Mechanical Systems and Signal Processing 98 (2018) 281–291

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing

journal homepage: www.elsevier.com/locate/ymssp

Shaper-Based Filters for the compensation of the load cell

response in dynamic mass measurement
Dario Richiedei ⇑, Alberto Trevisani
Università degli Studi di Padova, Stradella San Nicola 3, 36100 Vicenza, Italy

a r t i c l e i n f o a b s t r a c t

Article history: This paper proposes a novel model-based signal filtering technique for dynamic mass mea-
Received 29 December 2015 surement through load cells. Load cells are sensors with an underdamped oscillatory
Received in revised form 1 February 2017 response which usually imposes a long settling time. Real-time filtering is therefore neces-
Accepted 29 April 2017
sary to compensate for such a dynamics and to quickly retrieve the mass of the measurand
(which is the steady state value of the load cell response) before the measured signal actu-
ally settles. This problem has a big impact on the throughput of industrial weighing
Keywords:
machines. In this paper a novel solution to this problem is developed: a model-based filter-
Load cells
Signal processing
ing technique is proposed to ensure accurate, robust and rapid estimation of the mass of
Dynamic mass measurement the measurand. The digital filters proposed are referred to as Shaper-Based Filters (SBFs)
Model-based filtering and are based on the convolution of the load cell output signal with a sequence of few
Shaper Based Filters impulses (typically, between 2 and 5). The amplitudes and the instants of application of
such impulses are computed through the analytical development of the load cell step
response, by imposing the admissible residual oscillation in the steady-state filtered signal
and by requiring the desired sensitivity of the filter. The inclusion of robustness specifica-
tions tackles effectively the unavoidable uncertainty and variability in the load cell fre-
quency and damping. The effectiveness of the proposed filters is proved experimentally
through an industrial set up: the load-cell-instrumented weigh bucket of a multihead
weighing machine for packaging. A performance comparison with other benchmark filters
is provided and discussed too.
Ó 2017 Elsevier Ltd. All rights reserved.

1. Introduction

Load cells are sensors widely employed for performing dynamic weighing, i.e. to compute accurately and rapidly the
value of the measurand while the sensor output is still in oscillation and before stable equilibrium is achieved. There are
several industrial applications where dynamic and automatic weighing is requested. Challenging examples are multihead
weighing machines, extensively employed, for example, in the food industry, where load-cell-instrumented weigh buckets
should measure accurately and very quickly the weight of parts falling from accumulation chambers or conveyors above the
buckets. To cope with the underdamped oscillatory response of load cells and with the presence of impulsive and step-like
excitations, signal filtering is essential since the use of active or passive control techniques to increase damping is very com-
plicate, expensive and often not effective. Indeed, active control would require control forces and hence actuators (see e.g.
[1]), while passive damping through dampers increases the rise time of the measured signals. Effective filters should be

⇑ Corresponding author.
E-mail addresses: [email protected] (D. Richiedei), [email protected] (A. Trevisani).

http://dx.doi.org/10.1016/j.ymssp.2017.04.049
0888-3270/Ó 2017 Elsevier Ltd. All rights reserved.
282 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291

