4.state of Health Estimation

4.1.1: Welcome to the course!
Welcome to the course!
key off: store data

key on: initialize
meas. voltage estimate estimate balance compute
Welcome to Battery State of Health (SOH) Estimation!

current state of state of power
temperature charge (SOC) health (SOH) cells limits
loop once each measurement interval while pack is active
This course is the fourth in a specialization that investigates

the proper management and control of battery packs, usually comprising many cells
BMS must estimate nonmeasurable quantities that describe battery pack condition
key off: store data

key on: initialize

Dr. Gregory L. Plett j University of Colorado Colorado Springs Battery State-of-Health (SOH) Estimation | How does lithium-ion cell health degrade? j 1 of 4
What topics will we study in this course?
key off: store data

key on: initialize
In this course, you will learn:

The primary ways that lithium-ion battery loop once each measurement interval while pack is active
cells degrade over time, motivating a need to be able to adjust

cell-model parameter values
How to estimate cell resistance as it changes over time
How to estimate cell total capacity as it changes over time using
several different methods (and their tradeoffs)
How to co-estimate the state and parameters of a cell model as
both change with different timescales (honors)
What skills will you gain in this course?

key off: store data
key on: initialize
After completing the course, you’ll be able to:

Identify primary degradation mechanisms loop once each measurement interval while pack is active
in lithium-ion cells and understand how they work

Execute provided Octave script to estimate total capacity using
WLS, WTLS, and AWTLS methods and lab-test data, and to
evaluate results
Compute confidence intervals on total-capacity estimates
Compute estimates of a cell’s equivalent-series resistance using
lab-test data
Specify the tradeoffs between joint and dual estimation of state
and parameters, and steps that must be taken to ensure robust
estimates (honors)
Prerequisites; for further study
key off: store data

key on: initialize
Prerequisites: loop once each measurement interval while pack is active
Course 1: Intro. to Battery Management Systems

Basic background in BMS requirements, sensing
Course 2: Equivalent Circuit Cell Model Simulation
Specific background in ESC cell model
Course 3: Battery State of Charge (SOC) Estimation
Probability theory, nonlinear Kalman filters (honors)
Resource:
We’ll study Ch. 4 of Battery Management Systems,
Vol. 2, Equivalent-Circuit Methods, Artech House
For further study, you can confer this optional resource
4.1.2: What changes as a cell ages?
Performance changes as cells age
key off: store data

key on: initialize
cells
We know that a BMS must estimate certain quantities

temperature charge (SOC) health (SOH) limits
indicative of its cells’ states and parameters

We have now seen a number of ways to estimate battery cell
state—the quickly changing quantities
Now, we turn our attention toward estimating battery cell
parameters—the slowly changing, pseudo-static quantities
In particular, we are most interested in those quantities that
reflect a change in the performance that the pack can deliver
These are indicators of battery pack “state-of-health” (SOH)
SOH most often summarized by estimating present total capacity, series resistance
Loss in total capacity versus time
key off: store data

key on: initialize
As a battery cell ages, its total capacity decreases, primarily

due to unwanted side reactions, structural deterioration

◆❡❣❛t✐✈❡ P♦s✐t✐✈❡ ◆❡❣❛t✐✈❡ P♦s✐t✐✈❡ ◆❡❣❛t✐✈❡ P♦s✐t✐✈❡
❡❧❡❝tr♦❞❡ ❡❧❡❝tr♦❞❡ ❡❧❡❝tr♦❞❡ ❡❧❡❝tr♦❞❡ ❡❧❡❝tr♦❞❡ ❡❧❡❝tr♦❞❡
❉✐s❝❤❛r❣❡ ❉✐s❝❤❛r❣❡ ❉✐s❝❤❛r❣❡
❈❤❛r❣❡
❈❤❛r❣❡ ❙✐❞❡ ❈❤❛r❣❡
r❡❛❝t✐♦♥
Ideally, lithium moves back and forth between electrodes; none is lost
Side reactions can consume lithium, primarily when charging cell
Structural deterioration can remove storage sites for lithium from one or both
electrodes
Capacity fade
key off: store data
key on: initialize
Phenomena often referred to as capacity fade

Knowing present value of total capacity

important because it is:
Major contributing factor to energy calculation
Major contributing factor to SOC estimation if
coulomb counting is used (else, minor
contributing factor)
(Insignificant in power estimation)
Power fade
key off: store data

key on: initialize
As a cell ages, its series resistance increases, also primarily

due to unwanted side reactions and structural deterioration

Knowing present ESR is important because:
Major contributing factor to power calculation
Major contributing factor to SOC estimation
for some voltage-based methods (else, it is a
minor factor)
(Insignificant role in energy estimation)
Since resistance and power are so tightly
coupled, resistance rise in a cell is commonly
referred to as power fade
Other cell parameter values
key off: store data

key on: initialize
The OCV relationship changes as the electrodes lose capacity

For most chemistries in common use, unlikely that this

change is large enough to be readily detectable
Other cell parameters almost certainly change 4.2
OCV after Aging
as well; however, I’m not aware of BMS that 4
make efforts to estimate them 3.8
Voltage (V)
ESR, total capacity have dominant impact 3.6
KF-based methods in this course could be 3.4 0% capacity loss

used to estimate the others, if application 3.2
10% capacity loss
20% capacity loss
demands it 3
0 20 40 60
30% capacity loss
80 100
State of charge (%)
Summary and next steps

key off: store data
key on: initialize
Cell parameter values change over time as cell ages

Total capacity and ESR are the most important to track

In this course, we first look at some qualitative explanations for cell aging
Next, we explore a simple method to estimate ESR
We then look at ways to estimate total capacity
We finally look at KF-based methods to estimate up-to-date values for any desired
set of cell parameters (honors)
4.1.3: Negative-electrode aging processes at particle surface
Lithium-ion aging: Negative electrode
key off: store data

key on: initialize
In the next lessons, we’ll seek to describe aging qualitatively.1

In negative electrode, aging observed at three scales:

At the surface of the electrode particles

Within the electrode particles themselves
Within the composite electrode structure (active
materials; conductive additives; binder; current
collector; porosity; etc)
In this lesson, our focus is on aging mechanisms that take place at particle surfaces
We will assume that a graphitic carbon is used as the negative-electrode active
material, although similar results hold for silicon as well
1
Much of this is from: Vetter et al., “Ageing mechanisms in lithium-ion batteries,” Journal of Power
Sources, 147, 2005, 269–281.
Negative electrode aging at surface of particles
key off: store data

key on: initialize
Graphitic negative electrodes operate at voltages outside

electrochemical stability window of electrolyte components

Reductive electrolyte decomposition takes place at the electrode/ electrolyte
interface when the electrode (and cell) is in the charged state
❋✐❧♠
This occurs mainly—but not only—at beginning of ❣r♦✇t❤ ❙♦❧✈❡♥t
cycling, especially during “formation” cycle ❙❊■
✂❧♠
The decomposition products form a “solid-electrolyte
interphase” (SEI) surface film covering the electrode’s
❙♦❧✐❞ ♣❛rt✐❝❧❡ ♦❢
surface ❛❝t✐✈❡ ♠❛t❡r✐❛❧
Properties of SEI
key off: store data
key on: initialize
The SEI is a passivating layer, which slows/prevents further

reaction between graphite and electrolyte

Lithium is consumed when SEI film forms, lowering capacity of cell
SEI is porous, allowing de/intercalation of Li from/to graphite, but increases
resistance of ion transfer
SEI composition is complex and not uniform
Various products form, then decompose, then form more stable products
Breakdown of SEI; regrowth
key off: store data

key on: initialize
High temperatures contribute to breakdown of SEI, which

can lead to new SEI forming on exposed graphite

Pores in SEI also allow some solvent to penetrate to ▲✐t❤✐✉♠ ♣❧❛t✐♥❣
graphite surface, reacting and growing more SEI film ❆♥♦❞❡ ♣♦✐s♦♥✐♥❣ ❙❊■
Trace water in electrolyte combines with ionized ✂❧♠
fluorine to form hydrofluoric acid (HF), which attacks ❆❝✐❞
❛tt❛❝❦
the SEI—new SEI forms
Whenever SEI forms/grows, lithium is lost and so cell ❙♦❧✐❞ ♣❛rt✐❝❧❡ ♦❢
❛❝t✐✈❡ ♠❛t❡r✐❛❧
capacity decreases
Anode poisoning; lithium plating
key off: store data

key on: initialize
meas. voltage estimate estimate compute
Products of positive-electrode degradation also end

balance
up as part of SEI (anode poisoning):

