You are on page 1of 6

ScienceDirect

Available online at www.sciencedirect.com

Available ScienceDirect
Procedia
online
Available
CIRP 00 (2019) 000–000
atonline
www.sciencedirect.com
at www.sciencedirect.com
www.elsevier.com/locate/procedia

ScienceDirect
Procedia CIRP 00 (2019) 000–000
ScienceDirect www.elsevier.com/locate/procedia

Procedia CIRP 00 (2017)


Procedia 000–000
CIRP 82 (2019) 302–307
www.elsevier.com/locate/procedia
17th CIRP Conference on Modelling of Machining Operations
17th CIRP Conference on Modelling of Machining Operations
A hybrid approach using machine learning to predict the cutting forces
A hybrid approach 28thunder
using
CIRP consideration
machine
Design Conference, of the
learning
May totool
2018, wear
predict
Nantes, the cutting forces
France

Penga*, Thomas under consideration


a of thea tool Klockewear a, Benjamin Döbbelera
A Bingxiao
new methodology toBergs
analyze , Daniel theSchraknepper
functional, Fritz and physical architecture of
existing
Bingxiao products
Peng for an
*, Thomas
Laboratory for Machine
Bergs assembly
a Tools and Production Engineering
a
, Daniel oriented product
Schraknepper a
familya
, FritzCampus-Boulevard
(WZL) of RWTH Aachen University,
Klocke identification
30, 52074 Aachen, Germany a
, Benjamin Döbbeler
a

* Corresponding author. Tel.: +49 241 80 28181; fax: +49 241 80 22293. E-mail address: b.peng@wzl.rwth-aachen.de
a
Laboratory for Machine Tools and Production Engineering (WZL) of RWTH Aachen University, Campus-Boulevard 30, 52074 Aachen, Germany
Paul Stief *, Jean-Yves Dantan, Alain Etienne, Ali Siadat
* Corresponding author. Tel.: +49 241 80 28181; fax: +49 241 80 22293. E-mail address: b.peng@wzl.rwth-aachen.de

Abstract École Nationale Supérieure d’Arts et Métiers, Arts et Métiers ParisTech, LCFC EA 4495, 4 Rue Augustin Fresnel, Metz 57078, France

