See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/221786413

Adaptive Control Optimization of Cutting Parameters for High

Quality Machining Operations Based on Neural Networks and
Search Algorithms

Chapter · October 2008

DOI: 10.5772/5539 · Source: InTech

CITATIONS READS
19 2,167

5 authors, including:

Jose Vicente Abellan-Nebot Hector R. Siller

Universitat Jaume I University of North Texas
62 PUBLICATIONS   649 CITATIONS    92 PUBLICATIONS   558 CITATIONS   

SEE PROFILE SEE PROFILE

Antonio Estruch Carlos Vila

Universitat Jaume I Universitat Politècnica de València
20 PUBLICATIONS   131 CITATIONS    53 PUBLICATIONS   351 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Sustainable Manufacturing View project

Scanning near-field optical microscopy View project

All content following this page was uploaded by Antonio Estruch on 28 May 2014.

The user has requested enhancement of the downloaded file.

 1

Adaptive Control Optimization of Cutting

Parameters for High Quality Machining
Operations based on Neural Networks and
Search Algorithms
J. V. Abellan, F. Romero, H. R. Siller, A. Estruch and C. Vila
Department of Industrial Systems Engineering and Design
Castellon, 12071, Spain

1. Introduction
In traditional Computer Numerical Control (CNC) systems, machining parameters are
usually selected prior to machining according to handbooks or user’s experience. These
practices tend to select conservative parameters in order to avoid machining failure and
assure product quality specifications. Less conservative practices try to find optimal
machining parameters off-line to increase process productivity after conducting
experimentation (Chien & Chou, 2001). However, variations during the machining process
due to tool wear, temperature changes, vibrations and other disturbances make inefficient
any off-line optimization methodology, especially in high quality machining operations
where product quality specifications are very restrictive. Therefore, to assure the quality of
machining products, reduce costs and increase machining efficiency, cutting parameters
must be optimised in real-time according to the actual state of the process. This optimization
Open Access Database www.i-techonline.com

process in real-time is conducted through an adaptive control of the machining process.

The adaptive control applied in machining systems is classified as (Liang et al., 2004; Ulsoy
& Koren, 1989): Adaptive Control with Constraints (ACC), Geometric Adaptive Control
(GAC), and Adaptive Control with Optimization (ACO). In the ACC systems, process
parameters are manipulated in real time to maintain a specific process variable, such as
force or power, at a constraint value. Typically, ACC systems are utilized in roughing
operations where material removal rate is maximized by maintaining the cutting forces at
the highest possible cutting force such that the tool is not in danger of breaking (Zuperl et
al., 2005). In the GAC systems, the economic process optimization problem is dominated by
the need to maintain product quality such as dimensional accuracy and/or surface finish
(Coker & Shin, 1996). GAC systems are typically used in finishing operations with the
objective of maintaining a specific part quality despite structural deflections and tool wear.
Sensor feedback is often employed to measure surface roughness and dimensional quality
between parts and adjustments, so tool offsets and feed overrides can be adjusted for the
next part. In the ACO systems, machine settings are selected to optimize a performance
index such as production time, unit cost, etc. Traditionally, ACO systems have dealt with
adjusting cutting parameters (feed-rate, spindle speed and depth of cut) to maximise
Source: Advances in Robotics, Automation and Control, Book edited by: Jesús Arámburo and Antonio Ramírez Treviño,
ISBN 78-953-7619-16-9, pp. 472, October 2008, I-Tech, Vienna, Austria
2 Advances in Robotics, Automation and Control

material removal rate subject to constraints such as surface roughness, power consumption,
cutting forces, etc (Venu Gopal & Venkateswara Rao, 2003). Other ACO systems optimise a
multi-objective function which are more practical in industrial applications (Zuperl & Cus,
2005). For example, it is quite often to search the optimal cutting parameters to minimize the
cost of the operation, maximize the production rate and maximize the part quality. ACO
systems are basically composed of several units which integrate the machine-tool system
and the equipment required for acquiring real-time process measurements and adjusting the
cutting parameters. Fig. (1) shows a simplified scheme of a basic ACO system presented in
(Koren, 1983). Basically, the ACO system requires a sensor system which provides real-time
data for tool wear diagnosis and part quality prediction. The real-time data are used by
process models previously obtained from experimental data. Tool wear and part quality
models are used in the multi-objective function together with cutting parameters. An
optimizer unit is then applied for searching optimal cutting parameters, and the selected
parameters are sent to the CNC system.

Fig. 1. Adaptive Control Optimization (ACO) scheme adapted from (Koren, 1983).
Interesting works related to ACO systems can be found in (Liu & Wang, 1999; Liu et al.,
1999; Chiang et al., 1995). Liu (Liu & Wang, 1999) proposed an adaptive control system
based on two neural network models, a Back-Propagation Neural Network (BP NN) and an
Augmented Lagrange Multiplier Neural Network (ALM NN). The BP NN was used for
modeling the state of the milling system, using as a single input the feed parameter and
sensing the cutting forces on-line. The ALM NN was used for maximising the material
removal rate which it was carried out adjusting the feed rate. Chiang (Chiang et al., 1995)
presented a similar work for end-milling operations, but surface roughness was also
considered as constraint. Both research works were based on theoretical formulas for
training the neural networks, and both applied an ALM NN for optimization, which it is
claimed to be an approach that can greatly reduce processing time in comparison to
conventional optimal algorithms and make real-time control possible. Liu (Liu et al., 1999)
also extended his previous work with a new optimization procedure based on a Genetic
Algorithm (GA).
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 3