If not electrically conductive, increases resistance ▲✐t❤✐✉♠ ♣❧❛t✐♥❣

❆♥♦❞❡ ♣♦✐s♦♥✐♥❣
Can plug pores, blocking lithium from cycling, ❙❊■
✂❧♠
causing capacity fade ❆❝✐❞
At low temperatures, diffusion in particles slows: ❛tt❛❝❦
If charging is forced, local overpotential can attain ❙♦❧✐❞ ♣❛rt✐❝❧❡ ♦❢

level that causes lithium plating on particle surface ❛❝t✐✈❡ ♠❛t❡r✐❛❧
Capacity irreversibly lost; dendrites can form and

grow, eventually leading to internal short circuit
Summary of surface effects

key off: store data
key on: initialize
Table summarizes surface effects (bold = more severe):

Cause Effect Leads to Enhanced by

Continuous low-rate electrolyte Li loss, Capacity fade, High temperatures,
decomposition builds SEI impedance rise Power fade high cell SOC
Decrease of accessible surface Impedance rise Power fade High temperatures,
area due to SEI growth high cell SOC
Changes in porosity due to Impedance rise Power fade High cycling rate,
volume change and SEI growth high cell SOC
Metallic Li plating, subsequent Lithium (electro- Capacity Low temperature,
electrolyte decomposition lyte) loss (power) fade high charge rates
Credits
key off: store data

key on: initialize

Image of lithium-ion electrode on slide 1 courtesy Sangwoo

Han, used with permission
4.1.4: Negative-electrode aging processes in bulk and composite electrode
Negative electrode aging in bulk
key off: store data

key on: initialize
Aging/degradation occurs inside negative-electrode particles

Dis/charging causes (anisotropic) particle volume ❙♦❧✈❡♥t

❝♦✲✐♥t❡r❝❛❧❛t✐♦♥
changes (usually less than 10 %)
❙❊■
●❛s ❡✈♦✲
Strain can lead to cracking of particles, more SEI ❧✉t✐♦♥
✂❧♠
formation on exposed graphite ❙♦❧✐❞
♣❛rt✐❝❧❡
Cracking of SEI itself, and more SEI formation on ♦❢ ❛❝t✐✈❡
exposed graphite ❈r❛❝❦✐♥❣ ♠❛t❡r✐❛❧
Graphite exfoliation (layers flaking off) due to solvent cointercalation with lithium is
considered to have a bigger impact
Solvent reacting with graphite inside particles releases gasses, accelerates cracking
Negative electrode aging in composite electrode
key off: store data

key on: initialize
Stresses and strains within electrodes can cause

balance
mechanical and electronic contact loss:

❆❝t✐✈❡ ♠❛t❡r✐❛❧s
Between graphite particles ❈♦♥❞✉❝t✐✈❡ ❛❞❞✐t✐✈❡
Between current collector and particles ❇✐♥❞❡r
Between binder and particles
Between binder and current collector
Results in higher impedance and can result in ❈✉rr❡♥t ❝♦❧❧❡❝t♦r
capacity loss if particles become electrically
disconnected from current collector
Porosity of electrode can be reduced by volume changes and growth of SEI,
impeding movement of lithium ions in electrolyte, increasing resistance
Current-collector corrosion
key off: store data
key on: initialize
At low voltages (near 1:5 V) copper current collector can

corrode, releasing Cu2C into electrolyte:

Reduced current-collector/particle contact, higher cell resistance

Corrosion products that deposit on electrode particles have poor electronic
conductivity, giving higher film resistance
Leads to nonuniform current and potential distributions across cell plate area,
accelerates aging in parts of the cell, gives preference toward lithium plating
Copper also makes a metallic annealing site that can accelerate lithium plating,
dendrite growth, and hence short circuits
Summary of aging in bulk, composite electrode
key off: store data

key on: initialize
Table summarizes bulk/composite effects (bold = more severe)


Solvent cointercalation, gas Loss of active Capacity Overcharge
evolution, graphite exfoliation material, Li loss fade
Contact loss b/w particles Loss of active Capacity High rate, low
due to volume changes material fade cell SOC
Decomposition of binder Li loss, loss of mech- Capacity High SOC,
anical stability fade high temperatures
Current collector corrosion Impedance rise; Pwr. fade; Overdischarge,
nonuniform current accelerates low cell SOC
and potentials aging
4.1.5: Positive-electrode aging processes
Lithium-ion aging in the positive electrode
key off: store data

key on: initialize
In previous lessons, you learned about physical degradation

mechanisms that occur in the negative-electrode region

of a lithium-ion battery cell
In positive electrode, as with negative electrode, aging occurs in three regions:
At the positive-electrode particle surface
Within the “bulk” of the active materials themselves
In the composite positive electrode
Positive electrode aging at surface of particles
key off: store data

key on: initialize
Electrolyte oxidation and LiPF6 decomposition can form

surface layer on positive electrode materials as well

This is not as pronounced as for negative electrodes
A bigger factor is the dissolution of metals from the electrode into the electrolyte,
and products formed from these metals which can re-precipitate on the surface as
high-resistance film
Dissolution of Mn or Co into electrolyte results in capacity loss (fewer lithium
storage sites), can poison negative electrode
Mechanism depends on which oxide is used, but tends to happen predominantly
at low/high states of charge, and can be greatly accelerated by high temperature
Positive electrode aging in bulk

key off: store data
key on: initialize
Phase transitions (distortions in shape of crystal structure,

without changing structure itself) can lead to cracking

Transitions caused by presence/absence of Li in storage sites, leading to different
local molecular forces
Some phase transitions are normal and reversible
Others cause collapse (e.g., overcharge layered structure), rapid capacity loss
Can also lead to “structural disordering.” when electrode crystal structure breaks
down (bonds broken, reformed to different atoms, collapsing the tunnel-like
structures that allow Li movement: Li sites are lost and, Li can be trapped)
Phase transitions near surface can lead to permanent sub-surface layers forming
that do not allow Li to move freely as in unaltered crystal structure
Positive electrode aging in composite electrode
key off: store data

key on: initialize
The positive electrode also experiences degradation
of the inactive components of the cell:

❆❝t✐✈❡ ♠❛t❡r✐❛❧s
Binder decomposition ❈♦♥❞✉❝t✐✈❡ ❛❞❞✐t✐✈❡
Oxidation of conductive particles (e.g., carbon ❇✐♥❞❡r
black)
Corrosion of the current collector
Loss of contact to conductive particles due to ❈✉rr❡♥t ❝♦❧❧❡❝t♦r
volume changes
Summary of aging at positive electrode (1/2)
key off: store data

key on: initialize
Principal aging mechanisms at positive electrode (1/2)


Phase transitions Cracking of Capacity fade High rates, high/low
active particles SOC
Structural Li sites lost and Capacity fade High rates, high/low
disordering Li trapped SOC
Metal dissolution Migration of Capacity fade High/low SOC, high
and/or electrolyte soluble species temperature
decomposition Reprecipitation of Power fade
new phases, form
surface layer
Summary of aging at positive electrode (2/2)

key off: store data
key on: initialize
Principal aging mechanisms at positive electrode (2/2)


