Hybrid virtual metering points – a low-cost, near real-time energy and resource flow monitoring approach for production machines without PLC data connection

Abstract

Transparent energy flows within a factory are the prerequisite for energetic improvements of the involved production machines. With the ongoing digitalization of industrial production, innovative ways of creating energy transparency on the shop floor are emerging. Virtual energy metering points predict the power consumption of a regarded entity and can therefore enable a cost-effective increase in energy transparency on machine level. However, many machines, especially in small and medium-sized enterprises (SMEs), have no external data connection, which prevents the use of data-based energy prediction models. In this paper, a near real-time deployable approach to predict the current energy consumption of production machines without a programmable logic controller (PLC) data connection is presented. By using a Raspberry Pi as low-cost edge analytics device, its integrated camera films the optical signals from light-emitting diodes (LEDs) of different PLC modules, which display the switching state signals of various machine sub-units. In a next step, the filmed PLC information is translated into state signals, which are correlated with temporarily measured electric energy data of the production machine as well as its principal sub-units. After an automated model training and hyperparameter optimization process, the empirical black box model is deployed in a near real-time environment on the Raspberry Pi. Thus, a hybrid virtual energy and resource flow metering point of the production machine as well as its sub-units is generated. In addition, challenges like model training for predicting different production processes as well as the necessary data set size for VMP model generation are addressed. The approach is tested and validated for a metal cutting machine tool and a cleaning machine of the ETA Research Factory at the Technical University of Darmstadt.

Full text

ScienceDirect
Available online at www.sciencedirect.com
Available online at www.sciencedirect.com
ScienceDirect
Procedia CIRP 00 (2017) 000–000
www.elsevier.com/locate/procedia
2212-8271 © 2017 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of the 28th C IRP Design Conference 2018.
28th CIRP Design Conference, May 2018, Nantes, France
A new methodology to analyze the functional and physical architecture of
existing products for an assembly oriented product family identification
Paul Stief *, Jean-Yves Dantan, Alain Etienne, Ali Siadat
École Nationale Supérieure d’Arts et Métiers, Arts et Métiers ParisTech, LCFC EA 4495, 4 Rue Augustin Fresnel, Metz 57078, France
* Corresponding author. Tel.: +33 3 87 37 54 30; E-mail address: paul.stief@ensam.eu
Abstract
In today’s business environment, the trend towards more product variety and customization is unbroken. Due to this development, the need of
agile and reconfigurable production systems emerged to cope with various products and product families. To design and optimize production
systems as well as to choose the optimal product matches, product analysis methods are needed. Indeed, most of the known methods aim to
analyze a product or one product family on the physical level. Different product families, however, may differ largely in terms of the number and
nature of components. This fact impedes an efficient comparison and choice of appropriate product family combinations for the production
system. A new methodology is proposed to analyze existing products in view of their functional and physical architecture. The aim is to cluster
these products in new assembly oriented product families for the optimization of existing assembly lines and the creation of future reconfigurable
assembly systems. Based on Datum Flow Chain, the physical structure of the products is analyzed. Functional subassemblies are identified, and
a functional analysis is performed. Moreover, a hybrid functional and physical architecture graph (HyFPAG) is the output which depicts the
similarity between product families by providing design support to both, production system planners and product designers. An illustrative
example of a nail-clipper is used to explain the proposed methodology. An industrial case study on two product families of steering columns of
thyssenkrupp Presta France is then carried out to give a first industrial evaluation of the proposed approach.
© 2017 The Authors. Published by Elsevier B.V.
Peer-review under responsibility of the scientific committee of the 28th CIRP Design Conference 2018.
Keywords: Assembly; Design method; Family identification
1. Introduction
Due to the fast development in the domain of
communication and an ongoing trend of digitization and
digitalization, manufacturing enterprises are facing important
challenges in today’s market environments: a continuing
tendency towards reduction of product development times and
shortened product lifecycles. In addition, there is an increasing
demand of customization, being at the same time in a global
competition with competitors all over the world. This trend,
which is inducing the development from macro to micro
markets, results in diminished lot sizes due to augmenting
product varieties (high-volume to low-volume production) [1].
To cope with this augmenting variety as well as to be able to
identify possible optimization potentials in the existing
production system, it is important to have a precise knowledge
of the product range and characteristics manufactured and/or
assembled in this system. In this context, the main challenge in
modelling and analysis is now not only to cope with single
products, a limited product range or existing product families,
but also to be able to analyze and to compare products to define
new product families. It can be observed that classical existing
product families are regrouped in function of clients or features.
However, assembly oriented product families are hardly to find.
On the product family level, products differ mainly in two
main characteristics: (i) the number of components and (ii) the
type of components (e.g. mechanical, electrical, electronical).
