Engineering Honors Thesis Fatigue Analysis of a Gallop ...
Transcript of Engineering Honors Thesis Fatigue Analysis of a Gallop ...
1
Engineering Honors Thesis
β
Fatigue Analysis of a Gallop-based
Piezoelectric Wind Energy Harvester and
Strategies for Long-Term Performance
Optimization
December 3, 2021
By Andrew Trostle
2
Abstract
With great demand for novel forms of renewable energy, there has been substantial interest
in wind energy harvesting based on aeroelastic excitations with the aid of piezoelectric
materials. While most research has been dedicated to optimizing power output [17],
relatively little research has been done on the deviceβs behavior in response to fatigue,
which may be a key failure mode given its persistent cyclic loading. A Python program was
created to simulate a galloping piezoelectric wind energy harvesting (GPEH) under varying
wind conditions in real time, monitoring performance metrics until fatigue failure. This can
generate Monte Carlo distributions for optimizing GPEH system parameters for durability.
This could then determine if such a form of renewable energy is currently feasible and if so,
provide a database of effective materials for the deviceβs implementation. In tandem with
this, analytical optimization methods were tested, but remain partially inconclusive,
requiring better dimensional analysis to simplify the myriad of system parameters.
However, initial runs of the code still provided useful data, identifying an optimal
performance state of all those tested, indicating that the tradeoff between performance and
durability is not either-or.
3
Table of Contents
1.0 Introduction 4
2.0 GPEH Mechanics and Theoretical Background 4
2.1 GPEH in Previous Literature 4
2.1.1 Aerodynamic Model 5
2.1.2 Structural Model 6
2.1.3 Results 7
2.2 Fatigue Analysis 9
2.2.1 Intro 9
2.2.2 Minerβs Rule 10
2.2.3 Crack Growth Equations 11
2.2.4 Additional Fatigue Effects: Crack-Induced Compliance 12
3.0 Numerical Simulation Setup 13
3.1 Assumptions and Default Model Features 13
3.2 Probabilistic Wind Model 15
3.3 Workflow Algorithm and Code Features 16
3.3.1 General Layout 16
3.3.2 Newmark-Beta Method 19
3.3.3 Adaptive Time-Stepping 20
3.3.4 Fracture Toughness Modeling 20
4.0 Numerical Simulation Results 21
4.1 Qualitative Results 22
4.2 Quantitative Results 23
5.0 GPEH Analysis and Optimization Strategy 27
5.1 Analytical Model for GPEH Performance 27
5.2 Optimization Strategy 29
5.2.1 Effect of Material Choice on Efficiency and Durability 29
5.2.1.1 Piezo-ceramics versus Piezo-polymers 29
5.2.1.2 Performance Optimization of a Piezo-Ceramic-Polymer 30
Composite
5.2.2 Methods of Analytically Correlating Design Parameters and
Performance Variables 32
5.2.2.1 Probability-Based Model 32
5.2.2.2 βLifetime Efficiencyβ-Based Model 37
6.0 Conclusion 41
7.0 Bibliography 42
4
1.0 Introduction
An emerging design concept in the field of small-scale renewable energy systems is the
Galloping Piezoelectric Wind Energy Harvester (GPEH). This design is based on two
phenomena: the aeroelastic phenomena of gallop, a circumstance in which lift dynamics
result in a net negative damping of the lifting body subjected to a quasi-steady fluid flow,
and the piezoelectric effect, exhibited in certain crystalline materials, where vibratory
oscillations in the material can result in an alternating current. Many researchers have
explored various configurations of the basic GPEH mechanic, resulting in gradual
improvements in system performance, achieving a maximum power of 53 mW [17].
With prior research proving that the performance of such a device can be optimized enough
to power low-power sensors and similar instruments, it can now be determined whether
such a system is feasible for long-term use. Sirohi & Mahadik [17] note visible fatigue
damage to their galloping device after testing, implying rapid degradation which must be
addressed. In this paper, we intend to test the fatigue life of a basic GPEH device subjected
to a realistically varying wind signal through numerical simulation and experimental
testing, as well test the effects of material properties on performance and endurance.