therefore synthesized to ensure a short settling time, no steady-state error, robustness with respect to the unavoidable
uncertainty and variability of the system physical parameters, and finally ease of implementation, small computational effort
and memory allocation to guarantee real-time processing.
In dynamic mass measurement, the synthesis of filters is a critical and not trivial task. Indeed, by solely decreasing filter
bandwidths to cut high frequency disturbances one usually increases settling times and hence downgrades the measure-
ment speed and the throughput of the machine using such filters. Besides the general-purpose filters developed in the field
of signal processing (such as the well-known Bessel or Butterworth filters), some filtering techniques have been specifically
addressed to load cell response correction. Model-based filtering schemes have been often synthesized in the literature.
Adaptive schemes have been proposed in [2,3] to perform pole-zero cancellation or to optimally trade-off between band-
width and settling time [4]. The use of a Kalman-filter has been exploited in [5] for objects moving on a weighing table with
known speed. A filtering technique based on an array of accelerometers has been proposed in [6] to remove the low fre-
quency components in load cell signals that are due to the frame flexibility. A different type of approach is the synthesis
of non-model based filters, such as those using neural networks [7], recursive least-square procedures [8], model-free
discrete non-stationary or time-variant filters [9–12].
In the light of the previous requirements, this paper proposes a novel model-based filtering technique which is devoted
for dynamic weighing through load cells. The proposed method takes advantage of the idea of the Input Shaping feedforward
control (IS), that has been developed for the optimal planning of command position references in flexible systems (see e.g.
[13–15]). The original application of IS is motion control of trolley-pendulum systems, such as cranes. IS physically modifies
the actuator force or the trolley displacement to ensure zero residual load oscillations after the end of the crane motion.
Hence it is a motion planning technique, that can be thought of as an open-loop control. Motion planning through IS relies
on the convolution of the actuator force or of the position reference of the crane with a sequence of a few impulses, referred
to as the shaper, to modify the crane trajectory in time so that load swing is controlled. The idea of convolving a signal with a
baseline of few impulses to compensate for elastic oscillation is here assumed as the starting point to develop a new filtering
technique for load cells. Rather than being convolved with a physical input, as in the IS, the sequence of impulses is here
convolved with the system output, that is the load cell measured signal. These filters are denoted Shaper-Based Filters (SBFs)
and, to the best of the Author’s knowledge, no extension of the IS theory in the field of signal processing can be found in
literature. SBFs are online signal processing schemes employed to quickly retrieve the steady state value of the load cell sig-
nal before it settles. In contrast with the original idea of IS, the filters proposed in this work are not employed to perform a
physical modification of the input exciting the flexible system, which is not allowed in weighing operations.
The filtering scheme proposed fits well the requirements of short settling time, accuracy and robustness. The potential
benefits of this kind of filters includes also ease of tuning, low computational effort and small memory required, since filter-
ing relies on the convolution of the measured signal with just a few impulses. All these features make the proposed SBFs
suitable for real-time signal processing, as required in industrial weighing machines.
The paper is organized as follows: Section 2 proposes the synthesis of the simplest form of a SBF, starting from load cell
dynamic model. The robustness issue is discussed in Section 3 and robust filters are synthesized. An experimental
verification of the effectiveness of the proposed filters is proposed in Section 4 by means of an industrial set up: the load-
cell-instrumented weigh bucket of a multihead weighing machine for food packaging. In the same Section a comparison with
the results obtained employing other popular filtering techniques is provided. Concluding remarks are finally given in
Section 5.

2. Basic theory of Shaper-Based Filters

2.1. Load cell model

The filter synthesis takes advantage of load cell dynamic model. For this purpose, the dynamic behavior of load cells com-
monly used in weighing devices, such as strain-gauge load cells, can be modeled through a single degree-of-freedom lumped
model (whenever they are fixed on a rigid frame). If the sensor transverse sensitivity is neglected and the elastic behavior is
linear, as it is common in practice, the second-order spring-damper-mass system model can be assumed (see e.g. [6]):

€eðtÞ þ 2nxN e_ ðtÞ þ x2N eðtÞ ¼ cx2N MðtÞ ð1Þ

where eðtÞ is the load cell response (i.e. the measured signal), c is the load cell static gain, and MðtÞ is the measurand mass. By
assuming a lumped parameter model the natural frequency xN , and the damping ratio n are defined as:
sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
k c
xN ðtÞ ¼ ; nðtÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ð2Þ
Mc þ MðtÞ 2 kðM c þ MðtÞÞ
In Eq. (1), k and c are, respectively, the constant load cell equivalent linear stiffness and damping coefficient in the mea-
suring direction. Mc is the equivalent mass of the empty load cell, which comprises the contribution of the uniformly
distributed mass of the sensor and the mass of the bucket. Fig. 1 sketches the system studied.
 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291 283

M(t)g

ε (t)
Load cell
Mc measuring
direction
Fig. 1. Scheme of the studied system.

