulation.
simulation.

The reduced value of this peak current does not lead to overheating of the filter com-
The
The reduced
reduced value
value of
of this
this peak
peak current
current does
does not
not lead to overheating of the filter com-
ponents since the associated dissipated power is reduced. Furthermore, such verification
ponents
ponentssince
sincethe
theassociated
associateddissipated
dissipatedpower
powerisisreduced.
reduced. Furthermore,
Furthermore, such
suchverification is
verification
is pejorative for the following reasons:
pejorative
is pejorativeforfor
thethe
following
followingreasons:
reasons:
(1) The proposed system adopts a Digilent Analog Discovery 2 board, which has a limi-
(1) The proposed
(1) proposed system
systemadopts
adoptsa aDigilent Analog
Digilent Analog Discovery
Discovery2 board, which
2 board, has ahas
which limi-
a
tation on the maximum output current that can be supplied by the DAC channels at
tation on
limitation
4 mA.
the
on maximum
the maximum output current
output that
current can
that be
can supplied
be by
supplied the
by DAC
the DAC channels
channelsat
(2) In
at 4our
4 mA.mA. simulation, the measurement system is only connected to the device being
(2) In our In
(2) tested. simulation, thethe
the real case, measurement system is only
generator is connected only
to connected
connected
a generic socketto the
to of
the device
thedevice being
being
electrical
tested. In the real case, the generator is connected to to aa generic
generic socket
socket ofof the
the electrical
electrical
system; therefore, the current that can be supplied (4 mA) is distributed in the various
parallel branches of the other connected devices, greatly reducing the intensity of the
portion that could affect the EMI filters.

5. Conclusions and Final Remarks

Modern home, energy, and assisted environment management systems require non-
intrusive monitoring of the power supply status of the various loads, regardless of infor-
mation related to their consumption. This parameter is not easy to obtain from NILM
systems. The SFRA technique, already widely used in the diagnostics of transformers and
asynchronous motors, has been applied here to characterize household appliances from
the point of view of their influence in modifying the frequency response of the electrical
system. The obtained signature, influenced by the physical characteristics of the loads, has
been used as input for a machine-learning algorithm, the SVM. The proposed algorithm
has been implemented in Python’s open-source development environment, thus reducing
the cost of the system.
Sensors 2023, 23, 5226 22 of 25

A large campaign of measurements was carried out on a test facility, during which
eight different electrical loads were powered individually and simultaneously. In particular,
variable consumption loads, such as a drill and a laptop, were considered, which are
generally among the most difficult for NILM systems to discriminate. The proposed
system demonstrated excellent performance, even when trained with a minimum number
of samples. In order to provide a comparison against other pre-published literature in
the field, works that used similar metrics [62–64] were considered. The performances
achieved by the cited works, by evaluating the F1-Score, were 91.5%, 93.2%, and 98.0%,
respectively. The proposed system outperforms all three systems, as when all training
data were provided (20 training samples for each scenario), the F1-Score achieved was
99.0%. It is important to note that the systems proposed in previous studies [62,63] were
outperformed, even when the system was trained with the minimum number of samples
when the system performance was 94.0%. The system is designed for local operation
and is thus oriented toward edge implementation. The final system can be conveniently
installed at any household outlet by detecting the presence of appliances connected to
the system autonomously and providing data externally, for example, through wireless
communication or the ability to download data histories via an SD card. The latter part will
therefore be the subject of future research developments. Furthermore, the proposed system
allows us to obtain information on which loads are powered in extremely short times (6.09 s
in the final configuration of the system). These times were evaluated by considering both
the time required to perform the measurement through the SFRA instrument and the time
required to perform the prediction via the SVM classifier. Therefore, to ensure real-time
operation, the edge system must incorporate multitasking capabilities. Two main tasks can
be identified: in the first task, the system acquires and process the data to obtain the SFRA
signature; in the second task, the system executes the SVM classifier and become ready
to transfer the data over the WiFi network. The task of acquiring data and obtaining the
signature, or SFRA trace, takes approximately 6 s, while the time required for processing
the signature using the SVM classifier and transferring the data over WiFi (e.g., via an
ESP32 module) is negligible and estimated to be around 10 ms based on experimental
evaluations. This second task can be performed during the acquisition time of the first task.
In fact, considering the first two signature-defining frequencies in the final configuration
of the proposed system, namely 10,000 Hz and 11,350 Hz, the time needed for acquiring
these initial points of the signature, as described in Section 2, amounts to 12 ms. Thus,
under these conditions, the system can maintain real-time operation while meeting the
requirements for post-processing and data transmission.

