The International Journal of Advanced Manufacturing Technology (2020) 108:539–552

[Link]

ORIGINAL ARTICLE

Self-optimizing grinding machines using Gaussian process models

and constrained Bayesian optimization
Markus Maier1 · Alisa Rupenyan1,2 · Christian Bobst3 · Konrad Wegener4

Received: 20 January 2020 / Accepted: 28 April 2020 / Published online: 16 May 2020
© Springer-Verlag London Ltd., part of Springer Nature 2020

Abstract
In this study, self-optimization of a grinding machine is demonstrated with respect to production costs, while fulfilling
quality and safety constraints. The quality requirements of the final workpiece are defined with respect to grinding burn and
surface roughness, and the safety constrains are defined with respect to the temperature at the grinding surface. Grinding
temperature is measured at the contact zone between grinding wheel and workpiece using a pyrometer and an optical fiber,
which is embedded inside the rotating grinding wheel. Constrained Bayesian optimization combined with Gaussian process
models is applied to determine the optimal feed rate and cutting speed of a cup wheel grinding machine manufacturing
tungsten carbide cutting inserts. The approach results in the determination of optimal parameters for unknown workpiece
and tool combinations after only a few grinding trials. It also incorporates the uncertainty of the constraints in the prediction
of optimal parameters by using stochastic process models.

Keywords Grinding · Process optimization · Performance measurement · Bayesian optimization · Gaussian process models

1 Introduction of grinding machines has been the subject of extensive

research in industry and academia. Already in 1994, Rowe
The selection of optimal grinding parameters is an important et al. [1] reviewed artificial intelligence systems applied to
task because grinding is often one of the last manufacturing grinding and listed various approaches such as knowledge-
steps where failures lead to high financial losses. Today based systems, expert systems, fuzzy logic systems,
parameter selection is mainly performed by experienced neural net systems, and adaptive control optimization
operators using a trial and error approach. Optimization techniques. Another review focusing on modelling and
simulation in grinding is conducted by Brinksmeier et al.
[2]. They categorize empirical models using regression
analysis, artificial neural networks, and rule-based models
 Alisa Rupenyan as applicable for process control, whereas other models
rupenyan@[Link]
such as molecular dynamic models, kinematic models, finite
Markus Maier element models, and first principle models are categorized
maier@[Link] as not applicable to process control. Over time, various
models have been proposed to control specific grinding
Christian Bobst
[Link]@[Link] tasks. For example, [3] reports an adaptive grinding force
control. These types of models are limited to optimize
Konrad Wegener partial aspects of the grinding process.
wegener@[Link]
A comprehensive overview on optimization techniques
and applications to grinding is given in [4]. An approach
1 inspire AG, Technoparkstrasse 1, 8005, Zurich, Switzerland to general process optimization is to mimic the action
2 Automatic Control Laboratory, ETH Zurich, Physikstrasse 3, of the operators using rule-based systems. Most often
8092, Zurich, Switzerland the rule-based systems are enhanced by partial process
3 Agathon AG, Gurzelenstrasse 1, 4512, Bellach, Switzerland modelling. For example, [5] combines different rule-based
4 Institute of Machine Tools and Manufacturing (IWF), ETH subroutines to adaptively optimize the grinding cycles and
Zurich, Leonhardstrasse 21, 8092, Zurich, Switzerland prevent grinding burn. Another example for centerless
540 Int J Adv Manuf Technol (2020) 108:539–552

grinding [6] describes the combination of models with rule- optimal parameters, using auditory, or olfactory informa-
based optimization for precision, productivity, and stability tion, as well as the output of the experiment as feedback.
increase. The main drawback of rule-based approaches Imitating this approach, in grinding sensory feedback dur-
is that generalization based on rules is very difficult ing processing can come from various sensors, such as force
and limits optimizations to very specific applications. or power sensors, acoustic emission, and temperature sen-
Furthermore, autonomously adapting rule-based system sors [11]. For design of experiments (DOE), standard DOE
using new observations is challenging. methods such as full or fractional factorial designs can be
Instead of rule-based approaches, one can optimize the used; however, they require a large number of experimen-
objective directly, where the goal is to find the process tal trials. Alternatively, to reduce the number of trials, the
parameters x, which minimize or maximize an objective Taguchi method has been very efficient in process opti-
function f (x) while fulfilling constraints c(x). For a mization, as demonstrated in [12]. The disadvantage of the
minimization problem, this can be written as follows. Taguchi method is that knowledge gained during testing is
not directly incorporated in the optimization, which reduces
x min = argminf (x), s.t. c(x) ≤ cmax (1)
data efficiency. Furthermore, one needs to make assump-
Typically the objective function is the cost of the grinding tions about the final shape of the optimization objective a
operation but it is also possible to consider energy efficiency priory, which is often infeasible in general optimization of
or environmental impact as an objective. For simplified grinding.
optimizations, the objective might be the grinding time In the present study, Gaussian process (GP) regression is
only. The constraints are mainly introduced by the process used for process modelling. Gaussian process regression is a
and the final workpiece requirements such as a desired powerful stochastic model, which is used widely to provide
surface roughness. Lee et al. [7] present a typical model- uncertainty estimation when dealing with stochastic data
based optimization approach, where empirical models are [13]. The main advantage of a Gaussian process regression
fitted based on available experimental data. Afterwards, is its flexibility, the in-built uncertainty prediction, and the
the models for the objective and constraint functions are non-parametric nature of the model. A detailed introduction
optimized based on an evolution strategy. This approach to Gaussian process regression can be found in [13]. Based
is extended in [8] for the optimization of multiple on a Gaussian process prior and available measurements,
grinding cycles using an evolution strategy and a dynamic Gaussian process regression can be used to predict the
programming approach. Another extension of [7] is given mean and the variance for every arbitrary combination
by [9], which couples the optimization to a database and of input parameters. The estimated variance is useful to
a graphical user interface. The optimization can also be incorporate uncertainty in the process optimization based on
extended to multiple objective functions, which is known the predictions from the Gaussian process regression.
as multi-objective optimization. Gholami and Azizi [10] Gaussian process regression is used in combination with
demonstrate constrained multi-objective optimization of Bayesian optimization (BO), where it provides a surrogate
grinding with Pareto fronts using genetic algorithms. These model for the unknown objective function. Bayesian
methods achieve the optimization by determining models optimization is a sequential design strategy for global
for the objective and constraint function, which are fixed optimization of an unknown objective function, which is
after an initial validation. While such an approach is well- expensive to evaluate [14]. Bayesian optimization uses
suited for specific tasks and niche applications where the predicted mean and variance of the Gaussian process
extensive training data is available, it is not applicable for model of the objective to determine the next experimental
general process optimization in grinding because grinding point with high information content; afterwards, the GP
processes are very sensitive to environmental conditions, model is updated with the newly available data [14]. This
and optimizing a grinding process on one machine does not procedure is repeated until suitable process parameters
necessarily result in the same performance on a different are found. Using constrained Bayesian optimization in
machine of the same type. Another obstacle is the lack of the sampling of successive trials outperforms random
information on the physical properties of the grinding wheel search, as demonstrated in [15] for two simulated functions
and of the workpiece, because they are trade secrets of the and for several applications related to hyperparameter
respective manufacturer. selection of learning algorithms. According to [16],
These challenges in process optimization can be over- random search can be considered a natural baseline for
come by interaction with the grinding machine, where hyperparameter optimization because it is more data-
sensory feedback is used to improve the models of both the efficient than the commonly used grid search approach.
objective and the constraint function. In fact, when experi- Bayesian optimization is not limited to hyperparameter
enced operators optimize the process, they devise a limited tuning of learning algorithms; it is also successfully used
number of carefully selected experiments to determine the in other areas such as photovoltaic power plants [17] and
Int J Adv Manuf Technol (2020) 108:539–552 541