2.2. Theoretical development of the zero vibration SBF

In this section the transfer function of a SBF is obtained. A step excitation of the measurand mass is initially considered to
simplify the development of the equations. The pole-zero analysis in Section 2.3 will prove that the filters obtained can be
applied to arbitrary types of excitation generated by the measurand mass.
Let us consider the unitary step response of a load cell modeled through Eq. (1):
8
< 0   t < 0
eðtÞ ¼ nxN t pffiffiffiffiffiffiffiffi sinðxD tÞ þ cosðxD tÞ ð3Þ
:c 1  e n
2
tP0
1n

pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
where xD ¼ xN 1  n2 is the damped natural frequency. Without lack of generality, it is assumed that the step is applied at
the time instant t = 0.
Filtering is performed by convolving the sensed signal eðtÞ with the filter impulse response FðtÞ, the so called shaper,
which is made by the sequence of N > 1 Dirac impulses dðt  t i Þ exerted at time ti and whose amplitude is Ai:
X
N
FðtÞ ¼ Ai dðt  t i Þ ð4Þ
i¼1

The filter transfer function FðsÞ can be therefore expressed in the Laplace domain as follows:
X
N
FðsÞ ¼ Ai esti ð5Þ
i¼1

Let us first assume N = 2, for clarity of explanation, which is the simplest type of SBF. The extension and the implications
on assuming more than two impulses are discussed in Section 3. The convolution of the sensed signal with the filter in Eq. (4)
leads to the filtered signal eF ðtÞ. The filter response for any t P t 2 (i.e. after the last impulse of the shaper) is
!
n
eF ðtÞ ¼ cðA1 þ A2 Þ  cA1 enxN ðtt1 Þ pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi sinðxD ðt  t1 ÞÞ þ cosðxD ðt  t1 ÞÞ
1n
!
nxN ðtt2 Þ n
 cA2 e pffiffiffiffiffiffiffiffiffiffiffiffiffiffi sinðxD ðt  t2 ÞÞ þ cosðxD ðt  t2 ÞÞ ð6Þ
1  n2
Eq. (6) can be expressed in a more compact form:
eF ðtÞ ¼ cðA1 þ A2 Þ  c sinðxD tÞ UðtÞ  c cosðxD tÞHðtÞ ð7Þ
The two terms UðtÞ and HðtÞ are defined as follows:
n n
UðtÞ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi A1 enxN ðtt1 Þ cosðxD t 1 Þ þ A1 enxN ðtt1 Þ sinðxD t 1 Þ þ pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi A2 enxN ðtt2 Þ cosðxD t 2 Þ þ A2 enxN ðtt2 Þ
1n 1n
 sinðxD t 2 Þ ð8Þ

n n
HðtÞ ¼  pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi A1 enxN ðtt1 Þ sinðxD t 1 Þ þ A1 enxN ðtt1 Þ sinðxD t 1 Þ  pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi A2 enxN ðtt2 Þ sinðxD t 2 Þ þ A2 enxN ðtt2 Þ
1n 1n
 sinðxD t 2 Þ ð9Þ
The values of the suitable impulse amplitudes ðA1 ; A2 Þ and time locations ðt 1 ; t 2 Þ can be calculated by requiring the
residual vibration to be zero after the last impulse, i.e. eF ðtÞ ¼ c for any t P t 2 , which holds if and only if
U ¼ 0 and H ¼ 0 ð10Þ
Additionally, the unitary steady-state gain condition is posed to ensure correct estimation of the measurand mass after
the transient:
284 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291

2
Load cell response
Filtered signal
weight 1.5

0.5

0
0 1 2 3 4 5 6 7 8
t [s]

A 1 response
1
A 2 response
0.8 Filtered signal
weight

0.6

0.4

0.2

0
0 1 2 3 4 5 6 7 8
t [s]

Fig. 2. Principle of the ZV SBF.