In spite of the potential application of ACO systems, their use in industry is limited due to
the non-existence of reliable on-line monitoring systems for tool wear diagnosis and quality
prediction (Azouzi & Guillot, 1997; Liang et al., 2004). Therefore, the optimal selection of
cutting parameters is usually done off-line for the cutting-tool life-cycle (Ghani et al., 2004;
Chien & Chou, 2001). The off-line parameters optimization is usually carried out through
short cutting experiments which are later used to obtain an empirical model which could be
optimized subjected to some constraints. Ghani (Ghani et al., 2004) optimized cutting
parameters using a Taguchi's Design of Experiments in end milling operations. With a
minimum number of trials compared with other approaches such as a full factorial design,
the methodology presented reveals the most significant factors and interactions during
cutting process which leads to choose optimal conditions. A similar methodology is
described in (Zhang et al., 2007). However, both methodologies do not permit to evaluate
quadratic or non-linear relations between factors, and the analysis is restricted to the levels
analysed in each factor. A more generic approach although more costly in experiments is
based on Response Surface Model (RSM) and Response Surface Model Optimization
(RSMO). Suresh (Suresh et al., 2002) used RSM for modeling the surface roughness as a first
and second-order mathematical model and the surface roughness optimization was carried
out through GA. Cus (Cus & Balic, 2003) also applied GA for optimising a multi-objective
function based on minimum time necessary for manufacturing, minimum unit cost and
minimum surface roughness. All the process models applied in his research were empirical
formulas from machining handbooks which were fitted through regressions. More complex
models have also been applied for surface roughness and tool wear modeling to optimise
off-line cutting parameters. Zuperl (Zuperl & Cus, 2003) also applied and compared feed-
forward and radial basis neural networks for learning a multi-objective function similar to
the one presented in (Cus & Balic, 2003). Choosing the radial basis networks due to their fast
learning ability and reliability, he applied a large-scale optimization algorithm to obtain the
optimal cutting parameters. Chien (Chien & Chou, 2001) applied neural networks for
modeling surface roughness, cutting forces and cutting-tool life and applied a GA to find
optimum cutting conditions for maximising the material removal rate under the constraints
of the expected surface roughness and tool life.
These previous works are off-line optimization methodologies which can be efficient
enough if tool wear effects have a minimal impact to surface roughness and/or a high
surface roughness quality is not required. Otherwise, an on-line optimization methodology
should be applied since optimal cutting conditions may vary during the cutting-tool life-
cycle due to tool wear effects on surface roughness. In this chapter, an ACO system is
presented for optimising a multi-objective function based on material removal rate, quality
loss function related to surface roughness, and cutting-tool life subjected to surface
roughness specifications constraint. The proposed system adjusts the cutting parameters
during the cutting-tool life-cycle in order to maximise in real-time the multi-objective
function. The core of the system is composed of three process models: a cutting-tool wear
model for diagnosing the state of the cutting tool, a surface roughness deviation model for
predicting the quality loss function and a cutting-tool life model. All models are developed
using artificial neural networks to model the non-linear relationships in machining
processes. Since the process models are black-box models, optimal cutting parameters are
obtained applying genetic algorithms and mesh adaptive direct search algorithms. The
proposed system is compared with 2 traditional methods for off-line cutting parameters
selection: (1) selection based on suggested cutting parameters from handbooks, and (2)
selection based on RSMO.
4 Advances in Robotics, Automation and Control

2. Experimental system
2.1 Machining process description
Machining hardened steels (hardness from 30 to 62 HRC) for moulds and dies with surface
roughness specifications less than 0.3 microns are commonly applied in industry, and
require costly and time-consuming traditional operations such as electro-discharge
machining or grinding. Recently, some research studies have reported the use of high
performance machining operations for these applications with important benefits as
reducing lead times and costs (Siller et al., 2008). However, tool wear process impacts
directly to surface roughness so optimal cutting parameters are difficult to obtain since they
vary according to cutting-tool state. Therefore, although high performance machining can
technically substitute grinding or electro-discharge machining, additional efforts should be
conducted in order to tune cutting parameters for an optimal machining. For these
applications, ACO techniques can improve the process significantly with respect to other
non-adaptive optimization techniques.
The machining process studied in this paper is presented in Fig. 2, and it consists of a face-
milling operation on workpieces of hardened AISI D3 steel (60 HRc) with dimensions
250x250 mm. The experiments were conducted on a CNC machining center suited for
mould and die manufacturing, and the cutting tool used was a face milling tool with Cubic
Boron Nitride (CBN) inserts. In order to generate a good surface finish and avoid run-out
problems, a single insert was mounted on a tool body with an effective diameter of 6.35 mm.

