2018 IEEE Transportation Electrification Conference and Expo, Asia-Pacific (ITEC Asia-Pacific)

Estimating Lithium-ion Battery State of Charge and

Health with Ultrasonic Guided Waves Using an
Efficient Matching Pursuit Technique
Purim Ladplia†, Chen Liua, Fotis Kopsaftopoulosb, and Fu-Kuo Changa
a
Department of Aeronautics and Astronautics, Stanford University, Stanford, CA, United States
b
Department of Mechanical, Aerospace and Nuclear Engineering, Rensselaer Polytechnic Institute, Troy, NY, United States
E-mail: †[email protected]

Abstract— This paper presents a novel guided-wave-based on-board BMSs due to the equipment’s complexity and size
framework for estimating state of charge (SoC) and state of [2].
health (SoH) of lithium-ion (Li-ion) batteries. The guided waves Recently, acousto-ultrasonic wave propagation has started
are propagated and sensed using low-profile, surface-mounted to gain interest in the research community potentially as an
piezoelectric transducers on off-the-shelf Li-ion pouch cells.
auxiliary, or even as an alternate, on-board method to probe a
Special emphasis is given to the development of an efficient
feature extraction strategy based on the Matching Pursuit (MP) battery’s mechanical behavior and ultimately improve the
technique. The proposed method decomposes complex estimation of SoC and SoH [3-7]. Most of earlier research
waveforms into constituent ‘atoms’, allowing the time-frequency efforts focus on applying through-thickness compression
information of the signals to be mined. The descriptive waves by using laboratory ultrasonic probes or large
parameters of the decomposed atoms can be extracted to show piezoelectric transducers [3-5, 7]. The energy of the waves is
strong and meaningful correlations with SoC and SoH. concentrated to allow them to transmit across interfaces
Statistical prediction models for estimating SoC/SoH are built between electrode layers and propagate in the direction
using the atom parameters as predictive features in conjunction perpendicular to the laminate, enabling the localized region
with the traditional voltage data. It is shown that the atom
immediately underneath the probe to be examined. In contrast,
parameters can augment the voltage data and statistically
improve the prediction of SoC by two-fold and, in particular, of we recently proposed a guided-wave-based technique wherein
SoH by twenty-fold. The results signify the importance of the stress waves take advantage of the geometric boundaries
encapsulating proper signal processing and feature extraction of the batteries to ‘guide’ wave propagation [6]. Using low-
into future guided-wave-based SoC/SoH prediction frameworks. profile, surface-mounted piezoelectric transducers, this
technique is not only more robust to variations in boundary
Keywords— state of charge, state of health, ultrasonic guided conditions and operator errors, but also allows the waves to
waves, matching pursuit, feature extraction cover larger propagation distances and areas with minimal
I. INTRODUCTION loss in energy.
Although promising results are demonstrated from both
Electric vehicle (EV) energy storage, particularly lithium- research thrusts, a handful of work has been done on
ion (Li-ion) batteries, are extremely complex systems with a expanding the concept towards practical, field-deployable
very narrow operating range and are prone to premature, SoC/SoH prediction. Davies et al. demonstrate that a library
unexpected failure. Optimizing battery performance, lifespan, of time-domain parameters of the compression waves
and most importantly, safety requires continuous, accurate (namely, time of flight and signal amplitude) may be created
monitoring of the battery state of charge (SoC) and state of from similar batteries, and then a regression model can be
health (SoH) [1]. However, relying only on the terminal built to predict SoC of an independent cell [3]. However, to
voltage measurement, on-board battery management systems obtain a satisfactory accuracy level with SoH, they show that
(BMSs) still critically lack a breakthrough technique for the complete waveform (data at all sampled times during a
directly probing the physical properties of the batteries to help ~10ȝs window, or as many as ~500 predictors) is needed. The
accurately estimate SoC and SoH [1]. BMS research seems to large dimensionality of predictors may require prediction
forgo the fact that as Li-ion batteries cycle and age, their models with model structures that are too computationally
electrochemical processes are closely intertwined with expensive for on-board BMSs. Our previous companion work
mechanical evolution of the electrodes – particularly the similarly shows that time of flight and signal amplitude of
distribution and redistribution of their mass densities and guided waves may serve as accurate predictors for SoC, as
moduli [2]. Even though characterization tools for measuring well as SoH, owing to guided waves’ good signal integrity
these mechanical quantities are widely utilized in a laboratory against noise and boundary condition variations [6].
setting, most of such techniques cannot be performed on Furthermore, we also use multiple transducers in a network
vehicle-scale batteries and cannot be readily implemented in configuration and pool data from multiple propagation paths
978-1-5386-5782-9/18/$31.00 ©2018 IEEE to greatly reduce the SoC/SoH prediction error.

