Professional Documents
Culture Documents
net/publication/327902760
Empirical modelling of hydraulic pumps and motors based upon the Latin
hypercube sampling method
CITATIONS READS
4 1,401
2 authors:
Some of the authors of this publication are also working on these related projects:
All content following this page was uploaded by Paul W. Michael on 10 November 2018.
1 Introduction
1.1 Background
ISO 4409 ‘hydraulic fluid power – positive displacement pumps, motors and integral
transmissions – methods of testing and presenting basic steady-state performance’
specifies circuit schematics, instrument accuracy ranges, test procedures, and reporting
requirements for characterising positive displacement machine performance (ISO 4409,
2007). It was promulgated by the International Organisation for Standardisation in 1986
to unify methods for testing hydraulic fluid power positive displacement pumps, motors,
and integral transmissions. The intent was to enable the performance of different
components to be compared. The document was revised in 2007 and 2017. The 2007
revision provided additional test method details. The 2017 revision removed the
calculation section and directed the user to reference ISO 4391 for engineering parameter
definitions (ISO 4391, 1986). It also provided additional details regarding hydraulic fluid
requirements. Sample size and test point selection are left to user discretion.
The procedures outlined in ISO 4409 are used for a variety of purposes including
production line quality inspections, commercial product comparison testing, design
validation, and performance mapping for simulation. Data collection, statistical
processing, and modelling strategies vary depending upon the analytical goal.
Well-defined sampling plans and control chart procedures have been developed by
quality scientists for use in production testing (Juran and Gryna, 1988). Confidence
interval charts have been used to compare the performance of fluids at specific speeds,
pressures, and temperatures (Michael et al., 2016). Complex fluid-structure models that
describe thermal conditions within tribological contacts have been created to optimise the
design of positive displacement machines (Murrenhoff et al., 2008; Schenk et al., 2013;
Dhar and Vacca, 2015). Simulations based on steady-state testing via ISO 4409 have
been used to optimise control systems and predict machine performance (Rahmfeld and
Skirde, 2010). Since performance data are needed to support quality control, component
design and machine simulations, it is a challenge for fluid power test facilities to keep up
with demand. In this investigation, we explore the use of Latin hypercube sampling
(LHS) to reduce the number of data points required to produce reliable empirical models
from ISO 4409 data.
1.2 Modelling
The performance of hydraulic pumps and motors is affected by friction, viscous drag,
fluid compressibility and pressure-driven flow losses. Analytical models have been
developed to characterise how these phenomena affect energy dissipation in positive
displacement machines (Wilson, 1950; Schlösser, 1961; Olsson, 1973; Manring, 2005a;
Jeong, 2007). Common physical terms for pump and motor models are listed in Table 1.
Some fluid-structure interactions within positive displacement machines do not readily
lend themselves to the analytical solution. Thermal effects, friction, pressure transients
and fluctuating gap heights create too much complexity. This complexity may be
overcome by using purely mathematical polynomial terms to approximate measured
curves (Ivantysynova, 2001). A variety of methods may be used to select terms and
determine coefficients for regression of ISO 4409 measurements. Ordinary least squares
regression is usually satisfactory though more sophisticated methods for selecting terms
Empirical modelling of hydraulic pumps and motors 3
may be used when higher accuracy models are desired (Tibshirani, 1996; Ulbrich, 2009).
In this investigation, empirical models based on physical terms were used to evaluate data
sets of varying size. The terms were selected, and coefficients were determined using
MATLAB code.