Repeated simulation with varying system parameters will lead to a brute-force optimization
of lifetime energy output for this highly nonlinear system.
2.0 GPEH Mechanics and Theoretical Background
2.1 Previous Literature
This analysis will be largely based on a model presented by Sirohi & Mahadik [17]. They
use a triangular prism tip body configured vertically, simply supported downstream by two
thin aluminum beams with a clamped support. On the clamped end, two strips of
piezoelectric material are attached to the top and bottom of each aluminum beam. The
device covers a total area of 16 x 25 cm. This model was simulated with numerical
integration in Matlab and then verified experimentally in a wind tunnel.
5
Figure 2.1.0: GPEH Diagram [17]
2.1.1 Aerodynamic Model [17]
The numerical method begins with the calculation of aerodynamic forces. For a prism of
chord length π and prism length πΏ, a windspeed πβ will result in a net vertical force πΉπ¦
equal to:
πΉπ¦ = πΏπππ πΌ + π·π πππΌ
Where:
πΏ =1
2πππ
2ππΏπΆβ(πΌ), π· =1
2πππ
2ππΏπΆπ(πΌ)
and
πΌ = πΌ0 β tanβ1 (οΏ½ΜοΏ½π‘ππ
πβ) + π€π‘ππ
β²
ππ 2 = πβ
2 + οΏ½ΜοΏ½π‘ππ2
6
π€ here denotes the beam displacement and πΌ0 is the default angle of attack. For a triangular
cross-section in this orientation with the oncoming wind gust, πΌ0 = 60β. Section lift and
drag coefficients are a function of angle of attack, here based on a lookup table provided in
[1].
Figure 2.1.1.1: Section Lift and Drag Coefficients for an Equilateral Triangle Cross-section
Additionally, Sirohi & Mahadik calculated the Den Hartog Criterion based on the above
coefficients. This criterion determines whether the aerodynamic state will result in a
negative damping, also known as transverse galloping. The criterion is defined as:
π»(πΌ) =ππΆβ
ππΌ+ πΆπ < 0
For a triangular cross-section the following angle-of-attack ranges are susceptible to
galloping:
7
Figure 2.1.1.2: Den Hartog Criterion for an Equilateral Triangle Cross-section [17]
2.1.2 Structural Model
With the aerodynamic model functioning, the structural model can now be constructed. For
this, Sirohi & Mahadik [17] used the Rayleigh-Ritz method, an approximate energy method
where mode shapes for the beam displacement are assumed and the weighted sum of these
modes is solved numerically. Beam deflection can be approximated as:
π€(π₯, π‘) = π1(π₯)πΎ1(π‘) + π2(π₯)πΎ2(π‘), π€βπππ π1(π₯) =π₯
πΏππππ, π2(π₯) = (
π₯
πΏππππ)3
π represents each shape function and πΎ represents the βweightβ of each mode on
displacement over time. The mass and stiffness matrices of this system of ordinary
differential equations are of the form:
πππ = β« ππ΄ππ(π₯)ππ(π₯)ππ₯πΏππππ
0
πΎππ = β« πΈπΌππβ²β²(π₯)ππ
β²β²(π₯)ππ₯πΏππππ
0
The paper goes into further detail on including electric displacements of the piezoelectric
material into this system of ordinary differential equations, which includes a coupling
matrix representing the resistance to stress introduced by the induced voltage field.
However, this project will be utilizing the coupled electromechanical model from [4].
The numerical method works by providing the beam tip an initial velocity, calculating the
vertical force generated at each moment in time, applying that load to the differential
8
equation, updating the beam displacement, and repeating this step in time steps via a
numerical integration method.
