Available online at www.sciencedirect.

com

ScienceDirect
Procedia CIRP 113 (2022) 582–587
www.elsevier.com/locate/procedia

ISEM XXI

Software-based setpoint optimization methods

for laser cutting machine tools
Raffael Amachera , Natanael Lanzb , Manuel Müllerc,∗, Ronny Bossartc
a Embotech AG, 8005 Zürich, CH
b ETH Zürich, 8092 Zürich, CH
c Bystronic Laser AG, 3362 Niederönz, CH

Abstract
The demands on laser cutting machine tools are increasing constantly. In this paper, we focus on increasing productivity while minimizing the
deterioration of part quality. Without specific countermeasures, increasing the dynamic settings leads to larger dynamic-induced contour errors
of the workpiece. Software-based methods offer potential to improve the contour tracking under varying dynamic settings without the need for
a mechanical redesign of the machine tool. Four such methods, which rely on different combinations of model- and/or sensor-based setpoint
compensation, are implemented and tested on a laser cutting machine tool. A comparison with respect to productivity and contour accuracy of cut
parts is presented, and the complexity of development as well as deployment for production are discussed.
© 2022 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 ISEM XXI

Keywords: contour accuracy; setpoint optimization; laser cutting

reachable tracking accuracy, showing an inverse proportional

1. Introduction dependence. In order to overcome the above limitations, either
the machine design would need to be adopted, or the setpoint
In high dynamic manufacturing in general and more specif- inputs would need to be reshaped/optimized to compensate for
ically for laser cutting machines a fundamental trade-off be- some of the induced dynamic deviations caused by an increase
tween achievable motion accuracy and productivity exists. The in dynamics and productivity. An improvement in the structural
productivity of a laser cutting machine is limited by the jerk, machine design is related to high costs, especially if it involves
acceleration and velocity limits of all axes, which must be sat- the manufacturing of new prototypes [5]. Recent developments
isfied by the setpoints supplied to the machine axis motion con- in simulation, leading to virtual machine prototypes [6] never-
troller. The reachable motion accuracy accordingly depends on theless enabled the reduction of the costs for structural machine
the mechatronic machine response to these setpoint inputs. The design optimization.
machine response itself is influenced by the structural machine Therefore, this work is focussing on the optimization of the
design and the motion control architecture and parametrization. machine accuracy by adopting the setpoint values, leaving ma-
Generally, if the machine productivity is increased by adopt- chine design and controller parameterization untouched. In lit-
ing the setpoints, the machine will respond with less accu- erature, different approaches for algorithms are described for
racy due to an increase of the tracking error, dynamic devi- this purpose involving either modelling or measurements of
ations caused by induced quasi-static deformations (Steinlin the machine response to setpoint inputs. Most important ap-
[1]) and vibrations of the machine structure. For a properly pa- proaches are based on (iterative) learning control (ILC) algo-
rameterised standard cascaded motion controller according to rithms, first introduced by Garden [7], that try to compensate
Batzies et al. [2] and Zirn [3], Spescha [4] derived a formula for for simulated or measured deviations by adopting the setpoints
the relation between acceleration or jerk setpoint limits and the accordingly. A good survey over the available algorithms is pre-
sented by Britsow [8]. They might rely on direct proportional
∗ Corresponding author. E-mail address: [email protected] feedback of the simulated or measured machine trajectory as

2212-8271 © 2022 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 ISEM XXI
10.1016/j.procir.2022.09.178

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Rafael Amacher et al. / Procedia CIRP 113 (2022) 582–587 583