Classical methodologies considering mainly single products
or solitary, already existing product families analyze the
product structure on a physical level (components level) which
causes difficulties regarding an efficient definition and
comparison of different product families. Addressing this
Procedia CIRP 98 (2021) 452–457
2212-8271 © 2021 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientific committee of the 28th CIRP Conference on Life Cycle Engineering.
10.1016/j.procir.2021.01.133
© 2021 The Authors. Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)
Peer-review under responsibility of the scientic committee of the 28th CIRP Conference on Life Cycle Engineering.
Available online at www.sciencedirect.com
ScienceDirect
Procedia CIRP 00 (2019) 000–000
www.elsevier.com/locate/procedia
2212-8271 © 2020 The Authors, Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer review under the responsibility of the scientific committee of the 28th CIRP Conference on Life Cycle Engineering.
28th CIRP Conference on Life Cycle Engineering
Hybrid virtual metering points – a low-cost, near real-time energy and
resource flow monitoring approach for production machines
without PLC data connection
Johannes Sossenheimera,*, Oliver Vetterb, Thomas Stahla, Astrid Weyanda, Matthias Weigolda
a
Technical University of Darmstadt, Institute of Production Management, Technology and Machine Tools (PTW), Otto -Berndt-Str. 2, 64287 Darmstadt, Germany
b
Technical University of Darmstadt, Institute of Software & Digital Business,
Hochschulstr. 1
, 64287 Darmstadt, Germany
* Corresponding author. Tel.: +49-6151-16-25852; fax: +49-6151-16-20087. E-mail address: j.sossenheimer@ptw.tu-darmstadt.de
Abstract
Transparent energy flows within a factory are the prerequisite for energetic improvements of the involved production machines. With the
ongoing digitalization of industrial production, innovative ways of creating energy transparency on the shop floor are emerging. Virtual energy
metering points predict the power consumption of a regarded entity and can therefore enable a cost-effective increase in energy transparency on
machine level. However, many machines, especially in small and medium-sized enterprises (SMEs), have no external data connection, which
prevents the use of data-based energy prediction models. In this paper, a near real-time deployable approach to predict the current energy
consumption of production machines without a programmable logic controller (PLC) data connection is presented. By using a Raspberry Pi as
low-cost edge analytics device, its integrated camera films the optical signals from light-emitting diodes (LEDs) of different PLC modules,
which display the switching state signals of various machine sub-units. In a next step, the filmed PLC information is translated into state
signals, which are correlated with temporarily measured electric energy data of the production machine as well as its principal sub-units. After
an automated model training and hyperparameter optimization process, the empirical black box model is deployed in a near real-time
environment on the Raspberry Pi. Thus, a hybrid virtual energy and resource flow metering point of the production machine as well as its sub-
units is generated. In addition, challenges like model training for predicting different production processes as well as the necessary data set size
for VMP model generation are addressed. The approach is tested and validated for a metal cutting machine tool and a cleaning machine of the
ETA Research Factory at the Technical University of Darmstadt.
© 2020 The Authors, Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer review under the responsibility of the scientific committee of the 28th CIRP Conference on Life Cycle Engineering.
Keywords: Virtual metering points; energy and resource monitoring; energy transparency
1. Introduction and Motivation
The issue of climate change and greenhouse gas emissions
is currently very present in the political and social debate.
Still, the global primary energy consumption has risen by an
average of 1.7 % per year over the last 10 years [1]. The
European Union's climate strategy aims to increase the share
of renewable energies and gradually reduce greenhouse gas
emissions by 2050, while improving the overall energy
efficiency in parallel [2]. The goal of the European Green
Deal is to achieve climate neutrality by 2050, which is to be
enforced by a European climate law that ensures that all
sectors make their contribution [3]. These ambitious targets
require significant industrial participation, whereby increasing
energy and resource efficiency plays a key role [4, 5].
Despite the great progress in research and development in
recent years, there is a gap between the available solutions for
energy efficiency and their actual implementation in industrial
companies [6, 7]. Various barriers hinder the implementation
of energy efficiency measures [8]. Frequently mentioned
causes include a lack of information about the respective
energy consumption and high investment costs for the
Available online at www.sciencedirect.com
ScienceDirect
Procedia CIRP 00 (2019) 000–000
www.elsevier.com/locate/procedia
2212-8271 © 2020 The Authors, Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer review under the responsibility of the scientific committee of the 28th CIRP Conference on Life Cycle Engineering.