Electrolyte Gas evolution High temperature
decomposition
Binder Loss of contact Power fade
decomposition
Oxidation of Loss of contact Power fade
conductive agent
Corrosion of Loss of contact Power fade High SOC
current collector
4.1.6: Sensitivity of cell voltage to changes in equivalent series resistance (ESR)
Sensitivity of cell terminal voltage to ESR
key off: store data

key on: initialize
A common approach to estimating SOH is to compute

estimates of equivalent-series resistance and total capacity

Estimating ESR R0 turns out to be relatively simple because it is highly
“observable” from voltage measurements
Consider vk D OCV.´k / C M hk
P
i Ri iRi ;k ik R0
Define the sensitivity of the voltage measurement to a change in resistance as
SvRk0 D Rv 0 ddRvk D R0
vk k
i
k 0
Since ik can be very large, absolute sensitivity is high
Sensitivity of cell terminal voltage to ESR
key off: store data

key on: initialize
One approach to estimating R0 is to compare voltages at

two adjacent time samples

P
vk D OCV.´k / C M hk i RP
i iRi ;k ik R0
vk 1 1D OCV.´k 1/ C M hk i Ri iRi ;k 1 ik 1 R0
vk 1 vk R0 .ik 1 ik / ,
where SOC, vCi and hk change relatively slowly compared with rate of change of ik
So, we can estimate
b0;k
R D vik vk
ik
1
.
k 1
Issues (1–2)
key off: store data
key on: initialize
Can compute R
ISSUE I : b0;k only when ik ¤ 0. So, we skip meas. voltage
current
temperature
estimate
state of
charge (SOC)
estimate
state of
health (SOH)
balance
cells
compute
power
limits
updates when jik j is small (avoids amplifying noise, as well)

Because ESC model has imperfect fidelity with respect to true cell, and
ISSUE II :
b0;k is quite noisy
because of inaccuracy introduced via specific approximations, R
Might consider using total least squares approach (see later re. total capacity
estimation) but can also simply filter. For example (where 0 < 1)
bfilt0;k
R bfilt0;k 1 C .1
D R b0;k
/R
Tends to work quite well
Example filtered resistance
key off: store data

key on: initialize
Cell current for resistance test Cell voltage for resistance test current
temperature
state of
charge (SOC)
state of
health (SOH) cells
power
limits
Terminal voltage (V)

50 4.2
Current (A)
0 4
Filtered resistance
−50
3.8 quite close to true
3.6 resistance
0 100 200 300 400 0 100 200 300 400
Time (min) Time (min)
R0 estimate (unfiltered) R0 estimate (unfiltered) R0 estimate (filtered)
4 4 4
R0 values (mOhm)
R0 values (mOhm)
R0 values (mOhm)
3 3 3
2 2 2
1 Estimate 1 Estimate 1 Estimate

Actual Actual Actual
0 0 0
0 100 200 300 400 80 100 120 140 160 0 100 200 300 400
Time (min) Time (min) Time (min)
Issues (3–4)
key off: store data

key on: initialize
ESR is SOC-dependent
ISSUE III :
SOC dependence could be handled by adapting resistance loop once each measurement interval while pack is active
vectors rather than scalars

ISSUE IV: ESR is temperature-dependent
Temperature dependence can be well modeled as

1 1
R0 D R0;ref exp ER0 ;ref
Tref T
but if pack dwells near one temperature for an extended period, results at other
temperatures may become biased
Can also adapt matrix of resistances vs. SOC and temperature
Summary
key off: store data
key on: initialize
An estimate of equivalent-series resistance R0

balance
is a component to most SOH estimates

The good news is that ESR is highly observable: R0 estimates compared

4
even simple methods can estimate its value well
R0 values (mOhm)
You learned a simple—but noisy—approach, 3
based on differencing voltages and currents 2

Filtering these results can produce good ESR
1
estimates Unfiltered
Filtered
You also learned some ways to handle SOC- 0
0 100 200 300 400
and temperature-dependence Time (min)
4.1.7: Sensitivity of cell voltage to changes in cell total capacity
Sensitivity to total capacity
key off: store data

key on: initialize
Estimating total capacity well turns out to be quite difficult

Consider the sensitivity of voltage measurement to capacity:

!
Q dvk Q d X
SvQk D vk dQ
D vk dQ
OCV.´k / C M hk Ri iRi; ;k ik R0
i
Notice that
1. Term in parenthesis is not directly a function of Q
2. But, we are evaluating a total derivative, not a partial derivative
So, there is still (some) hope that total capacity is observable through voltage
measurements
Evaluating sensitivity through OCV
key off: store data

key on: initialize
Consider the first term in the sensitivity equation:

dOCV.´k / @OCV.´k / d´k

dQ
D @´k dQ
For most cells, OCV slope is very shallow, so @OCV.´k /=@´k is very small. Further,
d´k d´k 1 d.1=Q/
dQ
D dQ
k 1 ik 1 t
dQ
D dd´Q
k 1
C k 1Qik2 1t
This total derivative can be calculated recursively
Considering recursive sensitivity through OCV

key off: store data
key on: initialize
From the previous slide, dd´Qk D d´dkQ 1 C k 1Qik2 1 t

Grows when ik is in the same direction for a considerable

amount of time and shrinks when ik changes direction
For random ik (e.g., HEV) it is around the same order of magnitude as the
second term, which can be computed from rms current
The second term has t factor, which is often on the order of 1=3600 or less
Summary to date: Sensitivity of voltage to capacity through the OCV term is small,
but can be maximized either by charging or discharging for long periods
Sensitivity through hysteresis
key off: store data

key on: initialize
Similarly, sensitivity of the voltage to capacity through

hysteresis term is small (it is zero through the other terms)

As a consequence, individual voltage measurements have very little information
regarding capacity
Must somehow combine many voltage measurements
Simple ideas, like used to estimate R b0 will not work well
So, we explore two basic approaches:
First, look at total-least-squares approaches, which are optimal
Next, look at KF-based approaches, which can work well (honors)
Summary
key off: store data

key on: initialize
An estimate of total capacity Q is a component of most SOH

estimates
The bad news is that Q is nearly unobservable: simple methods cannot estimate its
value well
The remainder of this course will explore reasons why one common approach to
estimating Q is incorrect, and how to estimate Q correctly
Next week, we explore why ordinary-least-squares regression is the wrong
approach, and why total-least-squares regression is correct
Then, we look at computationally efficient methods for computing the
total-least-squares solution
Examples will demonstrate these claims
4.1.8: Where from here?
Summary of this week
key off: store data

key on: initialize
This week, you learned that cell static and dynamic

characteristics change as the cells age

Generally, capacity is lost. . . “capacity fade”
Generally, resistance increases. . . “power fade”
You learned, qualitatively, many of the specific degradation mechanisms in both the
negative and positive electrodes
You learned that cell voltage is very sensitive to changes in resistance, so
resistance is relatively simple to estimate well
You learned that cell voltage is very insensitive to changes in total capacity, so total
capacity is very difficult to estimate well
Where from here?
key off: store data

key on: initialize
Next week, you will begin to learn ways to estimate cell total
capacity
First, as a baseline method, you will learn how to use

ordinary and weighted least-squares methods
You will learn that while these methods are very simple to
implement, they produce biased estimates of total capacity
You will learn improved total-least-squares methods that are
unbiased, but impractical to implement in a BMS
However, later in the course you will learn how to make
practical unbiased estimators
Credits
key off: store data
key on: initialize
Credits for photos in this lesson


“Climbing upward” on slide 2: Pixabay license

(https://pixabay.com/en/service/license/), from
https://pixabay.com/en/white-male-3d-man-isolated-3d-1871366/
4.2.1: What’s wrong with using ordinary least squares to estimate total capacity?
Linear structure of capacity equation
key off: store data

key on: initialize
Recall that the sensitivity of the cell voltage to cell capacity is

X
very low, so noise tends to bias results

k2 1
t
Consider the SOC equation, ´k2 D ´k1 kik
Q
X k Dk1

k2 1
We can rearrange its terms to get: t kik D Q .´k2 ´k1 / ,
k Dk1
y
x
where the obvious linear structure of y D Qx becomes apparent