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
 Nevertheless, neither effort has yet taken advantage of the (AA Portable Power Corp.) (Fig. 1). Guided wave signals
rich information contained in the dynamic, non-stationary were gathered at various battery SoC and SoH from surface-
nature of pulse ultrasonic waveforms [3, 6]. Therefore, mounted, small-footprint piezoelectric disc transducers
building upon our previous work, this study aims to develop a (6.35mm-diameter PZT-5A in the SMART Layer format;
signal processing and feature extraction framework to exploit Acellent Technologies, Inc.) at the locations shown in Fig. 1.
the feature-rich nature of guided waves in the context of One of the piezoelectric discs can be chosen as an actuator to
practical SoC/SoH prediction. Signal processing and feature generate acousto-ultrasonic guided waves. The other
extraction of guided wave signals have long been a topic of piezoelectric disc then serves as a receiver to record the
research in the non-destructive testing (NDT) and structural transmitted guided wave signals. The so-called “pitch-catch”
health monitoring (SHM) community [8]. For NDT/SHM, experiments used five-peak Gaussian-windowed tone bursts
time-frequency representation (TFR) techniques are often with center frequencies between 100 to 200 kHz. The
used to project non-stationary guided wave signals onto a piezoelectric transducers were actuated and sensed using
time-frequency plane, from which individual, localized wave anultrasonic data acquisition system (ScanGenie II; Acellent
components may be extracted to be used as predictors for Technologies, Inc.). Ultrasonic measurements were taken
structural damage. Herein, with reference to SoC/SoH every 1 minute during electrical cycling.
prediction, we develop the proposed framework based on a The ultrasonic data acquisition was synchronized with a
class of TFR, the Matching Pursuit (MP) decomposition, battery analyzer (BST8-3; MTI Corporation), which
which allows complex guided wave signals to be efficiently performed battery cycling and aging. The cells were
decomposed into a combination of simpler, localized atoms electrically discharged at an elevated temperature of 45°C
[9]. We then show via statistical analysis that the coefficients with a constant current rate of 3,000 mA from 4.2V (100%
describing these atoms are strongly correlated with battery SoC) to 3.0V (0% SoC). A total of 200 discharge cycles were
cycling and aging behavior and can be used as effective performed. The remaining capacity for each cycle, which is
predictors for SoC and SoH. The obtained results from this our definition of SoH, was calculated by evaluating the cycle
study form the basis for systematically and practically discharge capacity normalized with respect to the value from
applying the acousto-ultrasonic technique towards the the first cycle. The cycle-to-cycle terminal voltage of a
SoC/SoH prediction in on-board EV BMSs. representative cell and its capacity fading (SoH degradation)
characteristics are shown in Fig. 2.
II. PROBLEM STATEMENT AND METHOD OF APPROACH The guided wave responses could be seen to exhibit
Within the context of this paper, we consider pitch-catch repeatable time-domain variations with changing SoC and
guided wave data from indicative commercial Li-ion pouch SoH. Representative actuator-sensor waveforms from the
cells with pre-determined transducer configurations (Fig. 1). pouch cell experiment are shown in Fig. 3. Fig. 3A shows the
This serves as our problem setup to develop an efficient distinct evolution of the guided wave structure with respect to
strategy to process the complex guided wave signals and SoC within one discharge cycle. Fig. 3B, on the other hand,
extract predictive features that can lend themselves to demonstrates the changes in the time-domain waveforms due
enhancing SoC/SoH prediction. The obtained results can then to capacity fading (SoH degradation). These interplays
be generalized and scaled into a unified framework towards between guided wave propagation and electrochemical
other battery/transducer configurations, operating conditions, processes were shown to be the results of the distribution and
and prediction methodologies. redistribution of electrodes’ mechanical properties – namely,
We first experimentally gather ultrasonic guided wave moduli and densities – during cycling and aging [6].
data at various SoC and SoH from the indicative pouch cells
with surface-mounted piezoelectric transducers (Section III).
The waveform structure and its progression due to changing
SoC/SoH are studied in order postulate an appropriate MP-
based TFR technique to process and project the signals onto
basis functions (Section IV). Then, functional relationships
between the parameters of the basis functions and the changes
in SoC/SoH are established (Section V). A comparative
statistical analysis is conducted to evaluate the efficacy of the
proposed time-frequency feature extraction, in comparison
and in conjunction with the conventional voltage
measurement and time-domain features (Section VI).