Fig. 2. Machining process analysed

2.2 Monitoring system description

A monitoring system to estimate on-line tool wear and surface roughness is required to
select the optimal cutting parameters according to the actual state of the machining process.
In this chapter, the monitoring system implemented is a multi-component sensor system
composed of a piezoelectric dynamometer, accelerometers and signal conditioners (Fig. 3).
Two acquisition boards were used for data acquisition. The first board, an Iotech DaqBook
112, was used for acquiring cutting forces from the dynamometer and it was configured for
a sample frequency of 3 kHz. A second board, an Iotech DaqBoard 3000, was used for
vibration signal acquisition from accelerometers and it was configured for a sample
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 5

frequency of 100 kHz. Cutting forces were amplified and filtered by a Kistler 5405 amplifier
configured with a low-pass filter of 300 Hz cut-off frequency. Vibration signals were
amplified by a PCB 482A22 amplifier. Root-mean-square of forces and vibrations were
calculated for each cutting pass at the cutting-location x = 175 mm for a 2 seconds data
acquisition. Surface roughness (Ra) was measured by a Mitutoyo Surftest 301 profilometer
at the cutting-tool locations x = 40 mm, x = 110 mm, x = 175 mm every cutting pass
(sampling length λ = c/l = 0.8 mm and number of spans n = 5). Cutting tool wear (Vb) was
measured by a stereo-microscope Nikon MZ12 after each face-milling pass every 250 mm
length of cut. Fig. 4 describes the machining process with the Ra and Vb sampling
procedure.

Fig. 3. Multi-component sensor system

Fig. 4. Machining process and surface roughness and tool-wear sampling.

3. Design of experiments
In order to compare cutting parameters optimization by RSMO and AI approaches, it is
necessary to carry out a Design of Experiments (DoE) to be useful for both. RSMO requires
classical designs of experiments such as Box-Wilson Central Composites Designs (CCD) or
Box-Behnken designs (Nist, 2006), in case that it is only considered linear and quadratic
effects. On the other hand, AI approaches require enough data for training and testing,
varying the factors in all its domain, but it does not require any specific DoE design.
The factors considered in the experimentation were the feed per tooth (fz) and the cutting
speed (Vc). The radial depth of cut (ae) was considered constant, with a value of 31.25 mm to
maximize the material removal rate. The axial depth of cut (ap) was defined as a constant
(0.4 mm) since the machining operation studied was a finishing operation. The minimal
experimentation required to apply RSMO with two factors is a face centered CCD with one
center point which is equivalent to a 23 full factorial design. For each experiment, the face-
6 Advances in Robotics, Automation and Control

milling operation was carried out until the cutting tool edge was worn (Vb higher than 0.3
mm, usual value for finishing operations (ISO 8688-1, 1989)) or the surface roughness was
outside specifications. Fig 5 shows the cutting conditions analysed and the order of the
cutting experiments.

Fig. 5. Design of Experiments and run order. Face Centered Central Composite Design.

4. Definition of the optimization problem

The machining economics problem consists in determining the optimal cutting parameters
in order to maximize/minimize an objective function. Typical objective functions to
optimize cutting parameters are “minimize unit production cost”, “maximize production
rate”, “maximize profit rate”, etc. On the other hand, several cutting constraints have to be
considered in machining economics, such as tool-life constraint, cutting force constraint,
power, stable cutting region constraint, chip-tool interface temperature constraint and
surface finish constraint (Cus & Balic, 2003).

4.1 Objective functions

Typically, three objective functions are considered in a cutting parameters optimization
problem: (1) Material Removal Rate (MRR), (2) surface roughness and (3) cutting-tool life.
MRR is a measurement of productivity, and it can be expressed by analytical derivation as
the product of the width of cut (w), the feed velocity of the milling cutter (F) and the depth
of cut (ap) (Eq. (1)). Surface roughness is the most important criterion for the assessment of
the surface quality, and it is usually calculated empirically through experiments. Some
research works directly use the empirical relationship presented in Eq. (2), where Vc and f
are the cutting speed and feed rate respectively and k, x1, x2, x3 are empirical coefficients.
Cutting-tool life is the other important criterion for cutting parameters selection, since
several costs such as cutting-tool replacement cost and cutting-tool cost are directly related
with tool life. The relation between the tool life and the parameters is usually expressed by
the well-known Taylor's formula presented in Eq. (3), where KT, α1, α2, α3 are empirical
coefficients.

(1)

(2)

(3)
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 7

However, for high quality machining operations using CBN cutting tools, both traditional
surface roughness and tool life equations may not provide a good estimation. Machining a
very low feed speeds produce that additional mechanisms influence the surface roughness
generation such as vibrations, engagement of the cutting tool, built up edge, etc. (Siller et al.,
2008). On the other hand, CBN tools have a different wear process than traditional cutting-
tools such as high speed steels, so Taylor's formula may not be directly applied (Trent &
Wright, 2000). For both reasons, other empirical models based on experimental data must be
applied instead of Eqs. (2,3).
For the case study presented in this chapter which is a high quality face milling operation
based on CBN tools, two alternative objective functions were applied. Instead of Ra model,
it is applied the quality loss function described by Eq. (4). Considering a desired Ra value,
the quality loss function is usually applied to estimate the cost of manufacturing with a
quality variation. The loss function is defined as:

(4)

where Δ = Ramax - Ratarget with Ramax the maximum Ra defined by specifications and Ratarget
the Ra desired; V2 is the mean squared deviation as V2 = ((Ratarget - y1)2 + … + (Ratarget -
yn)2)/n , with n the number of samples; and Arework is the part cost if the part is outside
specifications. On the other hand, instead of the traditional Taylor’s formula, it is applied an
empirical model learnt from the experimentation which is defined by the Eq. (5), where f is
the function learnt.