X
N
lim FðsÞ ¼ Ai ¼ 1 ð11Þ
s!0 i¼1

The solution of the two equations in Eq. (10) in the presence of the constraint in Eq. (11) leads to the following results:

1
A1 ¼ t1 ¼ 0
pffiffiffiffiffiffi
pn ffi

1 þ e 1n2 ð12Þ
p
A2 ¼ 1  A1 t2 ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi
xN 1  n2
The nonlinear Eq. (10) admits an infinite number of solutions. In Eq. (12) it has been chosen the one with the smallest t 2 ,
i.e. with the earliest time at which the second impulse should be applied. This choice minimizes the settling time of the filter.
Such a time instant results at one half of the damped natural period of the load cell, 2p=xD .
Given the requirement of zero residual vibration, SBFs with two impulses are referred to as the Zero Vibration (ZV)
filter. ZV SBFs require the smallest number of impulses and therefore ensure the shortest time to settle down the filtered
signal in the ideal case (i.e. in the case of exact knowledge of the load cell frequency and damping ratio, and absence of
disturbances).
The graphical explanation of ZV SBFs is represented in Fig. 2 through a numerical example with c=1, xN ¼ 2p rad/s and
n ¼ 0:05. Fig. 2a shows the load cell response to a unitary step, i.e. the unfiltered signal, while Fig. 2b shows the filtered signal
and its components. The two signal components due to the convolution with the two impulses (denoted A1 and A2 responses
in Fig. 2), add destructively to yield zero residual vibration. The capability of SBFs to cancel the load cell signal oscillation can
be therefore interpreted as the destructive interference of damped sinusoidal waves.
The result obtained here agrees with the one usually proposed in the motion planning of flexible systems [13], where
optimal values of t1 , t2 , A1 and A2 are calculated for impulse excitations. It can be easily proved, by following with the same
approach proposed in this section that the baseline of impulses shown in Eq. (12) ensures zero residual vibration in the pres-
ence of ramp (or saturated ramp) excitations, which can be related to the dynamic measurement of bulk materials in weigh-
ing machines.
 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291 285

2.3. Pole-zero analysis

The analysis of the zeros and the poles of the transfer function in Eq. (5) provides another interesting interpretation of
SBFs and allows to extend their use for an arbitrary excitation. With reference to the ZV SBF introduced in Section 2.2, solving
for zeros and poles of the non-rational transfer function in Eq. (5) leads to the following result (see also Fig. 3):

FðsÞ ¼ 1 () s ¼ 1  jx ð8xÞ
ð13Þ
FðsÞ ¼ 0 () s ¼ nxN  jxd ð2n þ 1Þ n ¼ 0; 1; 2; . . .
Eq. (13) reveals that a ZV SBF places an infinite number of stable poles with infinite and negative real part, and an infinite
‘‘column” of pair of equally-spaced complex zeros having the same real part. If the load cell frequency and damping are
exactly modeled, a pair of these zeros will cancel the oscillatory poles of the load cell (see Fig. 3). This interpretation justifies
the use of SBFs for arbitrary type of measurand masses exciting the load cell.
The idea of pole-zero cancellation in load cell filtering has been already proved to be effective in the last years through
traditional low-pass or notch filters. However, at least two poles (with finite frequency) should be included in the filter to
ensure the filter to be proper. These additional poles increase the rise time and the settling time. In contrast, the not-
rational transfer function of the SBFs allows to place poles with infinite frequency. Additionally, as it is discussed in the next
section, SBFs can be designed to meet robustness specification for tackling the unavoidable uncertainty and variability of the
system model parameters.

2.4. Discrete time implementation of SBFs

The transformation in the discrete-time domain of the ZV SBF transfer function in Eq. (5) reveals that the filtered signal is
computed as the sum of the original signal, scaled through the gain A1, with a shifted signal, scaled through A2 and delayed by
t2:
 
t2
eF ðnÞ ¼ A1 eðnÞ þ A2 e n  ð14Þ
Ts

(T S is the sampling time and n denotes the time sample).

Therefore, by assuming a SBF with an arbitrary number of impulses N, the filtered signal is represented in the discrete-
time domain through the following scheme:

XN  
ti
eF ðnÞ ¼ Ai e n  ð15Þ
i¼1
Ts

Since SBFs just require the convolution with a small number of impulses, they are suitable for real-time implementation in
industrial hardware. In contrast, traditional filters usually impose the convolution with a much greater number of impulses
(the so-called filter taps), which makes their implementation more challenging and increases the filter duration (filter
length).