28th CIRP Conference on Life Cycle Engineering
Hybrid virtual metering points – a low-cost, near real-time energy and
resource flow monitoring approach for production machines
without PLC data connection
Johannes Sossenheimera,*, Oliver Vetterb, Thomas Stahla, Astrid Weyanda, Matthias Weigolda
a
Technical University of Darmstadt, Institute of Production Management, Technology and Machine Tools (PTW), Otto -Berndt-Str. 2, 64287 Darmstadt, Germany
b
Technical University of Darmstadt, Institute of Software & Digital Business,
Hochschulstr. 1
, 64287 Darmstadt, Germany
* Corresponding author. Tel.: +49-6151-16-25852; fax: +49-6151-16-20087. E-mail address: j.sossenheimer@ptw.tu-darmstadt.de
Abstract
Transparent energy flows within a factory are the prerequisite for energetic improvements of the involved production machines. With the
ongoing digitalization of industrial production, innovative ways of creating energy transparency on the shop floor are emerging. Virtual energy
metering points predict the power consumption of a regarded entity and can therefore enable a cost-effective increase in energy transparency on
machine level. However, many machines, especially in small and medium-sized enterprises (SMEs), have no external data connection, which
prevents the use of data-based energy prediction models. In this paper, a near real-time deployable approach to predict the current energy
consumption of production machines without a programmable logic controller (PLC) data connection is presented. By using a Raspberry Pi as
low-cost edge analytics device, its integrated camera films the optical signals from light-emitting diodes (LEDs) of different PLC modules,
which display the switching state signals of various machine sub-units. In a next step, the filmed PLC information is translated into state
signals, which are correlated with temporarily measured electric energy data of the production machine as well as its principal sub-units. After
an automated model training and hyperparameter optimization process, the empirical black box model is deployed in a near real-time
environment on the Raspberry Pi. Thus, a hybrid virtual energy and resource flow metering point of the production machine as well as its sub-
units is generated. In addition, challenges like model training for predicting different production processes as well as the necessary data set size
for VMP model generation are addressed. The approach is tested and validated for a metal cutting machine tool and a cleaning machine of the
ETA Research Factory at the Technical University of Darmstadt.
© 2020 The Authors, Published by Elsevier B.V.
This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/)
Peer review under the responsibility of the scientific committee of the 28th CIRP Conference on Life Cycle Engineering.
Keywords: Virtual metering points; energy and resource monitoring; energy transparency
1. Introduction and Motivation
The issue of climate change and greenhouse gas emissions
is currently very present in the political and social debate.
Still, the global primary energy consumption has risen by an
average of 1.7 % per year over the last 10 years [1]. The
European Union's climate strategy aims to increase the share
of renewable energies and gradually reduce greenhouse gas
emissions by 2050, while improving the overall energy
efficiency in parallel [2]. The goal of the European Green
Deal is to achieve climate neutrality by 2050, which is to be
enforced by a European climate law that ensures that all
sectors make their contribution [3]. These ambitious targets
require significant industrial participation, whereby increasing
energy and resource efficiency plays a key role [4, 5].
Despite the great progress in research and development in
recent years, there is a gap between the available solutions for
energy efficiency and their actual implementation in industrial
companies [6, 7]. Various barriers hinder the implementation
of energy efficiency measures [8]. Frequently mentioned
causes include a lack of information about the respective
energy consumption and high investment costs for the

Johannes Sossenheimer et al. / Procedia CIRP 98 (2021) 452–457 453
2 Johannes Sossenheimer et al./ Procedia CIRP 00 (2021) 000–000
efficiency measure itself and the metering equipment, since
the payback periods are sometimes unclear and difficult to
evaluate [5, 9–11]. Nevertheless, precise knowledge about the
own energy consumption is the prerequisite for identifying
energetic improvement and optimization approaches [12].
Thus, there is a need for better and low-cost energy
monitoring within the industrial sector [13].
Even though IoT-compatible production machines can
generate large quantities of data, especially small and
medium-sized enterprises (SMEs) often do not monitor and
analyze the production data [14]. Especially the machine-
internal data of programmable logic controllers (PLCs) as
well as the bus data can however be beneficial for condition,
process and low-cost energy monitoring [15, 16]. Previous
work [17] showed, how offline trained energy and resource
consumption models, also referred to as hybrid virtual
metering points (VMPs), can be set-up and deployed in near
real-time for IoT-compatible production machines, which
have full machine-internal data availability in terms of PLC
and NC-core data, as well as data from decentrally controlled
devices. Additionally, it is shown that the modeling accuracy
is significantly reduced if the required data is not available,
which is to be particularly questioned for brownfield
production machines. Therefore, the aim of this paper is to
transfer this low-cost monitoring approach to production
machines without PLC data availability. For this purpose, the
LED state signals of PLC input/output (I/O) modules are
captured and processed with a camera-based Raspberry Pi
edge computer, named energy technologies and applications
Edge Lense (ETA Edge Lense), as described in chapter 3.
Once the VMP models are trained, they can be deployed in
near real-time on the ETA Edge Lense.
The presented VMP approach is tested and validated on
different machines of the ETA Research Factory of the
Technical University of Darmstadt.
Nomenclature
ERDA energy and resource data acquisition
ETA energy technologies and applications
I/O module input/output module
IoT internet of things
LED light-emitting diode
NRMSE normalized root mean squared error
PLC programmable logic controller
SME small and medium-sized enterprise
VMP virtual metering point
2. Research Background
The following chapter gives insight into LED state signals
of I/O modules of PLCs. In addition, existing machine
consumption models and VMP approaches are presented.