Using a regression technique, for example, one may compute estimates of Q
One needs only to find values for “x ” and “y ”
Dr. Gregory L. Plett j University of Colorado Colorado Springs Battery State-of-Health (SOH) Estimation | Total-least-squares battery-cell capacity estimation j 1 of 5
Problem with ordinary least squares
key off: store data

key on: initialize
The problem with using standard (least squares) linear

regression techniques is that both summed current y

and difference between state-of-charge values x have ❨ ❉ ◗❜①
sensor noise or estimation noise associated with them
The least squares linear regression problem is a solution
to the equation .y y/ D Qx ; that is, there is noise
assumed on the measurements y , but not on the
independent variable x
A better way to think about the linear relationship

key off: store data
key on: initialize
Total-capacity-estimation problem is implicitly of form

balance
.y y/ D Q.x x/ since both the integrated

current and SOC estimates have noise ❨ ❉ ◗❜❳

That is, because estimates of SOC are generally
imperfect, there will be noise on the x variable, and using
standard least squares linear regression results in a
biased estimate of battery cell total capacity
Note that KF-based methods are (recursive) least
squares: they will tend to be biased by noise
What to do?
key off: store data

key on: initialize
Usual approach to counteract this problem is to try to ensure

that SOC estimates are as accurate as possible, then use

standard least-squares estimation anyway
For example, we might put constraints on how the capacity is estimated
Could force cell current to be zero before test begins and after the test ends (so
that cell is in an equilibrium state and SOC estimates are as accurate as possible)
Eliminates to a large extent (but not completely) the error in the x variable, and
makes the regression reasonably accurate
Still does not correctly handle the residual noise in x : minimizes the noise, but
never totally eliminates it
Solution: use “total least squares” instead of “(ordinary) least squares” estimation
Summary
key off: store data

key on: initialize
We can write an equation that involves measurable quantities

and is linear in Q
Tempting to use (ordinary least squares) regression techniques to estimate Q!
However, there are errors/noises in both the “x ” and “y ” terms in this regression
Ordinary-least-squares solutions will be biased by these noises
Instead, total-least-squares solutions should be used
That is what we study next
4.2.2: How to find the ordinary-least-squares solution as a benchmark
Derivation of weighted ordinary least squares
key off: store data

key on: initialize
b
Both ordinary least squares (OLS) and total least squares

b
(TLS), as applied to estimate total capacity, seek to find Q

such that y Qx using N -vectors of measured data x and y (N 1)
i th element .xi ; yi / corresponds to data collected from a cell over time interval i
xi is estimated change in state-of-charge over that interval
yi is net accumulated ampere hours passing through cell during that period
Specifically,
xi D ´k2;i ´k1;i for time interval i
yi D t
X
k2;i 1
kik
k Dk1;i
Major assumption of OLS
key off: store data

key on: initialize
OLS approach assumes that there is no error on xi and

balance
models data as y D Qx C y, where y is a vector of

measurement errors ❨ ❉ ◗❜①

Error bars on data points meant to illustrate
uncertainties, which are proportional to yi
We assume that y comprises zero-mean Gaussian
random variables, with known variances y2i (which are
not necessarily equal to each other)
OLS attempts to find an estimate Q of the true cell total b
capacity Q that minimizes sum of squared errors yi
Cost function
b key off: store data
key on: initialize
We generalize approach here slightly to allow for finding a Q

that minimizes sum of weighted squared errors
b
Weighting takes into account uncertainty of specific measurement
That is, we seek Q that minimizes weighted-least-squares (WLS) cost function
WLS2
D
X N
.yi Yi /2
D
X N
.yi b
Q xi / 2
y2i y2i
Db
i D1 i D1
where Yi is a point on the line Yi Qxi corresponding to measured data pair

.xi ; yi /, where yi is assumed to have noise but xi has no noise
Solution
key off: store data

key on: initialize
A number of solution approaches may be taken

b b
One that will serve us well is to differentiate cost function with

respect to Q and solve for Q by setting partial derivative to zero
X b N
b
@2WLS xi .yi Qxi /
D 2
yi2
D0
@Q
bX DX
i D1
N N
xi2 xi yi
Q
2
X
i D1 yi
,X 2
i D1 yi
bD Q
N
xi yi
2
N
xi2
2
i D1 yi i D1 yi
Summary
key off: store data

key on: initialize
Have now defined and solved weighted ordinary-least-squares

X
problem for estimating cell total capacity
Desired to find Q to minimize 2WLS D b N

.yi Yi /2 =y2i
i D1
Solution: Each time new data pair .xi ; yi / available, compute
c1 D
X N
xi2
and c2 D
X N
xi yi
2 2
b
i D1 yi i D1 yi
Then, Q D c2 =c1
4.2.3: Making the ordinary-least-squares solution computationally efficient
Recursive computation of weighted least squares

b
key off: store data

key on: initialize
WLS computes Q D c2 =c1 based on f.xi ; yi /gi D1:::N , where
X X
N N
xi2 xi yi
c1 D 2
and c2 D 2
i D1 yi i D1 yi
But, notice we can update summations recursively whenever new data pair available
c1;n D
X n
xi2
D c1;n 1 C xn2
and c2;n D
X n
xi yi
D c2;n 1 C xn2yn
2 y2n 2
b
i D1 yi i D1 yi yn
Then, we can write capacity estimate after n updates as Qn D c2;n =c1;n

Benefits of recursive approach:
Minimizes storage requirements
Evens out computational requirements when n gets large
Initializing recursion
key off: store data

key on: initialize
But, recursion requires initial estimates of c1;0 and c2;0

One approach is simply to set c1;0 D c2;0 D 0

Alternately, can recognize that cell having nominal capacity Qnom has that capacity
over a state-of-charge range of 1.0
Therefore, we can initialize with a synthetic zeroth “measurement” where x0 D 1
and y0 D Qnom
The value for y20 can be set to the manufacturing variance of the nominal capacity
That is, c1;0 D 1=y20 and c2;0 D Qnom =y20
Fading memory problem setup

key off: store data
key on: initialize
Can adapt method to allow fading memory of past data

Start by modifying WLS cost function to place more emphasis

on most recent measurements and less on measurements from distant past
We define the fading-memory weighted-least-squares
(FMWLS) cost function as
2FMWLS D
X N
N i .yi
b
Qxi /2
,
i D1
y2i
where the forgetting factor is in the range 0 1
Fading memory solution
key off: store data

key on: initialize
,X
Using same method as last lesson, can find

bDX
N N 2
N i xi yi N i xi
Q
i D1 i D1
y2i y2i
b
This solution may also easily be computed in a recursive manner
Compute Qn D cQ2;n =cQ1;n based on two running sums (initialized same as before)
cQ1;n D
X n
N i xi2
D cQ1;n 1 C x2n
2
y2i
X i D1 yn
n
N i xi yi
cQ2;n D y2i
D cQ2;n 1 C xn2yn
i D1 yn
Summary
key off: store data

key on: initialize
Have now derived WLS and FMWLS methods to estimate Q

In summary, both WLS and FMWLS have nice properties:

1. They give a closed-form solution for Q b
2. Only simple operations—multiplication, addition, and division—are required
3.
4.
The solutions can very easily be computed in a recursive manner
Fading memory can easily be added, allowing adaptation of Q to adjust for true b
cell total capacity changes
But, remember, that these solutions are biased
Will use as baseline against which to compare TLS-type solutions
4.2.5: Finding the solution to a weighted total-least-squares problem
Optimizing the WTLS cost function
key off: store data

key on: initialize
In the previous lesson, found WTLS cost function to optimize

bD0
b to minimize this cost function, set @2WTLS =@Q
To find Q
@2WTLS
N
X b xi
2.Q b yi x2
yi /.Q C xi y2 /
D i i
D0 (via Mathematica)
i D1
b
@Q b 2 x2
.Q i
C y2i /2
Unfortunately, this solution has none of the nice properties of the WLS solution:
1. There is no closed-form solution in the general case
2. There is no recursive update in the general case
3. There is no fading memory recursive update
1. No closed-form solution
key off: store data

key on: initialize
b
balance
A numerical method must be used to find Q

Optimization path using Newton−Raphson

2
Optimization path using Newton−Raphson
1
4000
x(2)
0 3000
Value
2000
−1 1000
0
Iterations: 36, Fn evals: 138 2 −2
−2 1 −1
−2 −1 0 1 2 0 0
−1 1
x(1) x(1) −2 2 x(2)
b
One possibility is to perform a Newton–Raphson search for Q
Newton–Raphson search
key off: store data
key on: initialize
Every time new .xi ; yi / available, perform several iterations of

