High Frequency Non-intrusive Electric Device Detection and Diagnosis

Roman Jonetzko, Matthias Detzler, Klaus-Uwe Gollmer, Achim Guldner, Marcel Huber,

Rainer Michels and Stefan Naumann

Institute for Software Systems, Trier University of Applied Sciences,

Environmental Campus Birkenfeld, Birkenfeld, Germany

Keywords:

Pattern Recognition, Non-intrusive Load Monitoring, Classiﬁcation, Fourier Descriptors.

Abstract:

The number of electronic devices in households as well as in industrial workplaces is continuously growing be-

cause of progress in automation. Identifying unusual operating behavior, detecting device failures in advance,

and recognizing energy saving potentials are key features to improve the reliability, safety, and proﬁtability

of those systems. Facing these tasks, todays research is focused inter alia on a non-intrusive load monitoring

approach, where the electrical signal is measured at a central point with modern hardware and processed by

pattern recognition algorithms. Thus, we developed a smart meter prototype with a high sampling frequency,

which allows for continuous measurement of the current and voltage from three-phase power lines. Besides

this, in this paper we describe the usage of current-only measurement data (simple and safe installation using

current transformers) with which we were able to classify state changes of a mobile air-conditioner with the

help of Fourier descriptors as well as with additional voltage measurement.

1 INTRODUCTION

So far, currently available smart meter technology is

used to measure inter alia electrical current, voltage,

and power for determining the electrical energy con-

sumption (or in-feed e.g. of photovoltaics) of a house-

hold or industrial building. For this purpose, it suf-

ﬁces to transfer the acquired data in intervals of sec-

onds to several minutes (calculated from data mea-

sured at sampling rates in the kHz range). This usage

is primarily focused on gaining knowledge about the

immediate energy demand and to visualize the power

curve. However, the analysis of signal data measured

at high sampling rates implies another possible ap-

plication of smart meters: the detection of switching

events and the speciﬁc states of electric devices in

a circuit by examining the signal for device speciﬁc

patterns to perform load monitoring e.g. in technical

work places.

In comparison to pure resistive devices such as

light bulbs, many modern devices consist of elec-

tronic components which produce harmonics (integer

multiples of the fundamental frequency) in the current

signal, because of a nonlinear current-voltage behav-

ior. With the calculation of these harmonics, which is

possible by applying Fast Fourier Transformation to

the measured signal, the frequency spectrum can be

used as a “ﬁngerprint” for different devices, as shown

in ﬁgure 1. For our application we need to transfer

measured data at appropriate speeds so that no events

in the signal are missed. Because we did not ﬁnd such

a smart meter available at a reasonable price for con-

sumers and small and medium enterprices on the mar-

ket, we had to develop a smart meter for our applica-

tion case.

Figure 1: Current signal and amplitude spectrum of an

LED-lamp (a), compared to a compact ﬂuorescent bulb (b).

A cost-efﬁcient method for load monitoring is to in-

stall a smart meter at a central point (e.g. distribution

209

Jonetzko R., Detzler M., Gollmer K., Guldner A., Huber M., Michels R. and Naumann S..

High Frequency Non-intrusive Electric Device Detection and Diagnosis.

DOI: 10.5220/0005434502090216

In Proceedings of the 4th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS-2015), pages 209-216

ISBN: 978-989-758-105-2

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

box). This removes the need for measuring every de-

vice individually but also makes disaggregation algo-

rithms necessary to detect individual devices from the

measured sum signal. Because no intrusion into ma-

chines or devices is needed, this method is called non-

intrusive appliance load monitoring (NIALM). The

aim of load monitoring is to get knowledge about

speciﬁc devices and their operation states, improve

fault detection and identify energy efﬁciency poten-

tials. In residential applications, this allows for identi-

fying devices, in industrial applications with complex

machines, it allows for identifying machine compo-

nents.

In our paper we ﬁrst consider the state-of-the-art

in NIALM research, focus in the following chapter on

our smart meter prototype. Afterwards, we describe

appropriate features for classiﬁcation, show inter alia

how to use Fourier descriptors for classiﬁcation with

current measurement data only, and how we realized

the classiﬁcation of the active devices.