Supplementary Materials: The following supporting information can be downloaded at: https://
www.mdpi.com/article/10.3390/s23115226/s1, Figure S1: The block diagram of the LabVIEW code.
Author Contributions: Conceptualization, S.M.; software, S.M.; validation, S.M.; investigation, S.M.;
writing—original draft preparation, S.M.; writing—review and editing, G.B.; visualization, A.F.;
supervision, F.C. and E.F.; project administration, G.B. All authors have read and agreed to the
published version of the manuscript.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Hill, J. The Smart Home: A Glossary Guide for the Perplexed. T3. Retrieved at 27 March 2017, 12 September 2015.
2. Hart, G.W. Non-intrusive appliance load monitoring. Proc. IEEE 1992, 80, 1870–1891. [CrossRef]
3. Bucci, G.; Ciancetta, F.; Fiorucci, E.; Mari, S. Load identification system for residential applications based on the NILM technique.
In Proceedings of the IEEE Instrumentation and Measurement Technology Conference I2MTC 2020, Dubrovnik, Croatia, 20–25
May 2020.
4. Dong, R.; Ratliff, L.J.; Ohlsson, H.; Sastry, S.S. Energy disaggregation via adaptive filtering. In Proceedings of the 2013 51st
Annual Allerton Conference on Communication, Control, and Computing (Allerton), Monticello, IL, USA, 2–4 October 2013;
pp. 173–180. [CrossRef]
Sensors 2023, 23, 5226 23 of 25

