JOEL M.G.

Modelling the suspension of a car with analogue

computers
Joel Martı́nez González

Abstract—Using operational amplifiers and some basic cir- In both cases, an illustrative solution can be obtained for an
cuitry, we built two analogue computers to solve in real-time the input of the form ℜ[X0 eiωt ] using a trial solution y(t) =
differential equations that describe a car’s suspension according ℜ[Ãeiωt ], so that for the basic system
to two different models. The analogue computer that considered
the force on the wheel as input suffered from design issues, but X0 /m
still managed to reproduce theoretical phenomena and solve for à = , (3)
iγω + ω02 − ω 2
different inputs. The other analogue computer, which considered
the height of the road as input, performed fully satisfactorily in and for the higher fidelity system
all regards within the uncertainty range.
X0 (ω02 + iγω)
à = . (4)
iγω + ω02 − ω 2
I. I NTRODUCTION
NALOGUE computers were the first computing devices This way, expressions for |Ã| and ϕ in à = |Ã|eiϕ can be
A developed to perform mathematical operations and solve
complex problems. These machines used physical quantities
obtained[1]. Both models show similar theoretical phenomena,
such as resonance or different damping regimes. In more
(e.g., voltage, current, etc) to represent and manipulate the general cases, the complexity of the solution will depend on
values of some corresponding quantities. They were later the complexity of the input.
replaced by digital computers, which offered more flexibility
and the potential to solve much more complex problems. In this experiment, we studied both systems making
Analogue computers have a precision limited by their use of four types of operational amplifiers to build an
components, a limited storage capacity, and range constraints analogue computer, which allowed us to manipulate the
in the physical quantities used (e.g., the voltage must remain voltage in different ways, controlling for the relevant resistor
within the use range of the components). However, in some and capacitor values. For convenience, the circuit modeling
cases, analogue computers can still prove useful. Such is (1) will be referenced as FIC (force-input circuit), and the
the case of electric analogue computers to solve differential one modeling (2) as HIC (height-input circuit).
equations. A real-life system modeled by a simple enough
ODE that may require quick solving is the suspension system III. P RACTICAL ASPECTS
of a car, which is precisely the one we studied. A. Circuit design and validation testing
We built an analogue computer for each theoretical model.
The aim of this project was to build and test an analogue An input voltage signal, representing either the force on
computer capable of this, obtaining a reasonable estimation the wheel or the height or the floor, was input into the
of its accuracy with the help of different simulations. corresponding of the two analogue computers we built. These
circuits then exploited the properties of different operational
II. T HEORY amplifiers (summing, inverting, integrating, and differentiating
The suspension system of a car can be simplified into four amplifiers) to output a signal representing the height of the
mass-spring-damper systems, each corresponding to one wheel wheel. In either case, the process essentially consisted of
and one-fourth of the total mass. These systems are effectively adding the different terms on the RHS of (1) and (2), and then
forced, damped, harmonic oscillators, which can be modeled integrating twice. A more detailed schematic can be found in
as Fig. 1. Before building the circuits, our design was tested in an
f (t) b k
ÿ = − γ ẏ − ω02 y, γ = , ω02 = , (1) LTspice simulation. This simulation was then matched against
m m m a numerical simulation in Python, which used a fourth-orther
where b is the viscous damping coefficient, k is the elastic Runge-Kutta[4] method to solve the constituting equations.
constant of the spring, f (t) is the force exerted by the floor,
and y is the displacement of the mass from the equilibrium
B. Building
position y0 = mg/k. In our case we chose, m = Mcar /4 =
360 Kg and ω0 = 2π · 1.18 ≈ 7.4 rad/s, while we varied the Once the design matched the numerical results, the circuit
damping coefficient. A system with higher fidelity to the real was built on a protoboard, using commercial resistors
system can be built, where instead of the input force f (t), the (±5% tolerance), LT1001 operational amplifiers, commercial
input is the road height h(t): electrolytic capacitors (±20% tolerance), and a 1N4148
diode. An oscilloscope was used to provide the input signal,
ÿ = −ω02 (y − h(t)) − γ(ẏ − ḣ(t)). (2) as well as measure the output through a scope probe to
JOEL M.G. 2

Fig. 1: Flow schematic of a naive approach to the two analogue

computers built. In blue, the section specific to the force-
input circuit. In yellow, the section specific to the height-
input circuit. Each circle represents an op-amp: inverting
(for scaling), summing, differentiator, or integrator. All the
terms that make the acceleration ÿ are added in the summing
amplifier, and then integrated twice to give y.

