Lopez de Lacalle L.N., Lamikiz A. Machine Tools for High Performance Machining

Подождите немного. Документ загружается.

5 Advanced Controls for New Machining Processes 193

mechanistic, and models based on observed data, which are empirical models that

could be referred to as observational. Mechanistic models tend to be more general.

Empirical models are more specialised and only operate under certain specific

conditions. Their properties depend heavily on the data observed (black box mod-

els), which include its abundance, integrity, completeness and timeliness. It is also

possible to develop semiempirical (grey box) models.

Mechanistic models used to predict surface roughness are of the geometric

type, based on the motion geometry of a metal cutting process, without taking

cutting dynamics into account. Hence their main disadvantage, namely, that these

mechanistic models do not include parameters related to cutting dynamics, such

as speed, depth of cut or the type of material used for the work-piece. A second

disadvantage is that they are purely theoretical and do not verify their models

with any experimental data. As a result, in-process prediction of surface rough-

ness tends today to involve the use of empirical models, and the design of the

experiments used to obtain observed data is of extreme importance. Here, Ta-

guchi techniques for the design of experiments (DoE) seem to be the most wide-

spread methodologies.

Generally speaking, the compromise that must be dealt with when building em-

pirical models is between the form and structure adopted and the way the model

depends on the selected data (e.g., variance). An incorrectly parameterised model

results in a biased model. An over-parameterised model will offer a high variance

which fits the construction sample well, but could be of little use in validating

a sampling. The bias/variance trade-off becomes particularly evident when work-

ing a small amount of data, where it is difficult to identify a smooth form due to

the variability of the data. The choice between the bias and variance in a neural

network model is more difficult to define.

Empirical models basically take one of two forms: There are statistical models,

based on non-linear multiple regression equations, and there are artificial intelli-

gence models, based above all on artificial neural networks (ANNs). A typical

example of a non-linear multiple regression equation appears in [7], which pro-

poses a functional relationship between surface roughness and some independent

variables:

c h

α1

α2

α3

α4

α5

(5.19)

where Ra is surface roughness (the arithmetic average), h is the hardness of the

part, f is the feed, r is the tool nose radius, d is the depth of cut and s is the cutting

speed.

Because there are equations of this kind, the results can be discussed from

a mechanical point of view, and the variables that are statistically more important

can be identified. These models show that machining merit variables offer con-

trasting contributions in areas such as productivity, surface finishing, dimensional

precision and machining costs, among other things. Considering only the most

representative interactions between independent variables and merit variables, the

following could be cited as being the most important:

194 J. Ramón Alique and R. Haber

• Productivity and surface roughness are contrasting figures of merit in the ma-

chining process.

• Feed has a notable influence on productivity and surface roughness. Productiv-

ity increases with increased feed, while surface quality diminishes.

• Cutting speed has a notable influence on the dynamic stability of the process

and surface roughness. Surface quality increases with greater cutting speed.

• Depth of cut has a notable effect on productivity, dynamic stability and surface

roughness.

• The hardness of the work-piece being produced and, above all, the radius of the

tool have an effect on surface roughness. Surface quality increases as the radius

of the tool is increased.

ANN-based models offer a very powerful alternative when there are no analyti-

cal models or when a low-grade polynomial is not appropriate. An essential step

when using an ANN for modelling is to use real observed data to find the synaptic

weight values for all neurons (learning algorithm). One of the algorithms used for

adjustment is the back-propagation algorithm, which is the most often-used algo-

rithm for modelling surface roughness.

Reference [38] proposed a model that used spindle speed, feed speed, depth of

cut and vibration average per revolution as the ANN input signals. According to

their results, using a 4-5-1 topology for the ANN, they could achieve an accuracy

of 95.87% precision. When using a more powerful topology, with two hidden

layers at 4-7-7-1, they obtained an accuracy of prediction of 99.27%. Training

times were not long, and operation in real time seemed to be guaranteed.