III. PRELIMINARY EXPERIMENTS & INDICATIVE RESULTS

The experimental results, which the MP feature extraction
and statistical analysis are based on, have been discussed in Fig. 1. A) Schematics of pitch-catch guided wave propagation. B)
our previous companion work [6]. Pitch-catch guided wave experimental pouch cell with built-in piezos.
propagation experiments were performed on 3,650mAh off-
the-shelf Li-ion pouch batteries (graphite/NCM chemistry)

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
Fig. 2. A) Cycle-to-cycle terminal voltage measurement from a
representative cell. B) Remaining capacity (SoH) -vs- cycle number showing
the indicative capacity fading characteristics of the batteries under testing.

IV. FEATURE EXTRACTION USING MATCHING PURSUIT

In our previous work, we showed the feasibility of the
guided wave method in estimating SoC/SoH by performing
state prediction using only two time-domain features – i.e.,
time of flight and signal amplitude (ToF and SA, respectively)
[6]. These limited features only represent the aggregate
behavior of the time-domain waveforms. However, pitch-
catch tone-burst guided wave signals are dynamic, non-
stationary, and multi-dimensional (in time and frequency). Fig. 3. Time-domain evolution of guided wave signals with respect to
This can be intuitively seen from Fig. 3 that different time- changes in SoC and SoH. A) Changes in the time-domain signals with respect
to SoC. B) Changes in the time-domain signals with respect to SoH.
axis portions of the waveforms (wave packets) show distinct Ultrasonic data from the diagonal path at 125 kHz center frequency.
characteristics with respect to the changes in SoC and SoH.
Instead, using an appropriate time-frequency representation We employ MP using the Gabor dictionary (a collection
method, numerous predictive features can be extracted from of scaled, translated, and modulated versions of Gaussian-
the information-rich guided wave signals. The multi- windowed tone bursts) to decompose the guided wave
dimensional time-frequency features may be pooled with the responses into a linear expansion of constituent tone-burst
one-dimensional voltage measurement to construct a new atoms (Fig. 4). The Gabor functions can be represented by the
class of SoC/SoH prediction models. This potentially allows following equation:
prediction accuracy to be significantly improved and/or
greatly simplifies the structure and complexity of the ݃ሺ‫ݐ‬ሻ ൌ  ݁ ିగ௧ ǡ
మ
(1)
prediction model.
In this current work, we propose an efficient feature- ‫ݐ‬െ‫ݑ‬
extraction algorithm based on the MP decomposition ݃ఊୀሺ௦ǡ௨ǡ௩ǡ௪ሻ ሺ‫ݐ‬ሻ ൌ ݃ ൬ ൰ ‘•ሺ‫ ݐݒ‬൅ ‫ݓ‬ሻ
‫ݏ‬
technique [9]. MP is a time-frequency representation method,
which is widely used in the field of guided-wave SHM for where the Gabor dictionary D is a collection of gȖ. Ȗ is a set of
analyzing guided wave signals with respect to the evolution of possible Gabor parameters s, u, v, and w, which are
structural damage. MP finds the best matching projections of coefficients of scaling, translation, modulation, and phase
the guided wave signals onto the span of a redundant change, respectively. The dictionary thus contains all possible
dictionary of waveforms or atoms. In essence, MP expected waveforms or atoms.
decomposes the signals into a linear combination of The Gabor dictionary is a natural choice for our
constituent waveforms that was found to best match the application as the actuation signal used in this study is also
original signal structure. designed to be a modulated Gaussian-windowed tone burst.