robotics [18]. Recently, it has also been successfully applied σt2 of the Gaussian process regression at an arbitrary
to the optimization of turning processes [19, 20]. location x can be calculated as follows [13],
In this study, an autonomous process setup is proposed μt (x) = k T (x)(K + σN2 I )−1 y t (5)
where the cutting speed and the feed rate in grinding are
optimized based on combined measurements of roughness,
temperature, and maximum dressing intervals. Constrained σt2 (x) = k(x, x) − k T (x)(K + σN2 I )−1 k(x) (6)
Bayesian optimization is applied to enforce quality and ⎛ ⎞
safety constraints. This paper is organized as follows: A k(x 1 , x 1 ) · · · k(x 1 , x t )
⎜ .. .. .. ⎟
short introduction to Bayesian optimization and Gaussian K=⎝ . . . ⎠ (7)
process models is provided first, followed by a methodology k(x t , x 1 ) · · · k(x t , x t )
section explaining the optimization objective, experimental where x t and y t are the vectors containing t pairs of
setup, and algorithm implementation. Afterwards, optimiza- measurements assumed to be corrupted with Gaussian noise
tion results performed on a cup wheel grinding process are N(0, σN2 ), I is the identity matrix, K is the covariance
reported and discussed. matrix, and k(x) is a covariance vector between x and x 1
to x t . Note that it is also possible to use different priors
than a Gaussian process such as a Student t distribution
2 Bayesian optimization and Gaussian [21]. However, a Gaussian process prior is a common choice
process models in Bayesian optimization. Its advantages are demonstrated
empirical modelling performance on several data sets and
A general introduction to Gaussian process models is given analytical traceability, such as shown in [22].
in [13] and an introduction to Bayesian optimization can The length scales li2 in Eq. (4), the signal variance σf2
be found in [14]. According to [13], a Gaussian process
in Eq. (2), and the signal noise σN2 in Eqs. (5) and (6)
represents a distribution over functions and is a collection of
are hyperparameters of the Gaussian process regression,
random variables, which have a joint Gaussian distribution.
which can be summarized in a hyperparameter vector θ .
A Gaussian process is fully defined by a mean function
It is possible to specify these hyperparameters before the
m(x) and a covariance function k(x, x  ) often referred as
Gaussian process regression based on expert knowledge.
the kernel. In this study, a Matern 5 kernel with automatic
Another method, as described in [13], is to determine the
relevance determination (ARD) [13] is used,
  hyperparameters θ ∗ based on available measurements by
√ 5 2 √

k(x, x ) = σf 1 + 5r + r exp(− 5r)
2
(2) maximization of the marginal log likelihood p(y t |θ).
3
1
 log p(y t |θ) = − y Tt (K θ + σN2 I )−1 y t
r = (x − x  )T P −1 (x − x  ) (3) 2
1 t
− log |K θ + σN2 I | − log 2π (8)
2 2
P = diag(l12 , l22 , ..., lD
2
) (4)
where σf2 is the signal variance, r is the distance between θ ∗ = argmax log p(y t |θ) (9)
the input data points x and x  , and P is a diagonal matrix A θ subscript is added to the covariance matrix in Eq. 8
containing characteristic length scale parameters li2 for to explicitly show its dependence on the hyperparameters
each input space dimension up to dimension D. In this θ . Figure 1 shows the Gaussian process regression results
paper, the input data points are process parameters that exemplarily for different hyperparameters. In the example,
have to be optimized, and the Gaussian process models the the true function is set to y = (x − 0.2)2 and
optimization objective, or the optimization cost function at displayed for illustration (dotted black line). The true
different candidate configurations of process parameters. function was evaluated at several process parameter values.
Additionally, the optimization constraint is modelled as The measurements are noisy observations of the true
another Gaussian process. The kernel function specifies the function ȳ = y + , where  ∼ N(0, 5e − 4). These
relation between the function values at input data points x measurements are the input data points to the Gaussian
and x  , where for a short distance between the input points, process regression. The red line displays the predicted mean
the corresponding function values are similar, whereas for of the Gaussian process model of the function y = (x −
larger distances, higher variations are observed. 0.2)2 given the observations, and the shaded green area
A Gaussian process is used as a prior for Gaussian shows the corresponding 95% confidence interval of the
process regression. In this study, the Gaussian process prior corresponding prediction. Figure 1 (a) shows the Gaussian
mean function m(x) is assumed zero, specifying no expert process regression for the hyperparameters maximizing the
knowledge. The resulting posterior mean μt and variance marginal log likelihood using Eq. (9). It can be seen that
542 Int J Adv Manuf Technol (2020) 108:539–552

