Accurate Interpolation of Machining Tool-paths Based on FIR

Filtering

Shingo Tajima Burak Sencer, Ph.D (corr. Eiji Shamoto, Dr.Eng.

Graduate Student Author) Professor
[email protected] Assistant Professor [email protected]
Mechanical Engineering [email protected] Graduate School of Engineering
Department Mechanical Engineering Department Nagoya University
Oregon State University Oregon State University Nagoya, JAPAN
Corvallis, Oregon, USA Corvallis, Oregon, USA

Abstract
This paper presents a novel real-time (online) interpolation algorithm based on Finite Impulse Response
(FIR) filters to generate smooth and accurate reference motion trajectories for machine tools and motion
systems. Typically, reference tool-paths are composed of series of linear (G01) or circular (G02) segments.
Basic point-to-point (P2P) feed motion can be generated by interpolating each segment with trapezoidal or S-
过渡
curved velocity profile. However, smooth and accurate transitions between path segments are necessary to
realize non-stop contouring motion. In this study, FIR filters are utilized, and the reference tool-path is
filtered to interpolate a non-stop rapid feed motion. By using a chain of FIR filters, acceleration and jerk
continuous motion profiles are generated from velocity pulse commands. A segment interpolation timing
technique is developed to control the contour errors during non-stop real-time interpolation of tool-paths.
Furthermore, by utilizing FIR filters for interpolation, frequency spectrum of the interpolated motion profiles
is controlled. The time constant (delay) of the filter is tuned to create notches around the lightly damped
vibration modes of the motion system, which allows mitigation of unwanted vibrations and thus enables
delivering accurate feed motion. Simulation studies and industrial scale experimental validations are
provided to illustrate effectiveness of the developed interpolation technique.

1. Introduction
Reference trajectory generation plays a key role in the computer control of machine tools and motion
systems. Generated trajectories must not only describe the desired tool path accurately, but must also have
smooth kinematic profiles in order to maintain high tracking accuracy, and avoid exciting natural vibration
modes of the mechanical structure or servo control system. As a matter of fact, most machining tool-paths
are defined in terms of series of linear (G01) segments or circular (G02) arcs [1][2]. This imposes serious
limitations in terms of delivering a non-stop smooth and rapid motion for productivity, and to achieve the
desired final part geometry.
There are several challenges associated with interpolating a smooth motion along these discrete tool-
paths. Consider interpolation on a single path segment; feedrate (tangential velocity) profile needs to be
planned with smooth acceleration and decelerations to avoid excitation of the machine tool’s structural
modes [3] and at the same time respect kinematic limits, i.e. torque, acceleration and jerk, of the drives
[4][5][6]. Polynomial based feed profiles, such as trapezoidal velocity [1], acceleration [4] and jerk profiles
[8] are well-known to the machine tool literature. They can be planned to fully exploit machine limits and
generate time-optimal feed motion along predetermined paths [5][6]. However, these methods suffer from
two bottlenecks. Firstly, they don't provide any quantitative means to control the frequency spectrum of the
interpolated acceleration commands. In practice, the jerk limit is used to mitigate any residual vibrations [3].
Note that, tuning the jerk limit smoothens acceleration profile. But, it does not directly control the frequency
spectrum. Robotics literature adapted exponential [9], trigonometric [10] or minimum jerk spline [6] based
acceleration profiles to help attenuate frequency spectrum of reference trajectories. In precision machine tool
literature, input shapers (IS) [11][12] and notch filtering are utilized to filter the reference motion commands
to attenuate the excitation around the lightly damped resonant frequencies of the machine. These techniques
are easy to implement and robust against parameter variations [11], which makes them suitable for practice.
However, input shaping distorts interpolated tool trajectories due to shaper dynamics and induces
interpolation errors. Either machining velocity (feedrate) has to be lowered to reduce those errors, which is
widely employed in practice; or, model-based compensation techniques that are mostly computationally
costly are proposed in the literature [13].
Another bottleneck is the computational load of reference trajectory generators. As the degree and
complexity of the acceleration profile becomes higher, computational cost to plan polynomial based
trajectory generation increases [14]. Recent efforts are directed towards generating online, real-time suitable
interpolation techniques [15][16][17]. These approaches essentially utilize a tuned dynamic system to filter
and smoothen velocity/displacement commands. They are designed with a chain of integrators and cascaded
feedback loops [18]. Online path smoothers can be implemented in the form of recursive difference
equations. To attain time-optimal motion, nonlinear feedback elements such as saturation blocks are also
introduced [19][20]. Nevertheless, unless combined with an input shaper, methods cannot control frequency
spectrum of generated trajectories.
Finite Impulse Response (FIR) filters provide a computationally efficient framework for online trajectory
generation. The use of FIR filtering for real-time interpolation and trajectory generation is known to the
machine tool literature [22]. Chain of 1st order FIR filters can be used to generate smooth reference
trajectories with trapezoidal acceleration and jerk profiles [23][24]. Time constants of filters can be assigned
to realize time optimal motion. Furthermore, frequency response of the filter can be tuned so that the
excitation of the reference trajectory is shifted away from the resonances of the machine tool. Finally, it can
be implemented as a moving average filter on modern micro-processors with minimum computational effort
[23].
Although FIR filtering is an effective technique for online interpolation of reference trajectories, so far,
its use is constrained in simple point-to-point (P2P) moves. If consecutive moves, e.g. linear or circular
segments, are interpolated continuously without a full stop at the segment junctions, large interpolation
contouring errors occur due to sudden change in the feed direction and the dynamics of the filter. Unless
these contour errors are confined, the use of FIR filtering for generating uninterrupted, rapid and accurate
feed motion in precision motion systems is limited. Recent literature recognized these shortcomings and
proposed compensation techniques[13][24][25]. However, these techniques are either computationally
expensive because they need to estimate errors through dynamic models. This greatly limits their application
in real-time implementation. Or, they consider contouring errors only around junction of linear segments
[24], which is not realistic since conventional machining tool-paths consist of both mixture of linear and
circular segments, and transitions in-between those segments must be considered for a non-stop high-speed
contouring motion.
This paper, for the first time, presents comprehensive interpolation techniques for generating
uninterrupted and accurate feed motion along multi-segmented machining tool-paths based on FIR filtering.
Contributions of the paper are laid out as follows. Section 2 first analyses high-order trajectory generation
based on FIR filtering technique. It is followed by the introduction of accurate interpolation of linear and
circular paths. A feedrate control technique is presented to control contour errors during interpolation of
circular paths. Section 3 presents online interpolation of multi-segmented toolpaths based on FIR filtering.
Dwell time control technique is presented to control the interpolation errors that occur during non-stop
transition between linear and circular segments. Finally, Sections 4 and 5 present illustrative examples and
rigorous experimental validations along complex tool-paths.