2 RELATED WORK

Using high-frequency sampled data for NIALM al-

gorithms was already mentioned by (Leeb et al.,

1995), where they describe transient event detection

via spectral envelopes. More recent research used up

to date high frequency measurement hardware avail-

able at the market (high-investment), which directly

outputs calculated complex Fourier coefﬁcients (FC)

and fed this data into neural networks (Srinivasan and

Liew, 2006). They indeed reached good detection-

rates, however they had no variable loads present in

their measurements. This presence of variable loads

is a widely stated problem for the algorithms, which

is described in various publications, e.g. (Lee et al.,

2005). With the ﬁnding of strong correlations be-

tween higher harmonics and the active power Lee et

al. subtracted the estimated power of the variable load

from the sum signal, showing that this applies if a sin-

gle variable load is present. (Zoha et al., 2012) gives

an overall overview of steady-state (using features

like power change) and transient-state (using features

like start-up current transients) disaggregation meth-

ods and points out that it is still a challenge for re-

searchers to develop a solution which is able to de-

tect all kinds of device types. A promising approach

could be an implementation of a multi-feature/multi

algorithm solution, which (Liang et al., 2010) pro-

pose, with the extraction of additional features, such

as the complete waveform of a cycle period and in-

stantaneous admittance.

So far, little research has been conducted on

low-investment non-intrusive measuring devices. In

(Guldner et al., 2013) we presented a centralized, low

investment data acquisition prototype based upon a

modiﬁed consumer energy meter by Reichelt Elec-

tronics Co. In this paper we further developed the

hardware (cf. section 3) and software (cf. section 5).

3 SMART METER HARDWARE

AND DATA PROCESSING

The purpose of the smart meter in our case is not to

use it for utility billing but rather for perfoming load

monitoring. Therefore no exact active power calcu-

lation but transferring measured data in real-time is

important. As mentioned in the introduction, we de-

veloped a smart meter to match these requirements,

which includes the following parts:

• Low-cost Measurement Instruments. Current

transformers and voltage transformers, which

make direct access to the power lines unnecessary

(direct connection to the power lines can be used

for voltage measurement, too)

• Poly Phase Measurement Chip. An ATMEL

90E36A chip reads current and voltage signals of

up to three phases at sampling rate of 7.324 kHz

• Real-time Processing of Measured Data. To

match the requirement of real-time processing,

we use a Teensy 3.1 microcontroller develop-

ment system that features a MK20DX256 ARM

Cortex-M4 with 72 MHz, 64 kB RAM memory, a

SPI and an USB Port

• Wireless Data Transfer to a Server. A Raspberry

Pi transfers the data via wi-ﬁ, which provides

galvanic isolation between potentially expensive

server hardware and the measured power lines of

up to 400 volts (in case of direct connection to the

power lines).

Figure 2 depicts the system structure. The data gen-

erated by the measurement chip is gathered by the

Teensy. Via the SPI bus, the Teensy sends the chip (in

slave-mode) into direct memory access mode (DMA).

Then - in master-mode - the chip sends the read cur-

rent and voltage measurements directly over the SPI

bus, where the Teensy collects them into two buffers

of 2,048 values. One buffer is ﬁlled with the alternat-

ing current and voltage readings of each phase, while

the second buffer is propagated to data preprocessing.

For synchronizing the data, the ATMEL chip pro-

vides a zero-crossing signal for each phase via a zero-

crossing detection pin, which propagates the start and

ﬁnish of a period in the current or voltage signals.

SMARTGREENS2015-4thInternationalConferenceonSmartCitiesandGreenICTSystems

210

Phases

A B C

Voltage

ADC

Current

ADC

ATMEL

90E36A

3 x

230 V~

3 x 1 V~

YHDC

SCT-013-005

split-core current

transformer

optional and alternative

Alternating

Buffer 1

Alternating

Buffer 2

FFT

Fourier

Descriptors

(optional)

Selection

of 50 Hz

multiples

Interrupt

Service

routine

Teensy 3.1 USB Board

Zero-

crossing

detection

Direct

Memory

Access

Triggers alteration

of buffers after 6

signal periods

SPI

PEBIS Smart Meter Board

Serial

port

Raspberry Pi

Wireless data

transfer to server

USB

Voltage

Transformer

Figure 2: Data acquisition and processing in the devised

smart meter board.