Fig. 1 Gaussian process

regression for different
hyperparameters. Panel (a):
hyperparameters determined by
maximizing the marginal log
likelihood using Eq. (9). Panel
(b): results for high signal noise.
Panel (c): results for short length
scale parameter. Panel (d):
results for long length scale
parameter

with these hyperparameters, the fit of the data is very the hyperparameters is that it naturally trades off model
good. A GP regression using the same data points but a complexity and goodness of model fit. The advantage for
high signal noise hyperparameter is shown in Fig. 1(b). Bayesian optimization is that it minimizes the number of
The prediction of the true cost function is still reasonable experiments while still providing a good model fit.
but shows a high overall uncertainty. Figure 1 (c) shows Once a GP model of the optimization cost function is
the Gaussian process regression for a short length scale available, in Bayesian optimization, an acquisition function
hyperparameter. In this case, more complex functions are is maximized to determine the candidate parameters where
also plausible candidates, which leads to high uncertainties the next experimental trial is conducted. In general, the
between the measurements, where data is missing. In the acquisition function provides a trade-off between process
Bayesian optimization setting, short length scales lead to parameters which are associated with a high uncertainty
additional experiments and slow down the optimization. On (exploration), and process parameters with a low predicted
the other hand, on Fig. 1(d), a Gaussian process regression value (exploitation) [14]. Various acquisition functions
for a very long length scale hyperparameter is shown. In this can be used in Bayesian optimization such as expected
case, the generalization from the observed measurements improvement [23], probability of improvement [24], upper
results in a simple (nearly linear) function for the mean confidence bounds [25], and predictive entropy search [26].
function of the Gaussian process and in low uncertainty In [15], good performance is demonstrated for constraint
over the whole range. In this case, the true function (black optimization based on an expected improvement acquisition
dashed line) is not always within the 95% confidence function. As shown in [27], predictive entropy search is
interval of the Gaussian process regression due to the able to outperform expected improvement for constraint
predominant generalization. In the Bayesian optimization optimization but this comes at the cost of increased
setting, very long length scales may lead to sub-optimal complexity, potential numerical instabilities, and analytical
parameter predictions. As pointed out in [13], an advantage intractability. Therefore, in a conservative approach, the
of maximizing the marginal log likelihood to determine expected improvement acquisition function is used in
Int J Adv Manuf Technol (2020) 108:539–552 543

this study. For the unconstrained case, the expected Figure 2 (b) shows exemplarily the result of a Gaussian
improvement function can be calculated as follows [23], process regression for a constraint quantity. In a real
application, the quantity might be surface roughness of the
aEI (x) = (Cmin − μt,co (x))F (Z) + σt,co (x)φ(Z) (10)
workpiece or the temperature during the grinding process
and must stay below a specified limit, which depends on
Cmin − μt,co (x) the requirements such as the final workpiece quality. In
Z= (11)
σt,co (x) this example, the limit is set to 250. The probability that
where μt,co (x) is the predicted mean cost function after the constraint is fulfilled is displayed in Fig. 2(c) and can
t iterations obtained by Gaussian process regression using be calculated using Eq. (13) and the Gaussian process
2 (x) is the predicted variance of the cost after
Eq. (5), σt,co regression of the constraint, as displayed in Fig. 2(b).
t iterations obtained by Gaussian process regression using Finally, the constrained expected improvement acquisition
Eq. (6), Cmin is the so far lowest measured cost, F (Z) function is displayed in Fig. 2(d) and can be calculated
is √the cumulative standard normal distribution F (Z) = according to Eq. (12), using the Gaussian process regression
Z
1/ 2π −∞ exp −t 2 /2)dt, and φ(Z) is the probability of the cost (displayed in Fig. 1(a)) and the Gaussian
√ function of a standard normal distribution φ(Z) =
density process regression of the constraint (displayed in Fig. 2(b)).
1/ 2π exp(−Z 2 /2). A standard normal distribution is a The constrained expected improvement acquisition function
normal distribution with zero mean and standard deviation is maximal for a process parameter value of 0.61,
of one. The acquisition function is displayed exemplarily which is the next experimental parameter. The constrained
for a cost minimization problem in Fig. 2(a) using the expected improvement acquisition function favors higher
Gaussian process regression from Fig. 1(a). Based on the process parameters than the acquisition function without
acquisition function, the next experiment is conducted at a considering constraints because according to the Gaussian
process parameter of 0.25, which maximizes the acquisition process regression of the constraint, the probability that low
function. process parameters fulfill the maximum allowed constraint
Most machining operations require fulfilling constraints of 250 is low. After the next experiment is conducted at a
such as workpiece quality, process, or safety constraints. process parameter value of 0.61, the optimization procedure
As proposed in [15], the expected improvement acquisition is repeated with the newly available measurements until a
function can be extended to constrained optimization, which maximum number of iterations or a stopping criterion is
is used in this study, reached.
n
aEI C (x) = aEI (x) pf,i (x) (12)
i=1 3 Methodology
where aEI (x) is the expected improvement without
constraints (Eq. (10)), n is the number of constraints, and 3.1 Grinding operation
pf,i is the probability that the constraint i is fulfilled.
Note that for the constrained case, the best observed value The grinding operation is performed on an Agathon DOM
Cmin is the minimal measured cost value, which fulfills all Semi grinding machine (see Fig. 3). An initial quadratic
constraints—and not the absolute minimal measured cost shaped insert (14.5 mm × 14.5 mm × 4.76 mm) made
as in the case without constraints. The probability that the of tungsten carbide (Tribo S25, grain size 2.5 μm, HV30
constraint i is fulfilled can be calculated as follows, 1470) is ground on two opposite sides, resulting in a final
  insert geometry of 10 mm × 14.5 mm × 4.76 mm. The
1 λi −(t − μt,ci (x))2 machine is equipped with a metal bonded diamond grinding
pf,i (x) = √ exp dt (13)
σt,ci 2π −∞ 2 (x)
2σt,c i
wheel with a mean diamond grain size of 46 μm and a
grain concentration of C100 (D46C100M717 from Tyrolit).
where λi is the maximum allowed value for constraint
During grinding, the grinding wheel is oscillated with 1 Hz
i, μt,ci (x) is the predicted mean of constraint i after
2 (x) is the to ensure a uniform wear of the grinding wheel. The spark-
t iterations calculated using Eq. (5), and σt,c i out time is set to 3 s and the oscillation during spark-out
variance of the constraint i after t iterations calculated using
is slightly increased to 1.5 Hz. After the grinding wheel is
Eq. (6). Here, the constraint i is modelled again using
worn out, it is conditioned by a dressing wheel (89A 240
Gaussian process regression and the predicted mean and
L6 AV217, geometry specification 150 × 80 × 32 W12 E15
variance are used in the calculation. The next experiment is
from Tyrolit). For all experiments, the dressing parameters
conducted for process parameters x ∗ , which maximize the
were constant (dressing infeed is 0.1 mm, grinding wheel
constrained expected improvement acquisition function.
speed is 6 m/s, dressing wheel speed is 12 m/s, feed rate is
x ∗ = argmax aEI C (x) (14) 0.5 mm/min, and spark-out time is 1 s). The used cooling
544 Int J Adv Manuf Technol (2020) 108:539–552