2. Online Trajectory Generation based on FIR Filtering

 2.1. Generation of high order kinematic profiles
Typically, “trapezoidal acceleration” or “trapezoidal jerk” based feed profiling is employed to generate
reference trajectories for high-speed and precision motion systems [4][8]. This section outlines the basic
methodology to generate high-order trajectories utilizing a chain of FIR filters [23].
A 1st order FIR filter is defined in Laplace (s) domain by the following transfer function [23][26]:

(1)

where Ti is the time constant (delay) of the ith FIR filter. Observed from Eq.(1), a FIR filter consists of an
integrator (1/s) and a pure delay resembling a simple moving average filter [26]. The impulse response is
evaluated by taking inverse Laplace transform of Eq.(1) as:

(2)

and as shown in Figure 1, it becomes a simple rectangular pulse with a duration of Ti having a magnitude of
1/Ti. This implies that for any Ti>0, the area underneath the impulse response is unitary. As a result, when an
arbitrary signal is convolved with the FIR filter, area underneath the original signal does not alter.
Furthermore, since the filter has a free integrator (1/s) it increases degree of the filtered (convolved) signal.
This property can be used to generate high order real-time motion trajectories as follows.
Amplitude

1
Unitary Area
Ti

Time
Ti
Figure 1: Impulse response of a 1st order FIR filter

Let us consider a simple linear movement for a length of L commanded at a velocity of F. This trajectory
can be commanded by a rectangular pulse for a duration of Tv=L/F as shown in Figure 2a. FIR filtering this
rectangular velocity pulse generates the well-known trapezoidal velocity profile [1] with piecewise constant
acceleration segments (See Figure 2b). Subsequently, another FIR filter can be convolved with the trapezoidal
velocity profile to generate smoother trapezoidal (jerk limited) acceleration profile [4]. As outlined in Figure 2,
high-order reference kinematic profiles can be generated by filtering a reference velocity pulse though chain
(series) of FIR filters [23][24]. Finally, the resultant velocity profile is integrated to obtain reference
displacement profile as shown in Figure 2c.
 T1 = 我们的T1+T2

T2 = 我们的T1
T2=T1则没有
A=Amax 匀加速
Figure 2: FIR filtering based smooth trajectory generation

Filtered kinematic profiles can be analyzed through analytical solution of convolution [26]. Consider a
simple trapezoidal velocity profile. Convolution of a rectangular velocity pulse command v(t), with the
impulse response of the FIR filter from Eq. (2) is written as:
t
v '(t )  v(t )  m(t ) 
1
 
v    v   Tv   u  t     u  t  T1    d
T1 0 
 t 
1  0 v   u  t    d  0 v   u  t  T1    d
t (3)
  t 
T1  v   T  u  t    d  t v   T  u  t  T    d 
 0 v 0 v 1 

and the filtered velocity signal v’(t) is derived by evaluating above integrals,

 F T1  t , 0  t  T1

F , T1  t  Tv
v '(t )   (4)
 F T1  t  Tv  T1  , Tv  t  Tv  T1
0 , Tv  T1  t

where the resultant acceleration signal is obtained by differentiation as:

(5)

Finally, smooth displacement profile s’(t) is generated by integration of the velocity profile.
 The use of convolution enables analytical derivation of filtered profile kinematics. Above trapezoidal
velocity profile is derived for the case of and also illustrated in Figure 3a. The peak acceleration
depends on the filter’s time constant and the commanded velocity, Apeak = F/T1. Commanded velocity F is
reached at filter’s time constant, t = T1, and the remaining cruise velocity duration becomes Tv-T1. On the
other hand, when the reference velocity pulse duration is equal or shorter than the filter’s time delay Tv≤T1,
motion kinematics alters. As depicted in Figure 3b, when Tv=T1, no velocity cruise section occurs. As
opposed, Figure 3c illustrates the case of Tv<T1. In this case, commanded velocity cannot be reached and peak
velocity becomes Vpeak=L/T1 with a cruise phase duration of T1-Tv. Although omitted here, convolution can
be used to analytically derive all these kinematic profiles. The overall motion duration is elongated by the
amount of filter’s time constant .