The cutting process


*Abstract
Corresponding author.isTel.:
a complex
+33 3 87nonlinear system.address:
37 54 30; E-mail Predicting such a system with conventional regression models is inefficient. In this paper, a
paul.stief@ensam.eu
hybrid approach using deep neural networks (DNN) is proposed to predict the specific cutting forces. With the aim of obtaining the hybrid training
data,cutting
The orthogonal cutting
process tests and nonlinear
is a complex 2D FEM system.
chip formation
Predictingsimulations have been
such a system performed under
with conventional diversemodels
regression cuttingisparameters,
inefficient.tool geometries
In this paper, a
and toolapproach
hybrid wear conditions.
using deep Predictive models (DNN)
neural networks using a isDNN and atoconventional
proposed linear regression
predict the specific method
cutting forces. Withwere
theestablished. In comparison
aim of obtaining the hybrid with the
training
Abstract
conventional
data, orthogonallinear regression
cutting method,
tests and 2D FEM the hybrid model using
chip formation the machining
simulations learning
have been is moreunder
performed accurate.
diverse cutting parameters, tool geometries
and tool wear conditions. Predictive models using a DNN and a conventional linear regression method were established. In comparison with the
In© today’s
2019 The
conventionalbusiness
linearenvironment,
Authors. theElsevier
Publishedmethod,
regression by trend
the towards more using
B.V. model
hybrid productthevariety and customization
machining learning is moreis accurate.
unbroken. Due to this development, the need of
agile and reconfigurable
Peer-review production
under responsibility systems
of the emerged
scientific to cope
committee of with various
The 17th CIRPproducts and product
Conference families.
on Modelling of To design and
Machining optimize production
Operations,
© 2019 The Authors. Published by Elsevier B.V.
systems
in as
the personwell as to choose
of the Conference the
Published Chairoptimal product
Dr Erdem matches,
B.V.Ozturk product
and Co-chairs analysis methods are needed. Indeed, most of the known methods aim to
© 2019 The Authors.
Peer-review under responsibility by ofElsevier
the scientific committee of The Dr
17thTomCIRPMcleay and Dr Rachid
Conference Msaoubi.
on Modelling of Machining Operations
analyze a product or one product family on the physical level. Different product families, however, may
Peer-review under responsibility of the scientific committee of The 17th CIRP Conference on Modelling of Machining Operations, differ largely in terms of the number and
nature
in the of
Keywords:components.
personhybrid Thismachine
approach,
of the Conference factChair
impedes
learning, an efficient
cutting
Dr Erdem comparison
process,
Ozturk FEM
and andDr
Co-chairs choice
Tom of appropriate
Mcleay and Dr product family combinations for the production
Rachid Msaoubi.
system. A new methodology is proposed to analyze existing products in view of their functional and physical architecture. The aim is to cluster
these products
Keywords: in new
hybrid assembly
approach, oriented
machine product
learning, cuttingfamilies
process,for
FEM the optimization of existing assembly lines and the creation of future reconfigurable
assembly systems. Based on Datum Flow Chain, the physical structure of the products is analyzed. Functional subassemblies are identified, and
a1.functional
Introduction
analysis is performed. Moreover, a hybrid functional and physical small architecture
undeformed chip
graph thicknesses.
(HyFPAG) In addition,
is the output specific
which depicts the
similarity between product families by providing design support to both,cutting production system
forces planners
are also and product
dependent designers.
on process An illustrative
parameters such
example of a nail-clipper
1. Introduction
Predicting is usedistoalways
cutting forces explainan theimportant
proposed methodology.
topic in the Ansmall industrial
as case study
undeformed
cutting speeds. Toonchip
two product
cover a wide families
thicknesses.
range ofInofprocess
steeringparameters,
addition, columns of
specific
thyssenkrupp Presta France is then carried out
design of machining processes. The mechanical energy, whichto give a first industrial evaluation
cutting of
determiningthe proposed
forcesthe approach.
areregression
also dependentmodelonwithprocess parameters
experimental such
data is
©can2017
be The
PredictingAuthors.
calculatedcuttingPublished
from forces
cuttingbyalways
is Elsevieran
forces, isB.V.
important
almost topic ininto
converted the as cutting
cost- speeds. To cover a wide range of process parameters,
and time-consuming.
Peer-review under responsibility
design of machining processes. of the
Thescientific
mechanical committee
energy, of the 28th CIRP
which Design Conference
determining 2018. model with experimental data is
the regression
the thermal energy. The thermal energy, which acts as a process
can be calculated
signature from cutting
[1], affects forces, is
the surface almost converted
integrity. In addition, into 𝐹𝐹𝑐𝑐 = 𝑘𝑘
cost- and ∙ 𝑏𝑏 ∙ ℎ1−𝑚𝑚𝑐𝑐 = 𝑘𝑘𝑐𝑐 ∙ 𝑏𝑏 ∙ ℎ
𝑐𝑐1.1time-consuming. (1)
Keywords: Assembly; Design method; Family identification
the thermal energy.
mechanical loads on The thermal
cutting energy,
tools whichthe
influence acts as alife.
tool process
The 1−𝑚𝑚𝑐𝑐
signature
deformation[1],and affects the surface
vibration of cutting integrity.
tools Incaused addition,by 𝐹𝐹𝑐𝑐 =In 𝑘𝑘addition
𝑐𝑐1.1 ∙ 𝑏𝑏 ∙ ℎto regression
= 𝑘𝑘𝑐𝑐 ∙ 𝑏𝑏 models,
∙ℎ analytical models, (1) for
mechanical loads on cuttingalso
influence toolsthe influence
workpiece the geometry
tool life. Theand example shear plane theory has been investigated. The theory
1.deformation
Introduction
surface and Therefore,
roughness. vibration of cutting cutting
predicting tools forces
causedis by an of In addition
assumes
the product todeformation
that therange regression models,
takes placeanalytical
and characteristics on only one
manufacturedmodels,
plane, for
the
and/or
mechanical
essential partloads influence also
for optimizing the workpiece
the cutting process. geometry and example
shear shear
plane. By plane theory
applying thehas been investigated.
plasticity theory, The
cutting theory
forces
assembled in this system. In this context, the main challenge in
assumes that the deformation
can be calculated takes place on onlymodels
[3; 4].is Although one plane, thea
surface
A lotroughness.
Due toof work
the fast Therefore,
has been predicting
conducted
development oncutting
in predictingforcescutting
the domain is an
of modelling and analysis now not analytical
only to cope with have single
essential
forces. The part forand
optimizing
Kienzle-force the cutting
model, astrend
given process.
in Eq.1, is one ofand the shear
significantplane. By
time applying
advantage, the
the plasticity
applicationtheory,
of cutting
analytical forces
models
communication an ongoing of digitization products, a limited product range or existing product families,
can
in also be calculated
reality [3; 4].byAlthough analytical models have anda
mostA lot of work
commonly
digitalization, has regression
used
manufacturingbeen conducted modelson
enterprises predicting
based
are facing cutting
on important
empirical but to beisable limited
to analyze several
and ideal products
to compare assumptionsto define
forces.
data. TheThe in Kienzle-force model, asrepresents
given in Eq.1,
modelenvironments: iscontinuing
one of the
an a approximate significant
simplifications. time advantage, the application of analytical models
challenges today’s market new product families. It can be observed that classical existing
in Numerical
reality is methods,
limited in byparticular
several theideal assumptions and
most
tendencycommonly
exponentialtowards used regression
relationship
reduction between
of product models basedforce
the development
specific on times
empirical
and andthe product families are regrouped in functionfinite element
of clients method
or features.
data. Theproduct
undeformed Kienzle-force
chip lifecycles.
thickness. model
König represents [2] an
et al.there approximate simplifications.
(FEM) in machining processes, have been developed for
shortened In addition, isdetermined
an increasing the However, assembly oriented product families are hardly to find.
exponential relationship
forces between the specific force Numerical
decades and methods,
have achievedin particular
several the finite
breakthroughs element
in method
modeling
specific ofcutting
demand customization, c/kc1.1 atin
kbeing theturning
same time in aand
processes the
with
global On the product family level, products differ mainly in two
(FEM)
material in
andmachining
friction (i) processes,
behavior have years
in recent been [5–8].
developed
With for
undeformed
competition chip competitors
different workpiece
with thickness.
materials König
alland
overet al. [2]
cutting
the determined
tool
world. geometries.
This trend, the main characteristics: the number of components and (ii) the
the
specific decades
help of and
FEM have achieved
chip formationseveral breakthroughs
simulations, in
specific modeling
cutting
However,
which is cutting
inducing forces
the extrapolations kc/kare
the development in allowed
c1.1 not turning
from macro processes
in the toregion with
microof type of components (e.g. mechanical, electrical, electronical).
different results
workpiece materials and cutting material and friction behavior in recent years [5–8]. With the
markets, in diminished lot sizes duetool geometries.
to augmenting Classical methodologies considering mainly single products
However, help of FEM chip formation simulations, specific cutting
product
2212-8271 ©the
2019extrapolations
varieties (high-volume
The areto not
Authors. Published allowedB.V.
low-volume
by Elsevier in the region[1].
production) of or solitary, already existing product families analyze the
Peer-review
To cope with under responsibility
this augmenting of the scientific
variety as committee
well as to of The 17th CIRP
be able to Conference on Modelling
product structure on of Machininglevel
a physical Operations, in the person
(components of the
level) which
Conference Chair Dr Erdem Ozturk and Co-chairs Dr Tom Mcleay and Dr Rachid Msaoubi.
identify
2212-8271 possible
© 2019 The optimization
Authors. Publishedpotentials in the existing
by Elsevier B.V. causes difficulties regarding an efficient definition and
Peer-review under
production system,responsibility of the scientific
it is important to havecommittee
a preciseofknowledge
The 17th CIRP Conference
comparison on Modelling of Machining
of different product Operations, in theAddressing
families. person of the this
Conference Chair Dr Erdem Ozturk and Co-chairs Dr Tom Mcleay and Dr Rachid Msaoubi.
2212-8271 © 2019 The Authors. Published by Elsevier B.V.
Peer-review
2212-8271 under
© 2017 responsibility
The of the by
Authors. Published scientific
Elseviercommittee
B.V. of The 17th CIRP Conference on Modelling of Machining Operations
10.1016/j.procir.2019.04.031
Peer-review under responsibility of the scientific committee of the 28th CIRP Design Conference 2018.
Bingxiao Peng et al. / Procedia CIRP 82 (2019) 302–307 303
2 Author name / Procedia CIRP 00 (2019) 000–000