Fig. 2 Panel (a) shows the

calculation of the acquisition
function without considering
constraints using Eq. (10) and
the Gaussian process regression
model of the cost as displayed in
Fig. 1(a). Panel (b) shows the
Gaussian process regression
model of the constraint
exemplarily, calculated by
maximizing the marginal log
likelihood. Panel (c) shows the
probability that the constraint is
fulfilled, calculated using
Eq. (13) and the Gaussian
process regression of the
constraint, as displayed in panel
(b). Panel (d) shows the result of
the acquisition function
considering constraints, which
can be calculated using Eq. (12)

lubricant is Blasogrind HC 5 (from Blaser Swisslube) is measured optically using an optical fiber embedded
supplied by a multi-needle nozzle. The flow rate of the inside the grinding wheel, which collects the emitted
cooling lubricant is controlled manually by a lever and the radiation at the contact zone of the grinding wheel and the
angle between the multi-needle nozzle and the grinding workpiece. Figure 4 shows a schematic of the measurement
wheel is also adjusted manually. In this study, the flow equipment. The optical signal is evaluated on the rotating
rate and the angle between the workpiece and the grinding
wheel were set before the optimization by an experienced
operator. During the optimization, these parameters were
kept constant. In future, it might be promising to use the
angle of the cooling lubricant and the flow rate as additional
optimization parameters.

3.2 Measurement setup and process constraints

High temperatures at the contact zone between grinding

wheel and workpiece result in severe thermal damage
of the workpiece and in grinding burn. Grinding burn
causes an undesired structure change of the workpiece
material and must be avoided. In addition, grinding
with high temperatures may result in ignition of the
grinding oil and results in potential hazardous situations.
In this study, grinding burn limited the maximum allowed
temperature because grinding burn was observed at lower
temperatures than ignition of grinding oil. The temperature Fig. 3 Overview of grinding kinematic and experimental setup
Int J Adv Manuf Technol (2020) 108:539–552 545

Fig. 4 Schematic of embedded

temperature measurement

grinding wheel (pyrometer), filtered and transmitted to a measurements. In [28], grinding burn is detected optically
receiver unit outside the grinding machine and converted by the black coloring of the workpiece surface. This is
to temperature. In total, four optical fibers are evenly a simple method for grinding burn detection which is
distributed on the grinding wheel circumference for the often used in an industrial environment. In this study, this
temperature measurement. For each grinding operation, approach is considered sufficient because it is only used
the final considered temperature is averaged twice. First, to specify the maximum allowed temperature and thereby
the average of the ten highest temperature readings is the desired workpiece quality. For workpieces with higher
calculated, followed by averaging the computed values for quality requirements, the maximum temperature might be
each sensor. Note that the evaluated temperature of the determined based on more sophisticated methods, such as
pyrometer depends on emissivity of the measured surface. Barkhausen noise measurements as applied in [29].
The temperature sensor was not calibrated, which allowed The surface roughness of the workpiece is measured
only for measurements of relative temperature. As the transversal to the grinding direction and orthogonal to the
temperature sensor is used to detect grinding burn with a cutting edge using a tactile measurement device (Form
burn threshold temperature determined in advance, having Talysurf 120 L from Taylor Hobson) with a measurement tip
access to relative temperatures proved sufficient for the radius of 2 μm. The evaluation length was reduced to 4 mm
goals of the grinding process optimization. because the total thickness of the insert was only 4.76 mm.
If the first ground insert side after dressing already The wavelength cutoff is set to λc = 0.8 mm. In general,
is exposed to high temperature and grinding burn, the the roughness of both ground sides per insert is measured
workpiece quality is not fulfilled for the current and and averaged. If grinding burn occurred on a ground side,
for the following workpieces, as in general the grinding the roughness was not measured on this side. Instead, the
temperature increases during grinding due to grinding wheel roughness measurement was repeated at a different location
dulling. The measured temperature of the first ground side on the other ground side, where grinding burn did not
after dressing is therefore a constraint in the subsequent occur. The roughness measurements change from insert
optimization procedure and should not exceed the allowed to insert with no clear trend. Therefore, in a conservative
threshold value. approach, the maximum measured roughness value of one
The temperature readings after dressing were used to optimization run, grinding with a freshly dressed wheel until
determine the maximum dressing interval. The maximum
dressing interval is defined as the number of grinding
cycles until the temperature exceeds the maximum allowed
temperature. The experimental run was stopped when this
temperature limit was exceeded. To avoid unnecessary
long experimental runs, the run was stopped after 8
inserts if the maximum temperature was not previously
exceeded. The maximum allowed temperature was set to
585 ◦ C because in preliminary experiments in [28] with
194 insert sides, grinding burn has been classified with
a 100% success rate using this temperature measurement
system and the same workpiece-wheel combination as in
this study. Figure 5 shows the classification of grinding Fig. 5 Grinding burn limit of preliminary experiments adapted from
burn based on a temperature threshold and temperature [28]
546 Int J Adv Manuf Technol (2020) 108:539–552

