CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

Contents lists available at ScienceDirect

CIRP Journal of Manufacturing Science and Technology

journal homepage: www.elsevier.com/locate/cirpj

Energy prediction for CNC machining with machine learning

Markus Brillingera,* , Marcel Wuwera , Muaaz Abdul Hadia , Franz Haasb
a
Pro2Future GmbH, Area 4.2 – Cognitive Production Systems, Inffeldgasse 25f, 8010 Graz, Austria
b
Institute of Production Engineering, Graz University of Technology, Inffeldgasse 25f, Graz 8010, Austria

A R T I C L E I N F O A B S T R A C T

Article history: Nowadays, the reduction of CO2 emissions by moving from fossil to renewable energy sources is on the
Available online 22 September 2021 policy of many governments. At the same time, these governments are forcing the reduction of energy
consumption. Since large industries have been in the focus for the last decade, today also small and
medium enterprises with production lot size one are increasingly being obliged to reduce their energy
Keywords: requirements in production. Energy-efficient CNC machine tools contribute to this goal. In machining
Energy prediction
processes, the machining strategy also has a significant influence on energy demand. For manufacturing
CNC machine tools
of lot size one, the prediction of the energy demand of a machining strategy, before a part is
Machine learning
CNC machining
manufactured plays a decisive role. In numerous previous studies, analytical models between the energy
NC code demand and the machining strategy have been developed. However, their accuracy depends largely on
the parameterization of these models by dedicated experiments. In this paper, different machine learning
algorithms, especially variations of the decision tree (’DecisionTree’, ’RandomForest’, boosted
’RandomForest’) are investigated for their ability to predict the energy demand of CNC machining
operations based on real production data, without the need for dedicated experiments. As shown in this
paper, the most accurate energy demand predictions can be achieved with the ’RandomForest’ algorithm.
© 2021 The Authors. This is an open access article under the CC BY-NC-ND license (http://
creativecommons.org/licenses/by-nc-nd/4.0/).

Introduction strategies with respect to geometry features to create awareness

for design engineers. Second, this helps determining the energy
Many government's environmental policies of the last decades consumption of a part before manufacturing and therefore
focus on decarbonizing of their energy power plants and the allowing to optimize in design phase for energy demand. Third,
reduction of energy consumption [31,37]. With a primary energy power demand prediction can help for peak power balancing to
consumption of more than 30%, industrial manufacturing is one of reduce the overall power spikes at factory level as a contribution to
the main sources of pollution [23]. As a consequence, attention for an increased electric grid stability [24].
energy-efficiency must be paid and is divided in two scopes:
Environmental friendly product design and energy-efficient State of the art
production. Hence, a comprehensive energy aware product design
and a low energy production strategy fosters governments In order to develop a prediction model for the energy
environmental policies [1,7]. Therefore, linking product geometry consumption during CNC machining, both are inevitable: the
features and energy consumption of production is inevitable for knowledge about characterization of the single elements respon-
this objective. With focus on CNC machine tools this can be sible for the energy consumption during machining as well as a
achieved by linking the NC code, which determines the geometry broader understanding of the already taken efforts in building
of the part, with the energy consumption of the CNC machine tool. predictive models for the energy consumption. Those two domains
The aim of this paper is contributing to a highly accurate NC are elaborated below:
code based energy consumption and power curve prediction for
CNC machine tool aggregates with variable power demand (x, y and Characterization of energy consumption in machining
z axis, spindle and tool change system) at CNC machine tool level.
First, this contribution will help to derive low energy machining In the past, research in machining was focused on developing
models to predict fundamental variables as stresses, strains, strain-
rates, temperatures, etc. [3]. Recently efforts were made to
understand the fundamental influences on the energy consump-
* Corresponding author. tion of a CNC machine tool.
E-mail address: [email protected] (M. Brillinger).

https://doi.org/10.1016/j.cirpj.2021.07.014
1755-5817/© 2021 The Authors. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