Figure 3: Trapezoidal velocity profile generated by Figure 4: Trapezoidal acceleration profile generated by
single FIR filter 2 FIR filters
 In precision motion systems jerk [4] or even snap [8] limited velocity profiles are favored to generate
smoother, more traceable reference motion profiles. As illustrated in Figure 2, utilizing 2 FIR filters with time
constants T1 and T2 generates the well-known trapezoidal acceleration (jerk limited) feed profile. Figure 4
illustrates the motion profile generated by filtering a trapezoidal velocity pulse. The profile kinematics can
be computed analytically by replacing rectangular velocity pulse command with the trapezoidal one in Eq.(3)
, and for the case of Tv>T1>T2, the generated velocity profile is derived as:
1 F 2
2 TT t 0  t  T2
 1 2

 1 FT2 F
 2 T  T  t  T2  T2  t  T1
 1 1


 T1  T2   t 
1 F 2
F  T1  t  T1  T2
 2 T1T2
 F T1  T2  t  Tv
v '(t )   , (6)
 F  1 F  t  T 2 T  t  T  T
 2 T1T2
v v v 2


 F  1 FT2  F  t  T  T   T  T  t  T  T
 2 T1 T1
v 2 v 2 v 1


1 F T  T  T  t 2
 2 T T  v 1 2  
Tv  T1  t  Tv  T1  T2
 1 2

0 Tv  T1  T2  t

and the corresponding filtered acceleration profile a’(t) becomes:

ì F
ï t 0 £ t < T2
ï T1T2
ï F
ï T2 £ t < T1
ï T1
ï
F F
ï
ï
-
T1 T1T2
( t - T1 ) T1 £ t < T1 + T2
ï
ï 0 T1 + T2 £ t < Tv
a'(t) = í (7)
F
ï
ï
-
T1T2
( t - Tv ) Tv £ t < Tv + T2
ï
ï F
ï - Tv + T2 £ t < Tv + T1
T1
ï
ï
ï -
F
+
F
T1 T1T2
(
t - ( Tv + T1 ) ) Tv + T1 £ t < Tv + T1 + T2
ï
ï 0 Tv + T1 + T2 £ t
î
As observed from Eq.(6), filtered (interpolated) motion profile is determined by the filter delays, T1 and
T2, and the length of the reference velocity pulse, Tv. There are in total 8 combinations. Owing to the linearity
of filtering operation, the order of filters in the chain does not matter, and 3 main cases characterize the
interpolated profiles; namely, Tv>T1+T2, Tv=T1+T2, or Tv<T1+T2. Figure 4a depicts the most common case
when the velocity pulse is longer than total sum of filter time constants, Tv>T1+T2 and T1>T2. Note that in
this case, a full 7-segmented jerk limited acceleration profile [4] with cruise phase can be generated (Eq.(6)).
However, if a rapid (high speed) move on a short travel distance is commanded, Tv may become smaller. For
instance, cruise velocity phase may disappear completely if Tv=T1+T2 as shown in Figure 4b. Furthermore, if
Tv<T1+T2, commanded path velocity cannot be reached. In this case, peak velocity is computed as Vpeak=L/T1
(See Figure 4c). The total motion duration is elongated by the filter delay as, Tv+ Td, where .

2.2. Frequency shaping of interpolated trajectories

The FIR filter structure also provides effective means to control frequency spectrum of the generated
trajectories. Filtered acceleration a’(t) profile controls the torque/force delivered by the feed drive, which
induces excitation to the overall motion system. If frequency spectrum of the reference acceleration profile
contains components near the lightly damped structural modes of the machine structure, it initiates forced
vibrations [11], [13].
For a rectangular pulse velocity input, the acceleration command consists of set of impulses separated by
Tv (See Figure 2a). Consider only a single acceleration impulse with a magnitude h

(8)

convolved with the chain of FIR filters; the frequency spectrum of resultant acceleration profile simply
becomes frequency response of the FIR filters in the chain, evaluated as:

(9)

and frequency ( ) response of a single FIR filter can be computed from Eq.(1) as:

(10)

Consequently, frequency spectrum of the acceleration profile becomes multiplication of sinc [27] functions
from Eq. (10) as:

(11)

The above property can be exploited to choose time constant of the FIR filter to avoid exciting lightly
damped structural frequencies of the machine tool. Every sinc function creates periodic notches (ripples),
which can be matched with the resonant frequency of the motion system by setting,

(12)

An example is presented in Figure 5. Simply setting time constant of the filters in the chain to the natural
periods of the resonant modes introduces shortest filter delay into the motion while
avoiding excitation of resonances. It is also notable that there is close resemblance between Input Shaper
[11][12] and the FIR filter. Input Shapers have the property to cancel any vibration at the half of the
vibration period, .
 Figure 5: Frequency response of FIR filter

2.3. FIR based interpolation of linear and circular paths

2.3.1. Linear interpolation
Interpolation of single axis motion based on FIR filtering is presented in the previous sections. This
technique can be extended to generate point-to-point (P2P) multi-axis linear motion. Figure 6 outlines the
process to interpolate planar P2P linear motion between two points, . Firstly, the
path length is computed from the Euclidean distance, , and the tangential feed pulse F for a
duration of Tv=L/F is generated as shown in Figure 6b. The feed pulse is dissolved into its Cartesian velocity
pulse components based on the path geometry,

(13)

and smooth axis velocity commands are interpolated by applying FIR filtering. Note that, time constants of
the filters are set identical so that the resultant motion is coordinated. Finally, the interpolated axis velocity
commands are integrated to interpolate position commands.
 ①先根据误差解 程求圆弧限速, 切向合速度F 虚线
② ×cos,sin 得到 各个采样/插补周期即Ts倍数的X/Y虚线速度
③ 对X/Y 分轴速度各 FIR。 成最后红蓝分轴实线。
1) 由方程(20)求出切向合速度F-黑色虚线
2) Fcos, Fsin得到平滑前的X/Y分轴速度-蓝红虚线
3) 对蓝红虚线的Fcos, Fsin分轴速度，分别做FIR
- 得到最后平滑的分轴速度蓝红实线