Table 1 Terms for empirical hydraulic pump and motor models
⎛ Δp ⎞⎛ ω ⎞⎛ Δp ⎞
Vi ⎜
2 πμ ⎟⎜ A + B ω ⎟⎜ p ⎟ Dorey (1988)
⎝ ⎠⎝ max ⎠ ⎝ case ⎠
( εVi )
2
Δp
Johnson (2018b)
μ
Pressure driven turbulent flow loss
2Δp
Vi 2/3 Schlösser (1961)
ρ
Δp
Manring (2005a)
μω
Δp
Jeong (2007)
ρ
ω3
ρ Jeong (2007)
Δp
p2 Manring (2016)
Compressibility loss
ΔpωViε Zarotti and Nevegna (1981)
( A − Bε ) Vi Δp Olsson (1973)
β
⎛ 1 + ε ⎞ ΔpωVi
⎜ Vr + β ⎟ β Dorey (1988)
⎝ ⎠
Δp
ω Jeong (2007)
β
4 P. Panwar and P. Michael
Table 1 Terms for empirical hydraulic pump and motor models (continued)
Compressibility loss
⎛ p2 ⎞
⎜1 − β ⎟Vi ωε Manring (2016)
⎝ ⎠
Viscous friction
µωVi Wilson (1950)
µω/Δp Manring (2005a)
µω Jeong (2007)
Turbulent friction
ρVi 5/3
ω2 Schlösser (1961)
4π
ρVi 5/3 2 3
ωε Thoma (1969)
4π
µω2 Ivantysyan and Ivantysynova (2003)
ρω2 Jeong (2007)
Coulomb or boundary friction
⎛ Δp ⎞
Vi ⎜ ⎟ Wilson (1950)
⎝ 2π ⎠
Vi p1 + ε′p2
Hibi and Ichikawa (1977)
1 + ( ω / ωo )
μω / Δp Manring (2005a)
2
Δp Jeong (2007)
( μω / Δp )
Michael et al. (2016)
A + ( μω / Δp )
2 Select one value from each stratum at random with respect to the probability density
within the interval.
3 Randomly pair the n values obtained for x1 with the n values of x2.
4 Randomly combine these n pairs with the n values of x3 to form n triplets.
5 Repeat this process until n k-tuplets are formed, thus creating an LHS set of n
samples with k variables from the multidimensional distribution.
Unlike in factorial testing, each sample in the resulting data set will be unique to its
strata. For example, if the range of rotational frequency is 1,000 to 2,200 rpm and the
number of data points n = 100, each strata will have a width of 12 rpm. Under the LHS
scheme, a single data point will be collected within the minimum frequency strata of
1,000 to 1,012 rpm.
Figure 1 Data points for orthogonal and Latin hypercube-based design of experiments (DOE)
(see online version for colours)
Empirical modelling of hydraulic pumps and motors 7
important to verify that the flow rate has stabilised because transient efficiencies greater
than 100% are possible when the pump is decelerating. It also should be noted that the
use of Vimax will yield at least one test point where ηV = 100% (Bramer et al., 2014).
The Toet method uses a two-step process to determine pump displacement (Toet,
1970). In the first step of the process, the ratio of pump outlet flow to rotational
frequency is plotted for each of several test pressures. In the second step, ∂Qout/∂N is
plotted as a function of pressure. The zero-intercept from a linear fit of the data equals the
derived displacement. Hence the derived displacement is 46.12 cc/revolution as shown in
Figure 3.
Figure 3 ∂Q/∂n is plotted versus differential pressure to determine derived displacement (Vi) via
the Toet method (see online version for colours)
multicollinearity results in models that over-fit data. Models that over-fit data can
generate high residuals when applied to new or independent data sets (Ulbrich, 2009). To
avoid collinear terms, the MATLAB code that was used to select model terms and
coefficients eliminated models that yielded high AIC and VIF values.
p ( εVi )
2
ωp
Q = β 0 + β1ωεVi + β 2 p2 + β3 + β4 (2)
K μ
The best five-term model based on R2 and S is shown in equation (3). Equation (3) is
similar to equation (2) except the displacement term (β1) includes the effect of fluid
compressibility. The mean value of (p/K) is 0.0123 so the effect of fluid compressibility
on (β1) is relatively small (1 – 0.0123 = 0.9877). Equation (3) also includes a viscous
flow loss term (β5). As can be seen in Table 2, the addition of this term reduced the
standard error of the model by 0.005 litres per minute. The effect of pressure-driven flow
losses on model fidelity was relatively minor because the leakage flow rate varied little as
the ratio of pressure to dynamic viscosity increased.
p ( εVi )
2
⎛ p ⎞ ωp p
Q = β 0 + β1 ⎜ 1 − 2 ⎟ ωεVi + β 2 p2 + β 3 + β4 + β5 (3)
⎝ K ⎠ K μ ωμ
Empirical modelling of hydraulic pumps and motors 11
As shown in Figure 5, the four-term and five-term models yielded a linear agreement
with measured pump flow rates. The maximum flow rate was approximately 100 lpm.