As one of the first, [13] investigated the energy demand for (windings and core) as well as the mechanical setup (friction,
different manufacturing processes (injection molding, machining, transmission) were added.
grinding, cutting, etc.) and characterized the energy needed for [12] formulated a multi-objective energy and product quality
machining with E = (P0 + kcQ)t. Where for a process duration of t, P0 model using multi-gene genetic programming. The chosen
is a constant power consumption required to support the process, approach includes both, statistical and classification strategies.
kc the specific cutting force and Q the material removal rate. The model was validated through analyzes of the turning process
[36] presented a way to calibrate a force estimation model using of steel and aluminium alloy. In the formulated model the cutting
the motor spindle power in flat and ball end milling processes. The speed was most influential on the energy consumption.
energy consumption was expressed in terms of material removed [22] studied the relationship between the parameters depth of
and contact area E = KTCQ + KTEA, with the tool/material cutting cut, spindle speed, feed rate, and nose radius of a dry milling
energy KTC and the tool/material edge force KTE. operation and modeled their highly nonlinear impact on the
[8] searched for possible improvement measures to reduce the specific cutting energy. The parameters with the most effect on the
constant part of the energy consumption of discrete part specific cutting energy were found out to be (in descending order):
production CNC machine tools. Also the effects of stand-by phases depth of cut, feed rate, spindle speed and nose radius.
in a lean production setting was found out to effect the energy
consumption. Prediction of energy in machining
[10] proposed a generic model to forecast the energy
consumption of manufacturing equipment by separation in After characterizing the energy consumption of CNC machine
statistical discrete events. The total energy consumption was tools, the question arose, if those gained insights could be used to
PP
calculated with E = c i(DtPc,i), whereas, Pc,i being the power create models to forecast the energy consumption during
consumption of a component c in state Si. Later [11] further machining of a part. As one of the first, [9] noticed the absence
developed this methodology for the usage in multi-machine of a suitable forecasting technique for the energy consumption of
manufacturing systems and expanded the model for different CNC machine tools and proposed a model formalism to predict and
energy carriers (electricity, pressured air and coolant). optimize the energy-efficiency: the process was divided into
[34] monitored the energy consumption of CNC machine tools operations, for each of which a static energy consumption was
in order to assign every operation performed by the CNC machine determined by measurements. In a case study simulating a milling
tool to the correlating energy consumption. In a case study a rule process, this approach was able to achieve an overall accuracy
based framework was developed to simulate the end milling of within 5%. Only a simple machining operation with a constant
aluminium with the operational states: startup, shutdown, idle, cutting depth was performed. Basic assumptions made in order to
and machining process. refine this simulation to achieve given accuracy stay unclear.
[27] modelled the energy consumption of a turning process in [4] developed a model to estimate the energy demand of the
order to optimize the machining process in regards to tool life and spindle and feed aggregates of a CNC machine tool based on the
energy consumption. The proposed model decomposes the total constant and variable power flows occurring in different oper-
energy consumption into the idle energy, the cutting energy as ations. In a case study aluminium milling with differing
proposed by [13], as well as the energy consumed by the tool parameters was modelled. An error of 3.8–18.1% was achieved
change system and the energy needed to produce the cutting tool. thereby. Only simple operations e.g. pocketing or contour milling
[17] discovered through experiments with a CNC lathe that the were tested. Also does this model require advanced measurements
additional energy losses in machining, which in previous energy in beforehand (masses of axis, idle power).
models were summarized in a constant energy loss, are on the [20] developed an empirical model to characterize the
contrary dependable on spindle speed and cutting torque. In the relationship between energy consumption and process variables
suggested model, energy losses from the electrical machine for different machining processes. In validation tests of the

Fig. 1. Comparison of measurements (a) and predictions (b) of an existing energy model by [2].

716
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

prediction model with eight different CNC machine tools an error [25] proposed an energy model based on the force estimation
of 2.37–8.05% was measured. This was possible by keeping the with finite element modeling simulation. The conducted experi-
parameters like feed rate, depth of cut and thus, material removal ments involving dry turning of a titanium alloy could be predicted
rate constant and neglecting auxiliary processes like the start-up. with an error in the range of 1–8%. The evaluation tests were
[26] trained an artificial neural network (ANN) model with the conducted by machining a fixed length (60mm) with constant
data from 250 high-speed ball end milling operations but did not cutting speeds.
provide evaluation measures for the predictions. Every sample [19] created an model by training an ANN to predict the cutting
consisted of seven inputs (spindle speed, feed rate, feed per tooth, energy while machining carbon steel. The training data was
axial depth of cut, radial depth of cut, tool radius, usage of coolant) obtained from 27 machining operations, with different input
and five power consumption outputs of the aggregates. parameters (spindle speed, feed rate, depth and width of cut). The
[14] analyzed the correlation between NC codes and energy difference between measured and predicted cutting energy was
consumption of single machine components. In the proposed 1.50%.
analytical method the energy needed for every component was [21] developed an analytical predictive model by decoupling
calculated based on the NC instruction and summed up to obtain the energy of the components of the CNC machine tool from the
the energy needed to execute given NC instruction. Considered cutting energy. On validation tests on two different CNC machine
components were the spindle, the feed axes, coolant pump and the tools the model differed from the measurements within an error of
tool change system as well as a constant energy consumption. One 0.36-0.55%. For the model extensive measurements of every
milled and one turned part were produced to evaluate the decoupled state were necessary, what limits the usage to
calculations. Despite the complex model and simple machining predefined operations.
operations, an error of 9.3% occurred. This model also requires [29] decomposed the NC code into single instructions and
process parameters, e.g., depth of cut, that were not included in the proposed a model to predict the energy consumption to further use
NC instruction. those results for online optimization. A polynomial regression
[16] proposed an energy-efficiency monitoring model that was model (PRM) and an ANN were tested with different pocketing,
not dependent on torque or force, but on the energy consumption slotting and drilling operations, where an error of 0.02–1.08%
of the CNC machine tool. In the several tests that were performed occurred. Only material removing NC instructions were taken into
on one of the CNC milling machines, the cutting power was account and the depth of cut was fixed.
predicted with an accuracy of 1.57–3.11%; by keeping the [5] generalizes an energy prediction model over multiple
parameters (such as, spindle seed, feed rate and cutting depth) process parameters and processes in order to optimize tool paths.
constant for each test. The model based on Gaussian process regression (GPR) predicted
[2] included auxiliary operations as start-up and non-cutting the energy consumption within an error of 3.288–5.744%. Only
movements into the proposed NC code based energy model. On a certain operations were studied (face milling, pocketing, plunge,
simple test piece (machining one pocket) the difference between air cut and rapid motion) and the depth of cut was fixed. It was
prediction and test piece was 5.23%. In Fig. 1 the plots of the noted, that the accuracy of the predictions fell if the machining
measurement and the prediction are compared: The power peak parameters of the test part were too different from the training
was neglected by the prediction and the prediction does not fit the data.
actual measurements but generalizes on a large scale (see Table 1). [30] developed a Therblig-based energy model to calculate the
[6] proposed an energy prediction model created with the help energy demand while machining. In a case study the model was
of an ANN. The network with two hidden layers was able to predict 2.23-10.22% off the actual measurements. Besides the simple test
the energy consumption in a not further specified test with an piece (face milling, slotting and drilling), extensive manual work
error of 2.46%. had to be done to assign the corresponding Therbligs to each of the
[18] developed a multi-objective predictive mathematical 215 NC instructions.
energy model for a turning process based on different methods [15] compared the ability of different algorithms to train energy
of data analysis (response surface methodology, grey relational prediction models from milling and grinding operations. It was
analysis and principal component analysis) to optimize machining shown that deep learning algorithms exceed in prediction accuracy
parameters. compared to the Gaussian process regression (GPR).