5.6 Open-Architecture CNC Systems

Computer numerical controls have evolved at extraordinary speeds in recent years,

due basically to the availability of powerful microprocessors, sophisticated soft-

ware and other developments in digital technology. Many manufacturers include

high performance processing blocks in their controls, with processing times of

under a millisecond. Extremely high speeds can now be attained in comparison

with the 70

ms that were available with classic CNCs. These processing-time

specifications cannot be looked at alone, but rather in relationship with the time

required to process, initiate and complete the required motion. This is the real

execution time, which defines the speed and accuracy of machining

Research on integrating personal computers (PCs) with open CNC systems has

continued in recent years, and this has necessarily led to superior properties in

terms of flexibility. The new open architectures are more reliable, to the extent

that they can operate the machine as a whole and open architecture also allows for

the real-time overall management of the whole flow of information concerning

a machine’s operation. This avoids the instability that is inherent in a link with

a PC or a connection with a PC operating system. Instability of this sort has been

particularly problematic lately with the Windows operating system.

5 Advanced Controls for New Machining Processes 195

Furthermore, results on the design of complex distributed-control architecture

are now available. In other words, in addition to the inclusion of high speed micro-

processors in CNCs that can share control tasks in real time, there have been ad-

vances in the development of digital and “intelligent” motor actuators for each

axis. Given the high speeds used, feedback loops are connected directly to the

actuators instead of going to the CNC, though the CNC must still monitor the

entire system as a whole.

A completely digital system, like some of those that are already commercially

available (e.g., the Heidenhain TNC 620

), provides an excellent platform for

achieving improved control. Some systems incorporate commands that switch to

“high precision” control modes in order to increase precision and reduce the cycle

time. These commands involve calculating different accelerations and decelera-

tions in real time based on feed speeds. Elsewhere, new intelligent filtering and

control strategies based on adaptive algorithms and fuzzy logic can now be used to

suppress vibrations in the system within a determined range of frequencies.

5.6.1 Networked Control and Supervision

In traditional CNCs, communications architecture is by point-to-point connection.

This architecture has been successfully applied in industry for many decades,

though it is not suitable for systems containing a large number of devices, as the

number of cables increases along with the number of elements. Nor does this kind

of architecture afford the possibility of dealing with new requirements, such as

modularity, decentralised control, simplicity and ease of maintenance [41].

A networked control system is one whose components (sensors, controllers, ac-

tuators, etc.) are distributed using some form of computer network technology. The

use of this technology brings with it important advantages, such as reliability, the

improved use of resources, ease of maintenance and error diagnosis and, above all,

the possibility of reconfiguring the different components. Nevertheless, there are

also disadvantages to this type of system: implementing closed-loop control in

a communications network leads to delays that inevitably can cause instability [14].

5.6.1.1 Communications in an Industrial Environment

The real-time communications market in the industrial sector has for a long time

been dominated by fieldbuses. Some of the most well-known standards are

Profibus, WorldFIP and Foundation Fieldbus. At present they have been standard-

ised but not unified. These technologies offer a number of advantages, one of

which is their deterministic behaviour. However, their disadvantages include the

high cost of hardware and the difficulty of integrating them with other products

[23]. To solve these problems, some computer network technologies, particularly

Ethernet, are being adapted for use in the industrial automation field, given their

simplicity, low cost, availability and high transmission speeds.

196 J. Ramón Alique and R. Haber

The main technical obstacle to Ethernet in industrial environments is its non-

deterministic behaviour, which in principle makes it unsuitable for applications

with real-time requirements. The reason for this is that, in classic Ethernet net-

works, the transmission medium is shared by all stations in the sub-network. The

CSMA/CD (carrier-sense multiple access with collision detection) protocol is used

to access the medium. This protocol cannot guarantee that the packets sent will be

received before a certain amount of time has elapsed, because a station will have

to wait until the transmission medium is not being used before being able to

transmit. Furthermore, if two stations on the network decide to transmit simulta-

neously, there will be a collision, which will mean that the message will not reach

the recipient correctly. When a collision is detected, the stations must wait for

a random amount of time, which is determined by the binary exponential backoff

algorithm. If a station detects 16 consecutive collisions, it will cease to transmit

and send an error message to the higher levels in the network protocol stack [34].

In short, Ethernet does not use up time in the bus, but collisions between packets

may lead to delays in information being sent and may even cause information to

be lost. It is therefore impossible to predict what delays will occur in the network.

Although it does not perform deterministically, under light traffic conditions

Ethernet sends information quite quickly and with little latency [19]. However,

with a greater load on the network, collisions are more frequent, thus causing

delays and reduced performance. One way of improving the performance of

a control system based on an Ethernet network would therefore be to reduce net-

work traffic, with the resulting reduction in the number of collisions.