forces, chip forms, tool wear and surface integrity can be experimental and numerical data, to overcome the limitation of
modeled and predicted. Due to the extremely long computing data volume. The ML model uses a batch-supervised learning
time, 3D FEM cutting simulations are not yet suitable for real- system that enables continuous training of the network with
time analysis to determine the cutting forces. A coupling new data sets. The specific cutting forces kc and kcN with
between the CAD/CAM system and the 3D FEM cutting respect to cutting parameters and tool geometries and tool wear
simulations for turning and milling is thus not yet possible. As VB can be obtained from a 2D FEM model and a test bench in
a compromise, a validated 2D FE-model is an optimal solution orthogonal cutting. An Artificial Neural Network (ANN) with
to predict the specific cutting forces in terms of time and cost. multiple hidden layers, also called Deep Neural Network
Regardless of experimental, analytical and numerical data (DNN), learns with the specific cutting forces. In this study, a
sources, there is no suitable method for handling such a batch learning system is proposed. With new data sets from
complex nonlinear system as cutting processes in terms of time simulations or experiments, the network learns alone in a few
and accuracy, which are influenced by materials, tool iterations in order to integrate the new data set into the network
geometries, cutting parameters, etc. A conventional regression and then batch them into the major data sets. The ML model is
model is not sufficient for such a complex problem. To evaluated by cross validations with an evaluation set. The
overcome these challenges, data-driven modeling is becoming evaluated ML model allows the prediction of specific cutting
increasingly popular and is being studied to model and predict forces. Using multiscale modeling technologies, the trained
the cutting process [9–11]. ML model will be used in future work to model machining
Wu et al. [12] conducted a study on machine learning processes such as turning and milling on the macroscopic level.
algorithms to predict the tool wear in the milling process. 2D FEM chip
Orthogonal
Artificial neutral networks (ANNs), support vector regressions formation
cutting
(SVRs) as well as random forest (RF) were compared using simulation
experimental data. The cutting forces, the vibrations and the Batch Training
AE signal were labeled as the input. The predictive models Ʃ
showed promising results with those machine learning ℎ
Ʃ Ʃ
Deep Ʃ 𝑘𝑘𝑐𝑐
algorithms. According to their study, RF provided more 𝑐𝑐
Add data

