Smith G.T. Cutting Tool Technology: Industrial Handbook

Подождите немного. Документ загружается.

Figure 258. Tool-condition monitoring on a turning centre. [Courtesy of Sandvik Coromant].

Machining and Monitoring Strategies 533

to the signal-processing device. Once the signal is re-

ceived, the processing device can immediately initiate

action by the machine’s CNC controller, if the tool is

either: worn, broken, or not in-cut. is situation is

all very well, but when should tool monitoring take

place and, what action should result? In Fig. 258a (i.e.

the inset diagram), the graphical depiction shows how

continuous monitoring of the axial force can be used

to triggered several alarm-states:

•

Level I – can be utilised to monitor whether a tool is

in-cut, or not-in-cut, as the situation arises, mean-

ing that either the tool, or component, or indeed

both, are missing,

•

Level II – can be used to detect tool wear, with the

alarm signal being used to initiate a tool change (i.e.

to a ‘sister tool’) on completion of the operation,

•

Level III – can be utilised for tool breakage, with

the signal being used to immediately stop the ‘feed-

ing-function’ of the machine tool, when breakage is

detected,

•

Tool crashes – a further level can also be employed

for crash protection, which acts in a similar man-

ner to ‘Level III’ ,

but this alarm immediately stops

all motions and in so doing, protects the machine

tool

In Fig. 258c, the schematic diagram illustrates typi-

cal monitoring positions on a two-axis turning centre,

showing potential sites to place the sensors, such as on

the ball-screw nuts of the recirculating ballscrew as-

semblies, for both the X- and Z-axes. Not shown here,

but normally also tted is a current sensor. us, these

signals are continuously monitored by either a single-

or multi-channel control unit, as will be the control

signals from the machine tool’s CNC controller. Any

alarm signals triggered, being passed back through a

closed-loop to the machine’s control unit for appropri-

ate action to be taken, or indeed if any. e function of

a typical commercially-available multi-channel signal

processing unit, might be to:

•

Sense – then process tool-cutting information from

signals at the various sensors and sites for the mul-

tiple channels of the unit,

76 ‘Tool breakage detection times’ , it is possible to vary the reac-

tion time, which is usually between: 0.1 to 1 second, but for

any form of tool breakage a shorter reaction time is desirable,

typically ranging from: 1 to 10 milliseconds.

•

Learn – by automatically memorising the signal

values obtained from the sharp cutting tools, whilst

in this ‘

learn-mode’

•

Stores data – for a signicant number of cutting

operations per channel in its memory for each cut-

ting operation, as well as automatically setting the

appropriate levels for each alarm signal from its

memory,

•

Reacts – by sending alarms to the machine’s control

unit, informing it if the tool is either: worn; broken;

or not-in-cut,

•

Coordinates – automatically, machining and moni-

toring on commands from the machine’s control

unit,

•

Adapts – to the particular machine and its cutting

environment: once installed and programmed to

suit the machine tool, with the setup parameters

being modied to adapt to any further machining

requirements,

•

Communicates – between the operator and the

machine via the control panel, informing the oper-

ating personnel about cutting tool conditions and

providing an interface for control of all functions.

In Fig. 258c, this line-diagram depicts a typical turn-

ing centre application of tool condition monitoring.

e machine is controlled on two axes, with sensors

on the feed-drive bearings of both the X- and Z-axes.

A representative nominal force for these sensors is 40

kN, but this rating will depend upon the end-user’s re-

quirements. e sensors can be designed for tapered,

or angular contact bearings, or for a combined axial

and radial bearing application – suiting the particular

machine tool. When tool monitoring is needed for a

four-axis turning centre, two tool monitoring units are

usually needed, since each turret (i.e. to and bottom)

can be operated both independently. e key elements

in any tool condition monitoring situation are the

sensor’s position and its design. For universal instal-

lation on a variety of machine tool congurations, the

positioning of the sensing devices is usually on the re-

circulating ballscrew nut assembly.

77 ‘System-learning’ , in the past this was somewhat a basic of

functional performance, but with the advent of ‘articial neural

networks’ , they have an ‘AI-ability’* to ‘mimic’ human involve-

ment and react to their environment – once ‘trained’. More

will be said on this ‘AI-topic’ in the nal section of this book.