先根据G01的X/Y占比, 拆解到
X/Y分轴的速度,再在分轴上各 滤波

X/Y分轴各 滤波
之后得到了红蓝

Figure 7: Circular interpolation and corresponding

Figure 6: Multi-axis interpolation based on FIR filtering velocity profiles

2.3.2. Circular interpolation

Next, the approach for linear interpolation is adapted to interpolate circular paths. During circular
interpolation, the total travel length becomes the arc length of the circular path where R is the arc
radius and is the difference between starting and ending angular positions (See Figure 7).
Omitting the arc center, rectangular feed pulse F is dissolved into its axis components based on the circular
geometry as:

(14)

Notice that as opposed to the linear motion, reference axis velocity commands during circular interpolation
are not in rectangular pulse form but rather sinusoidal signals. The reference axis commands are then filtered
through chain of FIR filters and integrated to interpolate the circular path. Figure 7 illustrates the reference
and interpolated velocity commands during circular interpolation.
 Notice that sinusoidal axis motion commands are generated at the rotational frequency of , and
they are modulated by the frequency response of FIR filters. At steady state, the filtered axis motion
commands can be written from Eqs.(14) and (1) as:

F 
s 'x (t )  R GFIR  c
cos  t  init  
R 
 (15)
F 
s ' y (t )  R GFIR sin  t  init  
 c
R  
N
where GFIR   M i ( j ) is frequency response function of FIR filter. The discrepancy between reference and
i 1

interpolated (filtered) circular motion commands result in an interpolation contour error as shown in
Figure 7. This contour error is measured normal to the commanded circle, and its steady state value can be
calculated from Eqs.(14) and (15) as:

  sx  s 'x    s y  s ' y 
2 2

2 2
 F   F 
  R 1  GFIR  cos  t   s     R 1  GFIR  sin  t   s   (16)
 R   R 
 R 1  GFIR 

The steady state value of the contour error is controlled by magnitude of the frequency response of the
FIR filter at the fundamental frequency of the circular motion, . By lowering the feedrate F, excitation
frequency can be altered, and can be confined by a user-specified tolerance value. Without losing
generality, let us consider a trapezoidal acceleration profile generated by 2-FIR filters. The magnitude of the
frequency response at can be evaluated from Eq. (11) as:
æTFö æT Fö
sin ç 1 ÷ sin ç 2 ÷
è 2R ø è 2R ø
GFIR w =w = (17)
c T1 F T2 F
2R 2R
4-term Taylor expansion can be applied to Eq.(17) to obtain a polynomial expression:

GFIR » ç 1- ( T1w c 2 ) + (T1w c 2 ) ÷ ç 1- ( T2w c 2) + ( T2w c 2 ) ÷

æ 2 4
öæ 2 4
ö
(18)
è 3! 5! øè 3! 5! ø

and substituting Eq.(18) into Eq. (16) yields the relationship between the contouring error and the feedrate
as:

  R 1  GFIR    0
c

 1 1   10 3 3   1 1  
 x 4  80  2  2  x3  640  2 2  4  4  x 2  153600  4 2  2 4  x  3686400 4 4  0
 T1 T2   T1 T2 T1 T2   T1 T2 T1 T2  T1 T2 R
(19)
where

(20)

Eq.(19) is a 4th order polynomial whose roots can be solved conveniently in real-time, and the maximum
feedrate to bound the contour error by a tolerance value can be obtained from Eq.(20). It should be noted that
the polynomial approximation used in Eq.(18) only approximates GFIR at low frequency; namely, below the
first notch (ripple) of FIR filters. The first notch is typically matched with one of the structural resonances to
mitigate vibrations [1][12][24] and to minimize overall filter delay. Hence, the 4-term Taylor expansion is
suitable for practice.

3. Multi-segmented tool-path interpolation strategy

Previous section discussed generation of basic linear and circular trajectories based on FIR filtering of
axis velocity pulse commands. This section focuses on continuous and accurate interpolation of multi-
segmented tool-paths.
Figure 8 presents the overall strategy for online FIR based interpolation of multi-segmented NC tool-
paths. Each segment of the tool-path in a NC block/G-code is represented by a timed feed pulse. As
introduced in the previous section, depending on the interpolation type, e.g. linear or circular, tangential feed
pulse is dissolved into its axis velocity pulses, which are then filtered through chain of FIR filters to generate
smooth axis reference motion commands. The NC part program can be interpolated based on two types of
motion; namely, “point-to-point” (P2P), or non-stop “contouring”. The motion type can be controlled by
adjusting the “dwell” time between consecutive feed pulses (See Figure 8). For instance, in P2P mode, an
instantaneous stop between the programmed segments is desired. This kind of motion strategy is typically
employed in pick-and-place operations, ultra-precision machining and measurement. P2P motion can be
achieved by simply accounting for the FIR filter delay and adding a dwell time between consecutive feed
pulses that is equal to the filter delay Td as outlined in Figure 8. Figure 9a shows an example P2P motion
generated along 2 linear segments. Consecutive feed pulses are filtered with a dwell time of Td=T1+T2 (in
case of 2 FIR filters) to generate a full stop at P1 and the corresponding velocity profile is shown in Figure 9b.
 Figure 8: Overall path interpolation strategy Figure 9: Kinematic profiles during contouring motion

3.1. Contour error control during non-stop linear interpolation

On the other hand, in high-speed machining, uninterrupted accurate contouring motion is desired. As
presented in Figure 8, non-stop “contouring” motion can be generated by interpolating consecutive feed
pulses without fully waiting for the filter delay to die out. This enforces convolution of consecutive feed
pulse to begin with non-zero initial conditions, and through precise control of the dwell time, contouring
errors [29][30] along segment transitions can be confined.
Firstly, let us define an “overlapping time”, Tk, to control the overlap of convolution of the consecutive feed
pulses:
(21)