Im(s)
Load cell poles
ZV SBF zeros

Re(s)

Fig. 3. Pole-zero analysis of a ZV SBF.

286 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291

3. Robust SBFs

3.1. Residual oscillations and filter sensitivity

If the actual natural frequency of the load cell deviates from the nominal one, the effectiveness of the ZV decreases since
the destructive interference between damped sinusoidal waves is only partially obtained. This results in residual oscillation
and hence large settling time of the load cell filtered signal.
The effect can be evaluated analytically by computing, through Eq. (3), the residual oscillation after the last impulse of the
shaper exerted at time tn (i.e. for any t  tn):
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
cenxN t XN 2 XN 2
eF ðtÞ  c ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffi2ffi sinðxD t þ acosðnÞÞ i¼1
A i e nxN t i sinðx t Þ
D i þ i¼1
A i enxN ti cosðx t Þ
D i ð16Þ
1n
In Eq. (16) an arbitrary N has been considered. By dividing ðeF ðtÞ  c Þ by the amplitude of the steady state response c,
which is the correct value of the mass to be estimated, the time-varying amplitude V of the percentage residual oscillation
is obtained:
rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
enxN t XN 2 XN 2
Vðt; n; xN Þ ¼ pffiffiffiffiffiffiffiffiffiffiffiffiffiffi i¼1
A i e nxN t i sinðx t Þ
D i þ i¼1
A i enxN ti cosðx t Þ
D i ð17Þ
1  n2
The study of the residual oscillation as a function of the deviation between the frequency adopted for synthesizing the
filter (which can be thought of as the load cell nominal frequency) and the actual one (xN ) allows for the evaluation of
the filter sensitivity and for the design of robust filters. A similar study can be carried out to investigate the effect of the
damping estimation error.

3.2. Synthesis of robust SBFs

The value of V computed in Eq. (17) for the step response of a load cell is similar to the sensitivity analysis proposed in
[15] for the impulse responses of a crane controlled through IS. Hence, the idea proposed in [15] for the synthesis of robust
motion profiles for cranes can be adapted for synthesizing robust SBFs, capable of tackling model uncertainty. Robust filters
also allow for compensating the variability in load cell model parameters due to the product being measured, which might
cause frequency shifts of modify damping.
Robustness constraints can be set by imposing the percentage residual vibration curve Vðt ¼ t n ; n; xN Þ to approach zero in
a neighbor of the nominal frequency adopted for the filter synthesis, i.e. to be flat in a neighborhood of the nominal model.
The following additional derivative constraint is therefore included in the filter synthesis:
@Vðt n ; n; xN Þ
¼0 ð18Þ
@ xN
To satisfy this additional requirement, together with the zero-oscillation specification set in Eq. (10), three impulses
should be included in the shaper. The solution is reported in Table 1, where, for brevity of notation, the coefficient k has been
introduced:
pffiffiffiffiffiffi
pn ffi

k¼e 1n2
ð19Þ
The pole-zero analysis of the ZVD SBF reveals that it places two couples of zeros for each couple of poles of the load cell,
plus an infinite ‘‘column” of equally-spaced pairs of complex zeros. Numerical simulation of the response of the ZVD SBF is
also shown in Fig. 4 through the same example previously discussed in Section 2.2.
Higher robustness, i.e. flatter sensitivity curves in a neighborhood of the nominal model, can be obtained by also requiring
the second or the third derivatives of Vðt n ; n; xN Þ with respect to xN to be zero. These additional requirements lead to four
and five impulses respectively, and the related shapers are referred to as ZVDD (zero vibration and double derivative) and
ZVDDD shaper (zero vibration and triple derivative) [15]. The amplitude and the instants of application of the impulses

Table 1
Robust SBFs.