(5)

4.2 Multi-objective function

The optimization problem for the case study is defined as the optimization of a multi-
objective function which is composed of the objective functions defined by Eqs (1,4,5). Since
these objective functions are conflicting and incomparable, the multi-objective function is
defined using the desirability function approach. This function is based on the idea that the
optimal performance of a process that has multiple performance characteristics is reached
when the process operates under the most desirable performance values (Nist, 2006). For
each objective function Yi(x), a desirability function di(Yi) assigns numbers between 0 and 1
to the possible values of Yi, with di(Yi) = 0 representing a completely undesirable value of Yi
and di(Yi) = 1 representing a completely desirable or ideal objective value. Depending on
whether a particular objective function Yi is to be maximized or minimized, different
desirability functions di(Yi) can be used. A useful class of desirability functions was
proposed by (Derringer & Suich, 1980). Let Li and Ui be the lower and upper values of the
objective function respectively, with Li < Ui, and let Ti be the desired value for the objective
function. Then, if an objective function Yi(x) is to be maximized, the individual desirability
function is defined as

(6)
8 Advances in Robotics, Automation and Control

with the exponent w is a weighting factor which determines how important it is to hit the
target value. For w = 1, the desirability function increases linearly towards Ti; for w < 1, the
function is convex and there is less emphasis on the target; and for w > 1, the function is
concave and there is more emphasis on the target. If one wants to minimize an objective
function instead, the individual desirability function is defined as

(7)

Fig. (6) shows the individual desirability functions according to different w values. The
individual desirability functions are combined to define the multi-objective function, called
the overall desirability of the multi-objective function. This measure of composite
desirability is the weighted geometric mean of the individual desirability for the objective
functions. The optimal solution (optimal operating conditions) can then be determined by
maximizing the composite desirability. The individual desirability is weighted by
importance factors Ii. Therefore, the multi-objective function or the overall desirability
function to optimize is defined as:

(8)

with k denoting the number of objective functions and Ii is the importance for the objective
function I, where i = 1,2,…,k.

Fig. 6. Desirability functions according to the type of objective function

4.3 Constraints
Due to the limitations on the cutting process, manufacturers limit the range of the cutting
parameters to avoid premature cutting-tool failures. Therefore, selected cutting parameters
according to manufacturer specifications are constrained to:

Vmin ≤ Vc ≤ Vmax (9)

fmin ≤ fz ≤ fmax (10)

ap ≤ amax (11)
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 9

Surface roughness specification is also considered a constraint that can be expressed as

Ra ≤ Rspec Æ V2 ≤ (Ratarget - Raspec)2 (12)

In addition, cutting power and force limitations are usual constraints, but they are
commonly applied only for roughing operations.

4.4 Summary of optimization problem and numerical coefficients

The weights and the individual desirability coefficients for each objective function were
chosen according to each objective function in the machining process. First the weights were
defined considering how the objective function increases/decreases as the ideal value is not
matched. Secondly, a comparison among individual desirability coefficients was done to
define how much more important is each objective function than the other one. For the case
study presented, the objective functions were considered linear (w=1) and the coefficient of
importance were chosen to prevail productivity and surface roughness quality than cutting-
tool cost and cutting-tool cost replacements. Therefore, importance factors I1 and I2 which
are related to material removal rate and surface quality loss function were chosen as 1,
whereas importance factor I3 which is related to cutting-tool life was chosen as 0.5.
Considering the maximum and minimum values of each objective function obtained
analytically, the desirability functions were defined as follows.
- MRR desirability function

(13)

MRRtarget = 2387 mm3/min. MRRminimum = 398 mm3/min. Importance factor I1=1.

- Desirability function of Ra deviation objective function

(14)

V2target = 0.0001 µm2. V2maximum = 0.012 µm2. Importance factor I2=1. Note that the desirability
function of quality loss W for surface roughness can be defined by the surface roughness
deviation V2 since Eq. (4) relates W with V2 by a constant coefficient of Arework/Δ2.
- Cutting-tool life desirability function

(15)

Ttarget = 46.7 min. Tminimum = 7.43 min. Importance factor I3=0.5.

The multi-objective function or the overall desirability function to be optimized is:

(16)

constrained to:

100 m/min ≤ Vc ≤ 200 m/min (17)

0.04 mm/rev ≤ fz ≤ 0.12 mm/rev (18)

10 Advances in Robotics, Automation and Control

ap = 0.4 mm (19)

V2 ≤ (0.1 – 0.2)2 µm2 (20)

5. Parameter optimization based on handbooks

5.1 Description
Cutting tool parameters are traditionally chosen according to handbooks and cutting-tool
data catalogs. For a given cutting-tool and workpiece material, a range of possible cutting-
parameters are provided. The machinist chooses the parameters within the ranges using
some well-known practices in shop-floor. Some of these practices are:
- Higher cutting speeds increase surface roughness quality but decrease cutting tool life.
- Higher cutting speeds decrease cutting tool life.
- Higher feed rates increase productivity as material removal rate is increased.
- Higher feed rates decrease surface roughness quality.
- Higher feed rates decrease cutting-tool life.
- Higher axial depth of cut increases productivity.
- Higher axial depth of cut decreases cutting-tool life.
- Very low axial depth of cut burns the workpiece surface and generates a low surface
roughness quality and decreases cutting-tool life.
According to the final goal of the machining process, the machinist selects the best cutting-
tool parameters combination. For example, if the only important constraint is a high cutting
tool life, the machinist will select a low cutting speed, low feed rate and low-medium axial
depth. Fig. 7a describes the typical optimization process based on handbooks.

Fig. 7. Off-line cutting parameters optimization.