for PID-ILC [9]. These algorithms were also applied to ma-

chine tools as presented by Kim [10] and Haas [11]. Or the
ILC algorithms might alternatively rely on solving an optimiza-
tion problem, by minimizing a definable cost function with the
help of a model of the machine response e.g. by Togai [12] and
Haas [11]. As all of the ILC algorithms use previously gathered
simulation or measurement information, they are not applica-
ble online. Therefore they fit especially well to applications,
where the same trajectories have to be followed repeatedly, as
e.g. the case for high volume production of identical parts. In
this paper, however, an optimization based learning algorithm is
presented, by which the time required for the learning process
can be reduced by using a very fast solver (Forces Pro [14]). By
following a trajectory with the laser switched off, the required
corrected setpoint values can be calculated in a short time and
without material loss, which makes the algorithm economically Fig. 1. Reference part used for measurements. The outer diameter is 120 mm.
applicable even for parts with less repetitions than before.
The required measurements for the learning algorithms may
further be derived from different sources. In the presented work, error. The cutting was always carried out in the center of the
the same optimization based learning algorithm is applied to machine to exclude the variation over the cutting area.
different measurement sources, which are: simulation results, The contour deviations and statistical distribution of the er-
internal encoder measurements, accelerometer data close to the ror were determined from the data of the optical measurement.
tool center point (TCP) and external optical measurements of Based on this analysis, the methods were tuned accordingly. It
the parts, where for the accelerometer motion estimation the was found that the median error is most suitable for compari-
algorithm described by Lanz [13] is applied. Divided into four son since it excludes any outliers that may occur from the laser
methods according to the measurement or simulation source, cutting process (e.g. splashes or unclean cuts).
this work implements and tests the optimization algorithm for
the first time on an industrial laser cutting machine.

2.3. Setpoint compensation

2. Methods
Figure 2 shows the architecture of the setpoint compensa-
2.1. Test machine tion used in this paper. The reference trajectory for the X- and
Y-axes of the machine, which is generated by the CNC-Kernel,
The work was carried out on a ByStar Fiber 3015 laser cut- is compensated based on an estimate of the TCP position. The
ting machine. Initially, the machine was recalibrated with the goal of the compensation algorithm is to reduce the expected
double ball bar method to ensure good calibration including contour deviation while respecting velocity and acceleration
scaling, backlash and squareness compensation. The machine constraints of the machine axes. The compensation algorithm
features an acceleration sensor located in the cutting head, i.e. is implemented using FORCES PRO [14]. The trajectory time
close to the Tool Center Point (TCP). During cutting of the ref- is left unaltered by the compensator, such that a decrease of ma-
erence part, the drive and laser process data were recorded. Af- chining time due to an increased acceleration limit is preserved
ter performing setpoint compensation, the data was applied to after compensation. For the TCP estimation step, four differ-
the machine. In this way, a part can be cut based on the setpoints ent methods were implemented that rely either on an identified
generated with different compensation methods while maintain- machine model, measured data, or a combination of both:
ing the laser process data.
• Model-based (Method 1): Compensation based on a cal-
ibrated machine model
2.2. Test procedure • Machine feedback-based (Method 2 light): Compensa-
tion based on measured encoder positions
The measurements of the cut parts were carried out on a 2D • Machine feedback-based (Method 2): Compensation
optical measuring device. The measured part contour was fitted based on acceleration sensor at the TCP
to the nominal contour using Gauss fitting which identifies the • External feedback-based (Method 3): Compensation
perpendicular deviations of the geometry, so-called “spikes”. In based on the measured part contour
the following the absolute value of these spikes are referred to
as the contour error of the part.
The used reference part is shown in Figure 1 which was cut 2.3.1. Model-based compensation (Method 1)
with a 1mm thick metal sheet. The slender contour with fast For each axis, a linear time-invariant dynamical black-box
direction changes poses a challenge to minimizing the contour model is used that predicts the TCP position given the refer-

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 584 Rafael Amacher et al. / Procedia CIRP 113 (2022) 582–587

estimation algorithm were identified based on optical measure-

CNC kernel
ments, similar to the model identification that was necessary for
Reference trajectory method 1.

TCP estimator
2.3.3. External feedback-based compensation (Method 3)
Measured encoder positions,

This method relies on using the measured contour of one

acceleration sensor data

Machine model or more cut parts. When cutting a part and measuring the re-

Reference contour (DXF)

Compensator sulting contour in high resolution, that reference data can be
mapped back to the reference trajectory of the CNC kernel to
Compensated reference
Measured cut part contour

get an accurate estimate of the deviation of the TCP from the

trajectory
reference trajectory. For this method, a part needs to be cut, its
Machine contour needs to be measured with an external measurement
system, and the resulting measurement data needs to be trans-
ferred back to the machine software ecosystem. It can then be
Cut part