Table 1
Comparison of the different forecasting models.

Reference Method Machining Op. Aggregates Accuracy

Axes Spindle Tool change Coolant pump Auxiliaries

[9] Rule based Milling



5%
[4] Rule based Milling



3.8–18.1%
[20] Rule based Milling &turning



2.37–8.05%
[26] Artificial Neural Network Milling

N/A
[14] Rule based Milling &turning



9.3%
[16] Rule based Milling



1.57–3.11%
[2] Rule based Milling




5.23%
[6] Artificial Neural Network Milling


2.46%
[18] Rule based Milling

4.79%
[25] Finite Element Method Turning


1–8%
[19] Artificial Neural Network Milling


1.50%
[21] Rule based Milling



0.36–0.55%
[29] ANN &Polynomial regression Milling



0.02–1.08%
[5] Gaussian process regression Milling



3.288–5.744%
[30] Rule based Milling




2.23–10.22%
[15] Deep Learning Milling &grinding




N/A

717
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

The shown literature study in the field of energy prediction of 5 The model then predicts the energy consumption.
CNC machine tools do not fully cover the topic with respect to the 6 The actual measurements obtained while machining the part
motivation of this paper. The missing links are elaborated in the are compared to the predicted values.
following chapter and an approach is developed.
Besides the restrictions met, the state of the art shown above
Research gap takes into account multiple parameters for the energy consump-
tion (e.g. material removal rate, spindle speed, feed rate, effective
Derived from the previous mentioned state of the art, the cutting, air cutting, idle, start-up and shut down, etc.). Therefore,
research gap can be summarized as follow: measures were taken to enable the Machine Learning model to
Previous research papers investigated the energy consumption distinguish between those operations.
of a CNC machine tool in comparison to simulated data like depth
of cut and production strategy. Furthermore, a reference part to Methodology and challenges
parameterize energy consumption models had to be produced.
Towards data acquisition, low-frequency data with a sampling In this chapter we describe each step individually. We elaborate
interval of 200ms or more were collected via the standard OPC-UA on how the study is carried out and the challenges occurring.
interface. In the evaluation state outliers in the gathered data were
excluded due to the negligible influence of single power spikes on Data acquisition
the overall energy consumption. [33] evaluated that besides
significant effort made to understand the complex interactions The data from the machining process is the foundation for
causing energy consumption in machining, the validity of developing a prediction model. In this chapter the procedure is
analytical energy models remain highly questionable. elaborated to derive a data set for the Machine Learning model out
Mitigating these disparities, our approach uses only direct of the obtained data from the experiments [35].
accessible data from CNC machine tool and neither simulation data
nor rule-based energy models are included, which is the origin of this Manufacturing process
paper. Furthermore, no special reference part for energy consumption Part For our approach two different parts, shown in Fig. 3, are
model parameterization is needed. Gathering high-frequency data machined. The raw material is an aluminium alloy (AlCuMgPb, ISO
with sampling interval of 2ms is done via a innovative SIEMENS Edge No. 1645) with the size 125.3mm  19.34mm  14.52mm.
Device. Outliers in the collected data set are included. Our approach is CNC machine tool The machining parts are machined on a
to investigate three different machine learning models, namely SPINNER U5630 5-axis simultaneous CNC machine tool with a
’DecisionTree’, ’RandomForest’ and boosted ’RandomForest’, for high SINUMERIK 840D SL v4.8 numerical control unit (NCU).
precision energy prediction of CNC machining strategies. Tools The tools being used in the machining process are given in
Table 2. While four tools are used in both manufacturing processes,
Approach of study the other seven are used in either of the two. Hence, the validation
data set contains tools which are not in the training set.
As previously shown, numerous parameters on the energy Data acquisition The power measurements are not performed
consumption of a machining process exist. In this study the with a conventional electricity meter but with a SIEMENS edge
parameters are limited to one CNC machine tool, one material and device. This high frequency software sensor installed onto the NCU
dry machining. After those preliminary restrictions the approach provides high resolution data with a sampling interval Dt of 2ms.
of study is depicted in Fig. 2: This has be shown by [32] to be beneficial compared to
conventional low frequency data. The NC code used in this study
1 First, a training part is machined and high frequency measure- is acquired in a retroactive readout process from the NCU. This NC
ments and the NC instructions are acquired. code is extracted after the machining and therefore, compared to
2 Based on this data, a Machine Learning model is trained. the output of a CAM software, contains additional internal machine
3 Post this, the validation part is machined. instructions and possible changes of the machine operator,
4 The NC instructions from the validation part are passed to the ensuring the usage of the accurate NC code responsible for the
already trained model. measurements [28].