Neural Ʃ
Ʃ Analyze
accurate predictions than ANNs and SVRs. However, the Network Ʃ
Ʃ Ʃ 𝑘𝑘𝑐𝑐 solution
predictive model did not consider the tool geometries and other (DNN) Ʃ
relevant cutting parameters. In addition, the used ANNs had Ʃ 1
1
only a single hidden layer, which are not sufficient. 1
Liu and Guo [13] presented an approach to integrate data-
Batch Training (a few iterations)
driven methods and process mechanics to predict the specific
cutting energy based on the experimental data from end More data from simulations and
Prediction
experiments
milling. The experiments were performed with different cutting
parameters and a single type of end mill. Different from classic
ANNs or SVRs, the authors used a decision tree as a machine Modeling machining processes on the
macroscopic level
learning method.
In summary, the related work presented in this section
shows a high potential of data-driven methods for predicting Fig. 1. Concept of the hybrid approach and the machine learning architecture
the machining system. The unique benefit of data-driven
methods is that a deep understanding of the complex nonlinear 3. Experimental and numerical data sources
machining system, which is extremely difficult to describe
using conventional regression models, is not a prerequisite. 3.1. Experimental inputs
However, the data volume in related works has significant
limitations. Without sufficient data and data volume, the data- In this paper, tests in orthogonal cutting were performed with
driven methods will not really be predictive. For this reason, the workpiece material Direct Aged 718 (DA 718) and the
this paper explored a hybrid approach combining the advantage cutting tool material EMT 815 (WC-15Co). The used
of experimental data and numerical data to predict specific workpiece had dimensions of 45 mm × 40 mm × 3 mm.
cutting forces considering the influence of the flank wear, Compared to the standard Inconel 718, DA 718 under the direct
cutting parameters and cutting tool geometries. A multi-layer aging heat treatment has the same chemical composition [14].
ANN architecture with four hidden layers was used for the However, the direct aging heat treatment leads to an increase of
machining learning algorithm. Furthermore, a batch learning tensile strength and improves the cyclic fatigue performance
system was used to integrate the numerical data continuously. [15]. There are many assumptions and theories to explain the
improvements. One of them is the finer grain micro-section due
2. Concept of the hybrid approach and the machine to the direct aging effect.
learning strategy The orthogonal cutting was performed on a test bench built
on the vertical broaching machine Forst RASX 8×2200×600
Fig. 1 shows the architecture of the hybrid approach using M/CNC, as shown in Fig. 2. Compared to the original vertical
machining learning (ML) algorithms. The novelty of this broaching machine, the workpiece was clamped in a
strategy is the application of hybrid data sources, including customized fixture. The grooving insert tool for the orthogonal
304 Bingxiao Peng et al. / Procedia CIRP 82 (2019) 302–307
Author name / Procedia CIRP 00 (2019) 000–000 3

cutting was clamped in the cutting tool holder, which was fixed and kcN is slightly different under wet and dry cutting
on a customized Kistler dynamometer Z21289. The cutting conditions. The cutting tool had almost the same or even
force Fc and the cutting normal force FcN were measured by the slightly better results in dry cutting condition than in wet
dynamometer. cutting condition. However, the characterization of tool wear is
more than the macroscopic flank wear VB. Using the SEM
1 6 analysis the micro chipping was identified in dry cutting
condition. Further tool life tests according to the Table 1 were
5 performed to obtain the data sources for the ML model.