. loop once each measurement interval while pack is active
bk D Q
Q bk 1 @2WTLS =@Qb b2
@2 2WTLS =@Q
Numerator is “Jacobian” of original cost function, given by earlier equation
Denominator is “Hessian” of original cost function, which can be shown to be
@2 2WTLS X N
y4 xi2 Cx4 .3Q
b 2 yi2 2Q
b 3 xi yi / x2 y2 .3Q
b 2 xi2 6Q
b xi yi C yi2/
b2
D2 i i
b 2 x2 C y2 /3
i i
@Q i D1 .Q i i
b , and is guaranteed to converge to

Search initialized with prior WLS estimate of Q
global solution since cost function 2WTLS is convex
Number of significant figures in the solution doubles with each iteration
Around four iterations produce double-precision results
2. No recursive update
key off: store data

key on: initialize
There is no recursive update in the general case: this has

storage implications and computational implications

To use WTLS, the entire vector x and y must be stored, which implies increasing
storage as the number of measurements N increases
Furthermore, the number of computations grows as N grows
Not well suited for an embedded-system application that must run in real time with
limited storage capabilities
3. No fading-memory recursive update
key off: store data

key on: initialize
No fading memory recursive update (no recursive update)

Non-recursive fading memory cost function may be defined

N
X b xi /2
.yi Q
2
FMWTLS D N i
i D1
b 2 x2
Q i
C y2 i
The Jacobian of this cost function can be found to be

@2 XN b xi yi /.Q
.Q b yi x2 C xi y2 /
FMWTLS
D2 N i i i
b
@Q i D1
b 2 x2
.Q i
C y2i /2
3. No fading-memory recursive update

key off: store data
key on: initialize
The Hessian of the cost function can be found to be

@2 2
N
X y4 xi2 C x4 .3Q
b 2 yi2 2Q
b 3 xi yi /
FMWTLS
D2 N i i i
b2
@Q i D1
b 2 x2
.Q C y2 /3
i
!
i
x2i y2i .3Qb 2 xi2 b xi yi C yi2 /

6Q
b 2 x2
.Q i
C y2 /3i
b
Can use Newton–Raphson search to minimize cost function to find Q
Summary
key off: store data

key on: initialize
WTLS has none of the nice properties of the WLS solution:

1. There is no closed-form solution in the general case

2. There is no recursive update in the general case
3. There is no fading memory recursive update
Can use Newton–Raphson search, but requires growing memory and computation
Next week, you will learn special case of WTLS that gives a closed-form solution,
with recursive update, and fading memory
Leads to approximate solution to general WTLS that also has these nice properties
But first, we consider an important property of both the WLS and WTLS solutions
4.2.6: Confidence intervals on least-squares solutions
Confidence limits on estimated total capacity
key off: store data

key on: initialize
When computing Q
b , important to know certainty of estimate temperature charge (SOC) health (SOH) cells limits
Would like to know variance 2 to compute confidence

Q
b
intervals within which the true cell total capacity Q lies, with high certainty
For example, three-sigma bounds: Q 2 .Q b 3b
Q
b C 3 /
;Q Q
b
To derive confidence limits, must re-cast least-squares type problem as maximum-
likelihood optimization problem
Then, Cramer–Rao theorem will give us confidence bounds
Recasting WLS as maximum-likelihood
key off: store data

key on: initialize
First, recast problem as maximum-likelihood optimization

Assuming all errors are Gaussian, this is straightforward

If we form a vector y comprising elements yi , and a vector x comprising
corresponding elements xi and a diagonal matrix y having corresponding
diagonal elements y2i , then minimizing 2WLS is equivalent to maximizing

1 1
MLWLS D .2/N=2j j1=2 exp 2
b x/T 1 .y
.y Q y Q
b x/
y

1 1 2
D .2/N=2j j1=2 exp
2 WLS
,
y
which is a maximum-likelihood problem

Constant multiplying exponential causes MLWLS to integrate to 1, yielding valid pdf
Recasting WTLS as maximum-likelihood

key off: store data
key on: initialize
If we form vector d concatenating y and x, and vector b

d temperature charge (SOC) health (SOH) cells limits
concatenating corresponding elements Yi and Xi , and

diagonal matrix d having diagonal elements y2i followed by x2i , then minimizing
2WTLS is equivalent to maximizing

1 1
MLWTLS D .2/ jd j1=2
N
exp
2
d/T d 1 .d
.d b d/
b

1 1 2
D .2/N jd j1=2
exp
2 WTLS
The maximum-likelihood formulation makes it possible to determine confidence
intervals on Q
b. . .
Application of Cramer–Rao theorem
key off: store data

key on: initialize
Cramer–Rao theorem, a tight lower bound on variance of Q

b is temperature charge (SOC) health (SOH) cells limits
given by the negative inverse of second derivative of argument

of the exponential function, evaluated at the Q
b that optimizes the cost function
That is, for the cost functions we have investigated so far,

1
@2 2WLS
2
Q
b
2 @Q2
for WLS
QDb
Q
1
@2 2WTLS
2
Q
b
2 @Q2
for WTLS
QDb
Q
Finding the Hessians
key off: store data

key on: initialize
The second partial derivatives (i.e., Hessians) of WTLS and

FMWTLS cost functions have already been described in

the context of a Newton–Raphson iteration
For WLS and FMWLS, the situation is easier:
N N
@2 2WLS X xi2 @2 2FMWLS X xi2
@Q2
D2 2
and
@Q2
D2 N i
y2i
i D1 i D1 yi
These may be computed using the previously defined recursive parameters as
@2 2WLS @2 2FMWLS
@Q2
D 2c1;n and
@Q2
D 2cQ1;n
Summary
key off: store data
key on: initialize
Desire to find 2 to compute confidence intervals on Q

b temperature charge (SOC) health (SOH) cells limits
Q
b loop once each measurement interval while pack is active
For example, three-sigma bounds: Q 2 .Q

b 3b
Q
;Q
b C 3bQ /
Recasting xLS problem as ML problem allows invoking Cramer–Rao theorem,
1
@2 2xLS
2
Q
b
2 @Q2
QDb
Q
Hessians for WTLS/FMWTLS were already known
Hessians for WLS/FMWLS trivial to compute using already-existing quantities
key off: store data

key on: initialize
This week, you learned that total-capacity relationship has a

linear form with respect to a change in SOC xi between two

operating points and net ampere hours passed through cell over that interval, yi
However, since both xi and yi have uncertainty, standard least-squares linear
regression tools are not correct to apply to problem
Derived LS and WLS solutions as baseline to demonstrate the problems
Also derived WTLS, which will give better solutions
And, showed how to compute error bounds using Hessian of cost function
Where from here?
key off: store data

key on: initialize
While WLS allows closed-form, recursive, and temperature charge (SOC) health (SOH) cells limits
fading-memory solutions, WTLS does not

Since WTLS cannot be computed efficiently in general case,
we seek specific scenarios where it can
Then, we approximate those specific scenarios to give a
near-optimal WTLS approximate method for estimating total
capacity
Credits
key off: store data
key on: initialize


“Climbing upward” on slide 2: Pixabay license

https://pixabay.com/en/career-head-rise-profession-chance-1019755/
4.3.1: Simplifying the total-least-squares solution for cases having proportional uncertainties
A path to solution
key off: store data

key on: initialize
General WTLS solution provides excellent results, but is

impractical to implement in an embedded system

Therefore, we search for cases that lead to simpler implementations
Here, we look at an exact solution when the uncertainties on the xi and yi data
points are proportional to each other for all i , which leads to a simple solution that
can easily be implemented in an embedded system
With insights from this solution we will next look at an approximate WTLS solution
that also has nice implementation properties
Dr. Gregory L. Plett j University of Colorado Colorado Springs Battery State-of-Health (SOH) Estimation | Simplified total-least-squares battery-cell capacity estimates j 1 of 6
Proportional confidence on xi and yi
key off: store data

key on: initialize
D kyi , then WTLS cost function reduces to
If xi
generalization of standard TLS cost function