5.2 Optimization results

As it was explained above (section 4.4), the MRR is the most important objective function
together with product quality whereas tool life is less important (I1=1; I2=1; I3=0.5). A high
feeding value increases MRR but decreases surface roughness, so it seems reasonably to fix
the feeding value to an intermediate value. On the other hand, as Vc increases, both MRR
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 11

and part quality increase, but cutting tool life decreases considerably. As both MRR and part
quality are much important than cutting tool life, it seems reasonably to fix Vc at its
maximum value. Therefore, the optimal cutting parameters based on catalogs and
handbooks can be defined as Vc=200 m/min and fz=0.08 mm. Experimentation was
conducted in order to check the overall desirability function at these cutting conditions. The
experimental results showed a cutting tool life T=10.8 min, an average surface roughness
deviation of V2=1.6·10-3 µm2, and a MRR=1592 mm3/min. The overall desirability was 0.472,
and the evolution of the overall desirability function due to surface roughness variability
along cutting tool life-cycle is shown in Fig (8).

Fig. 8. Overall desirability function along cutting tool life-cycle. Parameter optimization
based on handbooks. Vc=200m/min; fz=0.08mm; ap=0.4mm

6. Parameter optimization based on RSMO

6.1 Description
A less conservative method to optimize cutting parameters could be carried out through the
Response Surface Model Optimisation methodology (RSMO). RSMO is a collection of
mathematical and statistical techniques that are useful for the modeling and analysis of
problems in which a response of interest is influenced by several variables and the objective
is to optimize this response (Montgomery, 2001). In most RSM problems, the form of the
relationship between the variable to be optimized (the response) and the independent
variables is unknown. Thus, the first step in RSM is to find a suitable approximation for the
true functional relationship between the response and the set of independent variables. First
and second-order models are commonly applied to approximate the function through least
squares. Considering a response surface y defined by y=f(x1,x2…xk) where x1, x2, …, xk are
independent variables, the optimal value of the response surface will be the set of x1, x2, …,
xk for which the partial derivatives This point is called the stationary
point and it could represent a point of maximum response, a point of minimum response, or
a saddle point. A general mathematical solution for the location of the stationary point can
be obtained as follows. Given a second-order model in matrix notation as:

(21)
where

(22)
12 Advances in Robotics, Automation and Control

That is, b is a (k x 1) vector of the first-order regression coefficients and B is a (k x k)

symmetric matrix whose main diagonal elements are the pure quadratic coefficients
( βˆi i ) and whose off-diagonal elements are one-half the mixed quadratic coefficients ( βˆi j , i ≠ j).
The derivative of ŷ with respect to the elements of the vector x equated to 0 is

(23)

The stationary point is the solution to Eq. (23), or

(24)

To define whether the stationary point is a point of maximum or minimum response or

saddle point, it is usually examined a contour plot of the fitted model. Refer to
(Montgomery, 2001) for RSMO concepts.
In this section the RSMO methodology is applied to the high quality machining operation
studied. The experimentation conducted through the DoE is used to obtained first and
second order process models. The multi-objective function which is based on these models
is optimized and the optimal cutting parameters are defined. Fig. 7b describes the
optimization process based on RSMO in the machining process studied.

6.2 Machining process models based on first and second-order functions

RSM technique lets model the responses of interest W, T, and MRR as a first and second-
order functions. After conducting the DoE shown in Fig. (5), a RSM was fitted for each
response. The first and second order functions with its coefficient of determination were:
- Ra deviation (V2) response:

(25)

- T response:

(26)

- MRR response

(27)
R2adj=100%.Note that this is the exact analytical equation for MRR.
Fig. 9 shows the response surface models for each response analysed.

6.3 Optimization results

The optimum cutting parameters were obtained maximizing the overall desirability
function, which depends on the previous first and second order models and the individual
desirability and weights coefficients defined in section 4.4. The result of the optimization
procedure showed the optimal cutting parameters V=165m/min and fz=0.12mm, with an
overall desirability value of 0.61. To validate the RSMO result, experimentation was
conducted with the expected optimal parameters. The experimental results showed a
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 13

cutting tool life T=13.7 min, a surface roughness deviation function of V2=5.5·10-3 µm2, and a
MRR=1967 mm3/min. The overall desirability of the experiment was 0.495, and the
evolution of the overall desirability function due to surface roughness variability along
cutting tool life-cycle is shown in Fig (10). Note that the expected overall desirability was
0.610 whereas the real value was 0.495. This error is due to the inaccuracy of the process
models developed through this methodology.

Fig. 9. Response Surface Models for deviation of Ra, T and MRR. Units: Vc (m/min); fz
(mm); Ra deviation (µm2); T (min); MRR (mm3/min).

Fig. 10. Overall desirability function along cutting tool life-cycle. Parameter optimization
based on RMSO. Vc=164m/min; fz=0.12mm; ap=0.4mm.