Fig. 2. Approach of study. Fig. 3. Manufactured parts used for the study.

718
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

Table 2 ”GETSELT”, etc.). All those instructions that do not lead to an

Tools used to machine the parts.
significant energy consumption higher than the idle power
Tool Diameter and angle Part consumption can be identified and summarized into a negligible
Training Validation
category.
Ball Mill ;4 mm

Center Drill ;8 mm &90


Multiple elements in one instruction
Chamfer Mill ;8 mm &60
One NC instruction can contain more than one element,
Chamfer Mill ;10 mm &90
responsible for different, simultaneous CNC machine tool oper-
End Mill ;3 mm

ations. However, the energy consumption measurement for the NC
End Mill ;6 mm

instruction is conducted as a whole and retrospective separation
End Mill ;10 mm


End Mill ;16 mm
and allocation to the respective CNC machine tool aggregates is not
Reamer ;5 mm
possible.
Twist Drill ;2.8 mm

Twist Drill ;4.7 mm


Feature extraction and feature engineering

The input parameters, called features, are generated from the

data set with respect to their relevance for the machining process
Merge measurements and NC code according to the analytical and technological considerations. A part
The NC code, containing all the NC instructions to machine a of the needed information can be directly used from the data set in
given part, and the power demand measurements are matched via the given form (e.g. the spindle speed). In other cases, the
a time stamp. Dependent on the execution length of each information needs to be transformed prior the be ready for the
individual NC instruction i, m consecutive measurement results j Machine Learning model. In the feature engineering process,
have to be assigned to the NC instruction. To merge those two data domain knowledge or additional information, like the diameters of
sets, every power demand measurement Pi,j is multiplied by the the tools, is added and mathematical transformations, like the path
sampling interval of the measurement device Dt to obtain the between two positions, are implemented on the information prior
current energy consumption. The total energy consumption Etotal is to transferring it to the data set for the Machine Learning model.
then: The model is trained and validated with 37 input features, of which
a selection is deeper discussed hereafter.
n X
X m
Etotal ¼ Pi;j Dt ð1Þ
i¼1 j¼1 Length of tool path
As the literature above suggests, one main parameter of the
Syntax interpretation cutting power consumption is the material removal rate Q ¼ vc A.
Multiplied with the specific cutting force kc one obtains the power
In this step of the study, the NC code has to be processed to demand Pc = kcQ. Given those prerequisites, one can calculate the
make the hidden information accessible for further usage, e.g. the energy consumption of a milling process:
category of operation (rapid positioning, material removal, As the run time of one NC instruction Dti is not directly
miscellaneous function, etc.) or the go-to positions, in the text utilizable out of the data, one has to substitute it by the feedrate vc
of the unprocessed NC instructions, see Fig. 4. For this data and the length of the tool path lp. With further simplifications one
transformation task domain knowledge is necessary in order to finally obtains:
improve the data quality set containing information, that can be lp
used later by a ML algorithm. Following chapters are elaborating Ec ¼ kc Q 
vc ð2Þ
this issue. ¼ kc Alp

Negligible code lines where, kc is a material constant that can be neglected in this
A part of the retained machine instructions don’t contain regression study as only one material (aluminium alloy) is used. It
information in regard to the energy consumption (e.g. ”STOPRE”, is assumed that the cross section A is also constant for a given
operation. In Eq. (2), it is shown that the cutting lengths for every
feed axis ~
l is an important feature in the data set, which can be
i
calculated by the difference of the current position of the tool ~
xi
and the position ~
xiþ1 after execution of the NC instruction i:
~
li ¼ D~
xi ¼ ~
xi  ~
xiþ1 ð3Þ

The cutting lengths ~ li are calculated for every feed axis

separately and not with the amount of the vector j~ xi  ~
xiþ1 j. The
power consumption is measured and later predicted for every axis
separately, therefore also a separate cutting length for every axis is
needed.