grinding burn occurred, was considered a constraint for the cost of grinding wheel wear during dressing, and cost of
optimization. The roughness limit depends on the grinding dressing wheel wear during dressing.
application. In the special case of cup wheel grinding of s · apg
CT (vs , f ) = Cm
inserts, the allowed roughness limit is a trade secret of each f
 
insert manufacturer. In this study, the roughness was limited Cgw ad,gw Cdw ad,dw Vg
+ Cm td + + (15)
to Ra,max = 230 nm, which leads to typical optimal cutting agw adw Vw (vs , f )
parameters.
The error of the final workpiece geometry is influenced The feed rate f and the removed workpiece material until
by the thermal distortion of the grinding machine and the wheel is dull Vw depend on the selected input parameters
the abrasive layer thickness of the grinding wheel, feed rate f and cutting speed vs , which are the process
which experiences wear during the grinding operation. In parameters used in Bayesian optimization. The other cost
this study, the geometrical error of the workpiece was parameters are constants as summarized in Table 1. The
not considered a constraint in the optimization because grinding wheel wear during dressing was considered and
the grinding machine allows adaptive correction of the the wear during grinding was neglected. This assumption is
geometrical error. The grinding machine is equipped with a reasonable because it is known that metal bonded grinding
measurement sensor (see Fig. 3) which allows to measure wheels hold the diamond grains very strongly, resulting in
the final workpiece dimension. The sensor measurement a high abrasive wear resistance whereby self-sharpening is
after each workpiece serves as a corrective feedback for limited [30]. The limited self-sharpening of the grinding
the grinding action. With the grinding parameters used in wheel requires redressing of the grinding wheel, which is
this study and the automatic insert handling switched on, assumed to be the main cause of grinding wheel wear. The
the measuring interval is always less than 1 min, which grinding wheel wear during dressing ad,gw was assumed
allows frequent corrective adjustments. For workpieces with to be 1 μm. Grinding wheel wear during dressing is
a higher geometrical complexity or a long grinding time, determined by the dressing parameters, which were fixed
it might be beneficial to include the geometrical error of in this study and not considered in the optimization. In
the final workpiece in the optimization as an additional this study, grinding wheel wear during dressing can be
constraint. considered a weight in the cost function. The spark-out
time for each grinding operation was not considered for
3.3 Cost calculation cost calculation because it only shifts the cost function up
by a constant value and does not influence the optimal
The grinding operation cost cannot be measured directly parameters. As reported in [31], the total cost for cooling
with a single sensor as it is composed of both process- and lubrication comprises costs for waste disposal, cleaning,
related quantities such as grinding time and grinding wear service, electricity, staff, and laboratory equipment, which
and economic quantities such as machine hourly cost and can add up to 16.9 % of the total machining cost. In
cost of the grinding wheel. The total cost to grind one insert this study, the cooling and lubrication costs are neglected
is calculated based on grinding time cost, dressing time cost, because only the electricity costs depend directly on the

Table 1 Parameters for cost

calculation Parameter Description Value

Vg Removed material per workpiece 310.6 mm3

s Number of ground sides 2
apg Infeed per side 2.25 mm
td Dressing time 19 s
agw Abrasive layer thickness of the grinding wheel 4 mm
ad,gw Wear of the grinding wheel during dressing 1 μm
adw Thickness of the dressing wheel 65 mm
ad,dw Wear of the dressing wheel during dressing 0.1 mm
Cm Machine hourly cost 100 U/h
Cgw Cost of grinding wheel 1500 U
Cdw Cost of dressing wheel 95 U
Int J Adv Manuf Technol (2020) 108:539–552 547

optimization parameters cutting speed and feed rate, which

according to [31] account for only 0.5 % of the total costs.
The small reported contribution of the electricity cost is in
line with rough calculations performed for the setup used
in this study. The other cooling and lubrication costs are
mainly influenced by the accumulated removed material,
which cannot be optimized and would only result in a
constant upwards shift of the cost function without changing
the optimal cutting speed and feed rate. However, if one
decides to optimize for environmental impact instead of
costs, or to use different cooling types as optimization
parameters, the cooling and lubrication cost might influence
the optimal parameters. For generality, in this report, the
cost is specified with units U.

3.4 Optimization implementation

The objective in this study is to find the process parameters

minimizing the total production cost CT and to fulfill the
maximum temperature constraint Tmax and the maximum
surface roughness constraint Ra,max .