2.1. LED state signals of I/O modules
I/O modules are one of the main components of a PLC.
They connect the PLC system with the machine’s components
and enable the integration of external signals from field
devices, such as analog or digital sensor signals, as well as the
control of actuators [18]. Most I/O modules have status LEDs,
which are generally used to locate system and peripheral
errors, and for speeding up troubleshooting [19].
As status LEDs indicate the switching or operating status
of the respective I/O module’s in- or output, they can be read
out and analyzed for other purposes, like state-based energy
consumption modelling, continuous energy monitoring or
load forecasting. As these signals can be recorded with a
camera, no machine data connection is needed, which allows
the application to non IoT-compatible production machines.
In order to assess the impact of the approach, data sheets of
all currently available I/O module families from Siemens,
Schneider Electric, Bosch Rexroth, B&R Automation,
Phoenix Contact, WAGO, Weidmüller and Beckhoff have
been examined. It is found that all mentioned I/O module
manufacturers display the current channel status on the digital
I/O modules with status LEDs. However, only few analog
I/O modules have status LEDs.
2.2. Virtual energy and resource flow metering points
As described in [17], the required data for energy
monitoring can be acquired in numerous ways. One
possibility to monitor the energy or resource consumption are
sensor-based physical metering points. Alternatively, the
consumption data can also be calculated with VMPs. In
accordance with [10, 20] this work refers to a VMP as a
virtual image of the physical power consumption of a
regarded entity. In general, virtual sensors can be beneficial
when [20]:
 the costs for purchasing, implementing or maintaining a
physical sensor are too high,
 physical sensors cannot measure the desired state or
physical value,
 the designated space is limited or too hostile for physical
sensors,
 physical sensors have either a too high lag time, a too low
sampling rate or when they cannot maintain their
calibration.
Virtual energy metering points can be subdivided into
aggregation and disaggregation as well as physical and hybrid
model based approaches [21]. Due to the simplicity of the
empiric model generation, hybrid VMPs, which are based on
condition-dependent energy data linked either to fixed
machine condition data or process dependent empirical data
models [17, 21], are used in this paper. Hybrid models can be
subdivided into [10, 17]:
 simple VMPs,
 factorized VMPs,
 state-based VMPs and
 process- and state-based VMPs.
Simple VMPs model the electric power consumption of a
system with its switching state and a constant baseline value,
which is either derived from nominal values or a temporary

454 Johannes Sossenheimer et al. / Procedia CIRP 98 (2021) 452–457
Johannes Sossenheimer et al./ Procedia CIRP 00 (2021) 000–000 3
measurement [10]. Factorized VMPs estimate the power
consumption of a machine via a

base load and a
Ǧ or dynamic share, which is obtained by
multiplying empirically determined factors with an averaged
process parameter [10]. Both VMP types can be used to
roughly approximate the energy consumption of binary-
controlled constant consumers. However, varying operating
modes or a dynamic energy consumption behavior cannot be
modelled at high accuracy for short temporal prediction
intervals of a few seconds.
State-based VMPs use the current state signals of the
machine and its support units to calculate the machine’s
energy consumption. In its simplest form, a state-based VMP
outputs the corresponding mean energy consumption during
each respective state of the machine [10]. In more
sophisticated VMP models, the operating modes of the
machine is modelled as a state machine [22–26], which
simulates the energy consumption during all operating modes
and transitions. For the similar EnergyBlocks approach [27],
the user assigns suitable characteristic energy consumption
profiles (e.g. continuous or constant load profile) to all
possible operating modes of the considered machines,
depending on their measured energy consumption. Another
approach uses energetic function blocks on a PLC to monitor
the energy consumption in near real-time and to enable an
automated stand-by management [28].
As a summary it can be stated that simple modelling
approaches come with a high inaccuracy. Factorized VMPs
offer a higher accuracy but can only be applied to selected
processing steps. State-based VMPs present a more general
model of the machine’s energy consumption, however the
transferability to new machining processes or other machines
is limited.
In general, empiric hybrid models require little
computational resources and are thus suitable for near real-
time deployment, which is only shown in few of the above
mentioned works.
3. Virtual metering concept for brownfield production
machines
Existing metering concepts for continuous energy and
resource data acquisition (ERDA) that take VMPs into
consideration [10] use inaccurate models like simple,
factorized or basic state-based hybrid VMPs. Therefore,
VMPs are only utilized to monitor energy flows of little
relevance. According to [10], physical metering points should
be used for production machines which are either energy-
intensive, have fluctuating electrical load profiles, follow
variable production tasks, or have no machine data access.