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
Given that the incident pulse is a Gabor function, the sensing ȁ‫݂ۃ‬ǡ ݃ఊᇱ ‫ۄ‬ȁ (2)
signals can also be represented by one or a sum of Gabor
functions. This adaptively accelerates the MP algorithm and The operator <., .> represents a standard vector inner
allows meaningful pulses to be efficiently extracted out of product. The residual signal Rf after subtracting the
noisy signals. component along the best atom chosen per above can be
computed as:

ܴ݂ ൌ ݂ െ  ‫݂ۃ‬ǡ ݃ఊᇱ ‫݃ۄ‬ఊᇱ (3)

The process is then repeated on the residual Rf and is

terminated after the Kth iteration (after K atoms have been
extracted). In this study, the first ten constituent atoms were
extracted from the guided wave signals at a given SoC/SoH
state (K = 10), which allows more than 95% of the original
signal energy content to be captured in every case. We also
employ the following transformation to encode the phase and
amplitude of the Gabor atoms into two dependent variables,
ak and bk, following the equation below [10]:

ܽ௞ ൌ  ‫ܴۃ‬௞ ݂ǡ ܲఊᇱೖ ‫ۄ‬ǡܾ௞ ൌ  ‫ܴۃ‬௞ ݂ǡ ܳఊᇱೖ ‫ۄ‬ (4)

‫ݐ‬െ‫ݑ‬ ‫ݐ‬െ‫ݑ‬
ܲఊᇱ ሺ‫ݐ‬ሻ ൌ ݃ ൬ ൰ ‘•ሺ‫ݐݒ‬ሻǡܳఊᇱ ሺ‫ݐ‬ሻ ൌ ݃ ൬ ൰ •‹ሺ‫ݐݒ‬ሻ
‫ݏ‬ ‫ݏ‬

where Rkf is the residual signal after kth atom-decomposition

iteration.

1.2) Optimal Search Based on Sub-Optimal Parameters.

Note that the pre-defined sub-space in the earlier stage is a set
of discrete Gabor basis functions for the sake of efficient
computation. Therefore, only sub-optimal atoms are
determined. Using the sub-optimal parameters Ȗ’k obtained
earlier as the initial guess, non-linear least squares can be
employed to perform a local search and determine the optimal
Gabor functions [11]. The standard Gauss-Newton method is
implemented to numerically search for the optimal parameters
Ȗk = (sk, uk, vk) in the neighborhood of the sub-optimal sets so
Fig. 4. Functionality of Gabor-based MP decomposition. A) Original as to minimize the energy of the residual at kth iteration. (ak,
waveform at 100% SoC and 100% SoH in comparison with the reconstructed bk) are then re-calculated using the same equation above. This
waveform using a linear combination of the first 10 constituent Gabor atoms. double-search strategy is performed at every atom-
B) The waveform structure of the first 5 decomposed atoms.
decomposition iteration. Consequently, this allows us to
obtain an optimal set of parameters for the constituent atoms
The MP algorithms iteratively decomposes the original
by performing decomposition on signals at a reference state
waveform into K constituent atoms in the following way:
(i.e., SoC = 100%, SoH = 100%).
1. Free Baseline Decomposition
2. Constrained Decomposition of Signals at Subsequent States
1.1) First-Pass Sub-Optimal Search. Let f be the original
To address the computational expense requirements for
signal at baseline SoC and SoH (e.g. SoC = 100%, SoH =
on-board BMS applications, an efficient decomposition
100%). First, we pre-define a subspace ī(s, u, v), where the
method is required for the conceivably large stream of data.
parameter s is selected among a uniformly distributed set of
Therefore, a constrained MP decomposition is applied
expected receiving pulse widths by considering the incident
wherein signals at subsequent SoC/SoH states are
pulse width (as determined by the actuation center frequency
decomposed via MP by imposing constraints using the
and number of cycles); v is naturally selected to be
parameters of the preceding state (according to the time
concentrated around the center frequency of the actuation; and
sequence of the data acquired during experiment). The shape
u is selected among the discrete sampled time. Within the pre-
of each atom of the subsequently analyzed signals is
defined subspace ī, the sub-optimal set Ȗ’ which represents
constrained to be similar to that of the preceding one –i.e.,
the best atom in D can be chosen to maximize:
fixed scale and frequency (sk, vk) k = [1, 2,… K] – while the