Impulse i ti ZVD ZVDD ZVDDD

1 0 1 1 1
1þkþk2 1þ3kþ3k2 þk3 1þ4kþ6k2 þ4k3 þk4
2 p=xD k
1þkþk2
3k
1þ3kþ3k2 þk3
4k
1þ4kþ6k2 þ4k3 þk4
3 2p=xD k2 3k2 6k2
1þkþk2 1þ3kþ3k2 þk3 1þ4kþ6k2 þ4k3 þk4
4 3p=xD – k 3
4k3
1þ3kþ3k2 þk3 1þ4kþ6k2 þ4k3 þk4
5 4p=xD – – k4
1þ4kþ6k2 þ4k3 þk4
D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291 287

Fig. 4. Principle of the ZVD filter.

Fig. 5. Residual signal oscillations of SBFs.

288 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291

are reported in Table 1. Since the filter duration is determined by the time of the last impulse, the increase in the number of
impulses causes an increase in the filter settling time in the ideal case (i.e. with no modeling error). In contrast, it is expected
that increasing robustness could have several benefits in real weighing devices where uncertainty is unavoidable.
The residual oscillations in the presence of deviation between the load cell nominal frequency and the actual one are
shown in the upper plot of Fig. 5 for the four SBFs discussed. Once robustness with respect to frequency is improved, robust-
ness with respect to damping modeling error is improved too, as it is shown in the sensitivity curves in the lower plot of
Fig. 5. Indeed, @V=@ xN ¼ 0 also ensures that @V=@n ¼ 0 in a neighborhood of the nominal model.

4. Experimental results

The experimental assessment of the method effectiveness is provided through an industrial set up: the weigh bucket of a
multihead weighing machine for food packaging [6]. A picture of the setup is provided in Fig. 6 (some parts are masked to
hide some components of the commercial device) and in Fig. 7. The weigh bucket is instrumented with a strain-gauge single-
point industrial load cell. At one end the load cell is fixed to a rigid frame, at the other it is fixed to the bucket (see Fig. 7). The
load cell signal is conditioned and acquired through a 32-bit National Instrument board, performing analog-to-digital (ADC)
conversion and recording the measured signal at a 2000 Hz sampling rate. A scheme of the experimental setup is sketched in
Fig. 8.
As previously mentioned, this set up is part of an industrial multihead weigher for the food industry. The instrumented
bucket has an openable base, driven by a mechanism. The base is opened whenever the machine controller needs unloading
the weighted material. The presence of such a mechanism, typically made through kinematic pairs with clearance, intro-
duces some non-linear disturbances in the weighing device that allow testing the SBF effectiveness and reliability in a severe
test where the actual system does not match exactly the nominal model assumed for the filter synthesis.
A sample mass of 100 g has been made fall into the bucket and represents the measurand. The same test has been
repeated 10 times to assess the filter repeatability through the mean values and the standard deviations of the settling times
recorded.
Fig. 9 shows the recorded load cell response in one sample test. It clearly highlights the underdamped behavior of the
system (a value of n = 0.05 has been estimated for the empty load cell) and the large peak in the sensor measurement
due to the impact between the mass and the bucket base.
The results of just the three robust SBFs are discussed because the ZV is, in practice, ineffective. Indeed, it does not tackle
uncertainty on the model parameters and the shift on frequency and damping, unless very precise tuning parameter is
performed. The ZVD, ZVDD and ZVDDD filters have been tuned by assuming the frequency of the empty load cell (whose
measured value is 21.8 Hz), to test the severe case where the nominal frequency deviates from the actual one. In this
way, it is shown how robust filters compensate for the frequency shift and the damping modification due to the product
within the bucket, as required through the derivative conditions (see e.g. Eq. (18)).

Fig. 6. Overview of the experimental setup. A: connection of the load-cell to the frame; B: connection of the load-cell to the bucket holder.
 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291 289

Fig. 7. Picture of the load-cell adopted in the experimental set-up. A: connection of the load-cell to the frame; B: connection of the load-cell to the bucket
holder.

M (t ) Load ε (t ) ε (n) SBF εF (n)

Pre-
cell ADC
filter

Fig. 8. Scheme of the experimental setup.