However, if the energy and resource flows of a production
machine should be monitored continuously and no physical
metering point is installed, VMPs based on modern data-
based modeling strategies are well suited, which is illustrated
in figure 1. If sufficient state and process data is available,
even dynamic energy consumers like a machine tool with
variable production tasks can be modelled at a prediction error
of 5-15 % [17, 29]. The following sections describe how LED
state signals from I/O modules can be used for state-based
ERDA, VMP model generation and near real-time
deployment on the ETA Edge Lense for production machines
without PLC data availability.
Fig. 1: Flow chart as decision support for the installation of
physical or virtual metering points in production machines.
3.1. State-based energy and resource data acquisition with
ETA Edge Lense
To set up a hybrid VMP model, temporal ERDA of the
regarded production machine is conducted during operation
and communicated from a Janitza UMG 604 and UMG 20CM
energy meter via Modbus TCP to an edge device. In addition,
the machine’s state signals need to be recorded. As this paper
addresses non IoT-compatible production machines, which
have no possibility to publish machine-internal data in near
real-time, the ETA Edge Lense is used to record the LED
state signals of the machine’s I/O modules.
The ETA Edge Lense consists of a Raspberry Pi 4 with 4
GB RAM and a camera with a resolution of 5 Megapixels and
160° viewing angle. With additional components for power
supply, casing, external storage and CPU cooling, the overall
hardware costs for the ETA Edge Lense are 97 € without
value added tax. The operating system is Raspberry PI OS
Buster and the image processing framework is programmed
with the Python library OpenCV, while the Python library
opcua is used for setting up an OPC UA output server.
To film the I/O modules, the ETA Edge Lense is fixed in
the switching cabinet with the help of magnets. The LEDs to
be captured are manually marked within the initialization step.
During the recognition process, which runs cyclically, the
current image is first converted into a binary image using the
Otsu's thresholding method, which determines an optimal
threshold value for binarization. The intensity at each LED
position is then checked against a threshold value and the
LED variable is set to true or false accordingly, which is also
marked in the stored image with a green or red dot (see
figure 2). The image processing takes between 0.2 and 0.5 s,
physical
sensors
available
Start
LED
state-signals on
I/O modules
PLC
IoT-compatible
yes
no
no
yes
near real-time VMP model deployment
automated VMP model generation
continous energy and resource consumption data in near real-time
retrofit
physical
sensor
yes
no
machine
relevant for
ERDA
End
hybrid VMP
modelling
conduct condition-based ERDA based on:
LED state signals machine-internal data

Johannes Sossenheimer et al. / Procedia CIRP 98 (2021) 452–457 455
4 Johannes Sossenheimer et al./ Procedia CIRP 00 (2021) 000–000
even for large production machines with more than 300 LED
state signals. Finally, each LED signal is published on an
OPC UA server. For the generation of VMP models, the user
does not need to know which component’s state is visualized
by which LED signal, as the model generation framework
automatically selects well suited input features, which is
explained in section 3.2. However, the allocation between the
LED signal and the switched machine aggregate can be found
in the PLC program or in the electrical plan and can be used
for further use cases like condition monitoring.
Fig. 2: Close-up view of the LED state signals (left), after Otsu filtering
(middle) and LED-identification (right) of the digital I/O module from a
Bosch Rexroth PLC of a MAFAC KE A cleaning machin e.
3.2. Automated generation of the hybrid VMP model
As described in previous work [17], the second step of the
set-up process is the VMP model training. For that, the data
set is split into training (80 % of data set size) and test (20 %
of data set size) data. The prediction models are automatically
trained in Python through a framework which is explained in
[17, 30]. The framework optimizes the model’s
hyperparameters through Bayesian optimization with the
Python library hyperopt.
Before each evaluation, a new model is trained based on a
set of hyperparameters, which are selected via probabilistic
functions. The chosen hyperparameters include the selected
features and the regression model with the respective
parameters. In a next step, the model is trained on the training
data set with the selected hyperparameters. The trained model
is validated with 5 fold cross-validation upon the  score and
the evaluation results enter the probabilistic functions for the
hyperparameter selection for the next evaluation. After all
evaluations are calculated, the best-performing model is tested
upon the unseen test data set and is saved as VMP model.
All models are evaluated with the  score and a variant of
the normalized root mean squared error (NRMSE) as
measures for the model’s accuracy. The latter is calculated
according to equation 1, where  is the number of samples, 
the value of the target variable, 
 the respective predicted
value,  is the 95th percentile and  the 5th percentile of
the target variable. This variation offers greater comparability
between datasets of dynamic energy consumers, as it is not
distorted by outliers like load peaks.



  


ሺͳሻ
3.3. Near real-time VMP deployment on the ETA Edge Lense
As described in [21], VMP models can be executed online
in a near real-time deployment pipeline on machine
controllers, edge devices or on central servers. For brownfield
production machines, the presented approach is deployed on
the ETA Edge Lense computer in Python. While for IoT-
compatible production machines different types of model
inputs like PLC and NC core data, as well as data from
decentrally controlled devices can be used [17], brownfield
production machines without any data availability can use the
ETA Edge Lense to read out the machine’s LED state signals
in near real-time (as described in section 3.1). Based on the
selected input features, the model predicts the desired target
variables, like the currently consumed electric power on
machine and support unit level, the compressed air
consumption or other resource consumption flows for every
second and publishes the predicted values via OPC UA,
which is displayed in figure 3.
