IWE

Lin-KK Tool
26.11.2014

Operating Rules and Examples

Dipl.-Ing. Michael Schönleber

B. Sc. Ravindra Goyal
Prof. Dr.-Ing. Ellen Ivers-Tiffée

Institut für Werkstoffe der Elektrotechnik (IWE)

KIT – Universität des Landes Baden-Württemberg und

nationales Forschungszentrum in der Helmholtz-Gemeinschaft
0
Content

1. Introduction ....................................................................................................................... 2
2. Operating Rules ................................................................................................................ 4
2.1 Fitting Mode .............................................................................................................. 4
2.2 Model Order............................................................................................................... 5
2.3 Capacity ..................................................................................................................... 5
2.4 Inductivity .................................................................................................................. 5
3. Interpreting Test Results ................................................................................................... 6
4. Examples ........................................................................................................................... 7
4.1 Example 1: Valid Li-Ion Battery Spectrum (low noise situation) ............................. 7
4.2 Example 2: Valid Li-Ion Battery Spectrum (noisy situation) .................................... 8
4.3 Example 3: Valid Li-Ion Battery Spectrum (outlier situation) .................................. 9
4.4 Example 4: Invalid Li-Ion Battery Spectrum (time-variance) ................................. 10
4.5 Example 5: Valid SOFC Spectrum .......................................................................... 11
4.6 Example 6: Invalid SOFC Spectrum (time-variance) .............................................. 12
5. Literature ......................................................................................................................... 13
6. How To Cite.................................................................................................................... 13
7. Disclaimer ....................................................................................................................... 13

1
1. Introduction
The validity of all analysis of a measured impedance spectrum requires the spectrum to be a
representation of a linear, time-invariant and causal system. If one of these requirements is
violated, no meaningful information can be extracted from the measurement.

In order to check the fulfillment of these requirements, two basic principles can be
distinguished: The first is to verify them experimentally, for example by checking for
harmonics in the system response or repeating the impedance measurement using different
excitation signals. Also, specially designed broadband excitation signals can be used, as
outlined in [1] and [2]. However, most of these approaches require insight into excitation and
response signals during an impedance measurement, information which is not provided by the
majority of the commonly used Frequency Response Analyzers (FRA). Other than that, the
second principle is based on analyzing only the obtained impedance spectrum utilizing the
Kramers-Kronig (KK) relations. The latter relate the real part of the impedance spectrum of a
linear, time-invariant and causal system to its imaginary part and vice versa. The relevant
equations

2   Z Im   

Z Re     d  (1.1)
 0  2   2
2   Z Re   

Z Im     2 d  (1.2)
 0    2

were, in a slightly different form, first introduced by Kramers [3] and Kronig [4]. The
derivation, however, was physically motivated [5]. A brief derivation from a system-theory
point of view can be found for example in [6].

By applying Equations (1.1) or (1.2) to either the real part or imaginary part of a measured
spectrum, the remaining part can be computed. By comparing the latter to the according
measured part, the spectrum’s accordance to the KK relations can be judged. However,
directly evaluating Equations (1.1) and (1.2) for measured data turned out to be prone to
errors due to the required integration range from zero to infinity, which cannot be covered in
real measurements. Approaches were made to solve the missing-part-problem by proposing
extrapolation strategies [7] or even more sophisticated methods [8], but in the end the general
problem of missing information remained.

An approach that overcame this was first introduced in [9-11]: Instead of directly evaluating
Equation (1.1) and (1.2), the accordance of an impedance spectrum with the KK relations is
judged by its reproducibility by an appropriate KK compliant equivalent circuit model
(ECM). A very general model, namely a series connection of RC-elements, is chosen, where
the resistors as well as the time-constants of the ECM are fitted to a measured impedance
spectrum. The benefit of this simultaneous fit is the possibility of revealing the stochastic
error structure of the measurement, thereby assessing a frequency dependent confidence
interval for the predicted impedance. The latter allows to define a quantitative criterion of
which parts of the spectrum are to be disregarded and which parts are valid.
However, the main drawback of this method, as named in [12], is the nonlinear nature of the
according fit problem. Due to fitting the time-constants of the ECM (i) good initial values
have to be chosen and (ii) several solutions related to local minima of the according cost
2
function to be minimized exist. Hence, while this approach is indispensable for a deep-level
analysis of a measured impedance spectrum, if one just wants to quickly check for validity of
one or several impedance spectra, the effort is quite high.
In [12] a modified version of this test was proposed, where the problems connected with the
nonlinearity of the fit were solved by only fitting the ohmic resistors and pre-setting the time-
constants. Thus, the fit problem became linear.

The latter concept (referenced to in the following as the Lin-KK test) has been extensively
used by the authors and experienced users can obtain valid test results quickly and easily. In
contrast, inexperienced operators or operators not familiar with the field of model fitting,
however often achieve ambiguous or even wrong test results. The reason for this is the user’s
responsibility to manually pre-set the number M of RC-elements to be fitted, thereby easily
creating situations of under-fitting or over-fitting. A solution for this problem has been
proposed in [6], where a strategy is given on how to automatically choose the number of RC-
elements to be fitted for any given impedance spectrum. Thus, valid test results are obtained
irrespective of the operator’s expertise.