T (vs , f ) < Tmax

x min = argmin CT (vs , f ) s.t. (16)
Ra (vs , f ) < Ra,max

The algorithm was implemented in MATLAB using the

GPML library [32] for Gaussian process regression. A flow
diagram of the implementation is shown in Fig. 6. The
optimization is started with two experiments at random
process parameters within the input optimization domain.
The optimization domain for the Bayesian optimization was
set based on experience to a minimal feed rate of 10 mm/min
and a maximum feed rate of 40 mm/min. The minimal
cutting speed was set to 12 m/s and the maximum cutting
speed was set to 30 m/s. In this way, a typical cutting
speed range was covered. After calculating the cost per part
at the measured process parameter values and determining
the hyperparameters, three Gaussian process regressions are
calculated, to model the cost (optimization objective), and
the temperature and roughness (optimization constraints).
Based on the obtained probabilistic predictions at
different (not previously probed) values of the process
parameters, the optimal parameters that fulfill Eq. (16)
can be calculated based on the obtained probability Fig. 6 Flow diagram of optimization implementation
distributions for the cost, temperature, and roughness. Using
stochastic models allows for probabilistic statements about
the fulfillment of the constraint limits. As discussed in detail probabilities higher than user-defined values pRa,min and
in [33], the optimal parameters x opt = (vs,opt , fopt ) are pT ,min .
calculated as the parameters, which minimize the expected
cost μt,co (vs , f ) after t experiments calculated using Eq. (5) pf,T (vs , f ) ≥ pT ,min
x opt = argmin μt,co (vs , f ) s.t. (17)
and fulfill the roughness and temperature constraints with pf,Ra (vs , f ) ≥ pRa,min
548 Int J Adv Manuf Technol (2020) 108:539–552

The minimal probability that a constraint is fulfilled can

be chosen freely and depends on the requirements. For
example, many parts for the aviation industry will require
a high minimal probability that the constraints are fulfilled.
On the other hand, if one produces disposable products, it
might be reasonable to accept a lower minimal probability
that the constraints are fulfilled. The probability pf,i that
the constraint i is below its maximum allowed value λi
can be calculated for temperature pf,T and roughness pf,Ra
by using Eq. (13) and the maximum allowed temperature
value Tmax = 585 ◦ C and the maximum allowed roughness
value Ra,max =230 nm (specified in Section 3.2). In general,
it is possible to specify separate minimal probabilities for
temperature and roughness. In this study, the parameters
pT ,min and pRa,min were set equal for simplicity. Fig. 7 Typical temperature measurement as a function of ground
inserts for a cutting speed of 24.3 m/s and a feed rate of 11.7 mm/min.
Afterwards, the optimal parameters can be used to assess The last side which fulfills the temperature constraint is the first side
convergence as proposed in [20], of plate 8. Therefore, the maximum dressing interval is 7.5 inserts

2σt,co (x opt ) <  (18)

set is calculated as 7.5 inserts, which corresponds to an
where σt,co is the predicted standard deviation of the cost at accumulated removed material volume Vw of 2329.5 mm3 .
predicted optimal parameters x opt after t iterations and  is Depending on the grinding task, considering workpiece
the convergence limit. In this way, convergence is reached handling time may lead to positive natural numbers as
when 95% of the predicted cost values at optimal parameters optimal dressing intervals instead of positive real numbers
are within μc (x opt ) ±  . The convergence limit  must be due to timesaving of dressing and workpiece handling
higher than the combined cost measurement and process parallelization. In this study, handling time of workpiece
uncertainty; otherwise, convergence can never be reached. was neglected because it is considered very short compared
To increase the convergence robustness, full convergence is with the dressing time and strongly depends on the used
reached, when the stopping criterion is below 0.04 U over handling system and corresponding parameters.
three consecutive iterations, the optimal feed rate change Figures 8 and 9 present the on-machine optimization.
within an interval of 0.4 mm/min, and the optimal cutting Convergence was judged for a minimal allowed probability
speed change within an interval of 0.2 m/s. If convergence of quality defects of 0.5 (pRa,min = pT ,min = 0.5). The
is not reached, the next experimental parameters will be first two experiments are conducted at random points in
determined based on maximizing the constrained expected the parameter space to initialize the optimization. The first
improvement acquisition function and the optimization and the second experiments do not fulfill the constraints;
continues. An initial assumption for the likelihood of the therefore, the algorithm does not approach convergence
model cost and the constraints is that they are Gaussian, due after these experiments. However, the proposed trial for
to the favorable properties of this distribution. Once more
experimental data is available, other distributions can be
experimentally verified if needed.

4 Results

Figure 7 shows a typical temperature reading as a function

of ground inserts for a full experimental run with constant
input parameters. The grinding wheel was dressed before
each experimental run, making it sharp and clean. In
general, the temperature is strongly influenced by the
accumulated removed material volume, where towards the
end of an experimental run the temperature increases.
The second side of the 8th insert does not fulfill the
constraint, resulting in grinding burn. Therefore, the Fig. 8 Convergence of Bayesian optimization with a maximum
measured maximum dressing interval for this parameter allowed probability of quality defects pRa,min =pT ,min =0.5
Int J Adv Manuf Technol (2020) 108:539–552 549

