The physical parameter estimates for both methods mustbe equivalent to make a fair comparison. To accomplishthat, various tests are conducted on the simulated drives,including acceleration tests, constant speed (with no load)tests, a variety of light machining tests. The observerparameters
(Je,
be, T,Kt, Ks) are indirectly computed usingthe data obtained from these tests. The observer is thentuned to a bandwidth frequency of 100Hz. With respect tothe estimator, similar tests, which additionally includemedium and heavy machining conditions, are performed toobtain the additional friction process parameters.Fig.
6
illustrates the accuracy of both methods. Omittingthe characteristics of acceleration estimation for bothmethods, the only major difference between them is thatthe estimator incorporates a detailed friction model. As canbe seen, the estimator is significantly more accurate thanthe observer at almost every machining condition. Themean relativeerrors can be given as 8.14% and 3.29% forthe observer and the estimator respectively. Hence, theresulting accuracy improvement is by a factor of 2.5.However, it is critical to notice that the accuracy of bothtechniques heavily depends on the quality of estimates onthe physical system parameters [lo].database. The basis functions
Y
of the process beingcomputed by estimator can then be formed with theseparameters. The Recursive Least Squares (RLS) withexponential data weighting [I21 is hereby chosen as aweight adaptation (supervisory) algorithm:e,(k)
=
Fp(k)

W;(kl)Y(kl)
(9c)
where Wo(k) is the parameter vector
E
??I4'';
(k) is errorcovariance matrix;
h
is the forgetting factor (0.95
..
0.99)while k is the time index (k
=
1,2,
...).
The forgetting factoralong with the initial values of WOand P dramatically affectthe estimation performance of the algorithm.
Also
the initialspindle angle
(eo),
which aligns the basis functions
(Y)
iththe estimates, plays a critical role on the accuracy. Eventhough a special estimator topology can be designed todetect
00,
it is here assumed to be known beforehand forthe sake of convenience.
Xaxn
cuulng
Force
Estmator
Supervisor
(RLS)
Selector
I
i4
yaxlsc
I
Force
EstmaG
I
Figure
6:
Accuracy plots for various machining condition.
5
OVERALL ESTIMATOR TOPOLOGYThe main problem in estimating the forces using thetopology presented in the previous section is that whenone of the drive motors of the machine is stalled, thecutting force component in that direction could not beusually evaluated due to strong Coulomb friction on the FDmechanical system (guideways and transmission system)[9]. In such a case, one of the principal cutting forcecomponents is assumed to be missing and has to begenerated with the utilization of the available cutting forcecomponent at hand.To construct the unknown cutting force component, theestimator shown in Fig.
7
has been developed. In thistopology, the selector in this topology simply invokes thereference (endmilling) model whenever one of the FDmotors is stalled (e.g. machining is done along afundamental axis). Thus, the supervisor adjusts thecorresponding model coefficients by relying on the cuttingforce estimate of the running drive. Note that a crossreference process model proposed by [I ] s employed tocompute the model coefficients of the unknowncomponent. As a result, the missing process force F,(where p stands for either x or y) can be generated in astraightforward fashion.In this scheme, the cut geometry parameters
0
(i.e. radialand axial depth of cut) are to be retrieved from the CIMFigure
7:
General cutting force estimator topology
6
EXPERIMENTAL RESULTSA number of machining tests have been performed toillustrate the performance of the estimator when one of theFD motors was stalled. The relevant parameters ofmachining tests are given in the Appendix. With
h
=
0.95,the RLS algorithm is only allowed to run during onemachining period. Here, the initial covariance matrix istaken as P(0)
=
0.5.14~4,
I:
identity matrix) while WO sinitialized with the rough estimates on the cuttingforceintensity coefficients (see [Ill) to increase the speed ofconvergence. In the adaptation phase, the algorithmadjusts the corresponding parameters of the model byutilizing the force estimates of the running drive. To makea fair evaluation; the (reference) model "predictions" for astalleddrive are compared to the corresponding forcemeasurements. Figs. 8 and 9 demonstrate the results. Ascan be seen, there exist some offsets between the RLSpredictions and the measurements since the plowingforces at the flatend of the cutting tool have not beenincorporated to the reference model. Despite these offsets,the accuracy and bandwidth of the topology are found tobe acceptable for most CNC machine tool applications.
7
CONCLUSIONThis paper proposed a novel disturbance force estimatorfor CNC machining centers. Unlike the observer, theestimator is not sensitive to measurement noise due toRDFT based
filtering/differentiation
scheme. Since theproposed estimator incorporates a sophisticated frictionmodel for the hydrodynamic guideways, it can decouplethe friction forces from the principal cutting forcecomponents in an iterative fashion. Not surprisingly, itsestimation accuracy was found to be significantly better