If Tk=0, the dwell time is equal to the total filter delay Td, and as shown in Figure 9a and b, a P2P motion
is generated. Figure 9c illustrates the case when Tk=Td/2. In this case, consecutive segment interpolation is
initiated before feed motion of the 1st block comes to a full-stop. As a result, feed direction is altered
continuously. Due to this gradual change in the feed direction an interpolation error, , occurs around the
junction point of consecutive path segments (See Figure 9a). When the overlapping time Tk is increased to its
upper limit Tk=Td, no dwell time is inserted between consecutive feed pulses. Since change in the feed
direction is also initiated earlier, larger interpolation contour error occurs as shown in Figure 9d.
The contour error around the segment junction due to non-stop change in the feed direction can be
controlled analytically. Consider the generic path shown in Figure 9, the deceleration motion towards
midpoint, P1 starts at t=Tv. When a non-zero overlapping time is set Tk>0, feed direction towards the
endpoint P2 is initiated with the start of convolution of the consecutive segment at t=Tv+Td -Tk. Note that
convolution of the 1st segment finishes at t=Tv+Td, which marks the completion of feed direction change. If
the feedrate F at consecutive segments is identical, total axis (x and y) velocity traverse, i.e. change in feed
direction, is controlled directly by the angle between linear segments as:

(22)

where represent axis velocities along the 1st linear segment and represent velocities on the 2nd
(consecutive) segment, and is the angle between the linear segments. Thus, when feed pulses with
identical amplitude F are commanded, deceleration/acceleration kinematics around the bisector from the
segment junction P1 becomes mirror-imaged (See Figure 9a). As a result, tangential feedrate exhibits its
minimum in the middle of the segment transition at t=Tv+Td-Tk/2. Similarly, the interpolated trajectory also
becomes symmetric where the maximum deviation from the junction point occurs along the bisector at
t=Tv+Td-Tk/2.

Figure 10: Axis kinematic profiles during segment transition

Kinematic profiles during segment transition depend on the overlapping time, Tk, and the filter delay, Td.
They can be computed by superimposing filtered velocity profiles of consecutive segments as shown in
Figure 10. For a 7-segmented trapezoidal acceleration profile the x and y-axis velocities during segment
transition can be calculated from Eq.(6) as:
 ì -
1 Fx,y
ï ( t - Tv )
- 2
Fx,y - Tv £ t £ Tv + T2
ï 2 T1T2
ï - -
1 Fx,yT2 Fx,y
ï
ï
-
F -
x,y
2 T1
-
T1
(
t - ( Tv + T2 ) ) Tv + T2 £ t £ Tv + T1
ï
ï -
1 Fx,y
(
(T + T + T ) - t )
2
ï Tv + T1 £ t £ Tv + T1 + T2 - Tk
ï 2 T1T2 v 1 2
ï - +
1 Fx,y 1 Fx,y
ï
( ( ) ) (
t - ( Tv + T1 + T2 - Tk ) )
2 2
v ' x, y (t) = í T + T + T - t + Tv + T1 + T2 - Tk £ t £ Tv + T1 + T2
ï 2 T1T2 v 1 2 2 T1T2
ï +
1 Fx,y
ï (
t - (Tv + T1 + T2 - Tk ) )
2
Tv + T1 + T2 £ t £ Tv + T1 + 2T2 - Tk
ï 2 T1T2
ï + +
ï 1 Fx,yT2 Fx,y
ï 2 T1
+
T1
(
t - ( Tv + T1 + 2T2 - Tk ) ) Tv + T1 + 2T2 - Tk £ t £ Tv + 2T1 + T2 - Tk
ï
ï +
1 Fx,y
(
(T + 2T1 + 2T2 - Tk ) - t )
2
ï +
Fx,y - Tv + 2T1 + T2 - Tk £ t £ Tv + 2T1 + 2T2 - Tk
ï 2 T1T2 v
î
(23)
for Tk<2T2. Similarly, maximum contouring error along the bisector can be calculated by superimposing the
remaining distance towards the midpoint P1 during interpolation of the 1st segment, and the distance traveled
due to the convolution of the 2nd segment. Based on Figure 10 Cartesian components of the maximum contour
error can be written as:

(24)

and integrating axis velocity profiles and from Eq.(6) yields the maximum contour error during
uninterrupted interpolation of consecutive linear segments from Eq.(24) as:

(25)

Finally, for a predetermined contour error tolerance the overlapping time can be solved from Eq.(25) as:
ì 24T1T2e
ï 0 £ Tk £ 2T2
F sin ( b 2 )
3
ïï
Tk = í (26)
ï 4T1e T2
ï T2 + - 2
F sin ( b 2) 3
2T2 < Tk £ T1 + T2
ïî

As presented above, the dwell time between consecutive interpolation of feed pulses is controlled by the
amount of Tk from Eq.(26), and interpolation contour error is confined by a predetermined value efficiently.
 3.2. Control of contour errors during non-stop linear and circular interpolation
As shown in Figure 11, contour errors occur during non-stop interpolation through circular (G2/G3) and
linear (G1) segments as well. The dwell time control method presented in the previous section can be
adapted to confine these contour errors by approximating the change in the feed direction.
During circular interpolation, interpolated tool motion settles down on a circular path that has a smaller
radius than the reference one due to the FIR filter dynamics (See Figure 7). As shown in Figure 11, the feed
direction at the start of the circular interpolation is bounded between the tangent vector of the reference

path and the tangent vector of the interpolated path shown. is known from the reference path

geometry, and can be computed by the geometry through the following relationship:

Figure 11: Linear to circular interpolation transition

(27)

where
2 
tc  Tx  Ox  Ty  Oy  
2

 (28)
 Tx  Px   Ty  Py  
2 2
t filt


Note that norm of is known from Eq.(16),

Tx  Ox   Ty  Oy   R  
2
tc 
2
(29)

and substituting Eqs.(29) and (28) in (27) yields:

  R   2  P  O    R    R 2   R   2  Py  Oy  
  Ox  Px  i 
x x
t filt
 R 2

 
(30)
  R   2  Py  Oy   R    R 2   R   2  P  O  
 x x
 O  P j
 
y y
R2
 

where and are the unit directional vectors in x and y directions, respectively. Feed direction during
linear to circular segment G01→G02/03 transition is then bounded between the tref and t filt , and hence the
largest change can be approximated as shown in Figure 11 as:
   t ref  t ref t ref  t filt


  arccos  min  ,  (31)
  t ref t ref t ref t filt 
  

where is the feed vector along the linear path. Eq.(31) is used to calculate the overlapping time Tk and
control the maximum value of the contour error. In a similar fashion, transition from circular to linear
segment G02/03→G01 is depicted in Figure 12a, and Eq. (31) can be adapted for this case as:
  t ref  t ref t filt

 tref 
  arccos  min  ,  (32)
  t ref t ref t filt

t ref 
  
Finally, the transition between two consecutive circular segments G02/G03→G02/G03 is illustrated in
Figure 12b. In this particular case, Eqs.(31)-(32) needs to be expanded to contain all possible combinations to
bound the feed direction, and the largest angular change, i.e. worst case, is determined as:
  t ref  t ref t ref  t filt
 
t filt  tref t filt
 
 t filt 
  arccos  min    ,   ,   ,    (33)
  t ref t ref t ref t filt t filt tref t filt t filt 
  

Figure 12: Feed direction during circular and linear transitions

4. Illustrative example
This section demonstrates application of the proposed FIR based block timing control technique to
accurately interpolate machining tool-paths. The reference tool-path shown in Figure 13 is given in the
Gcode/CL program defined by two G01 linear segments followed by circular move (G02). In order to
interpolate it with a jerk limited trapezoidal acceleration profile, 2 FIR filters are used with time constants set
to T1=50 [msec] and T2 =30 [msec]. The command feedrate is set to F=200 [mm/sec], and the maximum
interpolation error tolerance is ε=100 [μm].
The path is interpolated based on the P2P and the proposed non-stop contouring type interpolation
techniques (See Figure 8). Interpolation results are summarized in Figure 13. Figure 13a and b compare P2P
and the contouring type interpolated tool trajectories, and resultant feedrate profiles. As shown, in case of
P2P interpolation, the motion undergoes a full stop at each segment junction. A dwell time identical to the
total FIR filter delay Td=30+50=80 [msec] is inserted between the blocks. The total cycle time for P2P
motion results to Ttotal=1.137 [sec]. Proposed FIR filtering based contouring type interpolation technique can
generate accurate non-stop feed motion. Contouring errors around transition of linear and circular segment
junctions as well as the circular contour are precisely kept at and below the ε=100 [μm] tolerance value.
The circular interpolation error is bounded by lowering the feedrate to 84.2 [mm/sec] from Eqs.(19) and (20).
The contour error around segment transitions are controlled by calculating the overlapping time Tk based on
the change in the feed direction from Eq.(26). The overall cycle time is reduced to Ttotal=1.074 [sec]. Next
section presents interpolation of a more complex tool-path on an actual motion stage.

Figure 13: FIR based interpolation of multi-segmented path.

5. Experimental validation
5.1. Setup and implementation
Experimental validation and benchmark comparisons of the proposed technique are performed on the
Cartesian X-Y motion system shown in Figure 14. The planar X-Y motion table is driven by 3 linear motors.
The heavier X-axis is designed as gantry and carries the lighter Y-axis. In order to implement proposed
algorithms, servo amplifiers are set to operate in torque (current) control mode. Closed loop control is
implemented in the Dspace DS1103® real time control system by reading linear encoder feedback at a
resolution of 0.1 [µm] and commanding torque signal to the servos at a closed loop sampling interval of
Ts=0.1 [msec]. Both X and Y drives are controlled by P-PI cascade motion controllers with velocity feed-
forward action. The position feedback control bandwidths of the axes are roughly matched at w n = 50 [Hz] to
ensure good motion synchronization and contouring [29].

Figure 14: Experimental XY motion platform

Next, to generate motion commands at discrete time instants kTs, FIR filter’s transfer function from
Eq.(1) needs to be discretized. A simple Euler’s backward differentiation technique [26] is applied to derive
the z-domain transfer function as:

(34)

where is the number of (delay) samples of the filter. The filtered velocity commands v ' x, y are
generated by implementing Eq.(34) through the following simple difference equation:

(35)

where k is the sample counter. Note that generation of the filtered velocity commands from above difference
equation requires only 2 additions and 1 multiplication for a single FIR filter.