5. Egarter, D.; Bhuvana, V.P.; Elmenreich, W. PALDi: Online load disaggregation via particle filtering. IEEE Trans. Instrum. Meas.
2015, 64, 467–477. [CrossRef]
6. Barker, S.; Kalra, S.; Irwin, D.; Shenoy, P. Powerplay: Creating virtual power meters through online load tracking. In Proceedings
of the 1st ACM Conference on Embedded Systems for Energy-Efficient Buildings, Memphis, TN, USA, 4–6 November 2014;
pp. 60–69.
7. Tang, G.; Wu, K.; Lei, J.; Tang, J. A simple model-driven approach to energy disaggregation. In Proceedings of the 2014 IEEE
International Conference on Smart Grid Communications (SmartGridComm), Venice, Italy, 3–6 November 2014; pp. 566–571.
[CrossRef]
8. Jia, R.; Gao, Y.; Spanos, C.J. A fully unsupervised non-intrusive load monitoring framework. In Proceedings of the 2015 IEEE
International Conference on Smart Grid Communications (SmartGridComm), Miami, FL, USA, 2–5 November 2015; pp. 872–878.
[CrossRef]
9. Kolter, J.Z.; Jaakkola, T. Approximate inference in additive factorial hmms with application to energy disaggregation. In
Proceedings of the 15th International Conference on Artificial Intelligence and Statistics (AISTATS) 2012, La Palma, Canary
Islands, Spain, 21–23 April 2012; pp. 1472–1482.
10. Bucci, G.; Ciancetta, F.; Fiorucci, E.; Mari, S.; Fioravanti, A. Measurements for non-intrusive load monitoring through machine
learning approaches. Acta IMEKO 2021, 10, 90–96. [CrossRef]
11. Bucci, G.; Ciancetta, F.; Fiorucci, E.; Mari, S.; Fioravanti, A. Deep Learning applied to SFRA Results: A Preliminary Study.
In Proceedings of the 7th International Conference on Computing and Artificial Intelligence ICCAI 2021, Tianjin, China,
23–26 April 2021.
12. Figueiredo, M.; Ribeiro, B.; de Almeida, A. Electrical signal source separation via non-negative tensor factorization using on site
measurements in a smart home. IEEE Trans. Instrum. Meas. 2014, 63, 364–373. [CrossRef]
13. Ciancetta, F.; Bucci, G.; Fiorucci, E.; Mari, S.; Fioravanti, A. A New Convolutional Neural Network-Based System for NILM
Applications. IEEE Trans. Instrum. Meas. 2020, 70, 9246573. [CrossRef]
14. Bucci, G.; Ciancetta, F.; Fiorucci, E.; Mari, S.; Fioravanti, A. Multi-State Appliances Identification through a NILM System Based
on Convolutional Neural Network. In Proceedings of the IEEE Instrumentation and Measurement Technology Conference I2MTC
2021, Glasgow, UK, 17–21 May 2021.
15. Cannas, B.; Carcangiu, S.; Carta, D.; Fanni, A.; Muscas, C. Selection of Features Based on Electric Power Quantities for Non-
Intrusive Load Monitoring. Appl. Sci. 2021, 11, 533. [CrossRef]
16. Ruzzelli, A.G.; Nicolas, C.; Schoofs, A.; O’Hare, G.M.P. Real-time recognition and profiling of appliances through a single
electricity sensor. In Proceedings of the 2010 7th Annual IEEE Communications Society Conference on Sensor, Mesh and Ad Hoc
Communications and Networks (SECON), Boston, MA, USA, 21–25 June 2010; pp. 1–9.
17. Athanasiadis, C.L.; Papadopoulos, T.A.; Doukas, D.I. Real-time non-intrusive load monitoring: A light-weight and scalable
approach. Energy Build. 2021, 253, 111523. [CrossRef]
18. Yang, C.; Soh, C.; Yap, V. A systematic approach in appliance disaggregation using k-nearest neighbours and naive Bayes
classifiers for energy efficiency. Energy Effic. 2018, 11, 239–259. [CrossRef]
19. Guedes, J.; Ferreira, D.; Barbosa, B. A non-intrusive approach to classify electrical appliances based on higher-order statistics and
genetic algorithm: A smart grid perspective. Electr. Power Syst. Res. 2016, 140, 65–69. [CrossRef]
20. Wang, A.L.; Chen, B.X.; Wang, C.G.; Hua, D. Non-intrusive load monitoring algorithm based on features of V–I trajectory. Electr.
Power Syst. Res. 2018, 157, 134–144. [CrossRef]
21. Hassan, T.; Javed, F.; Arshad, N. An empirical investigation of VI trajectory based load signatures for non-intrusive load
monitoring. IEEE Trans. Smart Grid 2013, 5, 870–878. [CrossRef]
22. Chang, H.H.; Chen, K.L.; Tsai, Y.P.; Lee, W.J. A new measurement method for power signatures of nonintrusive demand
monitoring and load identification. IEEE Trans. Ind. Appl. 2012, 48, 764–771. [CrossRef]
23. Chang, H.H.; Lian, K.L.; Su, Y.C.; Lee, W.J. Power-spectrum-based wavelet transform for nonintrusive demand monitoring and
load identification. IEEE Trans. Ind. Appl. 2014, 50, 2081–2089. [CrossRef]
24. Ducange, P.; Marcelloni, F.; Antonelli, M. A novel approach based on finite-state machines with fuzzy transitions for nonintrusive
home appliance monitoring. IEEE Trans. Ind. Inform. 2014, 10, 1185–1197. [CrossRef]
25. Gillis, J.M.; Morsi, W.G. Non-intrusive load monitoring using semi-supervised machine learning and wavelet design. IEEE Trans.
Smart Grid 2016, 8, 2648–2655. [CrossRef]
26. He, K.; Stankovic, L.; Liao, J.; Stankovic, V. Non-intrusive load disaggregation using graph signal processing. IEEE Trans. Smart
Grid 2016, 9, 1739–1747. [CrossRef]
27. Batra, N.; Dutta, H.; Singh, A. INDiC: Improved Non-intrusive Load Monitoring Using Load Division and Calibration. In
Proceedings of the 2013 12th International Conference on Machine Learning and Applications, Miami, FL, USA, 4–7 December
2013; pp. 79–84. [CrossRef]
28. Kolter, J.Z.; Batra, S.; Ng, A.Y. Energy Disaggregation via Discriminative Sparse Coding. In Proceedings of the Advances in
Neural Information Processing Systems, Vancouver, BC, Canada, 6–9 December 2010.
29. Parson, O.; Ghosh, S.; Weal, M.; Rogers, A. Non-intrusive load monitoring using prior models of general appliance types. In
Proceedings of the National Conference on Artificial Intelligence, Toronto, ON, Canada, 22–26 July 2012; pp. 1–8. Available online:
http://eprints.soton.ac.uk/id/eprint/336812 (accessed on 1 May 2022).
Sensors 2023, 23, 5226 24 of 25