Feed rate determination

The feed rate is included in the data set as variable parameters
and set for each tool and each operation mode. Hence, an external
database with those settings has to be build. Therefore, based on
the tool and operation mode for each NC instruction, the correct
Fig. 4. Syntax interpretation of the NC code. feed rate can then be assigned (see Fig. 5).

719
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

’M’ codes in the NC instruction. E.g. the command ’G01’ is not less
then ’G02’ but defines a different operating method of the machine,
therefore those variables cannot be handled in the same way as
continuous variables like e.g. the tool diameter discussed earlier.
Those variables are turned into an one-hot encoded feature matrix,
with every possible operating state as a separate feature. Than, for
every NC command the activated operating states are set to ’1’
while the others remain deactivated with a value of ’0’.

Fig. 5. Determination of feed rate through external database.

Model training and prediction

Internal machine variables Three different models have been trained using the 'scikit-learn’
Positions in the NC code can be given not by their numerical package: ’RandomForest’, ’DecisionTree’ and ’DecisionTree’ includ-
coordinates but by a variable name. While variables like X_HOME, ing the ’AdaBoost’ method. As the later does not support
Y_HOME, etc. can be replaces by their numerical value, others multioutput predictions, for every output a separate model was
remain unknown. trained. The models have been trained with all of the training data,
gathered during machining of the training part, while the data
Material engaging and effective machining obtained during machining of the validation part was used to
As the tool engages into the part, the material removal rate Q is evaluate the predictions of the three algorithms.
not constant. During engaging the material removal rate increases
with the length of the tool path lrun, from the beginning until Results and discussion
attaining a constant material removal rate at a specific length
lthreshold. The statistical measures, specifying the quality of the pre-
 dictions, are given in Table 3 for all five considered CNC machine
variable; if lrun < lthreshold : tool aggregates (x, y and z axis, spindle and tool change system) and
Q¼ ð4Þ
constant; otherwise: for the three trained algorithms. The total deviation of the energy
predictions from the measurements for a real part geometry
An additional feature marks all those NC instructions that meet
including auxiliary operations like tool change system and start up
the condition lrun < lthreshold. For the Machine Learning model those
phase is 7.16% for the DecisionTree model. The comparison
instructions will be recognizable as a group of instructions that
between the algorithms for one of those aggregates (y axis) is
take place at a beginning of every removal process. The value of
depicted in Fig. 6. Although there are exceptions for specific
lthreshold is retained by minimizing the error of the predictions for
NC instructions of the engaging process.

Tool diameter
The removed area A correlates to the tool diameter dtool as
bigger tools will be used to remove more material at a time. As the
removed area A correlates to the material removal rate Q and in
consequence to the energy consumption.
Ec ¼ kc f ðdtool Þlp ð5Þ

If material and tool path length lp are constant, the energy

consumption of a bigger tool is higher than the energy consump-
tion of a known tool.

Categorical variables
Of large importance for the energy consumption is the Fig. 6. Comparison of the prediction quality for the y-axis of part B of different
operating state of the machine. This is defined by the ’G’ and Machine Learning algorithms.

Table 3
Validation of the three investigated machine learning algorithms with the test data set.

Algorithm Measure x axis y axis z axis Spindle Tool change system

’RandomForest’ Total deviation (%) 0.2 2.56 3.29 11.66 46.14
Mean deviation (Ws) 5.02 1.62 3.1 199.62 2.38
RMSE 48.45 11.73 32.95 1301.92 18.03
Explained variance 0.72 0.54 0.9 0.42 0.52

’DecisionTree’ Total deviation (%) 9.12 18.47 3.96 8.9 11.75

Mean deviation (Ws) 7.44 3.77 5.69 258.76 3.21
RMSE 58.77 13.96 54.5 1368.99 17.72
Explained variance 0.59 0.35 0.73 0.36 0.54

’AdaBoost’ + ’DecisionTree’ Total deviation (%) 14.25 29.18 10.71 13.48 15.19
Mean deviation (Ws) 5.61 3.24 3.63 277.71 2.91
RMSE 51.98 12.96 53.1 1330.77 19.33
Explained variance 0.68 0.45 0.74 0.4 0.45

720
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

Fig. 7. Comparison of the prediction quality for the y-axis for the whole machining cycle.

measures and aggregates, the most accurate predictions can be better performed by the ’DecisionTree’ than by a model, that was
achieved with ’RandomForest’ algorithm. In Fig. 7, the plot of the build using the ’bagging’ method (e.g. ’RandomForest’). In ’bagging’,
predictions of the energy consumption of the y axis aggregate is the data set is split into smaller sub data sets and those are used to
compared to the measurements. Hence, the energy demand curve train sub models, which leads to an averaging effect or a smoothing of
of the machining process gets predicted accurately. Also the peaks, the regression function. For singular operations as the tool change
although differing in their value, are correctly predicted by the operation this leads to disadvantageous predictions caused by
trained model. underfitting in the single sub data sets.