The Lin-KK tool at hand is based on the Lin-KK test proposed in [12] and comprises the
automatic strategy for finding an appropriate model order as proposed in [6].
Thus, it represents the state of the art of quick impedance testing.

3
2. Operating Rules
2.1 Fitting Mode
The cost function J to be minimized can either be chosen to penalize errors between the real
parts of fitted and measured impedance, the corresponding imaginary parts or both. For fitting
the real part (real-fit), the according cost function is given by

2
N  Z    Zˆ Re i  
J    Re i  (2.1)

i 1  Z i  

with N the number of frequency points of the impedance spectrum to be tested. For fitting
both, real and imaginary parts (complex-fit), it is

2 2
N  Z    Zˆ Re i    Z Im i   Zˆ Im i  
J    Re i    (2.2)

i 1  Z  i    Z i  

If a model to be fitted does not include a capacitor, one might also consider fitting the ECM to
the imaginary part of the measured impedance spectrum (imag-fit). Especially for spectra
where the low frequent imaginary part shows no asymptotic behavior this was stated to be
beneficial [13]. However, as soon as a capacitor has to be added to the ECM – as it is the case
for Lithium-Ion batteries, for example – this method needs to be discarded in the proposed
test-framework. There, simultaneous fitting of the imaginary parts of RC-elements and a
capacitor yields an assignment of information to different parts of the ECM which is prone to
ambiguities: Parts of the low frequent imaginary part, which should be represented by RC-
elements, might be assigned to the capacitor. As the impedance of the capacitor has no
contribution to the real part, the misassigned portion of the imaginary part will cause an error
within the real part and hence will falsely indicate a valid spectrum as invalid.

Recommendation:

Choosing complex-fit yields most robust results and should therefore be used by default. Only
to increase the sensitivity of the test by distributing errors not into real and imaginary part of
the residual, but shifting all errors into the imaginary part, real-fit might be used. Imaginary-
fit is the least robust mode and is therefore not recommended.

4
2.2 Model Order
As mentioned in the introduction, a crucial shortcoming of the Lin-KK test as proposed in [12]
is the necessity to set the parameter M, the number of RC-elements in the ECM used to fit the
impedance spectrum. If this chosen number is too small, under-fitting will be the
consequence. This means, that the order of the ECM is too small to fit the measured
impedance spectrum, even though the latter might in fact be fully compliant with the KK
relations. For example, a spectrum consisting of three perfect RC circuits cannot be fitted with
M = 2, if the time-constants are distributed to some extent. Thus large residuals are detected,
thereby falsely indicating an invalid impedance spectrum. On the other hand, if the selected M
is too large, not only the useful information contained in the spectrum is fitted, but the
measurement noise as well. A computed imaginary part from an over-fitted real part therefore
contains, imprecisely speaking, KK transformed measurement noise. Therefore, a comparison
with the measured imaginary part of the spectrum will certainly fail, no matter if the spectrum
is valid or not. Hence, the according residuals will indicate invalidity, even though the tested
spectrum is fully valid.
The RC-auto strategy implemented in the Lin-KK tool overcomes that problem and guarantees
valid test results by automatically selecting an appropriate number of RC elements. The
scientific background of RC-auto is described in [6].

Recommendation:

The model order should be chosen automatically by the Lin-KK tool by selecting RC-auto.
Only if RC-auto fails, indicated by sinusoidal like residuals, RC-manual should be used. In
this case carefully increase the number of RC-elements manually and check if the sinusoidal
oscillations disappear. However, please do so only in complex-fit mode as this mode is most
robust against over-fitting.

2.3 Capacity
When testing an impedance spectrum of a capacitive system, a serial capacity has to be added
to the ECM to be fitted. This can be done by selecting Add Capacity (on). Not adding a
capacity might lead to residuals indicating invalidity even though the spectrum is valid.

2.4 Inductivity
In order to account for unavoidable inductive effects of the measurement setup, a serial
inductivity is automatically added to the ECM to be fitted.

5
3. Interpreting Test Results
The test result is given as residuals, which represent the relative deviation between the
measured figures and an according Kramers-Kronig-ideal fitted spectrum over frequency.
They are defined as

Z Re    Zˆ Re  
 Re   
Z  
(3.1)
Z Im    Zˆ Im  
 Im   
Z  

For interpreting these residuals, the following rules apply:

- The smaller the residuals, the better the spectrum’s quality. Non-biased residuals of
smaller than about 0.5% over the whole frequency range indicate a valid spectrum of
good quality.
- Noise-like residuals indicate that the spectrum contains measurement noise (which is,
to some extent, normal).
- Biased residuals, especially in the low frequency range, indicate time-variance. The
according spectrum is likely to be invalid.
- When obtaining sinusoidal like residuals, most likely under-fitting occurs due to a
failure of RC-auto. In this case carefully increase the number of RC-elements
manually and check if the sinusoidal oscillations disappear. However, please do so
only in “complex-fit” mode as this mode is most robust against over-fitting.