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
amplitude and translation parameters (ak, bk, uk) are allowed between these secondary atoms and SoC, their physical
to change. This also ensures that the decomposition is robust interpretation is subject of ongoing work. The analysis of
against external factors such as noise and data corruption, these complex waveforms requires high-fidelity
which would have otherwise caused invalid atoms to be computational modelling of wave propagation. Nevertheless,
extracted. Only a Gauss-Newton numerical search is required the essence of this work is captured: MP with the Gabor
to seek the optimal (a’k, b’k, u’k) parameters of the subsequent dictionary is capable of extracting waveform compositions
signals in the vicinity of the immediately preceding sets (ak, that are strongly correlated with battery SoC.
bk, uk). It is important to note that the double-search algorithm The correlations between the Gabor atoms and SoC can be
needs to be performed once for the initial state (SoC = 100%, phrased in a way that is useful for SoC prediction by
SoH = 100%). The subsequent data points only require the considering the functional relationships between the Gabor
computationally lighter constrained decomposition. We then parameters (ak, bk, uk) and SoC. These functional relationships
monitor the evolution of (ak, bk, uk) with reference to changing for the first three atoms are plotted in Fig. 5B showing (ak, bk,
SoC and SoH. uk) as a function of SoC. The overall trends of (ak, bk, uk) agree
with the time-domain observations at varying SoC as
V. EVOLUTION OF GABOR PARAMETERS WITH SOC/SOH presented above. Additionally, the (ak, bk, uk) trends also
The changes in the guided wave signals with respect to capture the non-linearity driven by intercalations and phase
SoC become more apparent when their constituent atoms are transitions at various stages along the discharge process. The
individually observed. Fig. 5A shows the changes in the correlations between (ak, bk, uk) and SoC established through
original guided wave signals and their first three constituent these functional relationships suggest that the Gabor
Gabor atoms at two SoC (100% and 0%) from the first parameters may be utilized in a collective fashion to aid the
discharge cycle (SoH = 100%). It is worth noting that MP is prediction of SoC.
able to decompose the entire complicated original waveform Independently of SoC, strong correlations between guided
into localized and more defined tone bursts, which are wave signals and SoH can also be seen in the extracted Gabor
centered at different points along the time axis. Each portion atoms as shown in Fig. 6A. The first three atoms are plotted
of the original waveform can now be analyzed separately by to show the time-domain shifts in the waveform after 200
considering the evolution of the constituent Gabor atoms with discharge cycles (SoH degraded to 96.5%) despite the same
respect to varying SoC. SoC of 100%. Similarly, functional relationships between the
The first atom, which represents the direct-path waveform Gabor parameters (ak, bk, uk) and SoH can be established at
and thus containing the majority of the energy content, shows any given SoC as illustrated in Fig. 6B. Comparing to the
the expected behavior of decreasing SA and increasing ToF at terminal cell voltage, which is minimally perturbed by aging
lower SoC. The behavior is reminiscent of the decrease in the (Fig. 2A), the Gabor parameters show a significantly larger
aggregate modulus to mass density ratio with decreasing SoC, net change with respect to the change in SoH. Therefore,
which results in decreasing propagation speeds and higher besides presenting more features to the SoH prediction, the
attenuation. The following atoms are known to originate from greater magnitude of the effects from the Gabor parameters
secondary wave modes that travel at different speeds, (relative to the variance of the variables themselves) may
boundary reflections, and scattered signals from interlaminar effectively contribute to improving prediction accuracy.
surfaces. While there are meaningful and strong correlations

Fig. 5. Behavior of Gabor atoms due to the changes in SoC. A) Time-domain behavior of the first three constituent atoms with respect to SoC during Cycle 1
(SoH = 100%). B) The extracted (ak, bk, uk) from the first three atoms and their functional relationships with respect to SoC.

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
Fig. 6. Behavior of Gabor atoms due to the changes in SoH. A) Time-domain behavior of the first three constituent atoms with respect to SoH while holding
SoC constant at 100%. B) The extracted (ak, bk, uk) from the first three atoms and their functional relationships with respect to SoH.