30. Kim, H.; Marwah, M.; Arlitt, M.; Lyon, G.; Han, J. Unsupervised disaggregation of low frequency power measurements. In
Proceedings of the 2011 SIAM International Conference on Data Mining, SIAM, Mesa, AZ, USA, 28–30 April 2011; pp. 747–758.
31. Paradiso, F.; Paganelli, F.; Giuli, D.; Capobianco, S. Context-based energy disaggregation in smart homes. Future Internet 2016, 8,
4. [CrossRef]
32. Bonfigli, R.; Principi, E.; Fagiani, M.; Severini, M.; Squartini, S.; Piazza, F. Non-intrusive load monitoring by using active and
reactive power in additive Factorial Hidden Markov Models. Appl. Energy 2017, 208, 1590–1607. [CrossRef]
33. Makonin, S.; Popowich, F.; Bajić, I.; Gill, B.; Bartram, L. Exploiting HMM sparsity to perform online real-time nonintrusive load
monitoring. IEEE Trans. Smart Grid 2015, 7, 2575–2585. [CrossRef]
34. Aiad, M.; Lee, P.H. Unsupervised approach for load disaggregation with devices interactions. Energy Build. 2016, 116, 96–103.
[CrossRef]
35. Li, Y.; Peng, Z.; Huang, J.; Zhang, Z.; Son, J.H. Energy disaggregation via hierarchical factorial hmm. In Proceedings of the 2nd
International Workshop on Non-Intrusive Load Monitoring, Austin, TX, USA, 3 June 2014; Volume 3, pp. 1–4.
36. Xia, M.; Liu, W.; Wang, K.; Zhang, X.; Xu, Y. Non-Intrusive Load Disaggregation Based on Deep Dilated Residual Network. Electr.
Power Syst. Res. 2019, 170, 277–285. [CrossRef]
37. Kaselimi, M.; Protopapadakis, E.; Voulodimos, A.; Doulamis, N.; Doulamis, A. Multi-Channel Recurrent Convolutional Neural
Networks for Energy Disaggregation. IEEE Access 2019, 7, 81047–81056. [CrossRef]
38. Zhang, C.; Zhong, M.; Wang, Z.; Goddard, N.; Sutton, C. Sequence-to-point learning with neural networks for non-intrusive
load monitoring. In Proceedings of the AAAI Conference on Artificial Intelligence, New Orleans, LA, USA, 2–7 February 2018;
pp. 2604–2611.
39. Ahmed, S.; Bons, M. Edge Computed NILM: A Phone-Based Implementation Using MobileNet Compressed by Tensorflow Lite.
In Proceedings of the 5th International Workshop on Non-Intrusive Load Monitoring (NILM’20), Association for Computing
Machinery, New York, NY, USA, 18 November 2020; pp. 44–48.
40. Kukunuri, R.; Aglawe, A.; Chauhan, J.; Bhagtani, K.; Patil, R.; Walia, S.; Batra, N. EdgeNILM: Towards NILM on Edge Devices.
In Proceedings of the 7th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation
(BuildSys ’20), Association for Computing Machinery, New York, NY, USA, 18–20 November 2020; pp. 90–99.
41. Bonfigli, R.; Felicetti, A.; Principi, E.; Fagiani, M.; Squartini, S.; Piazza, F. Denoising Autoencoders for Non-Intrusive Load
Monitoring: Improvements and Comparative Evaluation. Energy Build. 2018, 158, 1461–1474. [CrossRef]
42. Kong, W.; Dong, Z.; Wang, B.; Zhao, J.; Huang, J. A Practical Solution for Non-Intrusive Type II Load Monitoring Based on Deep
Learning and Post-Processing. IEEE Trans. Smart Grid 2020, 11, 148–160. [CrossRef]
43. Murray, D.; Stankovic, L.; Stankovic, V.; Lulic, S.; Sladojevic, S. Transferability of Neural Network Approaches for Low-Rate
Energy Disaggregation. In Proceedings of the ICASSP 2019 IEEE International Conference on Acoustics, Speech and Signal
Processing (ICASSP), Brighton, UK, 12–17 May 2019; pp. 8330–8334.
44. Noury, N.; Berenguer, M.; Teyssier, H.; Bouzid, M.; Giordani, M. Building an index of activity of inhabitants from their activity on
the residential electrical power line. IEEE Trans. Inf. Technol. Biomed. 2011, 15, 758–766. [CrossRef] [PubMed]
45. Bucci, G.; Ciancetta, F.; Fiorucci, E. Apparatus for Online Continuous Diagnosis of Induction Motors Based on the SFRA Technique.
IEEE Trans. Instrum. Meas. 2020, 69, 4134–4144. [CrossRef]
46. IEC 60076-18:2012; Power Transformers—Part 18: Measurement of Frequency Response. IEC: Geneva, Switzerland, 2012.
47. IEEE Std C57.149-2012; IEEE Guide for the Application and Interpretation of Frequency Response Analysis for Oil-Immersed
Transformers. IEEE: New York, NY, USA, 2013; pp. 1–72. [CrossRef]
48. Fioravanti, A.; Prudenzi, A.; Bucci, G.; Fiorucci, E.; Ciancetta, F.; Mari, S. Non intrusive electrical load identification through
an online SFRA based approach. In Proceedings of the 2020 International Symposium on Power Electronics, Electrical Drives,
Automation and Motion (SPEEDAM), Sorrento, Italy, 24–26 June 2020.
49. CEI EN 50160; Power Quality Standard. CEI: Blue Ash, OH, USA, 2020.
50. D’Innocenzo, F.; Bucci, G.; Dolce, S.; Fiorucci, E.; Ciancetta, F. Power line communication, overview of standards and applications.
In Proceedings of the XXI IMEKO World Congress “Measurement in Research and Industry”, Prague, Czech Republic, 30 August–
4 September 2015.
51. Géron, A. Hands-on Machine Learning with Scikit-Learn, Keras and TensorFlow: Concepts, Tools, and Techniques to Build Intelligent
Systems, 2nd ed.; O’Reilly Media, Inc.: Sebastopol, CA, USA, 2019.
52. Kelly, J.; Knottenbelt, W. Neural NILM: Deep Neural Networks Applied to Energy Disaggregation. In Proceedings of the 2nd
ACM International Conference on Embedded Systems for Energy-Efficient Built Environments (BuildSys ’15). Association for
Computing Machinery, New York, NY, USA, 4–5 November 2015; pp. 55–64. [CrossRef]
53. Bucci, G.; Ciancetta, F.; Fiorucci, E.; Mari, S.; Fioravanti, A. A Non-Intrusive Load Identification System Based on Frequency
Response Analysis. In Proceedings of the 2021 IEEE International Workshop on Metrology for Industry 4.0 & IoT (MetroInd4.0 &
IoT), Rome, Italy, 7–9 June 2021; pp. 254–258. [CrossRef]
54. Bishop, C.M. Pattern Recognition and Machine Learning; Springer: New York, NY, USA, 2006.
55. Hofmann, T.; Schölkopf, B.; Smola, A.J. Kernel methods in machine learning. Ann. Statist. 2008, 36, 1171–1220. [CrossRef]
56. Xu, X.; Tsang, I.W.; Xu, D. Soft Margin Multiple Kernel Learning. IEEE Trans. Neural Netw. Learn. Syst. 2013, 24, 749–761.
[CrossRef]
Sensors 2023, 23, 5226 25 of 25