* ‘AI’ is a term that is normally utilised when some form of ‘arti-

cial Intelligence’ is employed in the decision-making process.

534 Chapter 9

Much more could be said concerning the informa-

tion on their: operational setup; range; and adaptabil-

ity; for these tool condition monitoring systems. In

the interests of brevity, the reader should look to the

manufacturers of such equipment, or the references

and available literature for more specic in-depth in-

formation.

9.11.2 Adaptive Control and Machine

Tool Optimisation

Adaptive Control

Adaptive control systems have been utilised since their

introduction in the 1960’s, where their operational

performance and reliability was somewhat dubious,

because of the type of sensors utilised, the speed of

signal-data processing and their installation on the

machine tool. Many of these early systems attempted

to undertake many functions simultaneously and were

oen termed; ‘adaptive control optimisation’ (ACO),

but due to the problems mentioned above, they were

somewhat unreliable and as such, fell out of favour.

Later, a more pragmatic approach to adaptive control

constraint (ACC) was introduced called: ‘torque-con-

trolled machining’ (TCM), which oered a simpler

termed: ‘feed-only system’ – with a typical system be-

ing depicted in Fig. 259a. us, the operation of a TCM

system, involves unique sensory circuitry and compu-

tation methods that measure the net cutting torque,

then compares this value obtained, to that of the preset

torque limits – these previously being established for

the cutting tool and workpiece combination utilised.

e appropriate control actions, namely, a feedrate

reduction is then automatically taken, whilst keeping

within the maximum torque and power limits of the

spindle motor. If a condition arises where the feedrate

falls below a preset limit, a new tool (i.e. sister tool) is

called-up to complete the machining operation. is

feedback-loop in which continuous monitoring by the

sensors and updating the machine control unit – using

adaptive control, produces optimal cutting conditions

for the tool and workpiece combination.

Adaptive control via TCM (Fig. 259a), basically op-

erates in the following manner. Prior to its activation

and if for example, a variation of stock was present for

roughing operation with a large face-mill. e unpre-

dictability of the height of this stock if a TCM system

was not activated, might otherwise over-load the cut-

ting edges, possibly causing damage to the: cutter as-

sembly; workpiece; or even the machine tool. Once

the TCM has been correctly activated and preset to a

torque limit, then if the D

is large, the control sys-

tem senses a torque increase and simultaneously the

feedrate over-ride is initiated. is over-riding of the

programmed feedrate decreases the feed for this large

, it will then increase as the D

lessens, or rapidly

move over an ‘air-cut’ , thus producing optimal cutting

tool protection and eciency as the chip-load is more

uniform, regardless of the variable D

’s. Even if there

is no discernible dierence in the relative height of the

taken, but the bulk hardness of the part may vary

by up to 300% in some cases, machining with the TCM

activated will protect the tooling. So to mention the

some benets to be gained from TCM, they include:

extended tool life; optimised feedrates – without the

risk of tool damage; higher throughput of machined

parts; tool breakage minimised; quicker setup times;

and reduced operator intervention. Obviously very

small diameter tooling, may not respond to the torque

demands so readily, but for most machining opera-

tions and tool/workpiece combinations the system has

distinct benets to the overall machining production

process.

To summarise the principal benets of utilising

some form of adaptive control system, they are:

•

Main spindle motor is protected from overload,

•

Damage to the cutter and to the expensive value-

added workpiece are protected,

•

Optimal stock removal rates are possible, under

steady-state machining conditions,

•

Using a constant: cutting power; cutting force; and

feed force; optimises tool life,

•

If unpredictable air-gaps occur – whilst cutting, the

fastest tool travel is utilised,

•

Where workpiece hardness signicantly varies, tool

edges are protected by adjustments of the chip-

oads,

•

Where an operator’s experience, or the program’s

eciency may dier for varying cutting operations,

the adaptive control system eliminates this ‘techni-

al gap’ ,

•

ere is no over-shooting of the permitted cutting

power during re-entry into the workpiece material

whilst machining the part under regular condi-

tions.

Costs vary the for ‘post-installation’ of adaptive con-

trol systems to CNC machine tools, but at today’s

rices they range from: $ 9,000 to $ 15,000 (US). How-

ever, once installed they last the life of the machine

Machining and Monitoring Strategies 535

Figure 259. Either use: adaptive control or CNC program optimisation – for variable tool path trajectories.