Substitute xi D kyi into 2WTLS and associated results to get:
X
D .x
N
Xi /2
C .yi Yi /2
D
X N
.yi Qxi /2 b
b
2 i
TLS
i D1 .Q k C 1/yi
i D1
k 2 y2i y2i 2 2 2
Furthermore, Jacobian of WTLS cost function reduces to (again, via xi D kyi )
@2TLS Xb N
D 2 .Q x
b
yi /.Qk 2 yi C xi /
b b
i
@Q i D1 .Q2 k 2 C 1/2 y2i
b , without requiring iteration to do so
This may be solved for an exact solution to Q
Finding quadratic form

key off: store data
key on: initialize
We first collect terms

balance
Xb b
D 2 .Qxi 2yi2/.Qk 2yi 2C xi /

N loop once each measurement interval while pack is active
@2TLS 2
b b D0
@Q i D1 .Q k C 1/ yi
X k x y CQb X x k y C X x y D 0
b2
DQ
N
2 i i
N 2
i
2 2
i
N
i i
2 2 2
D D DD D D
i 1 yi i 1 yi i 1 yi
P D y = a k 2 c2;n b c1;n k 2 c3;n c c2;n
where we define c D
n 2 2
3;n
b using the familiar quadratic equation solution
i 1 i yi
Then, we can solve for Q
p
bD
Q
b b
2a
4ac 2
Routh test to choose root of solution
key off: store data

key on: initialize
In terms of familiar recursive quantities, specific solution is

q
k 2 c3;n /2 C
b .c1;n k2c 2
4k 2 c2;n
3;n / .c1;n
Qn D
2k 2 c2;n
Which of the two roots to choose? Can show via Routh array and test that this
b
quadratic always has one positive root and one negative root
Q2 k 2 c2;n c2;n
b
Q1 c1;n k 2 c3;n 0
b
Q0 c2;n 0
First column of the array always has exactly one sign change, so there is one root of
the polynomial in the right-half plane (and other in left-half plane or on axis)
Total capacity must be nonnegative!
key off: store data

key on: initialize
So there is one root of the polynomial in the right-half plane

(and other in left-half plane or on axis)

By the fundamental theorem of algebra, because the coefficients c1;n , c2;n , and c3;n
are real, the polynomial roots must either both be real or be complex conjugates
The fact that they are in different halves of the complex plane shows that they
cannot be complex conjugates, and therefore must both be real
We choose larger root of the quadratic equation, which corresponds to positive root
q
k 2 c3;n / C k 2 c3;n /2 C 4k 2 c2;n
b .c1;n .c1;n 2
Q n D 2k 2 c2;n
Summary
key off: store data
key on: initialize
You have already learned that WLS is biased but general

WTLS is computationally inefficient

But, if uncertainties on xi and yi are proportional for all i , then WTLS has simple
quadratic form that can be written in terms of recursively computed quantities
Notice that we refer to this specific condition as the TLS problem, as distinct from
the earlier more-general WTLS problem
Next lesson will summarize the TLS solution in more specific detail
4.3.2: Making simplified solution computationally efficient
Defining recursive quantities
key off: store data

key on: initialize
Solution to proportional-confidence TLS found to be


q
c1;n C k2c 3;n C .c1;n k 2 c3;n /2 C 2
4k 2 c2;n
Q
b
n D 2k 2 c2;n
; where
D c1;n 1 C xn2=y2 I c2;n D c2;n 1 C xnyn=y2 I c3;n D c3;n 1 C yn2=y2

c1;n i i i
Initialized by setting x0 D 1, y0 D Qnom , y2 to a value representing the uncertainty i

of total capacity with respect to nominal capacity
Therefore, c1;0 D 1=y2i , c2;0 D Qnom =y2i and c3;0 D Qnom
2 = 2
y i
Confidence interval of estimate
key off: store data

key on: initialize
The Hessian, required to compute estimate’s uncertainty,

may also be found in terms of the recursive parameters:

b 3 C 6.k 4 c b 2 C 12c k 2 Q
@2 2TLS 4k 4 c2 /Q c1 k 2 /Q 2 b C 2.c1 k 2 c3 /
b2
D. 3
b 2 k 2 C 1/3
@Q .Q
q
One-sigma bounds on Q
b are computed as 2=.@2 2TLS =@Q
b 2/
QDb
Q
Fading memory
key off: store data
key on: initialize
Fading memory may be easily incorporated recursively via


q
cQ1;n C k2c Q3;n C .cQ1;n k 2 cQ3;n /2 C 2
4k 2 c2;n Q
Q
b
n D 2k 2 cQ2;n
; where
cQ1;n D cQ1;n 1 C xn2=y2 I cQ2;n D cQ2;n 1 C xnyn=y2 I cQ3;n D cQ3;n 1 C yn2=y2

i i i
Initialization unchanged: cQ1;0 D 1=y2 , cQ2;0 D Qnom =y2 and cQ3;0 D Qnom2 = 2
i y i i
Can obtain Hessian in terms of recursive parameters cQ1 through cQ3

@2 2FMTLS . 4k 4 cQ2 /Q
b 3 C 6.k 4 cQ cQ1 k 2 /Q
b 2 C 12cQ k 2 Q
2 b C 2.cQ1 k 2 cQ3 /
b2
D 3
b 2 k 2 C 1/3
@Q .Q
Summary
key off: store data

key on: initialize
This TLS solution shares the nice properties of WLS solution:

1. It gives a closed-form solution for Q

b
No iteration or advanced algorithms required—only simple math operations

2. The solution can be very easily computed in a recursive manner
We keep track of the three running sums c1;n , c2;n and c3;n
When an additional data point becomes available, we update the sums and
compute an updated total capacity estimate
3. Fading memory is easily added.
Unfortunately, this solution does not allow x2 and y2 to be arbitrary—they must be
proportionally related by the scaling factor xi D kyi
i i
Next, we seek to approximate TLS to allow an arbitrary relationship
4.3.3: Defining geometry for approximate full solution to weighted total least squares
Desire a simple WTLS solution
key off: store data

key on: initialize
Know that WLS solution is biased, but WTLS demands too

much computation to use in practical BMS

Have seen proportional-uncertainty version of WTLS—which we call TLS—that is
feasible to implement: but, uncertainties are not necessarily proportional in practice
We desire solution that approximates WTLS problem but allows x2i and y2i to be
nonproportional, but which yields a recursive solution for feasible implementation in
an embedded system
In this lesson, you will learn an approach we can take, based on geometry of WTLS
solution
WTLS geometry (1/2)
key off: store data

key on: initialize
✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗
meas. voltage
current
estimate
state of
estimate
state of
balance compute
power
✭❳ ✐ ✱ ❨✐ ✮ ①✐
✧②✐
❘✐
②✐
★ ✾✁✂ ✄ ★
✭① ✐ ✱ ②✐ ✮
✧① ✐
Figure shows WTLS geometry, motivates approximate solution to be developed