This pin is also read by the Teensy in an interrupt,

which switches the data collection to the other buffer

after the maximum number of full signal periods that

ﬁt within the 2,048 value limit have been read (in Ger-

man power lines with 50 Hz, this results in six periods

being read). This approach ﬁxes the measuring win-

dow to a full multiple of a current- and voltage period,

allowing us to simply apply the Fast Fourier Trans-

form (FFT) to the signal as-is, without the need for

a window function and without distorting the spectral

estimate.

For processing the data, at ﬁrst, a complex, in-

place, ﬁxed-point Radix-4 algorithm is used to cal-

culate the FFT. Afterwards, the complex values of the

FC are extracted and the fundamental and ﬁrst 40 har-

monics selected. Thus, the data can be read contin-

uously (no ﬂuctuation in the current and voltage sig-

nals is missed) and is reduced from 8,192 bytes to

656 bytes (real and imaginary part of the 40 harmon-

ics plus the fundamental).

Optionally - if only the current signal is measured

- Fourier descriptors (description see chapter 4.2) are

applied. Finally, the data is sent to a Raspberry Pi (or

any other computer) over the USB-Port of the Teensy.

There, the data can be transfered to a database server

or directly be visualized, disaggregated or clustered.

4 CLASSIFICATION FEATURES

Previous research showed that so far, it is implau-

sible to detect all kinds of electric devices through

their electrical signal using only a single feature (Zeif-

man and Roth, 2011). The reason is a different

load behavior of different kinds of devices that cause

the extraction of speciﬁc features needed to detect

devices of speciﬁc load behaviors. Therefore, the

calculation of FC of the measured raw signal data

(at high sampling rate), provides information content

about signals shape and phasing. In the following,

a short overview is given of the appliance classiﬁca-

tions stated by (Hart, 1992) which is kept up by sub-

sequent NIALM research. Afterwards, the calculation

of FC and their subsequent processing is detailed.

Permanent Devices. Devices which are always ac-

tive and having a constant load (e.g. hard-wired

alarms, phones, routers).

On-off Devices. Devices with only one active state,

not operated constantly, and having constant load (e.g.

light bulbs, electric motors without speed control).

Finite State Machines. Devices with more than one

active state in between which can be switched and

having a constant load in each state (e.g. dryer, multi

state kitchen devices)

Continuously Variable Devices. Devices which can

have a continuously variable load behavior, having no

stepped variation of load (e.g. speed controlled elec-

tric motors like in power drills, modern energy efﬁ-

cient pumps and fans).

To depict the differences, the amplitude curve of the

FC envelope of the different device types is shown in

ﬁgure 3.

Figure 3: Amplitude curve of the FC envelope correspond-

ing to 50 Hz signal component.

HighFrequencyNon-intrusiveElectricDeviceDetectionandDiagnosis

211

4.1 Choice of Fourier Coefﬁcients and

Inﬂuence of Noise

It is widely known that, under certain assumptions,

the Fourier series of a signal s = s(t) with period T

and fundamental frequency f

= 1/T can be rewrit-

ten as its Fourier series. When a signal is in a steady

state, in our case, we can assume that the fundamen-

tal frequency is (nearly) 50 Hz, so that T = 0.02s

(since s = s(t) is given by the power supply). Be-

cause the observation time is given by a rectangular

window w, the result of the fast Fourier transform is

an approximation of the Fourier Transform of s · w

(we denote the Fourier Transform of s as F (s)). If

the length/time of w is chosen in a way that T | w,

meaning w = k · T with k ∈ Z, one can see, with the

choice of ω = ω

= m · 2π f

, where m ∈ N

that

F (s · w)(ω

) gives the calculation formula for the

(theoretical) m-th complex FC (noted as c

= c

(s))

of a T -periodic signal s = s(t). Since the signal s

is real (current, voltage, or power), it is also known

that c

−m

= c

. In steady state, by using the FFT,

we compute F (s · w)(ω), and choosing ω = m · 2π f

(m ∈ N

), we receive approximately the (complex) FC

of the Fourier series of s. Since there are ﬂuctuations

in f

(meaning that the fundamental frequency is not

stable at f

= 50 Hz permanently), in practice, we can

ﬁnd a good approximation to these complex FC. In

experiments we observed a variation of fundamental

frequency from the power supply in a range between

49.8 Hz and 50.2 Hz.

If the measured signal (the observation time is

given by window w) is nearly in steady state, these

complex FC can be used to characterize this signal.

