Metal Machining - Theory and Applications Episode 2 Part 7 ppt

As in the first example, wear was monitored indirectly by force measurements, using neural nets, but it was found that the axial and radial cutting position of the tool on the piece infl

As in the first example, wear was monitored indirectly by force measurements, using neural nets, but it was found that the axial and radial cutting position of the tool on the piece influenced the nets’ predictions: dynamic force signals were influenced by the work-pieces’ compliance One net was used to monitor wear while a second wear rate; both were trained by direct measurement (rather than, as in the first example, by model predictions)

Fig 9.29 Wear development estimated online for continuous change of both cutting speed and feed (Ghasempoor

et al., 1998)

Fig 9.30 Integration of monitoring, prediction and operation planning of cutting processes (Obikawa et al., 1996)

Because of this, a large amount of redundancy (robustness) was built into the nets, with 34 inputs to each net, as shown in Figure 9.31 Thirty of these were auto-regression (AR) coefficients of the feed force power spectrum (as much information as could be extracted from it), two were the total power of the spectrum of feed force and cutting force, and two were the axial and radial positions of the cutting tool, as already mentioned

Under the assumptions of the AR model, the power spectrum was defined as

1 +Σa k z –k

k=1

where p is its order (and also the number of peaks in the spectrum), Pn (jw) is the white noise power spectrum, a k is the kth AR coefficient and z = e j w In this case p was chosen

to be 30 by the Akaike Information Criterion (AIC – Akaike, 1974)

The outputs from the two nets, the flank wear (VB) t and its rate(V˘B) t, were combined

as follows, with Dt being the time interval between estimates, to give an even more robust

estimate:

Table 9.6 Operation planning conditions and initial turning conditions

Longitudinal cutting length of workpiece 150 mm

Tool geometry (–5, –6, 5, 6, 15, 15, 0.8)

Fig 9.31 Neural networks for predicting flank wear (Obikawa et al., 1996)

(VB—)t = — [(VB) t + {(V˘B) t Dt + (VB) t–Dt}] (9.43)

2 Figure 9.32 shows a comparison between estimated and measured flank wear in four different speed and feed cutting conditions Training the nets was carried out on one batch

of material and the estimates and measurements on another

The predominant wear in the conditions of this example could be modelled by equation (4.1c) The prediction element of Figure 9.30 was the physical model already described in Section 9.2.4, with an example of its outputs given in Figure 9.7 Precise prediction of

flank wear rate requires accurate values of the constants C1and C2in equation (4.1c): they can vary from batch to batch of the tool and workpiece Optimization needs them to be continually tuned and identified In this example, wear rate was calculated by the FDM simulator Q—FDM(Section 9.2.4) beforehand, for many combinations of C1and C2, cutting speed, feed rate and width of flank wear, to create a look-up table When the wear rate in

an actual turning operation was estimated by the monitoring system, the values of C1and

C2, which gave agreement with the estimate, were identified quickly by referring to the table

After tuning the constants, the cutting speed and feed could be optimized Figure 9.33

shows, for one batch, the width of flank wear VBendat the end of turning all the work-pieces, predicted for different speeds and feeds Under the constraint of maximum wear land length of 0.2 mm and shortest cutting time, a cutting speed of 130 m/min and a feed

of 0.225 mm would be chosen in this case These conditions could be set, adaptively, after tuning the constants while turning the first bar of the batch

9.4.4 The development of monitoring methods

The direction of development of monitoring methods during the 1990s can be understood from the list of reported studies in Table 9.7 Force continues to be the dominant signal to

be monitored In the area of signal processing, there is a slow growth in the application of

Fig 9.32 Flank wear development predicted by neural network (Obikawa et al., 1996)

wavelet transforms (wt), which translate a signal in the time domain into a representation localized not only in frequency but in time as well Neural networks are becoming a stan-dard method for the recognition of cutting states For pattern recognition, unsupervised ART 2 type neural networks (Carpenter and Grossberg, 1987) have been effectively used

(Tansel et al., 1995; Niu et al., 1998).The integration of wavelet transform coefficients as

Fig 9.33 Optimized cutting conditions using a tuned wear equation (Obikawa et al., 1996)

Table 9.7(a) Recent approaches to cutting state monitoring – abbreviations given in Table 9.7(b)

Signal processing Processes and Sensor features Recognition

monitored states signals and/or models methods References

Tapp: a, s, w F, Q cr, cv, me, pe, rm, va Pa, PV Chen et al (1990)

Turn: w A, C, F ar, rm, pd (FFT) Pa, NN Dornfeld (1990)