Left frame shows relationship between data point .xi ; yi / and its optimized map
b Xi when x2 and y2 are arbitrary
.Xi ; Yi / on Yi D Q i i
The error bars on each data point illustrate the uncertainties in each dimension,
which are proportional to x and y i i
WTLS geometry (2/2)

key off: store data
key on: initialize
✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗
meas. voltage
current
estimate
state of
estimate
state of
balance compute
power
✭❳ ✐ ✱ ❨✐ ✮ ①✐
✧②✐
❘✐
②✐
★ ✾✁✂ ✄ ★
✭① ✐ ✱ ②✐ ✮
✧① ✐
We see that the distance between xi and Xi is not necessarily equal to the
distance between yi and Yi —depends on respective error bounds
If quality of xi is better (poorer) than quality of yi , distance to its map Xi should be
shorter (greater) than the distance from yi to its map Yi
TLS geometry
key off: store data

key on: initialize
✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗
meas. voltage
current
estimate
state of
estimate
state of
balance compute
power
✭❳ ✐ ✱ ❨✐ ✮ ①✐
✧②✐
❘✐
②✐
★ ✾✁✂ ✄ ★
✭① ✐ ✱ ②✐ ✮
✧① ✐
Middle frame shows relationship between data point .xi ; yi / and its optimized map
b Xi when x2 and y2 are equal
.Xi ; Yi / on Yi D Q i i
Distance between xi and Xi is equal to distance between yi and Yi , and line

b Xi
joining data point .xi ; yi / and .Xi ; Yi / is perpendicular to the line Yi D Q
If x and y are unequal but proportional, x - or y -axis may be scaled to yield
i i
transformed data points with equal variances, and hence same idea applies
Approximated geometry (1/2)
key off: store data

key on: initialize
✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗
meas. voltage
current
estimate
state of
estimate
state of
balance compute
power
✭❳ ✐ ✱ ❨✐ ✮ ①✐
✧②✐
❘✐
②✐
★ ✾✁✂ ✄ ★
✭① ✐ ✱ ②✐ ✮
✧① ✐
Right frame illustrates definitions that will be used to derive an approximate

weighted total least squares (AWTLS) solution
Motivated by TLS, we enforce that line joining data point .xi ; yi / and .Xi ; Yi / be
b Xi —will result in a recursive solution
perpendicular to line Yi D Q
Approximated geometry (2/2)

key off: store data
key on: initialize
✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗ ✦❳
❨ ❂ ◗
meas. voltage
current
estimate
state of
estimate
state of
balance compute
power
✭❳ ✐ ✱ ❨✐ ✮ ①✐
✧②✐
❘✐
②✐
★ ✾✁✂ ✄ ★
✭① ✐ ✱ ②✐ ✮
✧① ✐
However, as with the WTLS solution, we weight distance between xi and Xi

differently from distance between yi and Yi in optimization cost function
This will give a better total capacity estimate than TLS when the uncertainties on xi
and yi are not proportional
Summary
key off: store data

key on: initialize
D Qb X
WTLS solution maps .xi ; yi / to .Xi ; Yi / on Y

with nonperpendicular line

Optimal but not practical to implement
TLS maps with perpendicular line
Optimal only for proportional xi and yi but practical
Will use observation of orthogonality to propose sub-
optimal mapping .xi ; yi / to .Xi ; Yi / that is perpendicular,
but also weights uncertainties in xi and yi
Will continue to use geometry from figure to right as we
proceed beyond this point
4.3.4: Finding appropriate cost function for approximate full solution to WTLS problem
Definitions arising from new geometry
key off: store data

key on: initialize
Recall the figure to the right, which will define the

balance
geometry of our approximate WTLS solution

Define xi be x -distance between data point i and line,

and yi be y -distance between data point i and line
b D yi =xi for all i
Slope of line is Q
b
Angle of line is D tan 1 Q
Shortest distance between line and any given data point:
Ri D yi cos
p q
b2
D yi = 1 C tan D yi = 1 C Q
2
First form of cost function
key off: store data

key on: initialize
Let xi D Ri sin and yi D Ri cos
These are the x - and y -components of the perpendicular

distance between data point i and the fitting line

We then weigh our fitting cost function according to
these variances
Therefore, we define the approximate weighted total
least squares (AWTLS) cost function as
X
N
xi2 2
2AWTLS D x2i
C y
2
i
i D1 yi
More useful form of cost function

key off: store data
key on: initialize
b 2 =.1 C Q
b 2 /:
Note that sin D 1 DQ

balance
2
cos2
! ! ! ! loop once each measurement interval while pack is active
y 2 b
Q 2 y 2 1
xi2 D i
I yi2 D i
1CQ b 2 1CQb 2 b
1CQ2 1CQb2
Since yi D yi Q b xi
!
X N
.y b
Q x /2 Q b 2 1
2AWTLS D
b 2 2 x2 C y2
i i
i D1 .1 C Q / i i
To verify that AWTLS approximates WTLS in some

cases, note both cost functions equal when xi D yi
However, they are not equal when x i
D kyi , but this
will be corrected in a later lesson
Summary
key off: store data

key on: initialize
Starting with geometry presented in last lesson, defined some

quantities describing line and distances from data point to line

in different directions
Proposed new cost function in terms of these new definitions
Then, rewrote cost function in terms of quantities available to us
We are now ready to derive the solution that optimizes this cost function
4.3.5: Finding solution to the AWTLS problem
Approximate full solution: Derivation
key off: store data

key on: initialize
In prior lesson, you learned new AWTLS cost function

!
N
X b xi b2
.yi Q /2 Q
2AWTLS D C 12
i D1 .1 C Q
b 2 /2 x2i yi
Jacobian of AWTLS cost function can be found to be
N
! !
@2AWTLS 2 X x y 2x 2 x 2 y 2
D b2 b
Q 4 i i
CQ
b 3 i i i
b
@Q .Q C 1/3 i D1 x2i x2i y2i x2i
! ! !
3x y 3x y x 2 2y 2 y 2 x y
CQb2 i i
y2i
i i
x2i
CQ b i
y2i
i
C 2i C 2i i
xi yi
Recursive solution
key off: store data

key on: initialize
Can be rewritten in terms of recursive summations

balance
@2AWTLS 2 loop once each measurement interval while pack is active
b
D b 2 C 1/3 5
c Qb 4 C .2c4 c1 c6 /Qb3
@Q .Q

C .3c2 3c5/Q b 2 C .c1 2c3 C c6 /Q
b c2
where
c1;n D c1;n 1 C xn2=y2n I c3;n D c3;n 1 C yn2=y2n I c5;n D c5;n 1 C xnyn=x2n
c2;n D c2;n 1 C xn yn =yn I
2
c4;n D c4;n 1 C xn =xn I
2 2
c6;n D c6;n 1 C yn =xn
2 2
b
Roots of the quartic equation, below, are candidate solutions for Q
b 4 C .2c4 c1 c6 /Q
c5 Q b 3 C .3c2 3c5 /Q b 2 C .c1 2c3 C c6 /Q b c2 D 0
Initializing recursive solution

key off: store data
key on: initialize
Initialized by setting
x0 D 1 and y0 D Qnom
y20 D uncertainty in Qnom versus Q
We assume that x0 D y0 to be compatible with TLS for k D 1 (will generalize for
different k later in code)
Therefore,
c1;0 D 1=y20 I c3;0 D Qnom
2
=y20 I c5;0 D Qnom =y20
c2;0 D Qnom =y20 I c4;0 D 1=y20 I c6;0 D Qnom
2
=y20
Selecting solution
key off: store data

key on: initialize
Any root the quartic equation found earlier is a possible

b that optimizes the cost function

solution for the Q

Roots may be found using closed-form Ferrari method
Roots may also be found via eigenvalues of a “companion matrix”
However, of the four roots only one is optimal, and no method to decide a priori
which to solve for
In my experience, with some sets of data all roots are real, but with other sets of
data some can be complex, and some can be negative
Only foolproof method to determine optimizing root is to evaluate 2AWTLS at each of
the four candidate solutions, retain the one that gives lowest value (skip for negative
and complex roots)
Efficiently computing cost function, Hessian
key off: store data

key on: initialize
Computing the cost function may be very readily done if we

rewrite it in terms of the running summations

1
2AWTLS D b4
c4 Q b 3 C .c1 C c6 /Q
2c5 Q b2 b C c3
2c2 Q
b 2 C 1/2
.Q
When the assumptions made in deriving AWTLS are approximately true, the
Hessian yields a good value for the error bounds on the total capacity estimate
After some straightforward but messy mathematics, we can find the Hessian to be