3
L = 0.45 m L = 1.8 m L = 11.25 m
2 VB = 50 µm VB = 85 µm VB = 169 µm

Cutting force Fc / N Cutting normal force FcN / N


800
2 4
600
1 Broaching machine
2 Force measuring platform 400
3 Workpiece (45 mm 40 mm 3 mm) L = 0.45 m
200 L = 1.8 m
4 Cutting tool and tool holder
L = 11.25 m
5 LED Ring light (GSVitec MultiLed R) 0
0 0.1 0.2 0.3 0 0.1 0.2 0.3
6 Highspeed Camera (Vision Research Phatom v7.3) Time t / s Time t / s
Workmaterial: Diret Aged 718 Undeformed chip thickness: h = 0.02 mm
Fig. 2. Experimental setup Process: Orthogonal cutting Depth of cut: ap = 3 mm
Cutting tool: WC-15Co Rake angle:  = 10
As shown in Table 1, tool life tests were performed in Cutting condition: Dry Clearance angle: α =3
orthogonal cutting to determine the cutting force Fc and the Cutting speed: vc = 10 m/min Cutting edge radius: rβ = sharp
cutting normal force FcN, depending on the undeformed chip
thickness h, the cutting speed vc and the flank wear VB Fig. 3. Cutting force components Fc / FcN and flank wear VB in tool life
measured by an optical microscope Keyence.
12000
Specific cutting forces

Table 1. Parameters for experiments and simulations in orthogonal cutting


kc and kcN / N/mm2

9500
Cutting speed vc / m/min 10 / 30
7000 kc in wet condition
Undeformed chip thickness h / mm 0.02 / 0.05
kc in dry condition
Rake angle /° -12 / -6 / 0 / 5 / 10 4500 kcN in wet condition
Clearance angle α/° 3 kcN in dry condition
2000
Cutting edge radius rβ / µm sharp / 20 / 35 250
Flank wear VB / µm 0 to 200 200
Flank wear
VB / µm

Fig. 3 shows the measured signal of Fc, FcN and the measured 150
flank wear VB at beginning, middle and end of the tool life. 100
The flank wear VB was measured with a digital microscope In wet condition
VHX-5000. The flank wear was distributed on the flank face 50
In dry condition
uniformly. In addition, the tool wear mechanisms were 0
analysed by SEM measurement. As expected, adhesion and 0 4 8 12
abrasion dominated the wear mechanisms of the cutting tools. Cutting length L / m
Due to the limitation of the number of pages, the tool wear Workmaterial: Diret Aged 718 Undeformed chip thickness: h = 0.02 mm
Process: Orthogonal cutting Depth of cut: ap = 3 mm
development will be discussed in detail in a separated paper. Cutting tool: WC-15Co Rake angle:  = 10
The cutting forces increased with increasing VB due to the Cutting condition: Dry / Wet Clearance angle: α =3
additional mechanical contributions at the worn flank surface. Cutting speed: vc = 10 m/min Cutting edge radius: rβ = sharp
The specific cutting forces kc and kcN were calculated with
respect to Fc and FcN in the stationary state as a function of VB: Fig. 4. Specific cutting forces kc / kcN and flank wear VB in tool life
𝐹𝐹𝑐𝑐 (𝑉𝑉𝑉𝑉) 𝐹𝐹𝑐𝑐𝑐𝑐 (𝑉𝑉𝑉𝑉)
𝑘𝑘𝑐𝑐 = ; 𝑘𝑘𝑐𝑐𝑐 = (2)
𝑏𝑏∙ℎ 𝑏𝑏∙ℎ 3.2. Numerical inputs
Fig. 4 shows the development of the specific cutting forces and
Compared to the FE-model based on the pure Lagrangian
flank wear VB depending on the cutting length L under finish
cutting parameters (h = 0.02 mm and vc = 10 m/min). The formulation, a FE-model based on the coupled Eulerian-
specific cutting forces increased with increasing VB. The kc Lagrangian (CEL) formulation was used for the chip formation
Bingxiao Peng et al. / Procedia CIRP 82 (2019) 302–307 305
4 Author name / Procedia CIRP 00 (2019) 000–000

simulation in this study, as shown in Fig. 5. The mesh for the function of flank wear VB. The FE-model shows a precise
workpiece is generated and fixed in the Eulerian domain. The prediction of the specific cutting forces depending on the tool
usage of the Eulerian method is based on the volume-of-fluid wear. Using the experimental and numerical inputs as the
method. By means of this method, the material point is tracked hybrid data sources, 540 data sets under different cutting
as it flows through the mesh by computing its Eulerian volume speeds, undeformed chip thicknesses, rake angles, clearance
fraction (EVF) within each element. angles, cutting edge radii and flank wear conditions were
generated for the ML model in the next section.
γ Cutting tool 8000