The mean standard error (S) for both flow models was approximately 0.43 lpm as shown
in Table 2. Hence, the mean standard error was 0.43% of the maximum flow rate.
Caution must be used when comparing the percent mean standard error (0.43%) and the
combination of error values (0.59%) because S is a global average while ∂ηV varies
proportionally with the measured flow rate. Nonetheless, it can be concluded that both
models describe the pump flow behaviour at a level of precision commensurate with the
measurement accuracy.
Table 2 Statistical measures of fidelity for four-term and five-term pump models
Figure 5 A comparison of the four-term and five-term models shows that the additional
parameter (5th term) had a minimal impact on the correlation with measured values
(see online version for colours)
the axial piston pump using the mathematical expressions in Table 1. The best four-term
model based on R2 and S is shown in equation (4). Equation (4) includes a constant (β0)
and coefficients for theoretical displacement (β1), pressure-driven flow losses (β2),
compressibility (β3), and displacement angle squared (β4).
p ( εVi )
2
⎛ p ⎞ p ωp
Q = β 0 + β1 ⎜1 − 2 ⎟ ωεVi + β 2 + β3 + β4 (4)
⎝ K ⎠ ωμ K μ
The best five-term model based on R2 and S is shown in equation (5). Equation (5)
incorporates the same terms as equation (4) plus a term for pressure-driven turbulent flow
losses (β5) that includes density (ρ) and rotational frequency cubed (ω3). According to
Jeong (2007), this expression describes piston port leakage flow. As shown in Table 3,
the addition of a fifth term had a negligible impact on the model error.
p ( εVi )
2
⎛ p ⎞ p ωp ⎛ ρ⎞
Q = β 0 + β1 ⎜ 1 − 2 ⎟ ωεVi + β 2 + β3 + β4 + β 5 ⎜ ⎟ ω3 (5)
⎝ K ⎠ ωμ K μ ⎝ p⎠
Table 3 The flow model error (S) was higher for the orthogonal dataset
Figure 6 Standard error for the pump flow model based upon random subsets of four replicate
100-point LHS and orthogonal runs (see online version for colours)
Four replicate tests were conducted using the orthogonal and LHS methods to yield a
total of 400 data points per sample plan. The sequence of the test results was re-ordered
by assigning a random number between 1 and 400 to each row. This process was repeated
Empirical modelling of hydraulic pumps and motors 13
until 1,000 randomly sequenced sets of test data were created. Twenty equally spaced
subsets were selected from each of the 1,000 randomised sets. For example, one subset
included all of the rows of data with a random number less than or equal to 30; the next
subset included all the rows of data with a random number less than or equal to 50, etc.
up to 400. Each of the resulting 20,000 subsets were regressed using equation (4). The
model coefficients were determined for each subset or iteration, as well as the standard
error. The standard error for both sample schemes was plotted versus the degrees of
freedom (DoF) as shown in Figure 6. Both sample strategies yielded normally distributed
errors. The standard error for the orthogonal data set was more than 30% higher. It is
interesting to note that the error distributions appear to evolve at the same rate, even
though the model accuracy is significantly different.
In order to investigate the effects of test replication, the procedure above was repeated
using two replicate tests rather than four. As shown in Figure 7, the mean standard error
and shape of the error distribution were relatively unchanged. It appears that the
distribution is broader; however, a comparison of the LHS error distribution in
Figures 6 and 7 at 100 degrees of freedom (DoF) reveals that the smaller data set has a
similar error distribution. Hence, a reduction the number of replicate tests did not
increase the uncertainty of the model for orthogonal or LHS datasets that incorporate
100 data points. It is noteworthy that for sample sizes greater than 120, LHS always
generated flow models of higher fidelity.