Fig. 2: Diagram of the height-input circuit (HIC), which

reduce loading. Two different power sources were combined encompasses the force-input circuit (FIC). In yellow are the
to power the operational amplifiers with ±15 V. components exclusive to the HIC (the rest are shared). The
thick arrows show the respective input and output points.
We aimed to reduce the number of op-amps used to 1. Summing amplifier. 2. First integrator. 3. Final integrator,
minimize the error. Since every type of op-amp in (6) allows gives the output. 4. Inverting amplifier with gain -1. 5. Ca-
to scale the signal and we can redefine the input as already pacitors connecting the ±15 V terminals to reduce coupling.
scaled, we can, looking at Fig. 1, effectively include the 6. Optional half-wave rectifier. 7. Differentiator amplifier. 8.
1
m in the input (in the FIC), and perform all other scaling Inverting amplifier with gain -1.
(for the FIC and HIC) inside the summing amplifier (though
two inverting op-amps were still needed in the positions of
the +ω02 and −γ 2 scalings to adjust the sign). To further multiple runs (in our case, N ≈ 1000 runs) according to
deal with the unit disparity between input and output in a uniform distribution (if more information was known
the FIC, the input was also assumed scaled by −10−1 , about the manufacturing process, a normal distribution would
giving the final scaling factors: f (t) = λF IC · Input(t), probably be more appropriate). This statistical approach allows
where λF IC = −3600 and h(t) = λHIC · Input(t), where obtaining both an estimation for the upper and lower bounds
λF IC = 1. This removes a total of three op-amps from and an estimation for the expected error from the standard
the naive design. While performing all scaling inside the deviation. The effect of noise was also accounted for through
summing amplifier is elegant, we decided to downscale the the standard deviation from the filtered signal (through a
output of the amplifier to then upscale it back in the first low-pass, first-order, Butterworth filter[5]). Errors in physical
integrator to reduce the chances of clipping, i.e., the voltage parameters were estimated through error propagation in their
being forced inside the ±15 V range given by the op-amps relations to resistor values (e.g., γ = Rf /R2 , ω02 = Rf /R1 ,
and therefore distorting the solution. This is because the where thePresistors correspond to the summing amplifier
R
summing amplifier was found to generally be the point of Vout = − i Rfi Vi ).
maximum voltage. A diagram of the circuit can be found in
Fig. 2, which shows how Fig. 1 was built in the breadboard IV. R ESULTS AND DISCUSSION
after the aforementioned optimizations. The analogue computers were tested against theoretical
predictions, as well as different input profiles.
C. Circuit testing and error estimation
Using different signal profiles, both builds were tested A. Contrasting with theory
against the expected results from the numerical and LTspice By setting the input signal as a long square wave, we can
simulations, as well as known theoretical results such as the test the effects of a sudden drop in voltage and the response
different damping regimes and resonance. to the subsequent constant input voltage. This is analogous
to setting the initial height of the wheel displaced from the
LTspice was used to estimate the systematic error originating equilibrium position, which depending on the damping regime
from the tolerance of the components. This was done through should show either underdamped, overdamped or critically
Monte Carlo tolerance analysis[2], randomly varying the damped oscillations. This will depend on the parameter
value of the components inside the tolerance range over ζ = 2ωγ 0 , so that ζ < 1 corresponds to underdamped
JOEL M.G. 3

Fig. 3: Different damping regimes in the FIC, depending (a)

on the parameter ζ. The theoretical equilibrium positions at
|yeq | = 0.458 ± 0.032 V are represented by the dashed lines.
The data were filtered with a first-order Butterworth filter. The
overdamped, underdamped, and critically damped regimes are
clearly visible, with the ζ = 0.451 ± 0.028 case corresponding
to adequate suspension.

oscillations, ζ = 1 to critically damped, and ζ > 1 to

overdamped. Additionally, for a suspension system, ζ should
ideally lie between 0.3 and 0.6[3], so that the movement of
the wheel is neither too rigid nor too loose.