The effects on the Gabor parameters due to changes in SoC

and SoH as observed independently can be combined to form VI. SOC/SOH PREDICTION MODELS
a basis for the multivariate SoC/SoH prediction. The Gabor In this section, we perform a statistical analysis to
parameters that encompass the entire range of SoC and SoH comparatively evaluate the efficacy of Gabor parameters as
within this experiment are analyzed collectively, and these predictors for SoC/SoH with reference to traditional voltage
results are shown in Fig. 7. The surface plots are the loci of measurements and time-domain features. This first-pass
extracted Gabor parameters at a given pair of state variables: statistical analysis attempts to create models that explicitly
SoC and SoH. In other words, the Gabor parameters can be describe SoC and SoH as a function of the predictive features
expressed explicitly as functions of both SoC and SoH. This and evaluate the model performance by batch regression. It is
illustrates the possibility of creating a framework to use the important to note that this comparative evaluation is model
Gabor parameters as predictive features in an inverse setting agnostic. For practical implementation, the most appropriate
to simultaneously estimate battery SoC and SoH. model needs to be identified from a plethora of machine
learning tools and state estimation techniques, with additional
work in optimizing the model structure.
The prediction methods are based on Generalized Additive
Models (GAMs) [12]. GAMs with regression splines, unlike
other statistical learning methods, are chosen because they are
easy to interpret and allow model structures of different
complexity levels to be systematically compared. GAMs are a
non-parametric regression technique, which allows dependent
variables to be described by smooth non-linear functions of
covariates. Thin-plate regression splines were employed as
smoothed non-linear fits of covariates using maximum
penalized likelihood to prevent overfitting. The GAM analysis
uses the ‘mgcv’ package (version 1.8-18) in R. A family of
multivariate GAMs, a multivariate normal model (multinom),
is used to fit the bivariate (SoC and SoH) response to multiple
predictors. Herein, only the smooth main effects were
included without interaction among predictors.

ܵ‫ ܥ݋‬ൌ  ߙଵ ൅ ݂ଵǡଵ ሺܺଵ ሻ ൅ ݂ଵǡଶ ሺܺଶ ሻ ൅  ǥ൅  ݂ଵǡ௉ ሺܺ௉ ሻ

Fig. 7. Gabor parameters of the first three atoms as a function of SoC & (5)
SoH. uk is the translation parameter. ak and bk are coefficient of the sine and ܵ‫ ܪ݋‬ൌ  ߙଶ ൅ ݂ଶǡଵ ሺܺଵ ሻ ൅ ݂ଶǡଶ ሺܺଶ ሻ ൅  ǥ൅  ݂ଶǡ௉ ሺܺ௉ ሻ
cosine terms in the Gabor function.
where fi,j is the regression spline relating the response Yi (SoC
and SoH) to the predictor Xj. Įi is the zero-offset for each of
the bivariate response.

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
 Each GAM involves P predictors Xj which are a subset of predictors. This is in stark contrast to our previous work and
predictors selected from the set of all available features (V, ak, recent efforts in the literature [3, 6], where SoC and SoH are
bk, uk); k = 1-10, where V is the cell terminal voltage. The estimated independently of each other – i.e., one state is held
different model structures analyzed in this section contain constant or not considered when estimating the other.
different subsets of predictors whose comparative statistical Therefore, the bivariate prediction results confirm that the
performance in predicting SoC and SoH is being evaluated. guided wave Gabor parameters, unlike cell voltage, have
The residual maximum likelihood (REML) scores and Akaike sufficiently meaningful correlations simultaneously to both
Information Criterion (AIC) scores [85] are calculated and SoC and SoH. Also considering that our statistical analysis
compared for each of the models. The fitted models are then was data-driven and was merely a comparative batch
applied to the validation dataset to compare the predicted and regression, there is room for improvement by, for instance,
measured SoC and SoH. From which, the 10-fold cross moving towards more efficient and robust model-based
validation (CV) error can be evaluated to reinforce the model approaches. Investigation is underway to develop physics-
comparison in addition to the use of classical statistical based models that functionally relate the Gabor parameters to
criteria. the SoC/SoH space. The model fidelity will also be improved
by experimentally exploring and taking into account other
VII. COMPARATIVE STATISTICAL RESULTS factors such as discharge profiles, temperature, cell-to-cell
The comparative performance of various GAM structures variability, etc.
with different subsets of features can be seen from the CV
errors of SoC and SoH prediction (root mean squared errors
(RMSE)) as shown in Fig. 8. The detailed statistical results of
the model structures considered in this study are summarized
in TABLE I. By using Gabor parameters in conjunction with
the conventional voltage measurement as predictors for SoC
and SoH, the prediction accuracy is significantly improved.
Using the entire set of the available features ((V, ak, bk, uk); k
= 1-10), the SoC prediction accuracy can be increased by
almost two-fold. Particularly remarkable is the SoH prediction
whereby augmenting the voltage measurement with guided-
wave Gabor parameters improves prediction accuracy by as
much as 20 times. It is well known that terminal voltage, while
being able to serve as a strong indicator for SoC, is almost
invariant with SoH. Given that this is a two-unknown problem
(SoC and SoH), relying only on voltage measurement alone
would intuitively result in an underdetermined system. Hence,
it is imperative and demonstrated herein that additional
parameters that are strongly correlated with SoH need to be
taken into consideration. Fig. 8. Prediction performance of different GAM structure with various
Our analysis also signifies the importance of signal subsets of predictive features. A) Average SoC prediction error (RMSE). B)
processing and feature extraction in using guided wave for Average SoH prediction error (RMSE).
SoC/SoH prediction. Our previous work only considered time
domain parameters (ToF and SA), which are indicative of the VIII. CONCLUDING REMARKS
global behavior of the waveform, or more specifically, the
This work established a framework for applying novel
behavior the wave packet with the highest energy content [6].
signal processing and feature extraction that enable the use of
The model that only uses the Gabor parameters from the first
acousto-ultrasonic guided waves for battery SoC/SoH
constituent atom ((ak, bk, uk); k = 1) exhibits almost equivalent
prediction. We experimentally demonstrated the SoC/SoH-
prediction performance to the time-domain-only model. This
induced evolution in the guided wave signals propagated
is because the first constituent Gabor atom is also by and large
through commercial Li-ion pouch batteries using low-profile,
reminiscent of the highest energy wave packet. By using
surface-mounted piezoelectric transducers. It was found that
Gabor parameters from subsequently extracted atoms the
the complex interplay between electrochemical processes and
SoC/SoH prediction accuracy can then be seen to
wave propagation physics resulted in meaningful SoC/SoH
precipitously improve. That is, a lot of meaningful
correlations to be encoded in these dynamic, non-stationary,
correlations with SoC and SoH are encoded in other portions
multi-dimensional waveforms. In this work, special emphasis
of the waveform beyond the primary wave packet, which can
was therefore given to developing an efficient time-frequency
only be extracted via an appropriate time-frequency
representation method to process the waveforms and extract
representation method.
predictive features that significantly improve the SoC/SoH
Moreover, it is noteworthy that this analysis treats SoC and
predictions. In essence, this study has contributed the
SoH as a bivariate response. The prediction models
following:
simultaneously estimate both SoC and SoH using pooled