Fig. 3: Schematic representation of the hybrid virtual metering model with
possible inputs from IoT-compatible (in gray) and brownfield (in black)
production machines.
4. Validation of hybrid VMP brownfield approach
Previous work has shown that the choice of the input
features, which are selected from the available machine-
internal signals, is of far greater importance for generating an
accurate VMP model than the choice of the regression
algorithm [16, 17, 25]. To further study the suitability of LED
state signals as input features to hybrid VMP models, several
production processes of an EMAG VLC 100Y machine tool
and a MAFAC KEA cleaning machine are recorded along
with their energy consumption data. Both machines are IoT-
compatible and have accessible machine data. However, in
order to simulate brownfield system behavior without
machine data availability, the model only uses LED state
signals read out by the ETA Edge Lense as input features.
4.1. VMP deployment and analysis for a machine tool
The energy consumption of dynamic consumers like
machine tools can be decomposed into an approximately
stationary or at least intervalwise stationary share of mainly
support units and a dynamic, process dependent share of the
machine drives [31]. Table 1 summarizes the prediction
results of the VMP model for both mentioned cases.
By separately measuring and predicting the energy
requirements of the machine and its components, the
advantages and disadvantages of the brownfield approach
become apparent. Because the LEDs only emit status signals,
but no process signals such as axis torques or spindle currents,
the model is more inaccurate for dynamic consumers. This
behavior is illustrated by the electrical energy consumption
predictions of the machine tool with and without the drives in
figure 4. In [17] the chip conveyor's energy consumption
prediction with machine-internal data still had a high error of
22.4 % due to the missing state data. By contrast, the ETA
PLC data
NC core data
data from decentrally
controled units
hybrid VMP
near real-time
deployment
pipeline
LED state signals

456 Johannes Sossenheimer et al. / Procedia CIRP 98 (2021) 452–457
Johannes Sossenheimer et al./ Procedia CIRP 00 (2021) 000–000 5
Edge Lense detected the chip conveyor's state signals via the
LEDs and therefore achieves significantly better results.
Following this finding, the PLC signals of the chip conveyor
were corrected, which is underlined by the improved
prediction result with machine-internal signals. Since the
hydraulic system is controlled decentrally via a frequency
inverter, relevant data cannot be obtained from the PLC or the
I/O LEDs, which explains the high prediction error.
Fig. 4: Measured and predicted power with and without PLC data availability
during operation of a machine tool.
Table 1. Accuracy of supply system’s electric energy and compressed air
consumption prediction models for an EMAG VLC 100Y machine tool.
Target variable
machine-internal signals
LED state signals
NRMSE
R²
NRMSE
R²
Main connection
0.105
0.91
0.222
0.604
Main connection
without drives 0.144 0.797 0.146 0.790
Cooling lubricant
system 0.151 0.805 0.148 0.813
Hydraulic system
0.36
0.389
0.375
0.336
Suction
0.121
0.887
0.138
0.855
Chip conveyor
0.08
0.944
0.063
0.965
Compressed air
consumption 0.207 0.319 0.114 0.792
4.2. VMP deployment and analysis for a cleaning machine
The electric power consumption of a water-based MAFAC
KEA cleaning machine is predicted on machine and support
unit level. Table 2 compares the prediction results from VMP
models which are trained with machine-internal signals and
thus model an IoT-compatible cleaning machine, to a
brownfield cleaning machine for which only LED state
signals are analyzed.
The prediction results show that the cleaning machine,
which has a relatively stationary energy consumption
behavior, can be modelled with a higher accuracy than the
machine tool with both approaches. The prediction error of
the approach with LED state signals is below 15 %, while the
approach based on machine-internal signals shows an error of
less than 5 %. As some support units, which are indicated
with a *-symbol in table 2, are only active for less than 5 % of
the measurement duration, their NRMSE error is calculated
with the full value range instead of the 95-5 % quantile.
Table 2. Accuracy of support unit’s electric energy and compressed air
consumption prediction models for a MAFAC KEA cleaning machine.
Target variable
machine-internal signals
LED state signals
NRMSE
R²
NRMSE
R²
Main connection
0.039
0.975
0.107
0.812
Central pump
0.054
0.988
0.092
0.965
Suction
0.027*
0.664
0.030*
0.591
Supply air fan
0.126
0.857
0.145
0.810
Air heater
0.130*
0.569
0.145*
0.463
Storage tank heating
0.027
0.985
0.035
0.975
4.3. Assessment of necessary data set size for VMP model
generation
The learning curve of the machine tool VMP model, which
is only based on LED input features, is shown in figure 5.