7. Parameter optimization based on AI

7.1 Description
Due to non-linearity, first and second order models from response surface methodology
cannot be enough accurate to model tool-wear and surface roughness in machining
processes. To overcome this limitation, AI techniques such as Artificial Neural Networks or
Fuzzy Logic can be applied in order to deal with non-linearity. However, these AI models
are called black-box models, and they cannot be optimized with conventional optimization
methods. Due to this limitation, a cutting parameter optimization methodology based on AI
models requires an advanced search methods for global optimization, such as Genetic
Algorithms (GA) and Mesh Adaptive Direct Search (MADS) algorithms.
GAs are search algorithms based on the mechanics of natural selection and natural genetics,
invented by (Holland, 1975), which can find the global optimal solution in complex
multidimensional search spaces. A population of strings, representing solutions to a
specified problem, is maintained by the GA. The GA then iteratively creates new
populations from the old by ranking the strings and interbreeding the fittest to create new
strings. So in each generation, the GA creates a set of strings from the previous ones,
occasionally adding random new data to keep the population from stagnating. The end
result is a search strategy that is tailored for vast, complex, multimodal search spaces. GAs
14 Advances in Robotics, Automation and Control

are a form of randomized search, in the way in which strings are chosen and combined is a
stochastic process. This is a radically different approach to the problem solving methods
used by more traditional algorithms, which tend to be more deterministic in nature, such as
the gradient methods used to find minima in graph theory. However, although GA is an
effective optimization algorithm, it usually takes a long time to find an optimal solution due
to its slow convergence speed (Cus & Balic, 2003).
On the other hand, MADS algorithms are iterative search algorithms where the optimization
is conducted through an adaptive mesh of points where the objective function is evaluated.
At the first iteration, the mesh is built according to an initial point of the objective function.
The algorithm computes the objective function at the mesh points until it finds one whose
value is smaller than the objective function evaluated on the initial point. If a mesh point has
a smaller value, the algorithm sets the next point in the sequence equal to this one and
multiplies the current mesh size by a mesh expansion factor. The mesh is then expanded
and the algorithm conducts a new iteration. In case none of the mesh points has a smaller
objective function value than the value at the best current solution, the algorithm does not
change the current point at the next iteration and the mesh size is contracted by a mesh
contraction factor. After the re-size, a new iteration is conducted. The iterations are
conducted until a stop condition is reached, typically when the mesh size reaches a
minimum value. Refer to (Audet & Dennis, 2004) for concepts related to MADS.
In this section it is proposed an optimization methodology based on Artificial Neural
Networks (ANN) models for modeling the machining process and GA-MADS algorithms to
optimize the multi-objective function defined by Eq. (16) and the ANN models. The
combination of GA and MADS algorithms lets reduce the computing time required for the
optimization. Basically, GA is firstly applied in order to find the region where the multi-
objective function is minimum. Then, the GA algorithm is interrupted and the MADS
algorithm refines the search using the GA solution as the initial point of the mesh. The
optimal cutting parameters are calculated when the MADS algorithm reaches the minimum
mesh size. Fig. 11 describes the procedure of the optimization methodology proposed, based
on process models using AI techniques and an initial Design of Experiments, and a GA-
MADS optimization. Table (1,2) defines the main characteristics of the ANN models applied
and the GA-MADS algorithms.

Fig. 11. On-line cutting parameters optimization based on AI

Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 15

7.2 Machining process models based on AI

After conducting the DoE, an experimental database composed of 188 samples of 9 different
cutting conditions was generated. Each sample was defined by: surface roughness deviation
(V2), cutting tool wear state (% of use), root-mean-square of cutting forces in X, Y direction
(RMS Fx, RMS Fy), root-mean-square of cutting forces in XY plane (RMS Fxy), root-mean-
square of vibrations in X direction (RMS Ax), root-mean-square of vibrations in Y direction
(RMS Ay), cutting time (Tc), cutting speed (Vc) and feed per tooth (fz). The experimental
database was used to learn three ANN process models: (1) Ra deviation model, (2) Cutting-
tool life model, (3) Cutting-tool wear state model. Before training the ANN, a statistical
study was conducted for each model in order to discard those inputs variables that were not
significant. The final inputs variables applied for each model and the main characteristics of
the ANN models are shown in Table 1. Note that vibrations signals were discarded for all
models since they were close related to the cutting-tool position, and they cannot provide
any objective information about the state of the process. Fig 12 shows the response of each
ANN model according to Vc and fz values.

Cutting-tool wear state

Ra deviation model Cutting-tool life model
model
Type Backpropagation Type Backpropagation Type Backpropagation
Inputs Vc, fz, State Inputs Vc, fz Inputs Vc, fz, RMS Fxy
Output Ra_deviation Output T_life Output State
Hidden Hidden Hidden
1 1 1
Layers Layers Layers
Neurons 3 Neurons 3 Neurons 2
Mapping Mapping Mapping
Logsig Logsig Logsig
functions functions functions
Training Training Training
Lev-Marq. Lev-Marq. Lev-Marq.
Method Method Method
Epochs 300 Epochs 300 Epochs 300
Learning Learning Learning
0.05 0.05 0.05
Rate Rate Rate
Table 1. Characteristics of ANN models

Fig. 12. ANN models of : (1) surface roughness deviation -V2- for a cutting tool wear of 50%;
(2) cutting tool life –T-; (3) tool wear state for a RMS Fxy of 60N.
16 Advances in Robotics, Automation and Control

7.2 Optimization results

The optimum cutting parameters were obtained maximizing the overall desirability
function every cutting pass of 250 mm length of cut. During the cutting pass, the cutting
forces were acquired and the cutting-tool wear state was predicted by the cutting-tool wear
state model. With the cutting-tool state prediction, the GA-MADS optimization procedure
was conducted to maximize the overall desirability function, which is based on the MRR
model, and the ANN models of Ra deviation and cutting-tool life. The optimal solution was
obtained after an average processing time of 6 seconds, and the new cutting parameters
were sent to the CNC controller for the next cutting pass. The process was repeated until the
cutting tool was worn or the surface roughness was outside specifications. The result of the
optimization procedure showed that the optimal cutting parameters vary considerably,
from 200m/min to 130m/min and from 0.12mm to 0.07mm for cutting speed and feed rate
respectively, with an expected overall desirability value of 0.545. The real overall desirability
value after measure surface roughness deviation and tool-life was 0.520. The evolution of
the overall desirability function and the variation of the optimal cutting conditions along the
cutting tool life-cycle are shown in Fig (13).