Tool change system Spindle

For the tool change system the ’DecisionTree’ algorithm The errors for the predictions of the energy consumption of the
achieves the lowest total deviation compared to the ’Random- spindle (see Fig. 8) are higher than for the remaining aggregates.
Forest’. This is caused by three main issues:
This is caused by performing the tool change operation only
several times during machining. Besides those singular events the
The acceleration and the deceleration of the spindle are
tool change system is turned off. In contrast to the other aggregates, responsible for exceptional high power peaks. This leads to a
where a regression problem is to be solved, in this case an algorithm high range of current energy consumption between those rare
has to distinguish between the NC instructions, where a tool change peaks and effective CNC machine tool operation. The algorithm
operation takes place, and those, where not. This decision problem is averaging those, leads to a high error.

Fig. 8. Comparison of the prediction quality for the spindle aggregate for the whole machining cycle.

721
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

Those power peaks occur rather rarely compared to effective Computational scope
machining. This leads to an underfitting of the models of this
power peaks. This problem of unbalanced data influences also Algorithms
other domains with power peaks, like rapid movement of the Similar studies often used artificial neural networks (ANN). It
axis. should be examined if more accurate predictions are possible,

Acceleration and the deceleration are not only energy intense although results will not be easily explainable, because of the
processes but also time consuming. The NCU moves ahead after hidden layers and the multiple connections between the neurons
initiating the spindle deceleration while the spindle is still of such a network.
decelerating and acts like a generator, thereby influencing the
energy consumption of the following NC instructions. Data processing
It poses a challenge that NC instructions can contain more than
one element responsible for energy consumption of the CNC
Peak consumption machine tool. The data could be processed in a way that those
elements will be separated and individual predictions for them will
The deviation of the predictions for the peaks in Fig. 7 are posed be computed. Another possibility in this domain is, to integrate a
on several reasons concerning especially the energy consumption duration component to the NC instructions, as the run time was
of CNC machine tools and the milling process: neglected in this study.

There are NC instructions, which are directly followed by an Data acquisition

identical copy of themselves, e.g. ’G0 G40 G60 G90 Z= In this study the instructions used were extracted from the CNC
$TC_CARR40[_TC1]-_TOOLL[2]*_FAK1’ or ’G90 D0 SPOS=0 POS machine tool after the machining. As this contains additional
[X]=_X FA[X]=RED_RAPID_SPEED POS[Y]=_Y FA[Y]=RED_RA- machine specific instructions, in a further study the input NC code
PID_SPEED POS[Z]=_Z FA[Z]’, whereas only one of those could be used.
instructions leads to an significant energy consumption of the
machining operation. This is a challenge in data acquisition, Expansion
which leads to high prediction errors: Having two identical
inputs generating highly different energy consumption, the Energy aware design
model has to average the output and produce high errors. Connecting this model with CAD-Systems would give the

Despite being similar, NC instructions can differ in energy design engineers feedback about energy-efficient design of their
consumption. For example the energy needed for the tool change parts.
system is dependent on the position in the tool magazine. This
information is not available in the data set and therefore it can Online optimization of parameters
not be considered. In CNC machine tools the parameters like feed rate and spindle
speed are usually fixed according to experience values. This model
with its high prediction quality of the energy consumption time
Outlook series could enable the energy prediction during the machining
process and an online optimization of the cutting parameters in
Besides of addressing the challenges discussed above, several respect to a constant load of the CNC machine tool.
opportunities for future research can be divided into two The results obtained in this study with ML provide accurate
categories: predictions of the power consumption for the whole machining
cycle. Those results prove that the developed energy model can be
Physical scope used to predict the energy consumption and is a valid tool to help
reduce cost in machining and to give direct feedback to part
Materials designers before machining as proposed in the introduction.
In this study, only one material is used. Therefore, training and
prediction is based on the same raw stock. Research question arises Declaration of interests
to create a model for multiple materials. A feature taken this into
account is e.g. the specific cutting force kc. It has to be investigated None.
to what extend it is possible to calculate accurate predictions if:
Authors’ contribution

A common model will be trained based on training data
containing different materials, Markus Brillinger: Conceptualization, Methodology, Writing –

Predictions will be made for materials, which were not in the Reviewing and Editing. Marcel Wuwer: Investigation, Writing –
training data set. Original draft preparation. Muaaz Abdul Hadi: Resources. Franz
Haas: Supervision.

CNC machine tools Acknowledgements

The data used in this research represents machining processes
on one specific CNC machine tool. Further studies should This work has partially been supported by the FFG, Contract No.
investigate the possibility to integrate multiple CNC machine tools. 854184; Pro2Future GmbH is funded within the Austrian COMET
Program Competence Centers for Excellent Technologies under the
Tools auspices of the Austrian Federal Ministry of Transport, Innovation
Linking data with simple geometries of the tools (e.g. tool and Technology, the Austrian Federal Ministry for Digital and
diameters), was found to deliver highly accurate results. Hence, Economic Affairs and of the Provinces of Upper Austria and Syyria.
including more information e.g. if the tool is used for face, chamfer, COMET is managed by the Austrian Research Promotion Agency
etc. milling might be advantageous. FFG.