Figure 7 Standard error for the pump flow model based upon random subsets of two replicate
100-point LHS and orthogonal runs (see online version for colours)
In LHS, data points are randomly selected within each stratum and therefore the
experimental set points vary from one experimental design to another. Experiments were
conducted to determine how randomisation within the LHS strata affected model
accuracy. Two new LHS sample plans were generated using MATLAB code and pump
performance data were collected using the new test points. Random subsets were
generated as previously described and the resulting 20,000 data sets were fitted to
equation (4). The coefficients from the previous LHS data set were used in order to assess
14 P. Panwar and P. Michael
model reliability. As shown in Figure 8 and Figure 9, the mean standard error was
essentially unchanged. Note however that the error distributions were skewed to the right,
and fewer instances of extremely low standard errors occurred at low DoF. This is to be
expected when an independent data set is fitted to a model without adjusting coefficients.
Figure 8 Standard error for the original four-term pump flow model from two replicates of an
independent 100-point LHS data set (DOE #2) (see online version for colours)
Figure 9 Standard error for the original four-term pump flow model from two replicates of an
independent 100-point LHS data set (DOE #3) (see online version for colours)
The standard error in Figures 8 and 9 appears to be normally distributed for DoF greater
than 100. In order to determine if fewer samples could yield models of similar fidelity, a
LHS test plan was created using 50 sample points. Performance data were collected for
two replicates of 50 data points using the new DOE. Random subsets were generated as
Empirical modelling of hydraulic pumps and motors 15
previously described and the resulting subsets were regressed using equation (4). The
coefficients from the previous data set were used in order to assess model reliability. As
shown in Figure 10, the mean standard error was essentially unchanged and once again
fewer instances of extremely low standard errors occurred at low DoF. Thus it is
hypothesised that less than 50 data points are required to create an accurate model.
Figure 10 Standard error for the original four-term pump flow model from two replicates of
50-point LHS data set (see online version for colours)
As described in Section 1.2, the frictional losses in a hydraulic motor are affected by fluid
properties and operational parameters such as the speed ω, differential pressure Δp,
displacement Vi, and dynamic viscosity μ. A steady-state nonlinear regression torque-loss
TL model for a radial piston motor was developed from the model terms show in Table 1.
Four criteria were used for assessing the model quality: the mean standard error, the
correlation coefficient, AIC, and the VIF. The analytical goal was to produce a model
16 P. Panwar and P. Michael
that had a standard error less than 10 Nm and a low VIF. As shown in Table 4, a
five-term LHS model fulfilled this requirement.
The best five-term model based on R2 and S is shown in equation (6). Equation (6)
includes terms for coulomb friction (β1), turbulent drag (β2), pumping losses (β3), viscous
drag (β4) and boundary or mixed-film friction (β5).
Vi p1 + ε ' p2 μω μω
TL = β1 α
+ β 2 μω2 + β3Vi μω + β 4 + β5 + β0 (6)
1 + ( ω ωo ) p p
As with the pump data, the sequence of the test results was re-ordered by assigning a
random number to each row. This process was repeated until 1,000 randomly sequenced
sets of test data were created. Twenty equally spaced subsets were selected from each of
the 1,000 randomised sets. Each of the resulting 20,000 subsets were regressed using
equation (6). The model coefficients were determined for each subset or iteration, as well
as the standard error. The standard error for both sample schemes was plotted versus the
DoF as shown in Figure 11. Both sampling strategies yielded normally distributed errors
that appear to evolve at a similar rate. As with the pump model, the experimental design
based upon orthogonal sample points produced models with a standard errors that were
more than 30% higher.