536 Chapter 9

tool, giving a superb pay-back on the original invest-

ment, when one considers the major benets listed

above.

Machine Tool Optimisation

If a company has signicant numbers of CNC machine

tools in their manufacturing facility, then it may not

be feasible to introduce an ‘adaptive control’ system

across all of these machines – despite the positive

merits described above, simply on nancial grounds

lone. Under such circumstances, perhaps a ‘soware-

approach’ by simulating the cutting operations to the

problem of machining optimisation, may be the way

forward? Some companies oer CNC programming

optimisation packages that are based upon literally

thousands of ‘man-hours’ of development and rene-

ment (i.e. Fig. 259b, shows a very sophisticated version

of such a tool verication and simulation system).

ese simulation systems are oen part of a larger: op-

timisation; verication and analysis product that can

be ‘tailored’ to suit a machining company’s product

range and manufacturing output. ese ‘knowledge-

based’ systems of the machining process, via previous

simulation, know the exact: D

; width of cut; and

angle of cut (i.e. for cutter orientation, when prol-

ing); for the machining process under consideration.

Further, the system also knows how much material

is to be removed by each cutting edge, as such, the

system also has information on the tooling available

from the magazine, therefore it selects correct tool and

assigns to it the optimum feedrate. Moreover, once

this information has been established for the new tool,

it outputs the tool path – which was identical to say,

that of the original tool, but now having signicantly

improved feedrates, although the system does not alter

its trajectory.

While setting up the system, it is usual for such

soware (Fig. 259b) to prompt the user for cutter set-

tings as the part simulation occurs, by in essence, add-

ing the user’s intelligence to that of the cutter’s opera-

tion. With these systems it is usual to have all cutter

settings stored in an optimisation library, thus the user

only has to dene the setting once. While, the more

sophisticated systems nd the maximum volume re-

moval rate and chip thickness for each tool, then it

employs them to determine the optimisation settings

for that tool.

In optimised roughing-out, the objective here is ob-

viously to remove as much stock material as possible

in the fastest time. Conversely, for nish-machining,

chip-loads may vary considerably, as the tool proles

through the workpiece material that was le behind

during previous roughing cuts over the contours

– to near-net shape. By optimising the tool’s path, the

soware adjusts the feedrates to maintain a constant

chip-load

(Fig. 259b). is cutter optimisation will

improve the tool life and give an enhanced machined

surface nish to the component. is fact is especially

ritical when ‘tip-cutting’ , with either a ball-nosed end

mill (Fig. 247b), or contouring over a surface with a

small step-over, such as when semi-nishing, or n-

ishing a steel mould cavity (Fig. 249b).

Summarising the advantages of utilising a simu-

lated optimisation cutter-path soware package, such

as the one in Fig. 259b which only illustrates some ba-

sic and simple tool paths. us, cutter-path optimisa-

tion oers the user the ability to:

•

Machine more eciently – cutting more parts in

the same amount of time, by signicantly reducing

the machined component’s cycle-time,

•

Reducing part cost thereby saving money – increas-

ing productivity by reducing the time it takes to cut

parts, will become a signicant saving per annum,

•

Improving part quality – by minimising the con-

stant cutting pressure, thus reducing cutter deec-

tion, with nished corners, edges and blend areas,

needing less subsequent hand-nishing,

•

Cutter life improved – because of optimised cut-

ting conditions are used, which prolongs tool life.

Moreover, with shorter in-cut time, this results in

less tool wear, also having the benet of reducing

down-time to change inserts, or tooling,

•

Reduction in machine tool wear – as a more con-

stant cutting pressure between the machine tool

and the workpiece reduces variable forces on the

axis motors, giving smoother machine operation,

•

Utilises time available more eectively – allowing

machinists to operate several CNC machine tools,

or setup the following job, etc., as they do not have

to be constantly ready to reduce/increase the ma-

chine’s feedrate over-ride.

By investing in suitable simulation and optimisation

soware of the tool’s path, enables a company that is

currently involved in a considerable amount of ma-

78 ‘Constant chip-loads’ , are normally recommended by cutting

tool manufacturers, as they reduce the eect of ‘chip-thinning’

somewhat.