Specific cutting force


h rβ
VB 6000
Lagrangian

kc / N/mm2
domain
w = -2.5 4000

2000
FcN vc Initial α
material Eulerian 0
Fc domain 12000

Specific cutting normal


Temperature T / C: Stress degradation D 9000

force kcN / N/mm2


610 1
6000

3000 Simulation
0 0 Experiment
0
0 50 100 150 200 250
Flank wear VB / µm
Fig. 5. Concept of the coupled Eulerian-Lagrangian (CEL) based FEM chip
formation simulation considering the flank wear Workmaterial: Diret Aged 718 Undeformed chip thickness: h = 0.05 mm
Process: Orthogonal cutting Depth of cut: ap = 3 mm
Cutting tool: WC-15Co Rake angle:  = -12
The concept of the FE-model was achieved with Cutting condition: Dry Clearance angle: α =3
ABAQUS/Explicit 6.14. The used Eulerian element type Cutting speed: vc = 30 m/min Cutting edge radius: rβ = 20 µm
EC3D8RT enables the existence of multi-materials within one
element and the thermo-mechanical analysis. Furthermore, the Fig. 6. Validation of the FEM simulation
cutting tool geometry was characterized with respect to the rake
angle , the clearance angle α and the cutting edge radius rβ. 4. Prediction of the cutting process using machine learning
Using the element type C3D8RT, the cutting tool was modeled
with the Lagrangian formulation and was constrained as a rigid 4.1. Machining learning with Deep Neural Network (DNN)
body. The cutting tool acting as an obstacle to force the flowing
workpiece separating into chip and machined workpiece. A In order to handle complex nonlinear systems such as the
temperature dependent friction model defined the tool-chip cutting process, building an intelligent machine using artificial
interface between the Lagrangian element and the Eulerian neutral networks according to the architecture of the brain was
element. Fig. 5 also shows the application of flank wear in the inspired [17]. In this study, a multi-layer perceptron (MLP)
FE-model. The worn tool topography was measured by the with multiple layers of linear threshold units (LTUs), also
laser scanning microscope Keyence VK-X150 for 3D imaging called hidden layers, was built. A LTU is an artificial neuron
and surface profiling. The flank wear profile was characterized and computes a weighted sum of inputs and a bias neuron.
by a width of flank wear land VB and a negative clearance
angle w, calculated by a script in the MATLAB program. The 𝑧𝑧 = 𝑤𝑤1 ∙ 𝑥𝑥1 + 𝑤𝑤1 ∙ 𝑥𝑥1 + ⋯ = 𝒘𝒘𝑡𝑡 ∙ 𝒙𝒙 (3)
negative clearance angle w occurred by non-constant wear
conditions along the cutting contact area, which was also The perceptron can be trained using a variant of the rule that
observed by other researches [16]. The negative clearance takes account of the error made by the network. For every
angle ensures the contact condition between the flank face and output neuron that produced a wrong prediction, it reinforces
the machined workpiece. In this study, a negative clearance the connection weights from the inputs that would have
angle w = -2.5° was implemented into the FE-model by directly contributed to the correct prediction, as shown in Eq. 4. The
changing the cutting tool geometry. Furthermore, the Johnson- factor η is the learning rate, which influences the convergence
Cook constitutive material model and the Johnson-Cook and stability of the learning process.
damage model have been used to model the serrated chip (𝑛𝑛𝑛𝑛𝑛𝑛𝑛𝑛 𝑠𝑠𝑠𝑠𝑠𝑠𝑠𝑠)
𝑤𝑤𝑖𝑖,𝑗𝑗 = 𝑤𝑤𝑖𝑖,𝑗𝑗 + 𝜂𝜂 ∙ (𝑦𝑦𝑗𝑗 − 𝑦𝑦̂𝑗𝑗 ) ∙ 𝑥𝑥𝑖𝑖 (4)
formation of DA 718. More details about the used material
models, the temperature dependent friction model as well as the In addition, the backpropagation training algorithm or called
validation of the FE-model can be found in previous work [8]. Gradient Descent using reverse-mode was applied for training
Fig. 6 shows the comparison of the experiments and MLP. The algorithm measures the output error of the network.
simulations with respect to the specific cutting forces as a
306 Bingxiao Peng et al. / Procedia CIRP 82 (2019) 302–307
Author name / Procedia CIRP 00 (2019) 000–000 5

Then it computes how much each neuron in the last hidden reaches the input layer. This reverse pass efficiently measured
layer contributed to each output error. It proceeds to measure error gradient across all the connection weights in the network
how much of these error contributions came from each neuron by propagating the error gradient backward in the network [17].
in the previous hidden layer and so on until the algorithm
18000 18000
Specific cutting force Specific cutting force
Predicted kc / N/mm2