57. Dolce, S.; Fiorucci, E.; Bucci, G.; D’Innocenzo, F.; Ciancetta, F.; Di Pasquale, A. Test instrument for the automatic compliance check
of cast resin insulated windings for power transformers. Measurement 2017, 100, 50–61. [CrossRef]
58. Anaconda Inc. Anaconda Software Distribution. 2020. Available online: https://www.anaconda.com/distribution/ (accessed on
13 February 2022).
59. Mari, S.; Bucci, G.; Ciancetta, F.; Fiorucci, E.; Fioravanti, A. Advanced Architecture for Training and Testing NILM Systems. In
Proceedings of the IEEE Instrumentation and Measurement Technology Conference I2MTC 2022, Ottawa, ON, Canada, 16–19
May 2022.
60. Makonin, S.; Popowich, F. Nonintrusive load monitoring (NILM) performance evaluation. Energy Effic. 2014, 8, 809–814.
[CrossRef]
61. Schaffner Group. Very High Performance Single-Phase Filters; FN 2090 datasheet; Schaffner Group: Luterbach, Switzerland, 2017.
62. Aiad, M.; Lee, P.H. Energy disaggregation of overlapping home appliances consumptions using a cluster splitting approach.
Sustain. Cities Soc. 2018, 43, 487–494. [CrossRef]
63. Jain, A.K.; Ahmed, S.S.; Sundaramoorthy, P.; Thiruvengadam, R.; Vijayaraghavan, V. Current peak based device classification
in NILM on a low-cost embedded platform using extra-trees. In Proceedings of the 2017 IEEE MIT Undergraduate Research
Technology Conference (URTC), Cambridge, MA, USA, 3–5 November 2017; pp. 1–4.
64. Cannas, B.; Carcangiu, S.; Carta, D.; Fanni, A.; Muscas, C.; Sias, G.; Canetto, B.; Fresi, L.; Porcu, P. Real-Time Monitoring System
of the Electricity Consumption in a Household Using NILM Techniques. In Proceedings of the 24th IMEKO TC4 International
Symposium and 22nd International Workshop on ADC and DAC Modelling and Testing, Palermo, Italy, 14–16 September 2020;
pp. 90–95.

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.