Machining and Monitoring Strategies 537

chining activities to become very cost-eective and

ecient when compared to their direct competition,

both nationally and internationally. One could cer-

tainly ask the question, under these circumstances

just mentioned: ‘Can a company aord not to be using

such soware, if their main competition – both here and

abroad have it available now?’

9.11.3 Artificial Intelligence:

AI and Neural Network

Integration

Introduction

Over the past decade and a half, some signicant ad-

vances in machining materials have occurred, while

complementary progress has also been made in the

machine tool’s CNC controllers, coupled to their faster

micro-processor speed and additional technological

renements. Many of these machine tools are inte-

grated into fully-automated systems machining lines

– for volume part production purposes, or into ex-

ible manufacturing cells/systems (FMC/S) – allowing

scope for mixing batch sizes and perhaps employing

a ‘Group Technology’ (GT) approach (i.e. see Footnote

24, Chapter 6). So that the full potential of these ma-

chine tools can be exploited, it is exceedingly impor-

tant that production processes are both monitored and

controlled in an ‘intelligent manner’.

Previously, when little cutting data and minimal

tooling-related behaviour had been established for a

new production run, it was necessary to instigate some

form of tool measurement procedure. So, aer operat-

ing a cutting tool for an extended time-period in-cut,

so that the tool’s wear pattern (Fig. 174) had begun to

reach the end of its productive life (Fig. 176), it was

necessary to exchange it for a new tool. is arbitrary

tool-changing strategy was at the discretion of the op-

erator, therefore it relied upon their past machining

experience to decide when it was advisable to instigate

the necessary down-time – for this tooling-related ac-

tivity. An alternative approach, was to employ some

form of condition monitoring procedure, by utilising

o-line direct measurements to ascertain the amount

of wear that had occurred so far. is assessment ac-

tivity entails a certain degree of operator competence

in a variety of disciplines, because the cutting tool’s

inspection required microscopical analysis by metro-

logical/metallographical techniques to determine the

current status of the tool’s cutting edge(s). is tool-

ing investigation necessitated that the tool be at rest

and out-of-cut, so that its life could be correctly estab-

lished, which can be a costly and time-consuming pro-

cess, diminishing the cost-eectiveness of the overall

production process.

One machining strategy that can be used to over-

come most production deciencies, is to have some

form of on-line, indirect system, which has the ad-

vantages of being benecial in terms of: improved

running costs; enhanced component quality; and e-

ciency in production performance. In order to achieve

such benecial tooling-related and part production

enhancements, it is necessary to utilise some form of

‘on-line tool condition monitoring’. So that this tool

monitoring objective can be successful, a number of

hard- and so-ware activities must be undertaken,

then integrated into a usable ‘workshop-hardened’

instrumental package. In the early-to-mid 1990’s a

novel approach to this problem, but also included the

some distinct renements by: ‘on-line tool condition

monitoring – using neural networks’ was developed by

Littlefair et al. (1995). is fundamental and applied

research work was fully-supported by a range of in-

dustrial companies, it was later also installed at sev-

eral widely-diering manufacturing companies. In or-

der to comprehend the complexity of such an on-line

tooling related activities, the following case-study has

been included (Littlefair, et al., 1995), as it succinctly

describes the hard- and so-ware issues that had to be

overcome.

9.11.4 Tool Monitoring Techniques –

a ‘Case-Study’

e technique of tool wear monitoring can be classi-

ed in two distinct manners, these are by either:

•

Direct monitoring – produce accurate results, but

they are dicult to fully-implement in a shop-oor

environment,

•

Indirect monitoring – considers various parameters

which change as a result of increasing tool wear.

e latter tool monitoring strategy was utilised in a

single-point turning operation on a CNC turning cen-

tre, by incorporating: tool force; vibration; and acous-

tic emission; by being integrated into a neural network;

and this theme will now be mentioned. Each of these

monitoring systems will be briey described, plus the

neural network – appropriate for complete sensor-fu-

sion, will then be described.

538 Chapter 9

Tool – Force Monitoring

In single-point turning, if one ignores the orthogonal

cutting condition, then for oblique cutting three re-

actionary forces are experienced by the tool, termed:

tangential; axial; and radial force components (Fig.