Fig 3 shows the expected regimes in the FIC for a

(b)
range of ζ values (the HIC circuit shows similar results), with
the critically damping correctly being the sharpest decline. Fig. 4: Resonance behavior in each analogue computer (a)
The adequate response to a constant force is also seen, where Force-input circuit (FIC) (b) height-input circuit (HIC). The
the position of the wheel converges to the new equilibrium FIC (both LTspice simulation and experimental data) presents
position |yeq | = |λF IC V0 |/k = 0.458 ± 0.032 V. signicant deviations from the theory, while the HIC shows
close agreement with the theory.
If the input is sinusoidal, the output’s amplitude and
phase will directly depend on its frequency (equations (3)
and (4)). Focusing on the amplitude’s frequency dependence B. Specific test cases
(resonance) we tested all our sources of data against the
theory for both models. In Fig. 4, it is shown how the Python We devised multiple road profiles to test each circuit.
simulations match the theory, as expected. The errors were However, here are only shown what we considered to be the
taken as a combination of noise effects and the standard most relevant ones. Some of the profiles we tested involved
deviation from the Monte Carlo analysis. The HIC circuit discontinuities and could not be tested against the Python
also shows close agreement between the LTspice simulation simulation (since it became unstable). The experimental data
and the theory, while the experimental values are within the in those cases are only approximations (the ”infinities” that
estimated error range. The FIC circuit shows that experimental would occur are clipped inside the ±15 V range). It is worth
and LTspice values are within the error range, and both are noting that in the regime of input frequencies and amplitudes
far from the theory for ω ∈ (5, 20) rad/s. This points to tested, the noise level stayed under 2% and did not play a
a fundamental flaw in the design of the circuit that was relevant role.
not flagged during validation testing: frequency-dependent
clipping. The amplitudes used to obtain Fig. 4 (a) were First, for the basic circuit, we tested a triangular and a
observed to not cause clipping, but this is really only the case square force profile (Fig. 5). The range determined by
when ω < ω0 , and despite our precautions, we found clipping the Monte Carlo analysis through LTspice is quite large
in the summing amplifier of the FIC in those cases. (especially in Fig. 5 (b)), evidencing the effects of clipping
JOEL M.G. 4

(a) (a)

(b) (b)

Fig. 5: Test cases for the force-input circuit (FIC). (a) Triangu- Fig. 6: Test cases for the height-input circuit (HIC). (a) Ramp-
lar force. (b) Square force. The dimmed blue profile represents shaped bumps. (b) Half-wave-shaped bumps. The dimmed blue
the upper and lower limits of a Monte Carlo simulation in profile represents the upper and lower limits of a Monte Carlo
LTspice. The units of the input (force) are shown on the left simulation in LTspice. Data were filtered with a first-order
and the units of the output (height) are on the right. Data were Butterworth filter.
filtered with a first-order Butterworth filter.

in the FIC circuit. The triangular profile shows a reasonable

of a car. We tested them against theoretical predictions
agreement between the Python simulation and experimental
and different input profiles. Due to unexpected clipping,
data, though a slight phase and amplitude difference is visible
the force-input circuit (FIC) was not as accurate as the
(still well inside the Monte Carlo range).
height-input circuit (HIC) but was still able to reproduce the
different damping regimes and solve for different input shapes.
Regarding the HIC circuit, we tested a road with bumps in
the form of ramps and half-waves. To achieve the half-wave
It is evident that the clipping in the FIC is the main
rectified input, we used the diode mentioned in the previous
area of improvement, which could be addressed by increasing
section together with a 1k Ohm resistor. Shown in Fig.
the input voltage to physical units ratio, or further downscaling
6, the cicuit reproduced the intuitively expected behavior.
the output of the summing amplifier. Both circuits would
The Python simulation and experimental data are close, and
benefit from components with lower tolerance (especially the
the visible deviation is mainly due to the imperfections in
capacitors). This project outlined the effectiveness of analogue
the experimental half-wave. The Monte Carlo ranges were
computers at solving ODEs, as well as the difficulties that
observed to be generally smaller than those in the basic
arise from the fact that magnitudes in the model must match
circuit, though some rare clipping cases were found to
with real physical magnitudes (in our case, voltage) and the
correspond to visible protuberances in the Monte Carlo limits.
practical limitations of the circuit (e.g., the voltage range).
Looking forward, other applications of analogue computers
V. S UMMARY AND CONCLUSION could be tested, with a special interest in possible use cases
We designed, built, and tested two analogue computers in neural networks[6], where a proof of concept one-neuron
that reproduced two different models of the suspension circuit might be a feasible extension.
JOEL M.G. 5

VI. B IBLIOGRAPHY

[1] J. B. Marion and S. T. Thornton, Classical Dynamics

of Particles and Systems, 5th. Fort Worth, TX: Saunders
College Publishing, 1995, ISBN: 978-0030985437.
[2] C. Li, X. Zhang, Y. Hu, F. Liu and L. Wang, “Monte
carlo tolerance analysis in electronic circuits: Theory and
applications,” Microelectronics Reliability, jourvol 61,
pages 28–39, 2016, ISSN: 0026-2714. DOI: 10.1016/j.
microrel.2016.02.006.
[3] C. M. Carr, Project manual: Analogue computer, march
2019.
[4] R. L. Burden and J. D. Faires, Numerical Analysis,
10th. Boston: Cengage Learning, 2016, ISBN: 978-1-305-
25333-0.
[5] S. Butterworth, “On the theory of filter amplifiers,”
Wireless Engineer, jourvol 7, pages 536–541, 1930.
[6] B. Widrow and M. A. Lehr, “30 years of adaptive
neural networks: Perceptron, madaline, and backpropa-
gation,” Proceedings of the IEEE, jourvol 78, number 9,
pages 1415–1442, 1990.