Using these FC, we can calculate a ﬁnite Fourier se-

ries which gives a good approximation to the ob-

served signal. These considerations motivate our

choice of feature selection: In the DFT-Calculation

(DFT: Discrete Fourier transform), we use these DFT-

values with ω, so that it is approximately ω = m ·

2π f

, where m ∈ N

. Using this approach, we also

have two restrictions to observe: First, the number

of interesting FC is bounded by the Nyquist-Shannon

sampling theorem. In the discrete form, we observe

this constraint also by the conjugate symmetry of a

DFT calculated vector. The second constraint is a lim-

itation given by the magnitude of the noise.

4.2 Fourier Descriptors

As mentioned in chapter 3, our system allows us

to classify the devices, measuring only the current

signal (cf. chapter 5.1). In this case a calculation of

Fourier descriptors is required, which the following

passage describes.

No Capturing of the Phase Shift Between Cur-

rent and Voltage. When the measurement of the

current signal s = s(t) is started at a random point

without measuring the voltage at the same time, we

have no knowledge about the phasing of the signal.

Having a continuously changing window position of

the Fourier Transform, this causes a continuous shift

in the real and imaginary parts of FC c

= c

(s),

which is not feasable for device detection. Calcu-

lating the amplitude spectrum, we can get a constant

set of frequency components of the signal, usable as

a “ﬁngerprint” of a decive. But it lacks informa-

tion about any angle relation of the frequency com-

ponents. Alghough we have no knowledge about the

phasing of the fundamental frequency related to the

voltage signal, we can relate the phasing of high-

order-harmonics of the current signal to the funda-

mental frequency of the current signal to get a de-

sciption about the signal shape, which represents ad-

ditional information for device detection. Calculation

of Fourier descriptors

(s) (cf. formula (1)) pro-

vides this and makes evaluation of periodicity of the

signal possible:

If c

6= 0 :

= |c

| · e

iϕ

; ϕ ∈ [0,2π) (1)

iϕ

)

;

iϕ

)

; ...;

iϕ

)

k : (positive) harmonic index

Figure 4: Illustration of the ﬁrst three FC before (l.) and

after applying Fourier descriptors (r.).

For two signals s

we have:

+ s

) 6=

) +

) (2)

Since Fourier descriptors are not additive (cf. formula

(2)) all possible states/combinations of measured de-

vices have to exist as reference patterns.

To show the additional information acquired by us-

ing Fourier descriptors in comparison to the ampli-

tude spectrum, we generated two signals of different

SMARTGREENS2015-4thInternationalConferenceonSmartCitiesandGreenICTSystems

212

Figure 5: Signals s

and s

of different shape (a) have the

same amplitude spectrum (b) but a different frequency spec-

trum where Fourier descriptors are applied to FC (c).

shape. The point is that these signals could not be dis-

tinguished by a pattern recognition algorithm, which

uses the amplitudes of the frequency components for

classiﬁcation. Although one can see that the two sig-

nals are clearly of a different kind of shape, it is recog-

nizable in ﬁgure 5 that they have the same amplitude

spectrum. In comparison to calculation of magnitudes

we keep the information about signal shape existing in

the FC when calculating Fourier descriptors and have

a constant “ﬁngerprint” for device classiﬁcation.

4.3 Ratios Between nth and 1st

Harmonic

The problem of detecting devices when variable loads

are present caused us to research features which re-

main constant for a variable load during their op-

eration. As a result, the ratio between high-order

harmonics and the fundamental frequency seems to

be relatively constant in initial tests (cf. ﬁgure 6).

However, the ratio is not an additive feature, so sum

signals with more than one active device cannot be

used. For generating a signal where only one device is

active, the delta-form has to be calculated. The delta-

form is the difference between the measured instanta-

neous data and the data at the previous measurement

point. In this period of time between two measured

data, we assume that only one device is switched.

Figure 6: The load variation of the hotplate (a); the variation

in amplitude spectrum (b); ratios diagram (c), where a good

stability of ratio between 3rd and 1st FC is recognizable

(150 Hz to 50 Hz component ratio of around 20%).

ock, 2009) states that devices of different elec-

tric function have speciﬁc values of those ratios. In

table 2 the classiﬁcation of devices used in our exper-

iments is listed.