Predicted kc / N/mm2
12000 12000
R2 = 0.971 R2 = 0.809

6000 6000
Experiment Experiment
Simulation Simulation
0 0
0 6000 12000 18000 0 6000 12000 18000
Experimental and simulated kc / N/mm2 Experimental and simulated kc / N/mm2
18000 18000
Specific cutting normal force Specific cutting normal force
Predicted kcN / N/mm2

Predicted kcN / N/mm2


12000 R2 = 0.977 12000 R2 = 0.767

6000 6000

0 0
0 6000 12000 18000 0 6000 12000 18000
Experimental and simulated kcN / N/mm2 Experimental and simulated kcN / N/mm2
a) Machine learning with Deep Neural Network b) Linear Regression

Fig. 7. Comparison of original and predicted specific cutting forces kc / kcN

In order for the gradients to work properly, especially for VB as discussed in section 3.1. The reason for this is that the
nonlinear cases, the activation function is used for the tool topography becomes unstable at higher tool wear due to
backpropagation algorithm. In this paper, the ReLU function the chipping, which also influences the simulative results. In
was used, which is continuous except at z = 0. addition to learning rates, activation functions, number of
layers, number of neutrons, weight initializations, scale
𝑅𝑅𝑅𝑅𝑅𝑅𝑅𝑅(𝑧𝑧) = max(0, 𝑧𝑧) (5) methods, mini-batch sizes, momentum and training iterations
are all relevant factors (hyperparameters) that influence the
The machine learning model in this study was achieved with learning process and thus the prediction. A sensitivity analysis
TensorFlow developed by Google. It offers plenty of low-level for part of the hyperparameters with respect to the Root Mean
and high-level Python APIs to build and train neural networks. Square Error (RMSE) was conducted in the following sections
Using the proposed ML architecture in Fig. 1, the DNN with
four hidden layers and forwards neurons 10 / 8 / 8 / 6 for each 4.2. Sensitivity analysis of the hyperparameters of DNN
layer learned with the hybrid data sources inclusive 15 % of
evaluation set was built. Instead of using the original data, the As explained in the previous section, the ML model with
data was pre-processed with a standard scaler in the input layer. DNN is not a black box. It can be affected by structures and
The ML model was trained with 10,000 iterations. The multiple hyperparameters. In this section, a sensitivity analysis
comparison of the hybrid data and the predict data (540 data has been performed with respect to the hyperparameters:
sets) using DNN is shown in Fig. 7. In addition, a reference evaluation set (5 % / 10 % / 25 %), number of hidden layers
group using a conventional linear regression model is also (3 / 4 / 5), iterations (5000 / 15000 / 20000) and number of
shown in Fig. 7. With respect to the coefficient of neurons per layer (6 / 8 / 10). The training plan was designed
determination R2, the prediction of the specific cutting forces and analyzed according to the Taguchi method. Since the
with the ML model (R2 = 0.971 / 0.977) is significantly better evaluation set was randomly selected, the ML model was
than with the conventional linear regression model trained three times repeatedly. The average RMSE of the scaled
(R2 = 0.809 / 0.767). It is also found that the deviation of ML data was used as the criterion. Fig. 8 shows a convergence
model increased with higher mechanical loads caused by larger curve of the ML model with respect to the RMSE of the scaled
Bingxiao Peng et al. / Procedia CIRP 82 (2019) 302–307 307
6 Author name / Procedia CIRP 00 (2019) 000–000

data. Regarding to the convergence, the RMSE did not change Furthermore, more FEM simulations are running to train the
much from the training iterations of 15,000. ML model continuously.
RMSE of scaled data / -

1.0 Acknowledgements
0.8
0.6 The authors would like to thank the German Research
0.4 Foundation (DFG) for the funding of the depicted research
within the project “Modelling of broaching processes by multi-
0.2
scale discretization” (KL 500/159-1).
0.0
0 5000 10000 15000 20000 References
Training iteration n / -
[1] Klocke, F. (2011): Manufacturing process. (Reihe: RWTH edition) Berlin:
Fig. 8. Convergence of the ML-model (5 % test set; 5 hidden layers; 20000 Springer
iterations; 10 neutrons per layer) [2] König, W.; König, W. (1973): Spezifische Schnittkraftwerte für die
Zerspanung metallischer Werkstoffe // Specific cutting force data for
According to the main effect diagram in Fig. 9, the increase
metal-cutting: Verein Deutscher Eisenhüttenleute Düsseldorf: Verl.
of hidden layers, iterations and neutrons can increase the Stahleisen
accuracy of the ML model. Although their influence on [3] Merchant, M. E. (1945): Merchant 1945 - Mechanics of the metal cutting
learning time is not significant for the case with 504 data sets process I. In: Journal of Applied Physics
in this study, the computing time may increase if the data [4] Oxley, P. L. B.; Oxley, P. L. B. (1989): The Mechanics of Machining: An
analytical approach to assesing machninability. (Reihe: Ellis Horwood
volume increases to a much higher level. In addition, the 15 %
series in mechanical engineering) Chichester: Horwood
evaluation set had the best result in terms of RMSE. However, [5] Gonzalo, O.; Jauregi, H.; Uriarte, L. G.; López de Lacalle, L. N. (2009):
there was no clear correlation between the RMSE and the test Prediction of specific force coefficients from a FEM cutting model. In: The
set. The selection of the test set must be careful, since a small International Journal of Advanced Manufacturing Technology. Jg. 43, Nr.
test set can lead to overfitting. 3-4, S. 348–356
[6] Arrazola, P. J.; Özel, T.; Umbrello, D.; Davies, M.; Jawahir, I. S. (2013):
0.2 Recent advances in modelling of metal machining processes. In: CIRP
RMSE of scaled data / -