3000
Load cell response
2000 Measurand weight

1000
ε [grams]

-1000

-2000

-3000
0 0.5 1 1.5 2 2.5 3
t [s]
Fig. 9. Load cell response to the weighed 100 g.

An additional low-pass filter (2nd order Bessel filter with 45 Hz bandwidth) is also applied after the SBFs, just to remove
quantization and electrical noise. The results obtained in a sample test are shown in Fig. 10 for the three SBFs. Such a figure
clearly proves that these novel filters provide a very effective and quick settling of the estimated weight within the accuracy
threshold, which is for instance reasonably assumed in Fig. 10 as ±3 g threshold about the exact weight).
To evaluate more clearly the results, the comparison with other benchmark filters is proposed in Fig. 11 and in Tables 2
and 3. Fig. 11 shows the application of four different Bessel low-pass filters to the same sample test shown in Figs. 9 and 10.
These filters have been assumed as the benchmark since they are often recognized as among the most effective in this kind of
applications and hence are widely employed for load cell filtering. Different combinations of filter order (2nd, 4th and 6th)
and bandwidth xb (which is parameterized with respect of the load cell natural frequency xΝ) have been investigated, and a
meaningful selection is shown in the paper. Indeed, the selection of the most suitable filter parameters, in the case of tra-
ditional filters, is not straightforward and requires a large amount of trial-and-error tuning to select bandwidth and order,
or to define the suitable pass-band and stop-band specifications.
The mean value and the standard deviation (r) of the settling time ts recorded by 10 repetitions of the test are shown in
Table 2 and in Table 3 in the cases of, respectively, ±3 g and ±1 g thresholds, for both the three SBFs and the four Bessel filters.
290 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291

250
ZVD
ZVDD
200
weight [grams] ZVDDD

150

100

50

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t [s]
Fig. 10. Filtered signals with three SBFs.

250
2nd Bessel ωb=0.1 ωn

200 4th Bessel ωb=0.1 ωn

4th Bessel ωb=0.2 ωn

weight [grams]

150 6th Bessel ωb=0.2 ωn

100

50

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
t [s]
Fig. 11. Filtered signals with four examples of Bessel low-pass filters.

Table 2
3-g settling time: summary of the results.

Settling time ts ZVD ZVDD ZVDDD Bessel 4th Bessel 4th Bessel 6th Bessel 2nd
xb = xN * 0.1 xb = xN * 0.2 xb = xN * 0.2 xb = xN * 0.1
Mean [s] 0.2750 0.2857 0.2978 0.5648 0.4755 0.4005 0.4475
r [s] 0.0406 0.0424 0.0417 0.0145 0.0133 0.0360 0.0186

Table 3
1-g settling time: summary of the results.

Settling time ts ZVD ZVDD ZVDDD Bessel 4th Bessel 4th Bessel 6th Bessel 2nd
xb = xN * 0.1 xb = xN * 0.2 xb = xN * 0.2 xb = xN * 0.1
Mean [s] 0.3304 0.3179 0.3324 0.6151 0.5716 0.4779 0.5738
r [s] 0.0376 0.0401 0.0373 0.0142 0.0394 0.0560 0.0319

The comparison of Figs. 10 and 11 and the results in Tables 2 and 3, clearly highlight that SBFs allow to settle the signals in a
much shorter time, because of their very short rise time (leading to a significant signal overshoot) and their capability to
compensate for the load cell underdamped poles. The effectiveness of SBFs is even greater if the stricter ±1 g threshold is
required.
Among the three robust SBFs tested, similar settling time values are obtained. Just a minor increase in the settling time is
required by the ZVDDD filter, because of its much higher robustness that makes it suitable in the case of relevant model
uncertainty and variability. Nonetheless, it is still more effective than all the benchmark filters.
 D. Richiedei, A. Trevisani / Mechanical Systems and Signal Processing 98 (2018) 281–291 291