19a). e tangential force is generated due to the

workpiece’s rotation, this being by far the greatest of

the three forces. An axial force component is the re-

sult of the applied feed force, while the radial force is

a function of, in the main, the inclination of the ap-

proach angle and to a lesser extent inuenced by that

of the tool nose radius – this radial component being

the smallest of the forces. Each of these component

forces in oblique cutting are inuenced by a range of

factors, such as: workpiece material and its condition;

; tool cutting insert geometry; and cutting data

utilised – speed and feed. In this case, a special-pur-

pose holder for a platform-based dynamometer was

manufactured (Fig. 261a).

Tool – Vibration Monitoring

In machining processes, the onset and subsequent

development of vibration orginates from the overall

dynamic behaviour of the tool-workpiece-machine

system. e anticipated vibrational causes can be both

cyclic in nature – resulting from changes due to com-

pression and sliding of the workpiece material in the

shear zone, and, changes in the frictional conditions in

the contact zones – between the tool and workpiece.

So that vibrational inuences during continuous cut-

ting could be monitored, accelerometers tend to be

utilised. Normally, accelerometers are situated as close

to the cutting edge as possible, usually at a convenient

position on the toolholder. e vibration parameters

monitored are usually related to either the toolholder’s

natural frequency, or the frequency of chip segmenta-

tion. Moreover, it is also possible to eectively utilise

that of a dynamometer’s ‘force signal’ for indirect vi-

bration monitoring.

Tool – Acoustic Emission Monitoring

Acoustic emissions (AE) are those high-frequency

stress waves generated due to the spontaneous energy

release in materials undergoing: deformation; fracture;

phase transformations; etc. us, AE signatures can be

divided into two distinct types: continuous – contain-

ing low-amplitude and high-frequency signals (i.e. in

the range: 100 to 400 kHz); burst – containing higher

amplitude and lower frequency signals (i.e. in the

range: 100 to 150 kHz). By the application of Fourier

transforms coupled to that of statistical analysis-based

techniques, it is possible to utilise both of them for the

analysis of AE signals. e root-mean-square (rms)

value has been shown to produce an increasing trend

with increased amounts of tool ank wear, further,

the combination of both skew and kurtosis of the AE

signal will also indicate a correlation with ank wear

rates.

Tool – Sensor Fusion and Multi-sensor Integration

e application of multiple sensors can be eectively-

employed in a complex tool-wear monitoring system

for machining environments, to obtain harmonizing

information about the turning production process.

is multi-sensor monitoring acts to rearm the ‘con-



dence factor’ , when dealing with the prospective di-

agnostics from the single-point turning process. How-

ever, the exercise of utilising multiple sensors, entails

integration and fusion of the sensory information, to

extract the essential features from the data, by remov-

ing the ‘redundancy’ present in this data. In this re-

gard, the application of articial neural networks, can

provide the solution to the sensor-fusion and auto-

matic decision-making processes for this tool-condi-

tion monitoring system.

Artificial Neural Networks (ANN)

Articial neural networks (Fig. 260a), are composed of

many simple processing nodes which operate simulta-

neously. ese ANN’s mimic the functional behaviour

of biological neural network systems, allowing them

to be utilised to integrate and fuse information from

multiple-sensor sources. e functional behaviour

of the overall system is primarily determined by the

pattern of connectivity of the nodes (Fig. 260a). As a

system, ANN’s are capable of performing some high-

level functions, such as: adaptation; generalisation and

target-learning. ese capabilities are particularly rel-

evant for any form of tool-wear monitoring applica-

tions. e advantages of employing ANN’s to integrate

and fuse data, are their inherent capabilities to: adapt

to instructed environments; robustness to noise; fault

tolerance; simultaneous processing; and feasibility of

on-line realisation (i.e. via hardware implementation).

Possibly the most widely used ANN and the one

reported in this section, is that of the ‘multi-layer per-

ceptron’ type, which uses an ‘error-back-propagation

Machining and Monitoring Strategies 539

Figure 260. Neural network architecture and tool condition monitoring system. [Source: Littlefair, Javed & Smith, 1995].