For our code simulation, two mode shapes are used which both satisfy the boundary
conditions:
π€(0) = 0, π€β²(0) = 0, π€β²β²(πΏ) = 0
πβ²β²1(π₯) = (1 βπ₯
πΏ) , π2
β²β²(π₯) = (1 βπ₯
πΏ)2
Note π€β²β²(0) = π1β²β²(0)πΎ1 + π2
β²β²(0)πΎ2 = πΎ1 + πΎ2. This yields the stiffness and mass matrices:
[π] = ππ΄πΏ5 [
11
420
173
20160173
20160
13
810
] , [πΎ] = πΈπΌπΏ [1/3 1/41/4 1/5
]
2.1.3 Results
Sirohi & Mahadikβs research [17] showed a rough correlation between the numerical and
experimental model. Experimentally, the device was able to achieve a maximum power of
53 mW from all four strips. While the numerical model predicted a purely linear
relationship between incident windspeed and the deviceβs output power, the experiment
demonstrated a sharp dropoff in output power after 12 mph winds. The paper attributes this
mismatch to the wind tunnelβs fans, which may have introduced turbulence or unsteady
flow, violating the quasi-steady flow assumption of the numerical model. However, further
investigation is required into the matter as this could present a maximum performance limit
to gallop-based energy harvesters. The paper also notes that the piezoelectric material had
visible signs of fatigue after the wind tunnel experiment, which may have inhibited proper
performance as well. Regardless, this represents the next critical performance metric for the
device.
9
Figure 2.1.3.1: Correlation between Measured and Predicted output power for an incident
wind speed of 10 mph. [17]
Figure 2.3.1.2: GPEH Performance vs Wind Speed and Load Resistance [17]
2.2 Fatigue
2.2.1 Introduction
The next step in evaluating the feasibility of a GPEH system is to understand its
susceptibility to fatigue. While most engineered systems are designed to minimize any
aeroelastic excitations, this device intends to amplify such excitations. Such load cycling is
detrimental to structures, accruing damage in the form of cracks which eventually lead to
failure.
10
Understanding the effective fatigue life of these devices gives us an insight into their
effective lifetime capacity for energy harvesting. As the GPEH is cited as a possible
alternative to chemical batteries in situations where the replacement of these batteries is
difficult or undesired, we must verify that such a device can provide more energy than a
standard chemical battery, meaning it would require less frequent maintenance.
The simplest model of fatigue is the S-N curve. This represents the relationship between a
given stress and the number of load cycles of that stress a material can withstand before
failure. These curves are created through extensive testing of materials. These curves tend
to follow a linear-log trend [24], with 1 cycle for stress equal to the ultimate tensile stress,
going all the way to a point known as the endurance limit which corresponds with many
orders of magnitude more cycles . The endurance limit is the minimum stress at which
fatigue failure is achievable. For example, a standard aluminum material has an endurance
limit of around 10 ksi, lasting about 107 cycles [9].
The reason for the limit is based in the yielding stress of materials. As stress increases and
the stress-strain relationship grows increasingly nonlinear, the material will begin to
experience permanent deformations which compound with each cycle. But at low stresses,
where the stress-strain relationship is effectively linear, no such permanent damage occurs.
This theory is supported by these experimental S-N curves, which tend to have higher
endurance limits for brittle materials and lower endurance limits for ductile materials. For
example, lead zirconate titanate, or PZT, has an ultimate tensile strength of 80 MPa and an
endurance limit of 60 MPa [13]. While aluminum is only free from fatigue failure for
stresses below one-seventh its maximum allowable stress, PZT is free from fatigue failure
for stresses below three-fourths its maximum allowable stress.
However, such S-N curves are only applicable for constant-amplitude load cycles. Given
that real-world wind scenarios will have variable wind speeds and thus will induce
variable-amplitude load cycles, we must use a new method of fatigue analysis. The two
most common are Minerβs Rule, which is a very approximate method based on the idea of
cumulative damage, and a variety of Crack Growth Equations, which use differential
equations to model the growth of cracks of an initial size.
2.2.2 Minerβs Rule
Minerβs Rule is based on the idea of cumulative damage, that each load cycle contributes
some fractional amount of βdamageβ to a parameter D, which starts at zero and finally fails
when π· = 1. To determine the amount of damage, this method uses S-N curves as
reference points and assumes that this accumulated damage is linear.