a trade-off between these two costs. After 12 experiments,

the model uncertainty in the high feed rate range (above
25 m/min) is higher than the uncertainty corresponding
to low feed rate. Considering the high predicted costs for
these parameters, the uncertainty is still moderate. The
roughness mostly depends on the cutting speed. An increase
in cutting speed reduces the surface roughness whereas
the feed rate only influences the roughness minimal. This
is in line with the findings of the preliminary study
[28] for cup wheel grinding with the same setup, where
an extensive series of experiments was conducted. The
predicted uncertainty of the roughness is between 19.7
and 33.4 nm, which is 9 and 15% of the maximum
allowed roughness. An increase in feed rate or cutting speed
leads to higher temperatures, again confirming the results
reported in [28]. The predicted temperature uncertainty is
between 59.5 and 76.5 ◦ C which is between 10 and 13%
of the maximum allowed temperature. The temperature
uncertainty is higher for high feed rates because only limited
number of experiments were conducted at high feed rates
Fig. 9 Conducted experiments after 12 iterations. Experiment with due to high total costs at high feed rates. In such cases, the
lowest cost and fulfilled constraints was 10th experiment
recommended process parameter configuration by the BO
algorithm for the next trials moves to process parameters
the third measurement is at high cutting speeds and low corresponding to lower costs, and the more unfavorable
feed rates, which reduces both surface roughness and regions of the parameter space are not extensively explored.
temperature. The third measurement fulfills the constraints. Figure 11 shows a comparison of the model predictions and
The blue line on Fig. 8 shows the current measured cost the measured values after 12 iterations. The model shows
values. Experimental points explored due to high model very good agreement with the measured data.
uncertainty often result in high costs, for example, iterations Table 2 shows optimal parameters for several different
4, 7, and 9. In such cases, often the constraints are not values of pRa,min and pT ,min computed from Eq. (17) after 12
fulfilled. Until the 8th iteration, the uncertainty of the iterations. The algorithm recommends an increase in cutting
prediction is reduced drastically. After the 8th iteration, the speed when the minimal probabilities that the constraints
uncertainty remained low until full convergence is reached are fulfilled pRa,min and pT ,min are increased, without a large
after 12 iterations. Experimentally, the best parameters modification of the feed. An increase in cutting speed leads
are obtained at the 10th iteration with a cutting speed of to lower surface roughness and less probability to violate the
24.3 m/s, a feed rate of 11.7 mm/min, and a dressing interval surface roughness constraint. As expected, increasing the
of 7.5 inserts. After the 10th iteration, a small increase in minimal probability that the constraints are fulfilled results
the cost and the uncertainty is observed, however within the in higher production costs.
convergence limits. In this study, convergence is reached
after 12 iterations. Note that the number of iterations is 4.1 Discussion
influenced by the complexity of the objective and constraint
functions. Simple functions follow global trends, which Note that the optimization problem may change depending
can be modelled with a large length scale hyperparameter, on the grinding operation and the use of the final workpiece.
resulting in a minimal number of iterations. On the other Specifically the cost calculation, the optimized parameters,
hand, complex functions will show very localized behaviors, and the considered constraints may need adjustment. For
which correspond to short length scales, and therefore example, grinding inserts made of polycrystalline boron
require more iterations. nitride (PCBN) instead of tungsten carbide may require
Figure 10 shows the predicted mean and confidence the measurement of grinding wheel wear. Therefore, the
interval after 12 iterations for cost, roughness, and optimization will benefit from additional sensory feedback
temperature. The predicted cost is high for low feed rates such as measurements of the grinding wheel wear and
because the operation is time consuming. On the other electrical power consumption to further generalize the
hand, higher feed rates lead to shorter dressing intervals, optimization objectives for multiple grinding applications.
which makes redressing expensive. Optimal parameters are Additional optimization parameters such as angle between
550 Int J Adv Manuf Technol (2020) 108:539–552

Fig. 10 Predicted mean cost

(top left), cost confidence
interval (top right), predicted
mean roughness (middle left),
roughness confidence interval
(middle right), predicted mean
temperature (bottom left), and
temperature confidence interval
(bottom right) after 12 iterations

Fig. 11 Comparison of model predictions with 95% confidence interval and measured temperature, roughness, and cost values after 12 iterations
Int J Adv Manuf Technol (2020) 108:539–552 551

Table 2 Optimal predicted

parameters after 12 iterations Minimal probability that con- Optimal parameters Costs (U)
with different minimal straints are fulfilled (%)
probabilities that the
temperature and roughness 50 f = 12.0 mm/min, v = 23.8 m/s 0.78 ± 0.04
constraints are fulfilled. The 84.1 f = 11.8 mm/min, v = 25.2 m/s 0.83 ± 0.05
predicted optimal costs and the 97.7 f = 11.4 mm/min, v = 26.8 m/s 0.92 ± 0.07
predicted 95% confidence
interval of the optimal costs are 99.9 f = 11.6 mm/min, v = 28.8 m/s 1.08 ± 0.13
also given