Genetic Algorithm Mesh Adaptive Direct Search Algorithm

Variables to optimise Vc, fz Variables to optimise Vc, fz
Population Size 10 Initial Mesh Size 0.05
Generations 15 Max. Mesh Size Inf
Crossover Frac. 0.8 Max. Func. Eval Inf
Elite Count 2 Expansion 2
Mutation function Gaussian Contraction 0.5
Selection function Stochastic Poll Method Positive Basis 2N
Vc=[100,200]
Initial ranges Polling order Consecutive
Fz=[0.04,0.12]
Stall Time: 6s
Stop criterium Stop criterium Tolerance Mesh: 5·10-4
Stall Generations: 7
Table 2. Characteristics of search algorithms

8. Results and discussion

The results reported in the optimization methodology based on handbooks and catalogs
show the most conservative cutting parameters as it was expected, where the overall
desirability function has a value of 0.472. A simple RSMO through a 9 experimental runs
provide enough information to improve cutting parameter selection and increase the overall
desirability function to 0.495 which is an improvement of 5% in the desirability. The surface
response models show that high cutting speeds increase surface roughness variability which
were not taken into account by the machinist in the first optimization methodology.
Although RSMO increases the overall desirability, the RSMO prediction is quite inaccurate
as it is shown in the experimental validation. The predicted overall desirability by RSMO
was 0.610 whereas the overall desirability function after the experimental validation was
0.495. The inaccuracy of the prediction is due to the inaccuracy of the process models, where
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 17

surface roughness deviation and cutting-tool life models have been fitted by least squares
with a low coefficient of determination, 89% and 80.4% respectively.
The methodology proposed in this chapter, based on AI techniques for modeling the
machining process and the use of search algorithms to optimize the overall desirability
function on-line, shows an overall desirability function of 0.520. This methodology
compared with handbook optimization and RSMO increases the overall desirability in 10%
and 5% respectively. The main benefits reported by this methodology are due to two factors.
A first factor, the ANN process models let deal with non-linearity so this models are more
accurate than response surface models for modelling high quality machining operations.
Unlike RSMO where the error between the predicted overall desirability value and the
experimental one was 19% (0.61 versus 0.495), the proposed methodology presents an error
of 5% (0.545 versus 0.520). As a second factor, the on-line nature of the methodology lets
adapt the cutting parameters every cutting pass so the system is more flexible to adapt any
change in the objective function during the cutting-tool life. However, the ANN model for
cutting tool state prediction based on cutting forces seemed to be low accurate due to the
high variability of the cutting forces during machining. This variation produces that the
cutting tool parameters selected each cutting tool pass were quite irregular. This effect is
reflected in Fig. 13b, where cutting speed often varies from 195 to 150-130 m/min. The
replacement of the direct measurement of cutting forces by indirect methods which are not
so sensitive to cutting mechanisms such as current or power sensors might increase the
overall desirability function and select optimal parameters with a more regular variation.

Fig. 13. (a) Overall desirability function along cutting tool life-cycle. Parameter optimization
based on AI. (b) Adaptive cutting speed. (c) Adaptive feed per tooth
18 Advances in Robotics, Automation and Control

9. Conclusions
In this chapter, an Adaptive Control Optimization (ACO) system was presented for
optimising a multi-objective function based on material removal rate, quality loss function
related to surface roughness, and cutting-tool life subjected to surface roughness
specification constraints. Unlike traditional optimization techniques, this methodology lets
adapt the cutting parameters during the cutting-tool life-cycle in order to maximise in real-
time the multi-objective function according to cutting-tool state. The core of the system is
composed of three process models: a cutting-tool wear model for diagnosing the state of the
cutting tool, a surface roughness deviation model for predicting the quality loss function
and a cutting-tool life model. All models were developed using Artificial Neural Networks
(ANN) to model the non-linear relationships in machining processes. The cutting parameter
optimization is obtained applying genetic algorithms and mesh adaptive direct search
algorithms. The system proposed was compared with 2 traditional methods for off-line
cutting parameter selection: (1) selection based on suggested cutting parameters from
handbooks, and (2) selection based on RSMO.
The results showed how conservative are the cutting parameters selected by off-line
methodologies. A cutting parameter optimization based on handbooks provided an overall
desirability function of 0.472, whereas cutting optimization through RSMO gave an overall
desirability value of 0.495. The main inconvenient of RSMO is that this methodology is
based on mathematical first and second order models which are not enough accurate for
high quality machining operations. The new methodology proposed based on AI techniques
increases the overall desirability function up to 0.520. The improvement is due to two
effects: (1) the ANN process models deal with non-linearity so these models are more
accurate than response surface models for modelling high quality machining operations; (2)
the on-line nature of the methodology lets adapt the cutting parameters every cutting pass
so the system is more flexible to adapt any change in the objective function during the
cutting-tool life-cycle.