5 CLASSIFICATION

To perform state and device detection from measured

signals, we used three different pattern recognition

approaches. We performed real time recognition of

device states of a mobile air-conditioner with a clus-

tering algorithm. Here we especially tested classi-

fying with current only measurement data, and also

tested artiﬁcial neural networks. The used setup is

described in section 5.1. With the setup described

in section 5.2 we measured combinations of differ-

end devices and applied a nearest-neighbor classiﬁer

to that historic dataset.

5.1 Benchmark Setup 1

In this setup we analyzed a mobile air-conditioner in-

dividually considering its different operation states,

which are speciﬁed in table 1. This means the focus

is not on detecting the single components but detect-

ing the different states of the device. Therefore, the

new Teensy smart meter was used to beneﬁt from its

continuous measurement method.

5.1.1 Applying Cluster Analysis

For creating clusters we collected the ﬁrst 17 FC in

a row vector. We measured 2,000 datasets for each

possible state (corresponds to approx. 4 minutes per

state). These historic datasets of all possible states

HighFrequencyNon-intrusiveElectricDeviceDetectionandDiagnosis

213

Table 2: Overview of Ratios for Measured Devices and Physical according to (H

ock, 2009).

Device

Ratio n

to 1

harmonic [%]

Cause

n = 2 n = 3 n = 4 n = 5 n = 7

Air-conditioner fan 4 2 2 saturation

Speakers 35 18 4 rectiﬁer

Pedestal fan 4 2 2 saturation

Compact ﬂuorescent lamp (CFL) without ECG 1 9 1 2 gas discharge

Halogen lamp

Mixer 62 25 5 5 5 rectiﬁer

Hotplate 20

Food processor 16 3 saturation

Electric knife 6 2 saturation

Table 1: Overview of the mobile air-conditioner states.

State Active device components

FanLow Evaporator fan performs

FanMed air-circulation in three different

FanHigh power states (42 W, 44 W, 51 W)

CoolLow Condenser fan (74 W), condenser

CoolMed water pump (9 W),

CoolHigh evaporator fan (42 W, 44 W, 51 W),

compressor (500...550 W)

are averaged (by components) and the calculated C

average is used as center point for the respective clus-

ter. For calculating cluster radii, as a ﬁrst approach

we used the following formula:

r =

0.25

+ Q

0.75

+ 2 · IQR (3)

Q : quartile

IQR : interquartile range

Figure 7 shows the data ﬂow of applying the cluster

algorithm to measured input data.

If a new measured vector is located in more than one

cluster, it is assigned to the cluster with the minimum

distance to its cluster center. Is the vector not located

in any cluster, it is deﬁned as an outlier, if it falls

below the 25% quartile (low-outlier) or exceeds the

75% quartile (high outlier), respectively by more than

2 times the IQR.

The detection of the three fan only switching

states (low/med/high) with the clustering algorithm

using only current measurement data (no capture of

phase shift of the signal, thus applying Fourier de-

scriptors) works nearly as well as with phase capture

(cp. table 3). In comparison to that, using only ampli-

tudes results in a much lower detection rate. During

compressor activity (cooling states) changes in evap-

orator fan states are not recognizable by the cluster-

ing algorithm, even when phase shift of the signal

is considered through zero-crossing detection of volt-

age signal. The high difference in load scale between

compressor and evaporator fan makes detecting small

changes of the fan’s power constumption difﬁcult, a

problem we also observed in setup 2.

Table 3: Detection rates applying cluster algorithm for three

fan states and different preprocessed measurement data.

Measurement data Detection rate

Current and voltage 93%

(phase capture)

Current only 82%

(Fourier descriptors applied to)

Current only (amplitudes) 39%

5.1.2 Applying Artiﬁcial Neural Networks

Besides the clustering algorithm, we tested if an arti-

ﬁcial neural network (ANN), which it is another ap-

propriate method to solve pattern recognition prob-

lems, is able to detect all states of the mobile air-

conditioner. Using the MathWorks Matlab software,

we chose a two-layer feed-forward network with sig-

moid transfer function and back propagation training

and used as input vector the result of a principle com-

ponent analysis (PCA) of the FC, which represents a

lower number of weighted and sorted features. The

target-vector contains whether or not the particular

device is active (element contains 1) or inactive (el-

ement contains 0).