Some researchers have suggested changing the medium access protocol so as to

achieve deterministic behaviour. PROFINET is one standard that does so. Al-

though these alternatives use Ethernet hardware, they are entirely incompatible

with standard Ethernet traffic. In recent times, the use of switched Ethernet (which

uses switch connections) has offered further improvements. Switched Ethernet

behaves differently from the classic model. Switches are intelligent devices that

identify the intended recipient of the information and only retransmit the informa-

tion via the required output ports. This differs from the way the older system

works, via a hub that retransmitted all the information received via all its ports

(Fig. 5.12). Use of the switch system on the Ethernet eliminates any potential

collisions in information transmission and thus increases network efficiency.

Fig. 5.12 Hub vs Switch

5 Advanced Controls for New Machining Processes 197

Because there are no collisions, the network does not become unstable under

heavy traffic loads, and delays are drastically reduced. The delays that happen

when switched Ethernet is used are sufficiently low for industrial networks work-

ing in real time, making switched Ethernet quite a promising alternative for net-

worked control systems [18].

5.6.1.2 The Structure of a Networked Control and Supervision System

Figure 5.13 shows the structure for a networked control and supervision system

run via switched Ethernet.

In its simplest form, the system is comprised of two stations, the first of which

handles control, and the second, acquisition. The two stations are connected via

a switched Ethernet network. The acquisition station is in direct contact with the

process.

The system works as follows: the acquisition station, which is in constant contact

with the process, acquires information from the process. At given intervals, depend-

ing on the process in question, the acquisition station sends the information it has

obtained to the control station. The control station makes a decision based on a con-

trol algorithm that has been specifically implemented for the process. This informa-

tion will be sent to the acquisition station so that the necessary action can be taken.

This is therefore a client-server architecture. The control station (server) re-

ceives requests from the acquisition station (client), makes a decision on the basis

of some form of control strategy and replies to the original station. The control

station is also responsible for carrying out supervision duties, and it is the point at

which all the system’s resources must be integrated and centralised. Its duties

particularly include monitoring process variables and plant conditions and config-

uring devices and control systems.

Fig. 5.13 Networked control and supervision via Ethernet

198 J. Ramón Alique and R. Haber

5.6.1.3 Networked Control and Supervision via Ethernet

Internal model control (IMC) is a technique that is widely used and well established

in controller design. This closed-loop control scheme explicitly uses a model (G

)

of the dynamics of the plant to be controlled, which is placed in parallel to the plant

in question (G

). Furthermore, it also contains another model of the inverse dynam-

ics of the plant (G’

) placed in series with the process, which acts as a controller.

(a)

(b)

Fig. 5.14 Output of TWNFI model. a Direct model. b Inverse model

5 Advanced Controls for New Machining Processes 199

One of the advantages of this type of control lies in the fact that its stability and

robustness can be guaranteed. However, the inversion of non-linear models is no

easy task, and there may be no analytical solution. In certain cases, a solution may

exist but be impossible to implement. Another associated problem is that the in-

version of the process model could lead to unstable controllers when the system is

in a non-minimal phase.

The following is a representation of the use of the TWNFI algorithm to create

on-line models (direct and inverse) [33]. Both models are calculated with each new

entry in the control scheme. Using this neurofuzzy inference technique, the creation

of the inverse model is much easier and always offers a solution. The direct model

should be trained to learn the dynamics of the process. In order to do this, a TWNFI

system is used with input-output learning data in which the input data corresponds

to feed speed while cutting force is used as the output variable (Fig. 5.14a).

In order to calculate the inverse model, instead of inverting the direct model ob-

tained analytically, another TWNFI system is used in which the learning data

contains cutting force values as the input and feed speed values as the output. In

this way, the system is able to learn the inverse dynamics of a high performance

drilling process (Fig. 5.14b).

The learning data for both the direct model and the inverse model have been

obtained from real drilling operations using GGG40 material tests. This set of data

does not need to be more extensive, as representative values from each area of

operation are sufficient.

Inverse model G’

and direct model G

use prior states for the input variables

in order to achieve a better approximation in the input-output relationship in the

training data for the TWNFI algorithms:

() ´ ( (), ( 1), ( 1))

fk G Fk Fk fk=−−

ˆˆ

() ((),(1),(1))

Fk G fk fk Fk=−−

(5.20)

where

()

k is the cutting force estimated by the model.