@2 2AWTLS 2
@Q2
D .Q2C1/4 2c5 Q5 C.3c1 6c4 C3c6 /Q4 C . 12c2 C16c5 /Q3

C. 8c1 C10c3 C6c4 8c6 /Q 2
C .12c2 6c5 /Q C .c1 2c3 Cc6 /
Summary
key off: store data
key on: initialize
Have now found solution to AWTLS cost function

b is root of a quartic equation

Optimizing Q
Can be found via Ferrari method or eigenvalue method
Must substitute solutions into cost function to choose best one
Made efficient via recursive summations
Hessian can also be computed using recursive sums to provide confidence bounds
4.3.6: Adding fading memory
Approximate full solution: Derivation
key off: store data

key on: initialize
Fading memory can be added with modified cost function

!
N
X b xi b2
N i .yi Q /2 Q
2FMAWTLS D C 12
i D1 .1 C Q
b 2 /2 x2i yi
The Jacobian is
N
" ! !
@2FMAWTLS 2 X xi yi 2xi2 xi2 yi2
b
D b 2 C 1/3 i D1
Q CQ
N i b4
x2i
b3
x2i y2i x2i
@Q .Q
! ! !#
3x y 3x y 2 2 yi2
CQ
b2 i i i i
C b xi 2yi
Q C C xi yi
y2i x2i y2i x2i y2i
Recursive solution
key off: store data

key on: initialize
Jacobian can be rewritten in terms of recursive sums

balance
@2FMAWTLS 2 loop once each measurement interval while pack is active
b
D b 2 C 1/3 5
cQ Q
b 4 C . cQ1 C 2cQ4 cQ6 /Q
b3
@Q .Q

C .3cQ2 3cQ5 /Q
b 2 C .cQ1 2cQ3 C cQ6 /Q
b cQ2
where
cQ1;n D cQ1;n 1 C xn2=y2n I cQ3;n D cQ3;n 1 C yn2=y2n I cQ5;n D cQ5;n 1 C xnyn=x2n
cQ2;n D cQ2;n 1 C xn yn =y2n I cQ4;n D cQ4;n 1 C xn2 =x2n I cQ6;n D cQ6;n 1 C yn2 =x2n
b
Roots of the quartic equation, below, are candidate solutions for Q
cQ5 Q
b 4 C .2cQ4 cQ1 cQ6 /Q
b 3 C .3cQ2 3cQ5 /Q
b 2 C .cQ1 2cQ3 C cQ6 /Q
b cQ2 D0
Initializing recursive solution

key off: store data
key on: initialize
Initialized by setting
x0 D 1 and y0 D Qnom
y20 D uncertainty in Qnom versus Q
and x20 D y20
Therefore
cQ1;0 D 1=y20 I cQ3;0 D Qnom
2
=y20 I cQ5;0 D Qnom=y20
cQ2;0 D Qnom =y20 I cQ4;0 D 1=y20 I cQ6;0 D Qnom
2
=y20
Determining which solution to choose
key off: store data

key on: initialize
Again, can be solved using Ferrari method, for example; the

four candidate solutions must be checked against the cost

function to determine which is optimal
The cost function in terms of these variables is
1
2FMAWTLS D b 2 C 1/2
cQ4 Q
b4 2cQ5 Q
b 3 C .cQ1 C cQ6 /Q
b2 2cQ2 Q
b C cQ3
.Q
The Hessian is

@2 2FMAWTLS 2
b2
D b 2 C1/4
2cQ5 Q
b 5 C .3cQ1 6cQ4 C3cQ6 /Q
b 4 C . 12cQ2 C16cQ5 /Q
b3
@Q .Q

C . 8cQ1C10cQ3C6cQ4 8cQ6/Q C .12cQ2 6cQ5/Q
b C .cQ1 2cQ3 CQc6 /
b2
Scaling for better approximation
key off: store data

key on: initialize
AWTLS and WTLS cost functions not equal when xi D kyi
This can be remedied easily: define yQi D kyi ; then yQi D xi
b and
Invoke the AWTLS or FMAWTLS methods to find total capacity estimate Q
Hessian H using input sequences comprised of the original x vector and the scaled
yQ vector (i.e., .xi ; yQi / with corresponding variances .x2i ; k 2 y2i /)
True slope estimate found as Qb corrected D Q=k
b , and Hessian as Hcorrected D H=k 2
This is the method used next week, where k is estimated as k D x1 =y1
This scaling improves results even when yi and xi not proportionally related, if k
chosen to give “order of magnitude” ratio between uncertainties of xi and yi
Summary
key off: store data
key on: initialize
AWTLS solutions share the nice properties of WLS solution:

b : No iteration or advanced
1. Closed-form solution for Q
algorithms are required—only simple mathematical operations.
2. Solution can be computed very easily in a recursive manner: keep track of six
running sums c1;n through c6;n ; when additional data become available, update
sums and compute updated total-capacity estimate
3. Fading memory can be added easily to allow Q b to place greater emphasis on
more recent measurements than on earlier measurements, allowing adaptation
b to adjust for true cell total-capacity changes
of Q
4. Furthermore, this method is superior to the TLS solution since it allows
individual weighting on xi and yi data points
key off: store data

key on: initialize
WTLS (last week) gave great results but was computationally

ill-conditioned for implementation in an embedded BMS

This week, started by looking at TLS with proportional confidence on xi and yi
Had simple closed-form recursive solution, fading memory
Generalized TLS to AWTLS
Does not optimize same cost function as WTLS, so not optimal in the same sense
But, has closed-form solution (Ferrari method), recursive and fading memory
Where from here?
key off: store data

key on: initialize
Next week, you will learn how to implement all the xLS
balance
methods in Octave code

Code has main xLSalgos.m component that does the

computations
Also has “driver” code to execute xLSalgos.m with
different input conditions
You will see how the different methods perform under three
different HEV and BEV operating scenarios
Examples will also show the importance of a good SOC
estimator
Credits
key off: store data
key on: initialize


“Blank slate” on slide 2: Pixabay license

https://pixabay.com/en/white-male-3d-model-isolated-3d-1992593/

4.state of Health Estimation

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

4.state of Health Estimation

Uploaded by

Copyright:

Available Formats

4.1.1: Welcome to the course!

Welcome to the course!

key off: store data

Welcome to Battery State of Health (SOH) Estimation!

loop once each measurement interval while pack is active

This course is the fourth in a specialization that investigates

key off: store data

meas. voltage estimate estimate balance compute

loop once each measurement interval while pack is active

4.1.1: Welcome to the course!

What topics will we study in this course?

key off: store data

In this course, you will learn:

cells degrade over time, motivating a need to be able to adjust

4.1.1: Welcome to the course!

What skills will you gain in this course?

meas. voltage estimate estimate balance compute

After completing the course, you’ll be able to:

in lithium-ion cells and understand how they work

Prerequisites; for further study

key off: store data

Prerequisites: loop once each measurement interval while pack is active

Course 1: Intro. to Battery Management Systems

Performance changes as cells age

key off: store data

We know that a BMS must estimate certain quantities

loop once each measurement interval while pack is active

indicative of its cells’ states and parameters

4.1.2: What changes as a cell ages?

Loss in total capacity versus time

key off: store data

As a battery cell ages, its total capacity decreases, primarily

loop once each measurement interval while pack is active

due to unwanted side reactions, structural deterioration

4.1.2: What changes as a cell ages?

meas. voltage estimate estimate balance compute

Phenomena often referred to as capacity fade

loop once each measurement interval while pack is active

Knowing present value of total capacity

key off: store data

As a cell ages, its series resistance increases, also primarily

loop once each measurement interval while pack is active

due to unwanted side reactions and structural deterioration

4.1.2: What changes as a cell ages?

Other cell parameter values

key off: store data

The OCV relationship changes as the electrodes lose capacity

loop once each measurement interval while pack is active

For most chemistries in common use, unlikely that this

ESR, total capacity have dominant impact 3.6

KF-based methods in this course could be 3.4 0% capacity loss

4.1.2: What changes as a cell ages?

Summary and next steps

meas. voltage estimate estimate balance compute

Cell parameter values change over time as cell ages

Total capacity and ESR are the most important to track

Lithium-ion aging: Negative electrode

key off: store data

In the next lessons, we’ll seek to describe aging qualitatively.1

In negative electrode, aging observed at three scales:

At the surface of the electrode particles

4.1.3: Negative-electrode aging processes at particle surface

Negative electrode aging at surface of particles