The trained ANN yields the same result as the

clustering algorithm: it was also able to correctly clas-

sify the fan states, but not able to distinguish between

cooling states due to the high differences in power

consumption.

5.2 Benchmark Setup 2

The second setup consists of nine (cf. table 4) de-

vices which were measured individually for creating

reference patterns and combined with an automated

SMARTGREENS2015-4thInternationalConferenceonSmartCitiesandGreenICTSystems

214

Neuer Input-Datensatz

Input-dataset (no. n+1)

Input-dataset (no. n)

Interval: approx. 1,0 sec.

Input-

Dataset

(no. n)

Input-

Dataset

(no. n)

I1_Re: 0.378

I1_Im: 0.253

I2_Re: -0.131

…

U1_Re: 0.634

U1_Im: 0.342

U2_Re: 0.001

…

IRMS: 2.312

URMS: 227.212

…

P_Real: 4.955

P_React: 2.241

...

Historical data

Analysis

Generated from

Known

cluster

New

cluster Cm+1

Outlier definition

Clustering-

algorithm

Output

data

sequence

Known

clusters

C1 C2

C3 C4

…

Figure 7: Outlier detection algorithm.

device switching system composed of power distribu-

tion units (Energenie EG-PMS2, multi sockets which

are switchable by PC through a USB connection).

The goal of this setup was to disaggregate the single

devices from a measured sum signal. For this mea-

surement, in an early project phase, we used the Zig-

Bee smart meter prototype which measures at 4 kHz

sampling rate current and voltage signal and does not

provide continuous measurement.

Table 4: Classiﬁcation of the devices via their load classes.

Device Nominal Load Class

Power [W]

Air conditioner fan 75 on-off

Speakers 3 on-off

Pedestal fan 30 on-off

CFL 25 on-off

Halogen lamp 30 on-off

Mixer 40 on-off

Hotplate 15 cont. variable

Food processor 130 cont. variable

Electric knife 65 on-off

5.2.1 Applying Nearest-neighbor Classiﬁer

To detect individual loads from of the aggregated sig

nals, a nearest-neighbor classiﬁer has been imple-

mented. At ﬁrst, the current signals of the switchable

distribution unit’s microcontrollers were subtracted

from the measured reference dataset (individual de-

vices) and combination dataset. Thereby, a distur-

bance in the calculation of combinations from the in-

dividual signals could be avoided. Using FC, calcu-

lated through discrete Fourier transformation of the

measured raw signals, we classiﬁed with a brute-

force algorithm, which means all possible combina-

tions were calculated from the reference pattern (indi-

vidual measurements of all existing devices) and the

sum of the features was compared to the feature of the

measured dataset signal (real measurement of combi-

nation). The calculated combination with the smallest

Euclidean distance to the real measured combination

was classiﬁed as the current combination.

Applying the brute force algorithm using nearest

neighbor classiﬁer to the complete dataset of 9 de-

vices (containing devices from different load behavior

classes) using the ﬁrst 39 FC as features a relatively

low recognition rate of 27% for this dataset was com-

puted (cf. table 5). Separating out combinations of

devices which include devices with load behavior dif-

ferent than on-off-devices signiﬁcantly increased the

detection rate.

Table 5: Detection rates of complete and reduced dataset.

Left out devices None Speaker, hotplate,

mixing machine

correctly classiﬁed 139 58

combinations

Total no. of 511 64

combinations

Percentage 27% 91%

We identiﬁed the appearance of variable loads as pri-

mary cause for confounding the NIALM algorithm.

The devices with variable loads in this dataset are

the hotplate where a temperature controller adjusts

the current demand and the food processor which

HighFrequencyNon-intrusiveElectricDeviceDetectionandDiagnosis

215

runs up to higher rotational speed meaning that cur-

rent demand decreases (whereby the efﬁciency factor

increases). A continuously variable load coincides

with continuous changes of the frequency spectrum.

This implies that the averaged values of FC cause a

high similarity of different compared combinations.

In section 4.3, we presented an appropriate feature

(ratios between FC) that is likely to solve this prob-

lem. Because we often switched more than one de-

vice at the same time during the measurement of the

dataset, we are not able to use this feature yet, because

it needs the delta-form (only one device may switch

at the same time).