10. References
Audet, C. and Dennis, J. E. (2004). Mesh adaptive direct search algorithms for constrainted
optimization. Technical Report G-2004-04, Les Cahiers du GERAD, Montréal.
Azouzi, R. and Guillot, M. (1997). On-line prediction of surface finish and dimensional
deviation in turning using neural network based sensor fusion. International Journal
of Machine Tools and Manufacture, 37, 9, 1201–1217, (1997), ISSN 0890-6955.
Coker, S. A. and Shin, Y. C. (1996). In-process control of surface roughness due to tool wear
using a new ultrasonic system. International Journal of Machine Tools and
Manufacture, 36, 3, 411–422, (1996), ISSN 0890-6955.
Cus, F. and Balic, J. (2003). Optimization of cutting process by ga approach. Robotics and
Computer-Integrated Manufacturing, 19, 1-2, 113–121, (2003), ISSN 0736-5845.
Chiang, S. T. and Liu, D. I. and Lee, A. C. and Chieng, W. H. (1995). Adaptive-control
optimization in end milling using neural networks. International Journal of Machine
Tools and Manufacture, 35, 4, 637–660, (1995), ISSN 0890-6955.
Chien, W. T. and Chou, C. Y. (2001). The predictive model for machinability of 304 stainless
steel. Journal of Materials Processing Tech., 118, 1-3, (2001), 442–447, ISSN 0924-0136.
Adaptive Control Optimization of Cutting Parameters for High Quality Machining Operations
based on Neural Networks and Search Algorithms 19

Derringer, G. and Suich, R. (1980). Simultaneous-optimization of several response variables.

Journal of Quality Technology, 12, 4, 214–219, (1980). ISSN 0022-4065.
Ghani, A. K. and Choudhury, I. A. and Husni. (2002). Study of tool life, surface roughness
and vibration in machining nodular cast iron with ceramic tool. Journal of Materials
Processing Technology, 127, 1, 17–22, (2002), ISSN 0924-0136.
Ghani, J. A. and Choudhury, I. A. and Hassan, H. H. (2004). Application of Taguchi method
in the optimization of end milling parameters. Journal of Materials Processing
Technology, 145, 1, 84–92, (2004), ISSN 0924-0136.
Holland, J. (1975). Adaptation in natural and artificial systems. The University of Michigan
Press, Ann Arbor.
ISO –Int. Standard Organization-, (1989). Tool Life Test in Milling, Part 1: Face Milling. 1st Ed.
Jain, L. C. and Martin, N. M. (1998). Fusion of Neural Networks, Fuzzy Systems and Genetic
Algorithms: Industrial Applications. Crc Press edition, ISBN 0849398045.
Koren, Y. (1983). Comp. Control of Manufact. Systems. Mac.-Hill, ISBN 0070353417, New York.
Liang, S. Y. and Hecker, R. L. and Landers, R. G. (2004). Machining process monitoring and
control: The state-of-the-art. Journal of Manufacturing Science and Engineering-
Transactions of the Asme, 126, 2, (2004), 297–310, ISSN 1087-1357.
Liu, Y. M. and Wang, C. J. (1999). Neural network based adaptive control and optimization
in the milling process. International Journal of Advanced Manufacturing Technology, 15,
11, 791–795,(1999), ISSN 0268-3768
Liu, Y. M. and Zuo, L. and Wang, C. J. (1999). Intelligent adaptive control in milling
processes. International Journal of Computer Integrated Manufacturing, 12, 5, 453–460,
(1999), ISSN 0951-192X.
Montgomery, D. C. (2001). Design and Analysis of Experiments. John Wiley & Sons, Inc. ISBN
0471316497.
Nist. (2006). Nist/sematech e-handbook of statistical methods. Electronic Citation, 2006.
Siller, H. R. and Vila, C. and Rodriguez, C. A. and Abellan, J. V. (2008). Study of face milling
of hardened AISI D3 steel with a special design of carbide tool. International Journal
of Advanced Manufacturing Technology, 2008.
Suresh, P. V. S. and Rao, P. V. and Deshmukh, S. G. (2002). A genetic algorithmic approach
for optimization of surface roughness prediction model. International Journal of
Machine Tools and Manufacture, 42, 6, 675–680, (2002), ISSN 0890-6955.
Trent, E. M. and Wright, P. K. (2000). Metal Cutting. Butterworth Heinemann, ISBN
075067069X.
Ulsoy, A. G. and Koren, Y. (1989). Applications of adaptive control to machine tool process
control. Control Systems Magazine, IEEE., 9, (1989), 33–37, ISSN 0272-1708.
Venu Gopal, A. and Venkateswara Rao, P. (2003). Selection of optimum conditions for
maximum material removal rate with surface finish and damage as constraints in
SiC grinding. International Journal of Machine Tools and Manufacture, 43, 13, 1327–
1336, (2003), ISSN 0890-6955.
Wang, J. and Kusiak, A. (2000). Computational Intelligence in Manufacturing Handbook. CRC
Press edition. ISBN 0849305926.
Zhang, J. Z. and Chen, J. C. and Kirby, E. D. (2007). Surface roughness optimization in an
end-milling operation using the Taguchi design method. Journal of Materials
Processing Technology, 184, 1-3, 233–239, (2007), ISSN 0924-0136.
 20 Advances in Robotics, Automation and Control

Zuperl, U. and Cus, F. and Milfelner, A. (2005). Fuzzy control strategy for an adaptive force
control in end-milling. Journal of Materials Processing Technology, 164, 1472–1478,
(2005), ISSN 0924-0136 .
Zuperl, U. and Cus, F. (2003). Optimization of cutting conditions during cutting by using
neural networks. Robotics and Computer-Integrated Manufacturing, 19, 1-2, 189–199,
(2003), ISSN 0736-5845.

View publication stats