Figure 11 Standard error for the torque-loss model [equation (6)] based upon random subsets of
two replicate lhs and orthogonal data sets (see online version for colours)
limit that is concerned with costs and time needed to conduct testing. If the theory of
limited sample size is true, then it follows that there must come the point where the
number of samples causes the coefficients and figures of merit to reach terminal values.
That is, both the set of regression coefficients and the chosen figures of merit do not
undergo significant changes with further samples. Beyond this point, the addition of new
samples does not contribute to the regression process and data saturation may be said to
occur.
PSR was used to track the evolution of model coefficients as the number observations
increased to identify the onset of data saturation (Johnson, 2018c). In PSR the source data
array (laboratory test data), is progressively increased, step-by-step, one observation
(sample) at a time. At each step in the progression, a complete regression model is
created, and the coefficients are recorded. Data saturation occurs when the model
coefficients and figures of merit stabilise, exhibiting random variation within a narrow
range of values (similar to a hydraulic control system).
Seven independent LHS data sets consisting of 10, 15, 20, 25, 30, 35, and 40 points
were collected on the pump. Two replicate runs were acquired for each data set. The data
were randomised and ten iterations of PSR were performed for each set. Figure 12 shows
a plot of the standard error versus DoF for 70 iterations. The results for the 2 × 15 and
2 × 25 data sets skew high, similar to Figure 9 and Figure 10. The evolutionary trajectory
of the standard error suggest that data saturation occurs after two replicate runs of 30 data
points. The standard error for LHS data sets that include 30 or more points is comparable
to the results shown in Figure 6 and Figure 7, where multiple runs of 100 data point were
collected. Hence, in this particular instance, additional testing beyond two replicates of
30 LHS test points added nothing to the understanding of the data population.
Figure 12 PSR plots for Latin hypercube experimental designs ranging from 10 to 40 set points
per test (see online version for colours)
Likewise, PSR was used to compare motor torque-loss model evolution. A five-term and
six-term model was compared for five replicates of 50 LHS test results. The entire
18 P. Panwar and P. Michael
sample set was randomised, test points were selected, model coefficients were regressed,
and the standard error was plotted for each degree of freedom. The entire data set was
randomised between each incremental increase of the sample set to ensure thorough
randomisation. As shown in Figure 13, the standard error fluctuated significantly over the
first 80 to 100 data points. Beyond 100 data points, PSR revealed that the standard error
randomly fluctuated within a range of ± 10% of the terminal value for (S). This pattern is
characteristic of data saturation. It is noteworthy that both models reached data saturation
at virtually the same time even though they have different standard errors. The practical
implication is that the method used to sample the population has a larger impact than the
quality of the model.
Figure 13 PSR plots for five- and six-term torque-loss models (see online version for colours)
The effectiveness of orthogonal and Latin hypercube experimental design strategies was
compared in axial piston pump and radial piston motor testing. Pump flow and motor
torque-loss models were developed via regression of the experimental results. The
effectiveness of the sampling plans was evaluated by comparing the standard errors of the
models. The standard error for models based on orthogonal sampling was > 30% higher
than models based on LHS designs. Higher fidelity models were produced by the LHS
method because it provided a greater uniformity of sampling distribution throughout the
performance range of the pump and motor.
PSR was used to identify the range in which data saturation occurs. This was
accomplished by plotting evolution of model errors as the number of experimental
observations was incrementally increased. Data saturation was observed in the pump test
upon completion of two replicate 30-point sets. Two replicate 50-point data sets were
required for data saturation in the motor test. Beyond the point of data saturation, model
terms exhibited a small amount of random variation and additional testing added nothing
Empirical modelling of hydraulic pumps and motors 19
to the understanding of the data population. The threshold of data saturation was found to
be independent of the sampling procedure and model accuracy. These results demonstrate
the merit of combining LHS with PSR to improve model fidelity and determine
sample-size requirements in hydraulic pump and motor testing.