Once the models were laid out in a diagram, it was also decided to include

a low-pass filter (G

). This was incorporated in the control system with the aim of

reducing high-frequency gain and improving the system’s robustness. It also

served to smooth out rapid, brusque changes in the signal, thus improving control-

ler response. The filter takes the following form:

()

−

(5.21)

where k

and k

are design parameters, and usually k

Figure 5.15 shows the internal model control scheme. The diagram includes

the delay (both the delay introduced by the network and the delay inherent in the

process) through block L. The figure also shows direct and inverse models G

and G’

, respectively, filter G

and the process represented by G

[12]:

0.4

( ) 10.26

()

( ) 0.005241 0.09376 0.5414 1

Gs e

fs s s s

−

+++

(5.22)

200 J. Ramón Alique and R. Haber

Fig. 5.15 Internal model control scheme

where f(s) is the feed speed, F(s) is the cutting force and G

(s) is the process trans-

fer function in the Laplace domain.

The issues concerning the delay, L, and the architecture of the networked con-

trol system are explained in the following section.

5.6.1.4 Networked Control Systems

All delays (both from the process itself and from the network) have been grouped

together as block L. The use of networked control systems undoubtedly brings sig-

nificant advantages in terms of reliability, better use of resources, ease of mainte-

nance and error diagnosis and, above all, the possibility of reconfiguring the differ-

ent components. Fieldbuses have become consolidated as the most commonly used

communications networks in industry. Indeed, it is normal practice in the process

control field for models to incorporate the delay displayed by these networks.

However, the development of information and communications technologies

has meant that control and supervision systems are increasingly required for proc-

esses that cover some geographical area. For this reason, the control processes

discussed here for high performance drilling use Ethernet network technology.

Ethernet is being used massively in the field of industrial automation, given its

low cost, its availability and its high transmission speeds [6]. The main difficulty

associated with networked control in general and control using Ethernet in par-

ticular lies in the fact that, while the control system is strong enough to withstand

the design changes considered, it may not be able to tolerate unmodelled commu-

nications delays.

Ethernet’s main technical disadvantage in an industrial environment is its non-

deterministic behaviour, which makes it unsuitable for applications that include

real-time requirements. Ethernet does not use up time in the bus, but packet colli-

sion may cause delays in information transmission and may even cause informa-

tion to be lost.

However, the use of switched Ethernet (which is connected by means of

switches) removes the potential for collision during data transmission, thus increas-

5 Advanced Controls for New Machining Processes 201

ing the network efficiency. As there are no collisions, the network does not become

unstable under heavy traffic, and delays are drastically reduced. Studies have shown

that switched Ethernet causes delays that are sufficiently small and deterministic for

the system to be used in real-time industrial networks, and it has therefore become

quite a promising alternative for networked control systems.

An in-depth study of the delays found in manufacturing processes where

machine tools are connected to computers in a local network using Ethernet can

be found in [21], along with some experimental results. Based on this work, as

shown in Fig. 5.16, the maximum delay in an Ethernet network is established as

5×10

–3

sec.

The architecture of the networked system is shown in Fig. 5.17. PC1 is con-

nected using Profibus to the open-architecture CNC of the machine tool. This is

where the measurement and processing of measured signal F takes place:

0.4

SC CA drilling

τττ

++ =

(5.23)

From a technical point of view, the existence of proprietary software places re-

strictions on connectivity to the open-architecture CNCs systems. However, it

offers security in handling the CNC. PC2 uses free software (RT-Linux) and an

intermediary real-time system (ACE-TAO). The internal model control system is

implemented here. Assuming that the maximum delay is known (0.4

0.005

s),

it is possible to implement networked control using Ethernet.

Fig. 5.16 Statistical distribution of delays

202 J. Ramón Alique and R. Haber

Fig. 5.17 Architecture of the networked system

5.7 Programming Support Systems: Manual Programming

Moving on from the operational function of a numerical control, the next issue is

to analyse how a numerical control is programmed, what language is used and

how a programmer enters all the data required for machining (i.e., the part pro-

gram) into the numerical control. In the simpler cases, the programmer works

from a drawing that identifies all the coordinates and the lines that connect them.

However, if a complex part is being made, and particularly if the part includes

free-form surfaces, the identification process is much more complicated. A solu-