Learning curves are used to examine the effects of smaller
data set sizes. Therefore, the training set is iteratively
shortened and the model is retrained. The original data set,
recorded at a sampling rate of 0.5 s, contains 33500 data
points per feature, which corresponds to 4.65 hours of
machine operation. During this time, the machine tool
executed two different production tasks several times.
When reducing the training data set to 5 % of its original
size, low training errors, but high test errors are obtained,
which indicates overfitting. With training set sizes over 7.5 %
of the original set, the test set error values improve
significantly and continue to improve slightly with growing
data set size. This indicates that even short temporal ERDA
periods are suited to train VMP models, based on LED state
signals recorded with the ETA Edge Lense.
Fig. 5: Learning curve of machine tool VMP model based only on LED input
features.
5. Summary and Outlook
The increasing digitalization of industrial shop floors
enables the integration of innovative ways to monitor energy
and resource flows in manufacturing factories. The presented
work extends previously published hybrid virtual metering
approaches for production machines, with unrestricted
machine data availability [17], to brownfield production
machines with no machine data availability.
To create a VMP, the ETA Edge Lense, which is a
Raspberry Pi with an integrated camera, films optical signals
from LEDs of the PLC’s I/O modules and translates them into
state signals. In a next step, an empirical black box model is
trained on a central server and deployed in a near real-time
environment on the ETA Edge Lense. Thus, a hybrid VMP of

Johannes Sossenheimer et al. / Procedia CIRP 98 (2021) 452–457 457
6 Johannes Sossenheimer et al./ Procedia CIRP 00 (2021) 000–000
the production machine as well as its supply units is generated
for a brownfield production machine.
We compare the VMP prediction results of several
production cycles of an EMAG machine tool and a MAFAC
KEA cleaning machine, simulating the cases of full machine-
internal data availability and no data availability. The results
show that the brownfield VMP approach, which is purely
based on LED state signals, predicts dynamic energy
consumers like machine tools with increased prediction errors
of 20-25 %, as no process data (e.g. spindle current) is
available. Consumers with a relatively stationary consumption
behavior like cleaning machines, can be modelled at a lower
prediction error of less than 15 %.
As the industrial brownfield often comprises a large
number of PLCs of different manufacturers and versions, the
main advantage of the presented approach lies not only in
reduced acquisition costs for physical metering equipment,
but also in the simplicity of the machine-internal data
acquisition via LED state signals and near real-time model
deployment on the ETA Edge Lense. Whether or not a VMP
should be used ultimately depends on the requirements in
terms of data accuracy. Due to the low hardware costs of
VMP, they should be considered as a low-cost alternative to
physical metering points for use cases which tolerate the
presented modelling errors, such as factory internal load
management, energy performance indicator assessment or
energy consumption allocations to departments, production
machines or products. The presented approach can easily be
transferred to various brownfield production machines.
Future work will investigate the application of the PLC
state signals to other use cases like condition monitoring.
Acknowledgements
The authors are grateful to the EU and the state Hesse for
funding the presented work in the ArePron project.
References
[1] BP Statistical Review of World Energy (2018) BP Statistical Review of
World Energy 2018. 67th ed., London.
[2] European Comission (2012) Energy: Roadmap 2050. Publications
Office of the European Union, Luxembourg.
[3] European Comission (2019) The European Green Deal. 2019th ed.,
Brussels.
[4] European Comission (2019) Going climate-neutral by 2050: A strategic
long-term vision for a prosperous, modern, competitive and climate-
neutral EU economy. Publications Office of the European Union,
Luxembourg.
[5] Edenhofer O, Pichs-Madruga R, Sokona Y, Minx J (2015) Climate
Change 2014: Mitigation of Climate Change: Working Group III
Contribution to the IPCC Fifth Assessment Report. Cambridge
University Press, Cambridge.
[6] Bunse K, Vodicka M, Schönsleben P, Brülhart M, Ernst FO (2011)
Integrating energy efficiency performance in production management –
gap analysis between industrial needs and scientific literature. Journal of
Cleaner Production 19(6-7):667–79.
[7] Lorenz M, Lüers M, Ludwig M, Rees S, Zelinger M, Stiller R (2020)
For Machinery Makers, Green Tech Creates Green Business.
[8] Denkena B, Abele E, Brecher C, Dittrich M-A, Kara S, Mori M (2020)
Energy efficient machine tools. CIRP Annals.
[9] Sorrell S, Mallett A, Nye S (2011) Barriers to industrial energy
efficiency: A literature review, Vienna.
[10] Posselt G (2016) Towards Energy Transparent Factories. Springer
International Publishing, Cham.
[11] Kara S, Bogdanski G, Li W (2011) Electricity Metering and Monitoring
in Manufacturing Systems. in Hesselbach J, Herrmann C, (Eds.).
Glocalized Solutions for Sustainability in Manufacturing. Springer
Berlin Heidelberg. Berlin, Heidelberg, pp. 1–10.