cooling lubricant and grinding wheel, cooling lubricant flow 5. Morgan MN, Cai R, Guidotti A, Allanson DR, Moruzzi J,
rate, dressing feed rate, dressing wheel speed, dressing Rowe W (2007) Design and implementation of an intelligent
grinding assistant system. Int J Abras Technol 1(1):106–
wheel type, grinding wheel oscillation frequency, grinding
135
wheel grain size, grinding wheel bond type, and grain 6. Barrenetxea D, Marquinez JI, Álvarez J, Fernández R, Gallego I,
concentration of the grinding wheel may also help to Madariaga J, Garitaonaindia I (2012) Model-based assistant tool
industrialize the optimization. for the setting-up and optimization of centerless grinding process.
Mach Sci Technol 16(4):501–523
7. Lee CW, Choi T, Shin YC (2003) Intelligent model-based
optimization of the surface grinding process for heat-treated 4140
5 Conclusion steel alloys with aluminum oxide grinding wheels. J Manuf Sci
Eng 125(1):65–76
8. Lee CW (2009) Dynamic optimization of the grinding process in
In this study, Bayesian optimization with on-machine
batch production. pp 485–494
measurements was used to obtain optimal parameters for 9. Choi T, Shin YC (2007) Generalized intelligent grinding advisory
grinding within a few iterations, while fulfilling quality system. Int J Prod Res 45(8):1899–1932
and safety constraints. Gaussian process regression models 10. Gholami MH, Azizi MR (2014) Constrained grinding optimiza-
tion for time, cost, and surface roughness using NSGA-II. Int J
were successfully applied to model costs, roughness, and
Adv Manuf Tech 73(5):981–988
temperature of the grinding process. The proposed approach 11. Tönshoff H, Friemuth T, Becker JC (2002) Process monitoring
is suitable to consider the probability that the constraints are in grinding. CIRP Annals-Manufacturing Technology 51(2):551–
fulfilled in the calculation of optimal parameters. Bayesian 571
optimization is very flexible in finding optimal parameters 12. Shaji S, Radhakrishnan V (2003) Analysis of process parameters
in surface grinding with graphite as lubricant based on the Taguchi
without specifying any domain-specific knowledge such method. J Mater Process Technol 141(1):51–59
as grinding wheel type, workpiece material, or cooling 13. Rasmussen CE (2006). In: Williams CKI (ed) Gaussian processes
lubrication type. This enables the optimization of various for machine learning. MIT Press, Cambridge
grinding operations or manufacturing processes in general. 14. Shahriari B, Swersky K, Wang Z, Adams RP, Freitas Nd
(2016) Taking the human out of the loop: a review of Bayesian
optimization. Proc IEEE 104(1):148–175
Funding information The project was funded by CTI Number 25411. 15. Gardner JR, Kusner MJ, Xu ZE, Weinberger KQ, Cunningham
JP (2014) Bayesian optimization with inequality constraints. In:
Compliance with ethical standards ICML, pp 937–945
16. Bergstra J, Bengio Y (2012) Random search for hyper-
parameter optimization. J Mach Learn Res 13(Feb):281–
Conflict of interest Agathon AG filed a patent on parts of this work. 305
17. Abdelrahman H, Berkenkamp F, Poland J, Krause A (2016)
Bayesian optimization for maximum power point tracking in
photovoltaic power plants. In: 2016 European Control Conference
References (ECC), pp 2078–2083
18. Lizotte DJ, Wang T, Bowling MH, Schuurmans D (2007)
1. Rowe WB, Yan L, Inasaki I, Malkin S (1994) Applications Automatic gait optimization with Gaussian process regression. In:
of artificial intelligence in grinding. CIRP Ann Manuf Technol IJCAI, pp 944–949
43(2):521–531 19. Maier M, Rupenyan A, Akbari M, Zwicker R, Wegener K
2. Brinksmeier E, Aurich J, Govekar E, Heinzel C, Hoffmeister (2019) Turning: autonomous process set-up through Bayesian
H-W, Klocke F, Peters J, Rentsch R, Stephenson D, Uhlmann optimization and Gaussian process models (in press). In: CIRP
E (2006) Advances in modeling and simulation of grinding ICME 2019
processes. CIRP Ann 55(2):667–696 20. Maier M, Zwicker R, Akbari M, Rupenyan A, Wegener K (2019)
3. Jenkins HE, Kurfess TR (1999) Adaptive pole-zero cancellation Bayesian optimization for autonomous process set-up in turning
in grinding force control. IEEE Trans Contr Syst Tech 7(3):363– CIRP. Journal of Manufacturing Science and Technology
370 21. Shah A, Wilson A, Ghahramani Z (2014) Student-t processes as
4. Venkata Rao R (2011) Modeling and optimization of machining alternatives to Gaussian processes. In: Artificial intelligence and
processes. In: Venkata Rao R (ed) Advanced modeling and statistics, pp 877–885
optimization of manufacturing processes: International Research 22. Rasmussen CE (1997) Evaluation of Gaussian processes and other
and Development. Springer, London, pp 55–175 methods for non-linear regression. University of Toronto, Canada
552 Int J Adv Manuf Technol (2020) 108:539–552

23. Mockus J, Tiesis V, Zilinskas A (1978) The application of 29. Karpuschewski B, Bleicher O, Beutner M (2011) Surface integrity
Bayesian methods for seeking the extremum. In: Towards Global inspection on gears using barkhausen noise analysis. Procedia Eng
Optimisation, pp 117–129 19:162–171
24. Kushner HJ (1964) A new method of locating the maximum point 30. Marinescu ID, Rowe WB, Dimitrov B, Ohmori H (2013) 9
of an arbitrary multipeak curve in the presence of noise. J Basic - Abrasives and abrasive tools. In: Marinescu ID et al (eds)
Eng 86(1):97–106 Tribology of abrasive machining processes. 2nd edn. William
25. Srinivas N, Krause A, Kakade S, Seeger M (2010) Gaussian Andrew Publishing, Oxford, pp 243–311
process optimization in the bandit setting: no regret and 31. Brinksmeier E, Walter A, Janssen R, Diersen P (1999) Aspects
experimental design. In: International Conference on Machine of cooling lubrication reduction in machining advanced mate-
Learning, pp 1015–1022 rials. Proceedings of the Institution of Mechanical Engi-
26. Hernández-Lobato JM, Hoffman MW, Ghahramani Z (2014) neers, Part B: Journal of Engineering Manufacture 213(8):769–
Predictive entropy search for efficient global optimization 778
of black-box functions. In: Advances in neural information 32. Rasmussen CE, Nickisch H (2010) Gaussian processes
processing systems, pp 918–926 for machine learning (GPML) toolbox. J Mach Learn Res
27. Hernandez-Lobato JM, Gelbart M, Hoffman M, Adams R, 11(Nov):3011–3015
Ghahramani Z (2015) Predictive entropy search for bayesian opti- 33. Dentcheva D 4. Optimization models with probabilistic con-
mization with unknown constraints. In: International Conference straints. In: Lectures on Stochastic Programming. pp 87–
on Machine Learning, pp 1699–1707 153
28. Maier M, Gittler T, Weiss L, Bobst C, Scholze S, Wegener
K (2019) Enhanced sensors as an enabler for self-optimizing Publisher’s note Springer Nature remains neutral with regard to
grinding machines. MM Science Journal jurisdictional claims in published maps and institutional affiliations.