5.2. Experimental Results

In the 1st experiment, the clover shaped tool-path shown in Figure 15 is interpolated with the P2P and
contouring interpolation techniques. As shown, this tool-path consists of 5 linear and 5 circular segments.
The feedrate along the tool-path is set to F = 200 [mm/sec], and the maximum interpolation contour
tolerance isε=100 [μm]. 2 FIR filters are used to interpolate the tool-path with time constants set to T1=30
and T2=25 [msec]. As shown in Figure 15, the tool-path is interpolated non-stop within desired contouring
tolerance. Maximum contouring errors both around linear and circular segment transitions as well as along
the circular sections are respected. Kinematic profiles are shown in Figure 16. As shown, feedrate is lowered
along circular sections and also at segment junctions to generate accurate transition. Figure 16b depicts the
feed pulse timing and the resultant feedrate profiles. The cycle time during contouring interpolation is clearly
shorter and axis motion profiles are acceleration and jerk limited (See Figure 16c and d). Contouring errors
are measured experimentally [29] and presented in Figure 17 as well. As shown, the tool-path is interpolated
within the given contour error tolerance. Experimentally measured contour errors show small discrepancy
from the interpolated ones due to feedback tracking dynamics of the servo system. These errors are not
accounted for in the proposed technique, and although small in this experiment (<15 [μm]), they can be
further improved by well-known feed-forward control techniques [26].
A 2nd experiment is performed to showcase vibration suppression capabilities of the proposed
interpolation scheme. The starfish shaped tool-path shown in Figure 18 is interpolated using the proposed FIR
filtering based technique and compared against another technique that fits small Bezier segments at segment
junctions to realize a continuous motion transition and thereby non-stop contouring motion [28]. The tool-
path consists of 125 short linear segments. The programmed feedrate is set to F=50[mm/sec], and the
contouring tolerance is ε=30[μm]. 2 FIR filters are used to interpolate the path with trapezoidal acceleration
profiles. The filter time constants are tuned to avoid unwanted vibrations. As shown in Figure 14, 2 flexible
beams are placed on the X-Y table. The beams are flexible in orthogonal directions, e.g. X and Y, where their
1st bending modes are identified with accelerometers (PCB-3711E1110G) mounted on the top of the beams
as: ωx ≈ 7.4[Hz] and ωy ≈ 9.2 [Hz]. In order to avoid exciting the lightly damped beam resonances, FIR
filter delays are set to T1=136.1[msec] and T2=110.9[msec], accordingly. Figure 19 shows feedrate profiles of
the interpolation techniques. Note that, due to short linear segments, the programmed feedrate is never
reached. As a result, interpolated feed profiles and the corresponding accelerations fluctuate aggressively.
Nevertheless, the proposed FIR based filtering technique delivers the fastest cycle time while respecting
desired contouring tolerance along the entire tool-path. Figure 20 compares frequency spectrum of the
interpolated acceleration profiles. As shown, FIR based interpolation technique exhibits attenuated frequency
spectrum especially around the resonances of flexible beam structures. Most of the excitation is kept in the
lower frequencies. However, the Bezier based technique simply spreads the excitation at a much wider
bandwidth. Figure 21 presents beam accelerations measured through the attached accelerometers. As shown,
the level of acceleration of the beams are significantly less as the motion stage tracks the FIR based
interpolated trajectory. The maximum acceleration is same as the interpolated one from the reference
trajectory. As compared, the acceleration level is significantly higher on the Bezier based trajectory
generation technique. Finally, Figure 22 shows the frequency spectrum of the measured beam accelerations.
As shown, the Bezier based technique clearly excites the resonances and causes beams to vibration heavily.
On the other hand, FIR based technique does not induce any unwanted vibrations and the beams move as a
rigid body with the motion table. This experiment clearly demonstrates that the proposed technique can
interpolate complex tool-paths accurately, deliver rapid non-stop contouring motion and at the same time
mitigate unwanted residual vibrations.
Figure 15: Clover shaped tool-path
Figure 16: Interpolated kinematic profiles along clover
shaped tool-path Figure 17: Contour errors during clover tool-path
Figure 18: Starfish shaped tool-path
Figure 19: Interpolated kinematic profiles along starfish shaped tool-path

Figure 20: DFT of interpolated axis acceleration profiles

Figure 21: Experimentally measured beam accelerations Figure 22: DFT of beam accelerations

6. Conclusions
A novel online trajectory generation scheme has been proposed for Cartesian machines and motion
systems to generate high-speed and accurate feed motion. Owing to its simple filtering structure, proposed
scheme can interpolate linear and circular paths with high kinematic continuity and minimum computational
load making it suitable for real-time processors. The proposed block timing technique considers the change
in the feed direction and the total delay in the filter chain to generate accurate non-stop rapid feed motion.
For the first time, interpolation errors that occur during both linear and circular segment transitions as well as
circular arcs are considered making the proposed scheme comprehensive for multi-segmented paths.
Furthermore, by tuning the filter delays with respect to the dynamics of the motion system, frequency
spectrum of the acceleration profile is shaped and unwanted residual vibrations are avoided. Experimental
results validated that the proposed technique can interpolate multi-segmented tool-paths accurately. As
compared to the state of the art technique, the proposed interpolation method can eliminate unwanted
vibrations and reduce the cycle time up to ~20% while utilizing same level of acceleration proving it to be a
practical and effective online interpolation technique form modern NC systems.
References