Turn: w A me, rm, sk, vc Pa, CL Moriwaki and Tobita (1990)

Turn: w A, F, T aw, fw Qv, NN, ST Chryssolouris et al (1992)

Turn: t, v, w A, F rf, va, vc Pa, NN Moriwaki and Mori (1993)

Drill: w F, Q me, pe, rm, ft, tt Pa, Qv, NN Liu and Anantharaman (1994) Turn: w A, F cs, fr, ku, me, sd, sf, sk Pa, NN Leem et al (1995)

Turn: b, c, r, w A ku, sk , fb, me, sd, wt Pa, NN Niu et al (1998)

inputs with neural networks as classifiers can be expected to lead to more detailed and reli-able recognition of cutting states in the future

9.5 Model-based systems for simulation and control of machining processes

In this final section, the application of machining theory to complicated machining tasks

is described As larger and larger applications, taking more time, or more and more complex components, requiring more operations, are considered, the need for more

ration-al planning and operation becomes greater A totration-al or globration-al optimization is needed, in contrast to optimizing the production of a single feature Optimization in such conditions needs machining times, machining accuracy, tool life, etc, to be known over a wide range

of cutting conditions If the machining process is monitored, for example based on cutting force, the expected change in force with cutter path (in the manner of Figure 9.25) must also be known over a long machining time Once the time scale reaches hours, force measurement and its total storage in a memory become unrealistic For these reasons, cutting process simulation based on rational models, namely model-based simulation, is expected to have a significant role in the design and control of machining processes and to give solutions to rather complicated processes

Table 9.7(b) Abbreviations used in Table 9.7(a)

Processes and monitored states

EndM: end milling b: tool breakage v: chatter vibration

Turn: turning s: hole size error

Sensor signals

A: accoustic emission F: cutting forces T: temperature

C: spindle motor current Q: cutting torque

Signal processing features and/or models

af: cutting force moving df: dispersion in frequency sd: standard deviation

average per revolution ranges sf: power spectrum feature am: AE mode (amplitude fb: frequency band power sk: skew

probability density) ft: force-time area tt: torque-time area ar: AR coefficients fw: force-wear model va: variance

aw: acoustic emission-wear ku: kurtosis vc: coefficient of variation

cr: correlation pd: power spectral density wt: coefficients of wavelet cs: cutting speed rf: ratio of force components transform

cv: covariance rm: root mean square FFT: fast Fourier transform Recognition methods

Pa: pattern recognition CL: Cluster analysis based NN: neural network

Qv: quantitative value on mean square distance PV: probability voting

TH: threshold

9.5.1 Advantages of model-based systems

Consider some of the optimization issues associated with the roughing of the aerospace

component shown in Figure 9.34 (Tarng et al., 1995) Figures 9.34(b) and (c) show end

mill tool paths that convert the block (a) to the rough shape (d) First, machining is

conducted smoothly along Y–Z plane tool paths, then along X–Z planes In the X–Z plane paths, the end mill must remove steps left by machining along the Y–Z plane paths, as

shown schematically in Figure 9.35: step changes in axial depth of cut are unavoidable

The major constraints to the roughing operation may be: (1) the peak cutting force, Fpeak,

must be less than a critical value, Fcritical, which causes the tool to fail and (2) the finish-ing allowance left on the machined surface must be less than a given amount (dependfinish-ing

Fig 9.34 Tool path for machining an aerospace component (Tarng et al., 1995): (a) original workpiece, (b) tool paths

in the Y–Z planes; (c) tool paths in the X–Z planes; and (d) machined workpiece

on the required finished accuracy): this constraint eventually determines the Y cross feed for the X–Z plane machining strokes The objective in selecting the cutting conditions may

be to find the minimum machining time under these constraints

To simplify the problem of cutting condition optimization, the axial depth of cut in each

Y–Z plane path and the cross feeds in the X and Y directions may be set constant If the cutting speed is also held constant, the feed speed (Ufeed, Chapter 2) becomes the single variable that controls the cutting states The feed per tooth may change in a specified range

with an upper limit fmax; that too is one of the constraints

There are two methods to find optimal feed changes in the above milling operation One

is online adaptive control; the other is model-based simulation and control Adaptive control (Centner and Idelsohn, 1964; Bedini and Pinotti, 1982) is a method that adjusts cutting conditions until they approach optimal, based on monitored cutting states However, it has some response time, reliability and stability difficulties Although tool wear rate, chatter vibration, chip form, surface finish and dimensional accuracy are all candidate states for control, they are seldom used in adaptive control because of insuffi-cient reliability Cutting forces and torque are usually the only states that are selected