5. Conclusions

A novel model-based filtering technique for load cells is proposed in this work to ensure short settling time in dynamic
weighing through load cells. The filters obtained, referred to as Shaper-Based Filters (SBFs), are based on the convolution of
load cell signals with a sequence of a few impulses, typically between 2 and 5 impulses. The amplitudes and the time
instants of these impulses are computed through the system dynamic model, the level of admissible residual oscillations
in the steady-state filtered signal and the desired robustness.
The effectiveness of the SBFs has been proved experimentally by applying them to the measured signals provided by the
load cell of an industrial weigh bucket of a multihead weighing machine, and by comparing their performances with those
provided by benchmark low-pass filters typically employed in this field. The analysis of the settling time shows that SBFs
outperform the benchmark filters, by requiring shorter settling time. Indeed, the synthesis of SBFs just relies on a few design
specifications expressed in the time domain, which allow short rise times and effective compensation of the load cell under-
damped dynamics.
Besides being effective, the proposed filter structure also ensures ease of implementation, small computational effort and
small memory allocation, which makes SBFs perfectly suitable for real-time filtering in industrial applications. Additionally,
SBFs are easy and intuitive to tune. Indeed, just the load cell frequency and damping factor should be estimated with a rea-
sonable degree of accuracy, for instance through any of the simple and well-established estimation techniques usually
adopted for single-dof vibrating systems. Therefore, no time-consuming trial-and-error tuning should be performed, as it
is often the case for the other filters.

References

[1] M. Ikeshima, Weighing apparatus with Roberval mechanism. U.S. Patent No. 7,091,428 B2, 2006.
[2] M. Jafaripanah, B.M. Al-Hashimi, Application of analog adaptive filters for dynamic sensor compensation, IEEE Trans. Instrum. Meas. 54 (2005) 245–
251.
[3] W.J. Shi, N.M. White, J.E. Brignell, Adaptive filters in load cell response correction, Sensor Actuators A-Phys. 37–38 (1993) 280–285.
[4] J. Piskorowski, T. Barcinski, Dynamic compensation of load cell response: a time-varying approach, Mech. Syst. Signal Process. 22 (2008) 1694–1704.
[5] M. Halimic, W. Balachandran, Kalman filter for dynamic weighing system, in: Proceedings of the IEEE International Symposium on Industrial
Electronics, Athens, 1995, pp. 787–791.
[6] G. Boschetti, R. Caracciolo, D. Richiedei, A. Trevisani, Model-based dynamic compensation of load cell response in weighing machines affected by
environmental vibrations, Mech. Syst. Signal Process. 34 (1) (2013) 116–130.
[7] S.M.T. Alhoseyni, A. Yasin, N.M. White, The application of artificial neural network to intelligent weighing systems, IEE Proc. Sci. Meas. Technol. 146
(1999) 265–269.
[8] W. Hernandez, Improving the response of a load cell by using optimal filtering recursive least-squares (RLS) lattice algorithm to perform adaptive
filtering, Sensors 6 (2006) 697–711.
[9] P. Pietrzak, Fast filtration method for static automatic catchweighing instruments using a nonstationary filter, Metrol. Meas. Syst. 16 (2009) 669–676.
[10] R. Kaszynski, J. Piskorowski, Selected structures of filters with time-varying parameters, IEEE Trans. Instrum. Meas. 56 (2007) 2338–2345.
[11] P. Pietrzak, M. Meller, M. Niedźwiecki, Dynamic mass measurement in checkweighers using a discrete time-variant low-pass filter, Mech. Syst. Signal
Process. 48 (2014) 67–76.
[12] M. Meller, M. Niedzwiecki, P. Pietrzak, Adaptive filter approach to dynamic weighing: a checkweigher case study, in: Proceedings of the 19th IFAC
World Congress Cape Town, South Africa, 2014.
[13] N.C. Singer, W.P. Seering, Preshaping command inputs to reduce system vibration, ASME J Dyn. Syst. Meas. Control 112 (1990) 76–82.
[14] N.C. Singer, W.P. Seering, Design and comparison of command shaping methods for controlling residual vibration, in: Proceedings of the IEEE
International Conference on Robotics and Automation, Scottsdale, 1989, pp. 888–893.
[15] J. Vaughan, A. Yano, W. Singhose, Comparison of robust input shapers, J. Sound Vib. 315 (2008) 797–815.