Annals - Manufacturing Technology. Jg. 62, Nr. 2, S. 695–718


[7] Melkote, S. N.; Grzesik, W.; Outeiro, J.; Rech, J.; Schulze, V.; Attia, H.;
0.175 Arrazola, P.-J.; M’Saoubi, R.; Saldana, C. (2017): Advances in material
and friction data for modelling of metal machining. In: CIRP Annals. Jg.
0.15 66, Nr. 2, S. 731–754
[8] Klocke, F.; Döbbeler, B.; Peng, B.; Schneider, S.A.M. (2018): Tool-based
0.125 inverse determination of material model of Direct Aged Alloy 718 for FEM
cutting simulation. In: Procedia CIRP. Jg. 77, S. 54–57
0 10 20 30 2 3 4 5 6 6000 180004 6 8 10 12
[9] Gouarir, A.; Martínez-Arellano, G.; Terrazas, G.; Benardos, P.; Ratchev,
Test set Hidden layers Iterations Neurons
S. (2018): In-process Tool Wear Prediction System Based on Machine
/% /- /- /- Learning Techniques and Force Analysis. In: Procedia CIRP. Jg. 77, S.
501–504
[10] Arisoy, Y. M.; Özel, T. (2014): Machine Learning Based Predictive
Fig. 9. Main effect diagram Modeling of Machining Induced Microhardness and Grain Size in Ti–6Al–
4V Alloy. In: Materials and Manufacturing Processes. Jg. 30, Nr. 4, S. 425–
5. Conclusion and future work 433
[11] Chen, J. C.; Chen, J. C. (2005): An artificial-neural-networks-based in-
process tool wear prediction system in milling operations. In: The
In this paper, a hybrid approach using a supervised machine International Journal of Advanced Manufacturing Technology. Jg. 25, Nr.
learning model (ML) for predicting the specific cutting forces 5-6, S. 427–434
was introduced. A batch system using a deep neutron network [12] Wu, D.; Jennings, C.; Terpenny, J.; Gao, R. X.; Kumara, S. (2017): A
(DNN) was used to realize the continuous training of new data Comparative Study on Machine Learning Algorithms for Smart
sets. Experiments in orthogonal cutting and 2D FEM chip Manufacturing. Tool Wear Prediction Using Random Forests. In: Journal
of Manufacturing Science and Engineering. Jg. 139, Nr. 7, S. 71018
formation simulations were utilized as hybrid data sources. In [13] Liu, Z.; Guo, Y. (2018): A hybrid approach to integrate machine learning
this study, the following scientific contributions were made: and process mechanics for the prediction of specific cutting energy. In:
1. Compared to the conventional linear regression model, CIRP Annals. Jg. 67, Nr. 1, S. 57–60
an accurate prediction of the specific cutting forces was [14] Krueger, D. D. (1989): The Development of Direct Age 718 for Gas
achieved by using the machine learning model. Turbine Engine Disk Applications. In: Superalloy 718: Metallurgy and
Applications. Jg. 1989, S. 279–296
2. Machine learning and DNN are not black boxes. The [15] Oberwinkler, B. (2016): Integrated Process Modeling for the Mechanical
selection of hyperparameters and structures has a Properties Optimization of Direct Aged Alloy 718
significant influence on the results. [16] Lorentzon, J.; Järvstråt, N. (2008): Modelling tool wear in cemented-
The ML model shows an enormous potential in predicting carbide machining alloy 718. In: International Journal of Machine Tools
cutting processes. In addition to the specific cutting forces, the and Manufacture. Jg. 48, Nr. 10, S. 1072–1080
[17] Géron, A. (2017): Hands-on machine learning with Scikit-Learn and
ML model can also be used to predict the tool wear in-process TensorFlow. Concepts, tools, and techniques to build intelligent systems
depending on the mechanical loads in future work. Beijing: O'Reilly

You might also like