Acknowledgements
This investigation was initiated by a request from ISO Technical Committee 131,
Subcommittee 8 (Fluid power systems, Product testing). The committee sought an
objective means of determining the minimum sample size requirements for ISO 4409
testing. We gratefully acknowledge Jack Johnson’s leadership of this effort. His insights
into empirical modelling and progressively sequenced regression served as the foundation
for our research strategy. We also thank Professor Ron Jorgensen (MSOE) and John
Montague for helpful discussions of design of experiments, sample size analysis, and
methods for selecting regression models.
References
Bavendiek, R. (1987) Verlustkennwertbestimmungam Beispiel von hydrostatischen Maschinen
in-Schrägachsenbauweise, Doctoral thesis, VDI-Fortschrittsberichte, Reihe 7 Nr. 122, VDI
Verlag.
Bramer, J., Puzzuoli, A., Michael, P. and Wanke, T. (2014) ‘Hydraulic Fluid efficiency effects in
external gear pumps’, Proceedings of the 53rd National Conference on Fluid Power –
National Fluid Power Association, Paper NCFP I14-13.3, Las Vegas, NV
Davey, K. (2008) ‘Latin hypercube sampling and pattern search in magnetic field optimization
problems’, IEEE Transactions on Magnetics, Vol. 44, No. 6, pp.974–977.
Dhar, S. and Vacca, A. (2015) ‘A novel FSI-thermal coupled TEHD model and experimental
validation through indirect film thickness measurements for the lubricating interface in
external gear machines’, Tribology International, Part A, Vol. 82, pp.162–175.
Dorey, R. (1988) ‘Modelling of losses in pumps and motors’, 1st Bath International Fluid Power
Workshop, University of Bath, UK.
Hibi, A. and Ichikawa, T. (1977) ‘Mathematical model of the torque characteristics for hydraulic
motors’, Bulletin of JSME, Vol. 20, No. 143, pp.616–621.
International Organization for Standardization ‘ISO 4392-1 (2002) (E) Hydraulic Fluid Power –
Determination of Characteristics of Motors – Part 1: At Constant Low Speed and Constant
Pressure, Geneva.
International Organization for Standardization, ‘ISO 4391 (1986) Hydraulic Fluid Power – Positive
Displacement Pumps, Motors and Integral Transmissions – Parameter Definitions and Letter
Symbols, Geneva.
International Organization for Standardization, ‘ISO 4409 (2007) (E) Hydraulic Fluid Power –
Positive Displacement Pumps, Motors and Integral Transmissions – Methods of Testing and
presenting Basic Steady State Performance, Geneva.
International Organization for Standardization, ‘ISO 8426 (2008) Hydraulic Fluid Power – Positive
Displacement Pumps and Motors – Method for Determining Derived Capacity, Geneva.
Ivantysyn, J. and Ivantysynova, M. (2003) Hydrostatic Pumps and Motors: Principles, Design,
Performance, Modelling, Analysis, Control and Testing, Tech Books International, New
Delhi, India.
20 P. Panwar and P. Michael
Wyss, G.D. and Jorgensen, K.H. (1998) A User’s Guide to LHS: Sandia’s Latin Hypercube
Sampling Software, Sandia National Labs. United States, DOI: 10.2172/573301.
Zarotti, G. and Nevegna, N. (1981) ‘Pump efficiencies approximation and modelling’, 6th Int.
Fluid Power Symposium, Cambridge, UK.
Nomenclature
Symbol Description Units
K Bulk modulus [GPa–1]
n Data points in a subset [-]
N Data points in the population [-]
µ Dynamic viscosity [Pas]
ρ Fluid density [kg/m3]
p1 Inlet pressure [Pa]
v Kinematic viscosity [kg]
S Mean standard error [L/m or Nm]
TL Torque loss [Nm]
p2 Outlet pressure [Pa]
p or Δp p2–p1 pump or p1–p2 motor [Pa]
Ω Rotational frequency [rad/s]
ε Swash plate position [%Vi]
Vi Theoretical displacement [m3/rad]
Qi Theoretical volumetric flow rate [m3/s]
Q Volumetric flow rate [m3/s]