[1] Altintas, Y., 2000, Manufacturing Automation: Metal Cutting Mechanics, Machine Tool Vibrations, and
CNC Design, Cambridge University Press.
[2] Choi, Y. K., & Banerjee, A. (2007). Tool path generation and tolerance analysis for free-form surfaces.
International Journal of machine Tools and manufacture, 47(3), 689-696.
[3] Barre, P. J., Bearee, R., Borne, P., & Dumetz, E. (2005). Influence of a jerk controlled movement law on
the vibratory behaviour of high-dynamics systems. Journal of Intelligent and Robotic Systems, 42(3),
275-293.
[4] Erkorkmaz, K., & Altintas, Y. (2001). High speed CNC system design. Part I: jerk limited trajectory
generation and quintic spline interpolation. International Journal of machine tools and manufacture,
41(9), 1323-1345.
[5] Erkorkmaz, K., Alzaydi, A., Elfizy, A., & Engin, S. (2011). Time-optimal trajectory generation for 5-
axis on-the-fly laser drilling. CIRP Annals-Manufacturing Technology, 60(1), 411-414.
[6] Flash, T., & Hogan, N. (1985). The coordination of arm movements: an experimentally confirmed
mathematical model. The journal of Neuroscience, 5(7), 1688-1703.
[7] Sencer, B., Altintas, Y., & Croft, E. (2008). Feed optimization for five-axis CNC machine tools with
drive constraints. International Journal of Machine Tools and Manufacture, 48(7), 733-745.
[8] Lambrechts motion, P., Boerlage, M., & Steinbuch, M. (2005). Trajectory planning and feedforward
design for electromechanical systems. Control Engineering Practice, 13, 145–157.
[9] Rymansaib, Z., Iravani, P. and Sahinkaya, M.N., 2013, July. Exponential trajectory generation for point
to point motions. In Advanced Intelligent Mechatronics (AIM), 2013 IEEE/ASME International
Conference on (pp. 906-911). IEEE.
[10] Visioli, A., 2000. Trajectory planning of robot manipulators by using algebraic and trigonometric
splines. Robotica, 18(6), pp.611-631.
[11] Singhose, W. E., Searing, W. P., & Singer, N. C. (1996). Improving repeatability of coordinate
measuring machines with shaped command signals. Precision Engineering, 18(2), 138-146.
[12] Singer, N. C., & Seering, W. P. (1990). Preshaping command inputs to reduce system vibration.
Journal of Dynamic Systems, Measurement, and Control, 112(1), 76-82.
[13] Altintas, Y. and Khoshdarregi, M.R., 2012. Contour error control of CNC machine tools with
vibration avoidance. CIRP annals-manufacturing technology, 61(1), pp.335-338.
[14] Jeon JW, Kim Y-K. FPGA based acceleration and deceleration circuit for industrial robots and CNC
machine tools. Mechatronics 2002;12:635–42.
[15] Jones S, Goodall R, Gooch M. Targeted processor architectures for high-performance controller
implementation. Contr Eng Pract 1998;6:867–78.
[16] Osornio-Rios, R.A., de Jesus Romero-Troncoso, R., Herrera-Ruiz, G. and Castañeda-Miranda, R.,
2007. Computationally efficient parametric analysis of discrete-time polynomial based acceleration–
deceleration profile generation for industrial robotics and CNC machinery. Mechatronics, 17(9), pp.511-
523.
[17] Kim, D.I., Jeon, J.W. and Kim, S., 1994. Software acceleration/deceleration methods for industrial
robots and CNC machine tools. Mechatronics, 4(1), pp.37-53.
[18] Bonfe , M., & Secchi, C. (2010). Online smooth trajectory planning for mobile robots by means of
nonlinear filters. In 2010 IEEE/RSJ international conference on intelligent robots and systems (IROS)
(pp. 4299–4304).
 [19] Zanasi, R., Bianco, C. G. L., & Tonielli, A. (2000). Nonlinear filter for the generation of smooth
trajectories. Automatica, 36, 439–448.
[20] Zanasi, R., & Morselli, R. (2003). Discrete minimum time tracking problem for a chain of three
integrators with bounded input. Automatica, 39(9), 1643–1649.
[21] Zheng, C., Su, Y., & Muller, P. (2009). Simple online smooth trajectory generations for industrial
systems. Mechatronics, 19, 571–576.
[22] Chen, C.S. and Lee, A.C., 1998. Design of acceleration/deceleration profiles in motion control based
on digital FIR filters. International Journal of Machine Tools and Manufacture, 38(7), pp.799-825.
Vancouver
[23] Biagiotti, L. and Melchiorri, C., 2012. FIR filters for online trajectory planning with time-and
frequency-domain specifications. Control Engineering Practice, 20(12), pp.1385-1399.
[24] Sencer, B., Ishizaki, K., & Shamoto, E. (2015). High speed cornering strategy with confined contour
error and vibration suppression for CNC machine tools. CIRP Annals-Manufacturing Technology, 64(1),
369-372.
[25] Tsai, M. S., & Huang, Y. C. (2016). A novel integrated dynamic acceleration/deceleration
interpolation algorithm for a CNC controller. The International Journal of Advanced Manufacturing
Technology, 1-14.
[26] Ogata, K. and Yang, Y., 1970. Modern control engineering.
[27] Wiener, N. (1988). The Fourier integral and certain of its applications. CUP Archive.
[28] Sencer, B., Ishizaki, K., & Shamoto, E. (2015). A curvature optimal sharp corner smoothing
拐角插入5次 algorithm for high-speed feed motion generation of NC systems along linear tool paths. The
塞尔过渡, 三 J International Journal of Advanced Manufacturing Technology, 76(9-12), 1977-1992.
[29] Erkorkmaz, K., Yeung, C. H., & Altintas, Y. (2006). Virtual CNC system. Part II. High speed
contouring application. International Journal of Machine Tools and Manufacture, 46(10), 1124-1138.
[30] Goto, S., Nakamura, M. and Kyura, N., 1995, May. Trajectory generation of industrial mechatronic
systems to achieve accurate contour control performance under torque saturation. In Robotics and
Automation, 1995. Proceedings., 1995 IEEE International Conference on (Vol. 3, pp. 2401-2406). IEEE.
[31] Yeh, S. S., & Hsu, P. L. (2004). Perfectly matched feedback control and its integrated design for
multiaxis motion systems. Journal of Dynamic Systems, Measurement, and Control, 126(3), 547-557.