In comparison with setup 1, we could observe the

same problem in device detection, when there is a

large difference between the load scale of present de-

vices. Variations of the power consumptions of high

load devices can exceed smaller devices power con-

sumption level. Here, the speakers load is as high as

the variation range of power of the mixing machine.

So far, we did not ﬁnd any feature which could solve

this.

6 SUMMARY AND OUTLOOK

Our developed smart meter prototype, which is cur-

rently in testing phase, is able to provide information

about the electric signal without missing any event

occurring during time of measurement. We pointed

out that with Fourier descriptors, classiﬁcation with-

out consideration of the voltage signal is possible for

observing state changes of ﬁnite state machines. For

speciﬁc machines which have a ﬁnite state load be-

havior and whose states are of a similar load scale,

this could be a low cost and safe method of supervis-

ing. Like previous research, we observed the problem

of variable loads for the disaggregation algorithms at

setup 2. Therefore, additional extracted features like

ratios between FC could help to identify the existence

of variable loads and will be tested in further setups.

Another approach is to improve transient signal de-

tection, where envelopes of FC during device start-

up and shut-down could be allocated with speciﬁc de-

vices. The goal is to make all information contained

in the FC available and combining features extracted

from short time windows and long time windows in

detection algorithms.

ACKNOWLEDGEMENTS

This paper evolved from the research and develop-

ment project “Process Monitoring and Improved En-

ergy Efﬁciency of Technical Workplaces via Smart

Meters” (PEBiS), which is sponsored by the founda-

tion “Stiftung Rheinland-Pfalz f

ur Innovation” grant

number 961-386261/1048. The contents of this docu-

ment are the sole responsibility of the authors and can

under no circumstances be regarded as reﬂecting the

position of the foundation “Stiftung Rheinland-Pfalz

ur Innovation”.

REFERENCES

Guldner, A., Arns, S., Schunk, T., Gollmer, K.-U., Michels,

R., and Naumann, S. (2013). Detecting consumer de-

vices by applying pattern recognition to smart me-

ter signals. In Page, B., Fleischer, A. G., G

obel, J.,

and Wohlgemuth, V., editors, EnviroInfo2013 - Envi-

ronmental Informatics and Renewable Energies. 27th

International Conference on Informatics for Environ-

mental Protection, Hamburg, September 2-4, 2013,

pages 198–204.

Hart, G. W. (1992). Nonintrusive appliance load monitor-

ing. In Proceedings of the IEEE, Vol. 80, pp. 1870-

1891. IEEE.

ock, G. (2009). Dirty Power Oberschwingungen

durch nichtlineare Verbraucher. http://www.gmc-

instruments.ch/src/download/dDirty Power.pdf. [On-

line; accessed 01-December-2014].

Lee, K. D., Norford, L. K., Armstrong, P. R., Holloway, J.,

and Shaw, S. R. (2005). Estimation of variable-speed-

drive power consumption from harmonic content. In

IEEE Transactions on Energy Conversion, Vol. 20, pp.

566-574. IEEE.

Leeb, S. B., Shaw, S. R., and Kirtley, J. L. (1995). Transient

event detection in spectral envelope estimates for non-

intrusive load monitoring. In IEEE Transactions on

Power Delivery, p. 1200, 1995. IEEE.

Liang, J., Ng, S. K. K., Kendall, G., and Cheng, J. W. M.

(2010). Load signature studypart i: Basic concept,

structure, and methodology. In IEEE Transactions on

Power Delivery, Vol. 25, pp. 551-560. IEEE.

Srinivasan, D. and Liew, A. (2006). Neural-network-based

signature recognition for harmonic source identiﬁca-

tion. In IEEE Transactions on Power Delivery, vol.

21, pp. 398-405. IEEE.

Zeifman, M. and Roth, K. (2011). Nonintrusive appliance

load monitoring: Review and outlook. In IEEE Trans-

actions on Consumer Electronics, Vol. 57, pp. 76-84.

IEEE.

Zoha, A., Gluhak, A., Imran, M. A., and Rajasegarar, S.

(2012). Non-intrusive load monitoring approaches for

disaggregated energy sensing: A survey. In Sensors

2012, 12. Sensors - Open Access Journal.

SMARTGREENS2015-4thInternationalConferenceonSmartCitiesandGreenICTSystems

216