540 Chapter 9

algorithm’ – in its ‘learning phase’ (Fig. 260a). Funda-

mentally, this network operates in the following man-

ner: the input layer neurons transmit signals via the

hidden layers to the ouput layer. At this juncture in

its operation, the desired and actual outputs are com-

pared to evaluate the system error – usually Euclidean

error. Hence, this error value is employed to adjust the

strengths of the connectivity amongst the network’s

neurons. e algorithm utilised to perform this un-

dertaking is normally known as ‘error-back-propaga-

tion’ ,

which is available in many forms. e hidden

layer’s nodes have their capabilities developed during

‘

training’ , this is achieved in such a manner, that the

extracted features are better suited for the ‘classica-

tion task’.

Tool – Condition Monitoring System

e ANN described above, was utilised on a two-

axis slant-bed turning centre (Fig. 261b), which was

equipped with sensors and associated equipment al-

lowing on-line data capture, during a comprehensive

run of machining trials. ree sensors were utilised

in this work for monitoring the cutting process, t-

ted onto a specially-manufactured platform, situ-

ated on the tool turret (Figs. 261a and c). e three

sensors were a: Kistler force dynamometer (model:

9275B) – which sensed/measured the cutting forces

in three perpendicular axes (i.e. X, Y and Z); AE

sensor – Physical Acoustics (type: WDI); vibrational

sensor – Vibrometer (type: CE501 M101) miniature

accelerometer. e force and acceleration signals were

amplied and then sampled at 50 kHz, while the AE

sampling was undertaken separately by a digital stor-

age adaptor at a sampling rate of 1 MHz. All of this

information was then stored on a ‘suitably- fast’ PC.

e schematic layout of the monitoring hardware is

illustrated in Fig. 260b.

Prior to utilising by the neural network, a pre-

processing procedure was operated to reduce what is

termed its ‘dimensionality’ of the signals. is action

was achieved by computing the ‘power spectral densi-

ties’ of the captured time-domain signals and equally

dividing the resulting spectrum into eight discrete fre-

quency bands. is number of bands has been shown

to be the optimum in terms of reducing dimensional-

ity, yet maintaining information integrity. e multi-

layered perceptron type of neural network was also

utilised to further process data and subsequently cor-

relate it with dierent cutting tool wear states. is

correlation was achieved by integrating and fusing the

data components in order to remove redundancy of sa-

lient information in the data. So, depending upon the

particular application, one channel of force was com-

bined with the acceleration and AE channels, in order

to create a 24-by-1 input vector for the neural network.

In a practical example of its use, for a plunge-grooving

operation, this system gave signicantly more detailed

information in the axial direction, whereas, in simple

longitudinal outside diameter turning, the tangential

force was found to be more useful and as such, appli-

cable.

e neural network architecture used in these

tool-condition monitoring machining trials, consist

of an: input layer (24 nodes); a single hidden layer (10

nodes); with and output layer (3 nodes). is particu-

lar ‘nodal-architecture’ was derived through arbitrary

experimentation, with ve and een nodes in the

hidden layer increasing the ‘convergence time’ with

no appreciable improvement in tool wear classication

performance. Finally, the output layer provided ve

tool wear categories

which depended on the particu-

lar machining application.

e tool-condition monitoring system – once de-

veloped at Southampton Solent University, was tted

onto a two-axis at-bed turning centre at the Atomic

Weapons Establishment (Aldermaston). e system

proved to operate successfully at both establishments,

which was congured to; capture cutting data; reduce

and process the sensor data; apply the ‘previously-con-

gured’ neural network, then correlate all of this in-

formation into dierent tool wear states – in a mean-

ingful manner.

y utilising these ‘AI – neural networks’ , they have

the potential to evolve into an: adaptive on-line fee-

drate control system, which could be integrated into

the tool-changing system for a CNC machine tool.

Moreover recently, similar sensors have been utilised

to monitor the machining process and data fusion/

analysis, through the application of neural networks

and/or fuzzy-logic-based techniques.

79 ‘Tool wear states’ , were achieved by the system output gener-

ated in the form of a [1 x 3] output vector, which is a binary

encryption of ve wear states – being dened as follows: 000

– minimal ank wear; 001 – minor ank wear; 010 – major

ank wear; 011 – minor tool damage (chipping); 111 – major

tool damage (chipping).

Machining and Monitoring Strategies 541

Figure 261. Tool condition monitoring using articial neural networks. [Source: Littlefair, Javed & Smith,

1995]

542 Chapter 9