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.
 • An efficient MP algorithm was proposed based on the Energy, Inc., as well as Mr. Raphael Nardari at Zenith
Gabor dictionary to decompose complex guided wave signals Aerospace, Inc. for feedback and suggestions. The authors
into a linear combination of simpler, constituent tone-burst would like to thank Dr. Surajit Roy at California State
atoms. We also employed constrained MP decomposition that University, Long Beach, for valuable advice on the matching
allowed guided signals at contiguous states to be efficiently pursuit algorithms. We would also like to thank Acellent
and accurately decomposed by exploiting the atom Technologies Inc. for providing necessary hardware support
information of the preceding states. for the experiments conducted in this research.
• The parameters that describe the Gabor atoms’
waveform structure, or Gabor parameters, were extracted and REFERENCES
collectively analyzed over the SoC/SoH space studied in this [1] Waag, W., Fleischer, C., and Sauer, D.U.: ‘Critical review of the
methods for monitoring of lithium-ion batteries in electric and hybrid
work. Functional relationships between the Gabor parameters vehicles’, Journal of Power Sources, 2014, 258, pp. 321-339
and SoC/SoH showed strong and meaningful correlations [2] Linden, D.: ‘Handbook of batteries’, in Editor (Ed.)^(Eds.): ‘Book
which substantiated the use of the Gabor parameters as Handbook of batteries’ (1995, edn.), pp. 265
predictive features for SoC/SoH prediction. [3] Davies, G., Knehr, K.W., Van Tassell, B., Hodson, T., Biswas, S.,
Hsieh, A.G., and Steingart, D.A.: ‘State of Charge and State of Health
• GAMs were generated with different model structures Estimation Using Electrochemical Acoustic Time of Flight Analysis’,
utilizing different subsets of the pool of predictive features. Journal of The Electrochemical Society, 2017, 164, (12), pp. A2746-
We showed that the prediction accuracy of SoC and A2755
[4] Gold, L., Bach, T., Virsik, W., Schmitt, A., Müller, J., Staab, T.E., and
particularly SoH was significantly improved when Sextl, G.: ‘Probing lithium-ion batteries' state-of-charge using
augmenting the conventional voltage measurement with the ultrasonic transmission–Concept and laboratory testing’, Journal of
guided-wave Gabor parameters. We also treated SoC and SoH Power Sources, 2017, 343, pp. 536-544
as a bivariate response to demonstrate that the Gabor [5] Hsieh, A., Bhadra, S., Hertzberg, B., Gjeltema, P., Goy, A., Fleischer,
J., and Steingart, D.: ‘Electrochemical-acoustic time of flight: in
parameters are simultaneously and strongly correlated to both operando correlation of physical dynamics with battery charge and
SoC and SoH. health’, Energy & environmental science, 2015, 8, (5), pp. 1569-1577
The current and future work addresses the improvement of [6] Ladpli, P., Kopsaftopoulos, F., and Chang, F.-K.: ‘Estimating state of
charge and health of lithium-ion batteries with guided waves using built-
the MP algorithm to include possible dispersion in piezoelectric sensors/actuators’, Journal of Power Sources, 2018,
characteristics in an efficient manner by coupling data-driven 384, pp. 342-354
decomposition with physical knowledge. Transitioning from [7] Sood, B., Osterman, M., and Pecht, M.: ‘Health monitoring of lithium-
a data-driven predictive approach to a model-based approach ion batteries’, in Product Compliance Engineering (ISPCE), 2013 IEEE
Symposium on, pp. 1-6
is also underway. This task requires a full understanding of the [8] Raghavan, A.: ‘Guided-wave structural health monitoring’, 2007
underlying physical phenomena that govern the SoC/SoH- [9] Mallat, S.G., and Zhang, Z.: ‘Matching pursuits with time-frequency
induced changes in the time-frequency space of guided waves. dictionaries’, IEEE Transactions on signal processing, 1993, 41, (12),
pp. 3397-3415
While further investigation is needed to understand the [10] Ferrando, S.E., Kolasa, L.A., and Kovaþeviü, N.: ‘Algorithm 820: a
influence of factors, such as temperature and operating flexible implementation of matching pursuit for Gabor functions on the
profiles, the heretofore obtained results clearly show the interval’, ACM Transactions on Mathematical Software (TOMS), 2002,
feasibility of using guided waves for SoC/SoH prediction and 28, (3), pp. 337-353
[11] Hong, J.-C., Sun, K.H., and Kim, Y.Y.: ‘The matching pursuit approach
lay groundwork for future studies. based on the modulated Gaussian pulse for efficient guided-wave
damage inspection’, Smart Materials and Structures, 2005, 14, (4), pp.
ACKNOWLEDGMENTS 548
[12] Wood, S.N.: ‘Fast stable restricted maximum likelihood and marginal
The work is supported by the Advanced Research Projects likelihood estimation of semiparametric generalized linear models’,
Agency - Energy (U.S. Department of Energy) [grant number Journal of the Royal Statistical Society: Series B (Statistical
DE-AR0000393] and the Stanford Precourt Institute of Methodology), 2011, 73, (1), pp. 3-36
Energy. The authors highly appreciate support from Dr. Keith
Kepler, Dr. Hongjian Liu, and Dr. Michael Slater at Farasis