As in the cornering cut described in Section 9.2.2, the cutting force is effectively controlled by feed Therefore, to minimize machining time, it might be decided, in an adaptive control strategy, to maximize the peak cutting force by adjusting the feed from an

initial value f, with a measured force Fpeak, to a new value fa,force:

f

Fpeak where Fcriticalis the largest safe value

If a model-based system is used to control f, force change with cutting time is simulated

based on one of the force models: generally equation (9.6) is recommended Then feed is adjusted to raise the simulated peak force to the critical level It may be necessary in prac-tice to allow for feed servo control delays that are inevitable in numerical control

If no trouble arises in a machining process, adaptive and model-based control should yield the same results However, if a sudden increase in the axial depth of cut or the effec-tive radial depth of cut occurs, as at steps in Figure 9.35 or at corners in Figure 9.6, adap-tive control may not function well, because of the response time limitation mentioned above Under adaptive control, with time minimization as its goal, an end mill is probably moving at its highest feed rate before it meets a step or a corner The sudden increase in

Fig 9.35 A schematic of a tool path and pre-machined steps in an X–Z plane

the axial depth of cut or effective radial depth of cut is likely to yield a very large cutting force, causing tool damage, before the adaptive controller can command the reduction of feed rate and the feed is actually reduced Tool damage due to sudden overloading is more likely to be avoidable if the force change is predicted by model-based simulation The cutting conditions may be optimally designed beforehand to decrease the feed to a value low enough to anticipate the changes at steps and corners

In short, the principal difference between the two control methods is that model-based simulation is feed-forward in its characteristics, whilst adaptive control is a feedback method Its feed-forward nature is one great advantage of model based simulation

A second advantage of model-based simulation is that prediction of change in cutting states can support monitoring and diagnosis of cutting state problems in complicated machining processes In the absence of an expected response, a monitoring system cannot distinguish a normal from an abnormal change A third advantage is that the machining time under optimized conditions is always estimated beforehand This helps the schedul-ing of machinschedul-ing operations

From all this, a model-based system is a tool for global optimization In this sense, adaptive control is a tool for local optimization

9.5.2 Optimization and diagnosis by model-based simulation

Model-based simulation has been applied to the end milling example of Figure 9.34 (Tarng

et al., 1995) Figure 9.36(a) shows the simulated resultant cutting force in fixed feed

condi-tions The detail force model of equation (9.6) and the specific cutting force model of equation (9.7b) (Kline and DeVor, 1983) were used The spindle speed selected was 1200

rev/min, the maximum axial depth of cut (the depth of cut in Y–Z plane paths) was 6 mm,

the maximum radial depth of cut was the full immersion of 12 mm, and the feed rate was fixed at 105 mm/min

Figure 9.36(b) shows a simulation under variable feed Compared with Figure 9.36(a), peak forces are more uniform; and the machining time has been reduced by about a third Furthermore, the simulated result was confirmed experimentally, when the operation was actually carried out with the planned strategy (Figure 9.36(c))

The strategy was to adjust the feed to

Fpeak

fa, force= (– 2 ——— +3)f (9.44b)

Fcritical where f is the original constant feed By this means, the feed rate fa ,force= f when Fpeak=

Fcriticaland rises linearly to 3f as Fpeakreduces to zero

Similar pre-machining feed rate adjustment in end milling and face milling has been applied to the control of the average torque, average cutting force, and maximum dimen-sional surface error caused by tool deflection, as well as the maximum resultant cutting force (Spence and Altintas, 1994) It is the Spence and Altintas (1994) model-based system that is illustrated in Figure 9.1(a)

Figure 9.1(b) shows a machining operation system that can generate a machining scenario for a given operation (Takata, 1993) The machining scenario describes changes

in cutting situations predicted by geometric and physical simulation Cutting situations include both machining operations and cutting states For end milling, five types of

operations are recognized: slotting, down-milling, up-milling, centring and splitting The machining scenario is used to control cutting force and machining error by pre-machining feed adjustment, and to diagnose machining states

Figure 9.37 (from Takata, 1993) shows an example of the effectiveness of pre-machining feed adjustment in controlling dimensional errors in end milling Figure 9.37(a) shows plan

Fig 9.36 Variation of resultant cutting force (Tarng et al., 1995)

Fig 9.37 Effectiveness of pre-machining feed adjustment in controlling dimensional error (Takata, 1993)