[12] Thiede S, Posselt G, Herrmann C (2013) SME appropriate concept for
continuously improving the energy and resource efficiency in
manufacturing companies. CIRP Journal of Manufacturing Science and
Technology 6(3):204–11.
[13] O’Rielly K, Jeswiet J (2015) The Need for Better Energy Monitoring
within Industry. Procedia CIRP 29:74–9.
[14] Bischoff J (2015) Erschließen der Potenziale der Anwendung von
,Industrie 4.0‘ im Mittelstand, Mühlheim an der Ruhr.
[15] Bauerdick CJH, Helfert M, Menz B, Abele E (2017) A Common
Software Framework for Energy Data Based Monitoring and
Controlling for Machine Power Peak Reduction and Workpiece Quality
Improvements. Procedia CIRP 61:359–64.
[16] Sossenheimer J, Walther J, Fleddermann J, Abele E (2019) A Sensor
Reduced Machine Learning Approach for Condition-based Energy
Monitoring for Machine Tools. Procedia CIRP 81:570–5.
[17] Sossenheimer J, Vetter O, Abele E, Weigold M (2020) Hybrid virtual
energy metering points – a low-cost energy monitoring approach for
production systems based on offline trained prediction models. Procedia
CIRP 82.
[18] Bolton W (2009) Programmable logic controllers. Newnes, Amsterdam,
Boston.
[19] Bosch Rexroth AG. Rexroth IndraControl S20: System und Installation.
https://www.boschrexroth.com/various/utilities/mediadirectory/downloa
d/index.jsp?object_nr=R911335987 (accessed on 19.06.2020).
[20] IntelliDynamics. Introduction Brief: Virtual Flow Metering.
http://www.intellidynamics.net/content/technologies/virtual-
sensors.html (accessed on 23.11.2019).
[21] Flick D, Kuschicke F, Schweikert M, Thiele T, Panten N, Thiede S,
Herrmann C (2018) Ascertainment of Energy Consumption Information
in the Age of Industrial Big Data. Procedia CIRP 72:202–8.
[22] Schmitt R, Bittencourt JL, Bonefeld R (2011) Modelling Machine Tools
for Self-Optimisation of Energy Consumption. in Hesselbach J,
Herrmann C, (Eds.). Glocalized Solutions for Sustainability in
Manufacturing. Springer Berlin Heidelberg. Berlin, Heidelberg, pp.
253–257.
[23] Bittencourt JL (2013) Selbstoptimierende und bedarfsgerechte
Steuerungsstrategien für Werkzeugmaschinen zur Steigerung der
Energieeffizienz. Zugl.: Aachen, Techn. Hochsch., Diss., 2012. 1st ed.
Apprimus Verl., Aachen.
[24] Jia S, Tang R, Lv J, Yuan Q, Peng T (2017) Energy Consumption
Modeling of Machining Transient States based on Finite State Machine.
The International Journal of Advanced Manufacturing Technology
88(5-8):2305–20.
[25] Mechs S, Lamparter S, Müller JP (2012) On Evaluation of alternative
Switching Strategies for energy-efficient Operation of modular Factory
Automation Systems. Proceedings of the 17th International Conference
on Emerging Techologies & Factory Automation. Kraków, Poland, 17.-
21. September 2012. IEEE. Piscataway, NJ, pp. 1–8.
[26] Mechs S, Müller JP, Lamparter S, Peschke J (2012) Networked Priced
Timed Automata for Energy-efficient Factory Automation. American
Control Conference (ACC) 2012. IEEE. Piscataway, NJ, pp. 5310–5317.
[27] Weinert N, Seliger G (2010) Vorgehensweise für Planung und Betrieb
energieeffizienter Produktionssysteme. Zugl.: Berlin, Techn. Univ.,
Diss., 2010. Fraunhofer-Verl., Stuttgart.
[28] Bosch Rexroth AG, (Ed.) (2017) Handbuch Energieeffizienz: Methoden
zur Steigerung der Energieeffizienz in Industrieunternehmen. Bosch
Rexroth, Würzburg.
[29] Sossenheimer J, Weber T, Flum D, Panten N, Abele E, Fuertjes T
(2019) Non-intrusive Load Monitoring on Component Level of a
Machine Tool Using a Kalman Filter-Based Disaggregation Approach.
in Armendia M, Ghassempouri M, Ozturk E, Peysson F, (Eds.). Twin-
Control. Springer International Publishing. Cham, pp. 155–165.
[30] Weber T, Sossenheimer J, Schäfer S, Ott M, Walther J, Abele E (2019)
Machine Learning based System Identification Tool for data-based
Energy and Resource Modeling and Simulation. Procedia CIRP
80:683–8.
[31] Kuhrke B (2011) Methode zur Energie- und Medienbedarfsbewertung
spanender Werkzeugmaschinen. epubli GmbH, Berlin.