used to compensate the reference trajectory. For further stabi-

lizing the method, always two measurements from two different
External cuts were used and averaged the estimated TCP position. The
drawback of this method is that impurities, or process-induced
measurement contour errors that are not caused by the machine’s position
system control, or errors in the optical measurement system can lead to
faulty compensations. The signal was therefore post-processed
to identify sections where one of these effects is suspected and
Fig. 2. The architecture of the setpoint compensation. compensation was switched off in these sections. Further, in the
lead-in region, there is a small section where it cannot be differ-
entiated whether contour deviations are caused at the beginning
ence trajectory from the CNC kernel as an input in an open- or at the end of the cut, and therefore, compensation was also
loop fashion. The model was identified by comparing the refer- switched off in these regions.
ence trajectory with optical contour measurements. Using this
method, an estimate of the TCP trajectory can be generated us-
2.4. Measurements
ing only the reference trajectory as an input, i.e., without any
feedback signal once the model is identified.
Table 1 lists all cuts that were made for this paper. Optimized
calibration “No” means that the machine’s calibration was in
2.3.2. Machine feedback-based compensation (Method 2 / factory settings (determined using double ball bar method),
Method 2 light) “Yes” means that it was specifically tuned for the part at hand.
These methods rely on machine feedback. The part of in- Compensation “None” means that the reference values from
terest is ”dry-cut” (laser switched off) once, and the encoder the CNC-Kernel were directly passed to the drives, in the other
positions of the X- and Y-axes, as well as the acceleration mea- cases, compensation was applied offline and the result was fed
sured close to the TCP are recorded. This information is used to the drive, bypassing the CNC-Kernel.
for computing an estimate of the TCP position. Two variants of
this compensation were implemented. Table 1. Measurements presented and discussed in this paper are summarized.
Method 2 light relies on the encoder feedback for the X- and
ID Optimized Calibr. Accel. Compensation
Y-axis only. The encoder data is directly used as an estimate for
the TCP position without further manipulation or model-based M1 No 4/10 None
manipulation. Effects that further influence the TCP position M2 No 4/10 Method 1 based on M1
relative to the encoder position due to e.g. the machine tool M3 No 4/10 Method 2 light based on M1
M4 No 4/10 Method 3 based on M1
mechanical behaviour are neglected. This method therefore fo- M5 Yes 4/10 None
cuses on reducing the residual errors from the position control M6 Yes 4/10 Method 1 based on M5
system. The advantage of this approach is that the method is M7 Yes 4/10 Method 2 based on M5
completely calibration free. M8 Yes 4/10 Method 2 light based on M5
Method 2 relies on the acceleration sensor data, which is M9 Yes 4/10 Method 3 based on M5
M10 Yes 9/10 None
physically closer to the TCP than the encoder measurements. M11 Yes 9/10 Method 1 based on M10
Therefore, it can potentially provide a more accurate TCP esti- M12 Yes 9/10 Method 2 based on M10
mate, as also e.g. mechanical vibrations of the machine tool can M13 Yes 9/10 Method 2 light based on M10
be taken into account. However, an estimation algorithm needs M14 Yes 9/10 Method 3 based on M10
to be parametrized and evaluated for converting the accelera-
tion data into a TCP position estimate. The parameters for this
3

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Rafael Amacher et al. / Procedia CIRP 113 (2022) 582–587 585