TABLE I
COMPARATIVE STATISTICAL RESULTS OF DIFFERENT GAM STRUCTURES USING GABOR PARAMETERS TOGETHER WITH CELL VOLTAGE AS PREDICTORS
Guided wave information only Guided wave + cell voltage

Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 8 Model 9 Model 10 Model 11
Parameters
(ak, b k, u k); (a k, b k, u k); (a k, b k, uk); (ak, bk, uk); (V, ak, bk, uk); (V, ak, bk, uk); (V, ak, bk, uk); (V, ak, bk, uk);
Predictors (ToF, SA) (V) (V, ToF, SA)
k=1 k = 1-3 k = 1-5 k = 1-10 k=1 k = 1-3 k = 1-5 k = 1-10

Number of predictors 2 3 9 15 30 1 3 4 10 16 31

Statistical Results

SoC CV10 RMSE [%] 5.24 11.96 1.65 0.94 0.55 0.67 0.57 0.57 0.44 0.39 0.36
SoH CV10 RMSE [%] 0.56 0.7 0.08 0.06 0.05 1.04 0.19 0.19 0.07 0.06 0.05
REML 24426 31541 -15311 -26096 -39329 11272 -17596 -21105 -34968 -46282 -49316
AIC 48539 62685 -31443 -53347 -80615 22412 -35616 -42677 -70666 -93772 -100561

Authorized licensed use limited to: ASTAR. Downloaded on February 22,2023 at 08:36:48 UTC from IEEE Xplore. Restrictions apply.