722
M. Brillinger et al. CIRP Journal of Manufacturing Science and Technology 35 (2021) 715–723

References [30] Sihag, N., Sangwan, K.S., 2019, An improved micro analysis-based energy
consumption and carbon emissions modeling approach for a milling center.
[1] AlGeddawy, T., ElMaraghy, H., 2016, Design for energy sustainability in Int J Adv Manuf Technol, 104:705–721.
manufacturing systems. CIRP Ann, 65:409–412. http://dx.doi.org/10.1016/j. [31] Stich, V., Hering, N., Starick, C.P., Brandenburg, U., 2013, Energy-efficiency
cirp.2016.04.023. concept for the manufacturing industry. in Prabhu VV., Taisch MM., Kiritsis
[2] Aramcharoen, A., Mativenga, P.T., 2014, Critical factors in energy demand DD., (Eds.) Advances in production management systems. Sustainable produc-
modelling for CNC milling and impact of toolpath strategy. J Clean Prod, tion and service supply chains. Springer Berlin Heidelberg, Berlin, Heidelberg,
78:63–74. pp. pp.86–93.
[3] Arrazola, P., Özel, T., Umbrello, D., Davies, M., Jawahir, I., 2013, Recent advances [32] Trabesinger, S., Butzerin, A., Schall, D., Pichler, R., 2020, Analysis of high frequency
in modelling of metal machining processes. CIRP Ann, 62:695–718. data of a machine tool via edge computing. Proc Manuf, 45:343–348.
[4] Avram, O.I., Xirouchakis, P., 2011, Evaluating the use phase energy require- [33] van Luttervelt, C., Childs, T., Jawahir, I., Klocke, F., Venuvinod, P., Altintas, Y.,
ments of a machine tool system. J Clean Prod, 19:699–711. Armarego, E., Dornfeld, D., Grabec, I., Leopold, J., Lindstrom, B., Lucca, D.,
[5] Bhinge, R., Park, J., Law, K.H., Dornfeld, D.A., Helu, M., Rachuri, S., 2017, Toward a Obikawa, T., Shirakashi. Sato, H., 1998, Present situation and future trends in
generalized energy prediction model for machine tools. J Manuf Sci Eng, 139. modelling of machining operations progress report of the CIRP working group
[6] Borgia, S., Pellegrinelli, S., Bianchi, G., Leonesio, M., 2014, A reduced model for ‘modelling of machining operations’. CIRP Ann, 47:587–626. http://dx.doi.org/
energy consumption analysis in milling. Proc CIRP, 17:529–534. 10.1016/S0007-8506(07)63244-2.
[7] Cai, W., Liu, C., hung Lai, K., Li, L., Cunha, J., Hu, L., 2019, Energy performance [34] Vijayaraghavan, A., Dornfeld, D., 2010, Automated energy monitoring of ma-
certification in mechanical manufacturing industry: a review and analysis. chine tools. CIRP Ann, 59:21–24.
Energy Convers Manage, 186:415–432. http://dx.doi.org/10.1016/j.encon- [35] Wuwer, M., Brillinger, M., 2021, Energy consumption for a CNC machining
man.2019.02.041. process. Mendeley Data. http://dx.doi.org/10.17632/7cghj7fffp.1.
[8] Devoldere, T., Dewulf, W., Deprez, W., Willems, B., Duflou, J.R., 2007, Improve- [36] Xu, M., Jerard, R.B., Fussell, B.K., 2007, Energy based cutting force model
ment potential for energy consumption in discrete part production machines. calibration for milling. Comput Aided Des Appl, 4:341–351.
Advances in life cycle engineering for sustainable manufacturing businesses, [37] Zhou, Z., Yao, B., Xu, W., Wang, L., 2017, Condition monitoring towards energy-
Springer: 311–316. efficient manufacturing: a review. Int J Adv Manuf Technol, 91:3395–3415. doi:
[9] Dietmair, A., Verl, A., 2008, Energy consumption modeling and optimization 10.1007/s00170-017-0014-x..
for production machines. 2008 IEEE International conference on sustainable
energy technologies IEEE, 574–579. Markus Brillinger leads the research group Cognitive
[10] Dietmair, A., Verl, A., 2009, A generic energy consumption model for decision Production Systems at Pro2Future, an Austrian Re-
making and energy efficiency optimisation in manufacturing. Int J Sustainable search Center investigating digitalization in future pro-
Eng, 2:123–133. duction systems. His research interests include
[11] Dietmair, A., Verl, A., Eberspaecher, P., 2011, Model-based energy consumption production engineering, cognitive production, 3D-
optimisation in manufacturing system and machine control. IJMR, 6:380–401. printing and sustainable as well as energy aware pro-
http://dx.doi.org/10.1504/IJMR.2011.043238. duction systems. His research has won numerous
[12] Garg, A., Lam, J.S.L., Gao, L., 2016, Power consumption and tool life models for awards and prizes. As educating future engineers is
the production process. J Clean Prod, 131:754–764. one of his passion, he is a lecturer at two universities.
[13] Gutowski, T., Dahmus, J., Thiriez, A., 2006, Electrical energy requirements for Furthermore, he is a member of different industrial
manufacturing processes. 13th CIRP International conference on life cycle networks supporting the application of research in