3. Results and Discussion mated path from the feedback data of the measurement system
(Method 2 light), an equally good, if not slightly better, correc-
3.1. Standard dynamics and standard calibration tion could be achieved. Method 3 worked worse compared to
the Methods 2 and 2 light, since with this method the dynamic
By default settings, the machine accelerates at 4 m/s2 dur- errors are rather difficult to compensate.
ing cutting and at 10 m/s2 during positioning which will be re-
ferred to as “acceleration 4/10” in the following. First, a series
of measurements was performed at acceleration 4/10 and with 4. Conclusion and outlook
standard machine calibration. The corresponding results can be
seen in Figure 3. Method 3 significantly yields the best result Setpoint compensation can be achieved with three funda-
with 43% less median contour error (w.r.t. reference measure- mentally different approaches. First, model-based compensa-
ment). Using Method 2 light a small improvement of 2% can be tion can be applied offline based on model-based knowledge of
achieved while Method 1 results in no improvement. Basically, the machine. Second, machine feedback-based compensation
it should be understood that the contour error consists of two relies on the measured position and/or acceleration feedback of
parts. On the one hand there are quasi-static errors and on the the machine. Third, external feedback-based compensation re-
other hand there are dynamically induced errors. With acceler- lies on a measurement of the contour of a cut part. In this paper,
ation 4/10 and standard machine settings, the quasi-static errors implementations of all three of those approaches are used to
dominate (compare Figure 4). Method 3 takes the actual result- generate an estimate of the true TCP position and to apply it
ing contour into account and therefore can identify and correct for setpoint compensation to reduce the contour error of the cut
a constant static-offset while the other methods cannot. parts.
Method 1: This method relies on a machine model that needs
3.2. Optimized machine calibration to be developed, identified, and validated. Once available, the
big advantage of this method is that it can be implemented up-
In a next step the default calibration values of the machine stream of the machine as soon as the cutting plan in the CAD-
were optimally adjusted to the reference part in order to investi- CAM system is available, and thus can be decoupled from ma-
gate the potential of the setpoint correction methods for reduc- chine operation. However, that method cannot account for vari-
ing the dynamically induced error. Instead of using the double ation between different machines. In this paper, a rather simple
ball bar method, the calibration was conducted based on an opti- model for implementing this method is chosen, which resulted
cal measurement of the cut part to reduce the quasi-static errors in unsatisfactory performance. More complex and computation-
as much as possible. The description of this process is beyond ally more expensive models could help improve the results from
the scope of this paper and therefore not described. In general, this approach.
the quasi-static errors can also be reduced with other correc- Method 2 / Method 2 light: These methods operate on ma-
tion methods (e.g. friction compensation). After optimizing the chine feedback. The data can be collected with the laser turned
calibration, the measurement series was repeated as before. In off in “dry-cut” mode to avoid rejected parts. There is however
Figure 5 the results before and after optimizing the calibration a time investment that needs to be made for collecting the data;
as well as the comparison between the different methods are If the goal is to increase productivity, these methods are only
shown. Again, it can be seen that Method 3 results in the small- valid if the same parts are cut several times, ideally in mass
est median contour error. However, the median error is only production. As these methods relies on data that is generated
reduced by 9%. All other methods are not able to significantly on the machine itself, it is possible to fully integrate them in the
reduce the median error at this dynamic. This result could be software ecosystem of the machine provider which potentially
expected since the quasi-static error was minimized and for ac- results in a very robust implementation. Method 2 light can be
celeration 4/10 only a relatively small dynamic error is present. realized without the need to tune parameters of the accelera-
tion based TCP estimation algorithm and is therefore easier to
3.3. Increased dynamics industrialize than Method 2.
Method 3: This method requires the most effort to operate.
As a last step, the acceleration dynamics were increased to Further, it relies on data that is generated by a measurement
investigate a possible productivity increase while maintaining system that is typically not integrated in the machine tool sys-
the same accuracy. In doing so, the proportion of the dynam- tem. A manual work step is thus introduced which is potentially
ically induced errors increases. In Figure 6, the results for ac- error prone. A further drawback of this method is that it is not
celeration 9/10 are presented whereat the dynamically induced always possible to determine the underlying cause of contour
errors predominate the quasi-static errors (see Figure 4). Re- deviation. There might be artefacts introduced by the measure-
ferring to the median form error Method 1, Method 2, Method ment system, or contour deviations that are process related and
2 light and Method 3 reduced it by 4%, 26%, 26% and 14% cannot be corrected by path compensation. If such a deviation
respectively. Thus, all the methods have achieved an improve- is wrongly used for compensating the reference trajectory, the
ment with Methods 2 and 2 light performing best. Interestingly, result could be worse in some sections than without compensa-
the data from the accelerometer in the cutting head was not nec- tion. However, in sections where the measured contour is valid,
essarily needed to estimate the TCP (Method 2). With the esti- this method offers the highest robustness, as it does not rely on

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 586 Rafael Amacher et al. / Procedia CIRP 113 (2022) 582–587

Fig. 3. Distribution of the contour error of the measurements M1 (median: 12.4 µm), M2 (median: 12.6 µm), M3 (median: 12.2 µm) and M4 (median: 7.0 µm) that
were cut with acceleration 4/10.