Warning:

Violations of the linearity constraint during the measurement of an impedance spectrum (for
example when selecting too large an excitation signal) are only barely detected. If in doubt,
check for harmonics in the response signal of your impedance spectroscopy measurement.
Alternatively, if you have an only slowly time-varying system, repeat your measurements with
decreasing excitation strength and check for changes in your results.

6
4. Examples
Typical results are shown for measured impedance spectra. All spectra were evaluated with
the Lin-KK tool using complex-fit and RC-auto mode. The according data was also provided
with the Lin-KK tool, so feel free to play around.

4.1 Example 1: Valid Li-Ion Battery Spectrum (low noise situation)

The residuals are small (<0.5%) and show no significant bias. Hence, the spectrum contains
low noise and describes a time-invariant system.

7
4.2 Example 2: Valid Li-Ion Battery Spectrum (noisy situation)

The residuals are greater than in Example 1, but still do not show significant bias. Hence, the
spectrum contains noise but still describes a time-invariant system.

8
4.3 Example 3: Valid Li-Ion Battery Spectrum (outlier situation)

The measurement outlier at 100 Hz (quite common for frequencies which are multiplies of the
power supply frequency) has increased the residuals in the high frequency area over a wide
range. The reason is that the model fit tries to cover the outlier, at the cost of corrupting the
frequency range around the outlier. Selecting and deleting the outlier using the relevant
function within the Lin-KK tool will yield the true residuals of the spectrum.

9
4.4 Example 4: Invalid Li-Ion Battery Spectrum (time-variance)

The residuals are strongly biased towards low frequencies. This indicates time-variance and
therefore an invalid spectrum. In fact, the ambient temperature was purposely altered during
measurement. Since measuring low frequencies takes longer than measuring high frequencies,
low frequencies typically contain a higher degree of time variance, exhibiting large, biased
residuals.

10
4.5 Example 5: Valid SOFC Spectrum

The residuals of this spectrum (from a symmetric Anode-Anode Solid Oxide Fuel Cell) are
small (<0.5%) and show no significant bias. Therefore, the spectrum contains low noise and
describes a time-invariant system.

11
4.6 Example 6: Invalid SOFC Spectrum (time-variance)

The imaginary part residuals are weakly biased towards low frequencies. This indicates time-
variance and therefore an invalid spectrum. In fact, the spectrum was measured during the
activation phase of a symmetric Anode-Anode Solid Oxide Fuel Cell, where the Anode
properties changed during spectrum measurement.

12
5. Literature
[1] E. v. Gheem, R. Pintelon, J. Vereecken, J. Schoukens, A. Hubin, P. Verboven,
O. Blajiev, Electrochim. Acta, 49 (2004) 4753-4762.
[2] E. v. Gheem, R. Pintelon, A. Hubin, J. Schoukens, P. Verboven, O. Blajiev,
J. Vereecken, Electrochim. Acta, 51 (2006) 1443-1452.
[3] H.A. Kramers, Transactions of Volta Centenary Congress, Vol. 2 (1927) 545.
[4] R. de L. Kronig, Journal of the Optical Society of America. 12, Vol. 6 (1926) 547-556.
[5] C. Bohren, Eur. J. Phys. 31 (2010) 573–577.
[6] M. Schönleber, D. Klotz and E. Ivers-Tiffée, A Method for Improving the Robustness
of linear Kramers-Kronig Validity Tests, Electrochimica Acta 131, pp. 20-27 (2014),
10.1016/j.electacta.2014.01.034.
[7] M. Urquidi-Macdonald, S. Real, D. Macdonald, Electrochim. Acta, Vol. 35,
10 (1990) 1559–1566.
[8] J. Esteban, M. E. Orazem, J. Electrochem. Soc., Vol. 138, 1 (1991) 67-76.
[9] P. Agarwal, M. E. Orazem, L. H. Garcia-Rubio, J. Electrochem. Soc.,
Vol. 139 (1992) 1917.
[10] P. Agarwal, O. D. Crisalle, M. E. Orazem, L. H. Garcia-Rubio, J. Electrochem. Soc.,
Vol. 142 (1995) 4149.
[11] P. Agarwal, M. E. Orazem, L. H. Garcia-Rubio, J. Electrochem. Soc.,
Vol. 142 (1995) 4159.
[12] B. A. Boukamp, J. Electrochem. Soc., 142 (1995) 1885.
[13] P. Agarwal, M. E. Orazem, L. H. Garcia-Rubio, J. Electrochem. Soc.,
Vol. 142 (1995) 4159.
[14] http://www.iwe.kit.edu/Lin-KK.php

6. How To Cite
When using results obtained with the Lin-KK tool: Please reference [6] and [12] for the
scientific background of the tool and [14] for its availability. Please do not distribute the
Lin-KK tool. Instead, please refer to our website [14].

7. Disclaimer
The authors do not assume any legal liability or responsibility for the accuracy, completeness
or usefulness of the Lin-KK tool, nor for any claims evolving from the use of this program.
Thank you!

13