engineering (Leuven, Belgium), pp.623–638. industry. Before joining Pro2Future GmbH Markus
[14] He, Y., Liu, F., Wu, T., Zhong, F., Peng, B., 2012, Analysis and estimation of energy was the CEO of a consulting company in production
consumption for numerical control machining. Proc Inst Mech Eng B, 226:255– optimization and successfully completed a PhD in pro-
326. duction engineering, in which he developed a funda-
[15] He, Y., Wu, P., Li, Y., Wang, Y., Tao, F., Wang, Y., 2020, A generic energy prediction mental new 3D-printing technology, which has
model of machine tools using deep learning algorithms. Appl Energy, disruptive potential in this field of technology.
275:115402.
[16] Hu, S., Liu, F., He, Y., Hu, T., 2012, An on-line approach for energy efficiency
monitoring of machine tools. J Clean Prod, 27:133–140.
[17] Hu, S., Liu, F., He, Y., Peng, B., 2010, Characteristics of additional load losses of Marcel Wuwer is member of the research group Cog-
spindle system of machine tools. J Adv Mech Des Syst Manuf, 4:1221–1233. nitive Production Systems at Pro2Future GmbH. His
[18] Kant, G., Sangwan, K.S., 2014, Prediction and optimization of machining research interest lies in production engineering, auto-
parameters for minimizing power consumption and surface roughness in mation and digitalization of production. Besides his
machining. J Clean Prod, 83:151–164. research activities he is working as a design engineer
[19] Kant, G., Sangwan, K.S., 2015, Predictive modelling for energy consumption in in plant construction. The overall goal of his research
machining using artificial neural network. Proc CIRP, 37:205–210. contributes to foster sustainability and circularity of
[20] Kara, S., Li, W., 2011, Unit process energy consumption models for material production systems.
removal processes. CIRP Ann, 60:37–40.
[21] Lee, J.Y., Shin, Y.J., Kim, M.S., Kim, E.S., Yoon, H.S., Kim, S.Y., Yoon, Y.C., Ahn, S.H.,
Min, S., 2016, A simplified machine-tool power-consumption measurement
procedure and methodology for estimating total energy consumption. J Manuf
Sci Eng, 138. Muaaz Abdul Hadi is a member of the research group,
[22] Nguyen, T.T., 2019, Prediction and optimization of machining energy, surface Cognitive Production Systems, at Pro2Future GmbH
roughness, and production rate in skd61 milling. Measurement, 136:525–544. and a doctoral student at TU Graz, Austria. His research
[23] Papetti, A., Menghi, R., Di Domizio, G., Germani, M., Marconi, M., 2019, interest and direction lie in e-mobility, digitalization of
Resources value mapping: a method to assess the resource efficiency of production systems, cognitive production, and energy-
manufacturing systems. Appl Energy, 249:326–342. http://dx.doi.org/ aware production systems. His publications in the field
10.1016/j.apenergy.2019.04.158. of e-mobility have won numerous accolades at confer-
[24] Pechmann, A., Shrouf, F., Chonin, M., Steenhusen, N., 2017, Load-shifting potential at ences. During his master's study at TU Graz, he has
SMES manufacturing sites: a methodology and case study. Renewable Sustainable traveled to various locations, including a semester
Energy Rev, 78:431–438. http://dx.doi.org/10.1016/j.rser.2017.04.081. abroad at The University of Kansas, USA, as an exchange
[25] Pervaiz, S., Deiab, I., Rashid, A., Nicolescu, M., 2015, Prediction of energy and research student. His research direction paves the
consumption and environmental implications for turning operation using way to future sustainable production systems.
finite element analysis. Proc Inst Mech Eng B, 229:1925–2193.
[26] Quintana, G., Ciurana, J., Ribatallada, J., 2011, Modelling power consumption in
ball-end milling operations. Mater Manuf Process, 26:746–756.
[27] Rajemi, M., Mativenga, P., Aramcharoen, A., 2010, Sustainable machining: Franz Haas is head of the institute of production engi-
selection of optimum turning conditions based on minimum energy consid- neering at Graz University of Technology and key re-
erations. J Clean Prod, 18:1059–1065. searcher at Pro2Future GmbH. His research interest lies
[28] Schmid, J., Schmid, A., Pichler, R., Haas, F., 2020, Validation of machining in production, especially in grinding processes, additive
operations by a virtual numerical controller kernel based simulation.1478– manufacturing and development of smart and digital
1483. http://dx.doi.org/10.1016/j.procir.2020.03.094. production networks with respect to sustainability and
[29] Shin, S.J., Woo, J., Rachuri, S., 2017, Energy efficiency of milling machining: circularity. Moreover, he is member of many scientific
component modeling and online optimization of cutting parameters. J Clean advisory boards in industries and academia.
Prod, 161:12–29.

723