Fig. 4. Measured contours are shown 5 times superelevated with a ±30 µm tolerance band. Left: M1, M2, M3 and M4 with acceleration 4/10; Right: M10, M11,
M12, M13 and M14 with acceleration 9/10.

models and other assumptions about the machine tool but op- rectly online during operation using additional online data. For
erates on the real contour. Our results however indicate that the this, the architecture and parametrization of the motion con-
method works well in regions with low acceleration, i.e. it com- troller could be extended and adopted to consider more sensor
pensates quasi-static errors very well, but it does not perform as information as currently possible in a standard machine feed
well as Method 2 and Method 2 light in case of highly dynamic drive controller. Industrial applicability has to be ensured by re-
sections. specting the conservative requirements concerning robustness,
In conclusion, method 1 requires more work to demon- safety and simplicity of such future approaches.
strate the benefits using more complex, digital-twin approaches.
Method 2 light offers the highest potential for industrialization
in case of high productivity requirements to counteract the loss References
of contour accuracy. Method 3 could be industrialized for use [1] M. Steinlin, S.Weikert, K.Wegener, 2010, Open loop inertial cross-talk
cases of mass production in cases where the highest accuracy is compensation based on measurement data, in: Proc. of the 25th Annual
desired and the machine operator is willing to accept an addi- Meeting of the American Society for Precision Engineering (ASPE’10).
tional calibration step for each new part. [2] E. Batzies, T. Schöller, V. Welker, O. Zirn, 2007, Optimal Control of Di-
All presented methods as well as the presented state of the art rect Driven Feed Axes with Flexible Structural Components, in: 7th Inter-
national Conference on Power Electronics and Drive Systems. IEEE, pp.
are based on offline a posteriori machine simulation or measure- 1127–1131.
ment data. Future research and development could in addition [3] O. Zirn, 2008, Machine tool analysis–modelling, simulation and control of
investigate the possibility to compensate for dynamic errors di- machine tool manipulators, A Habilitation Thesis, ETH Zürich.

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.
 Rafael Amacher et al. / Procedia CIRP 113 (2022) 582–587 587

Fig. 5. Distribution of the contour error of the measurements M1 (median: 12.4 µm), M5 (median: 7.7 µm), M6 (median: 8.6 µm), M7 (median: 8.0 µm), M8
(median: 8.3 µm) and M9 (median: 7.0 µm) that were cut with acceleration 4/10.

Fig. 6. Distribution of the contour error of the measurements M10 (median: 18.2 µm), M11 (median: 17.4 µm), M12 (median: 13.5 µm), M13 (median: 13.4 µm)
and M14 (median: 15.7 µm) that were cut with acceleration 9/10.

[4] D. Spescha, S. Weikert, O. Zirn, K. Wegener, 2017, Synchronisa- [10] D. Kim and S. Kim, 1993, An iterative learning control method with ap-
tion of feed axes with differing bandwidths using set point de- plication for CNC machine tools in Industry Applications Society Annual
lay, International Journal of Automation Technology, 11, pp.155–164, Meeting, Conference Record of the 1993 IEEE.
doi:10.20965/ijat.2017.p0155. [11] T. Haas, N. Lanz, R. Keller, S. Weikert, K. Wegener, 2016, Iterative Learn-
[5] Altintas, Y., Brecher, C.,Weck, M.,Witt, 2005, S.. Virtual machine tool. ing for Machine Tools Using a Convex Optimisation Approach, in: Proce-
CIRP Annals - Manufacturing Technology; pp. 115–138. dia CIRP, volume 46, pp. 391–395, doi:10.1016/j.procir.2016.04.033.
[6] D. Spescha, 2017, Framework for Efficient and Accurate Simulation of the [12] M. Togai and O. Yamano, 1985, Analysis and design of an optimal learning
Dynamics of Machine Tools, Ph.D. thesis, TU Clausthal. control scheme for industrial robots: A discrete system approach, Decision
[7] M. Garden, Learning control of actuators in control systems, Google and Control, 1985 24th IEEE Conference.
Patents, 1971. [13] N. Lanz, 2021, Framework for Overcoming Structural Limitations of Ma-
[8] D. Bristow, M. Tharayil and A. Alleyne, June 2006, A survey of iterative chine Tools using Additional Tool Center Point Measurements, Ph.D. the-
learning control, Control Systems, IEEE, Bd. 26, Nr. 3, pp. 96-114. sis, ETH Zurich.
[9] S. Arimoto, S. Kawamura and F. Miyazaki, 1984, Bettering Operation of [14] A. Domahidi and J. Jerez, 2014-2021, Embotech AG
Robots by Learning, Journal of Robotic Systems, pp. 123-140, 1984. (https://embotech.com/FORCES-Pro)

This is a resupply of March 2023 as the template used in the publication of the original article contained errors. The content of the article has remained unaffected.