For example, if a stress of amplitude π is applied for π load cycles, and that stress π is
correlated with π cycles until failure, then the amount of damage accrued is equal to [3]:
11
Ξπ· =π
π
This method is extremely simple, however this simplicity comes at the cost of theoretical
rigor. As cracks grow, the stress concentration near the propagating tip of the crack grows
exponentially [19], disproving the linear damage assumption of Minerβs Rule. However, no
current alternative method following cumulative damage has been proven to out-predict
Minerβs Rule [23]. Minerβs Rule has stood the test of time through its statistical accuracy.
However, the lack of fidelity in the model beyond being based in experimental data makes
this less preferred for our fatigue analysis.
2.2.3 Crack Growth Equations
This method provides a differential equation of crack growth per load cycle, from an initial
flaw size, based on the geometry of the loading scenario and experimentally derived curve-
fit parameters.
For a crack of length π and currently applied stress π, we can calculate a parameter known
as the stress intensity factor, πΎ = π(π, π). There are lookup tables for a given boundary
condition at the site of the critical plane, where stress is greatest and/or most persistent.
While a beam undergoing bending has a certain associated formula for stress intensity
factor, the individual flanges of the beam will effectively be under pure tension or
compression. Thus, the crack growth for a center crack of length 2a is calculated from the
basic formula [19]:
πΎ = πβππ
Given the loading cycle has a range ππππ₯ and ππππ, there is a corresponding range πΎπππ₯
and πΎπππ. The crack growth equation relies on ΞπΎ = πΎπππ₯ β πΎπππ.
Every crack growth equation correlates ΞπΎ with the per-cycle crack growth ππ
ππ through
varying curve fits. The simplest is the Paris-Erdogan Equation [15]:
ππ
ππ= πΆ(ΞπΎ)π
Where πΆ and π are material constants. However, this model disobeys the concept of a
material endurance limit, as a load cycle with ΞπΎ = 0, in other words, a load cycle of
amplitude zero, will still contribute a crack growth of πΆ.
Experimentally derived curves between ΞπΎ and ππ
ππ typically feature three key regions [20].
The first region, at low ΞπΎ values, in which ππ
ππ asymptotically approaches zero at the ΞπΎ-
value corresponding to the fatigue threshold, an analog for the endurance limit. The second
region, at medium ΞπΎ values, features power law crack growth, quantified by the Paris-
12
Erdogan formula. The third region, at high ΞπΎ values, where ππ
ππ asymptotically approaches
infinity at a ΞπΎ value known as the fracture toughness. While more involved formulae can
accurately account for all three of these regions and the transition between them, the more
involved curve-fit parameters are not so widely studied. Additionally, standard materials
such as metals and ceramics tend to have an insignificant fatigue threshold, and thus the
Region II behavior can be assumed all the way to ΞπΎ = 0. Likewise, the transition from
Region II to Region III is generally very abrupt and thus failure can simply be asserted once
ΞπΎ = 2πΎπΌπ. As crack growth is roughly exponential, a steadily growing crack will exceed
fracture toughness very rapidly, and thus failure is already inevitable regardless of curve
modelling.
2.2.4 Additional Fatigue Effects: Crack-Induced Compliance
Galloping problems are best understood as an enforced displacement problem, in which
beams oscillate in their natural frequency modes, to an amplitude determined by their own
stiffness, damping, and aerodynamics. As galloping induces a positive feedback, the tip
displacements will often be much greater than a simple quasi-static aerodynamic force
analysis would indicate.
One result of this model is that it introduces the possibility of stable crack growth. If a
steady load is applied to the device rather than a steady amplitude, the crack will grow due
to a high stress intensity factor, in turn causing an even higher stress intensity factor, and so
on as the crack grows exponentially. With a steady displacement, we observe that crack
growth will drastically reduce the stiffness of the device, decreasing the applied stress as
well as the stress intensity factor, potentially halting crack growth.
The stability of the crack is also dependent on aeroelasticity, as less stiff structures tend to
experience higher loads than those structures which are too stiff to gallop, flutter, etc. Thus,
crack-induced compliance may have the effect of increasing the effective load experienced
by the beam.
To determine the outcome of this multi-faceted problem, we integrate a crack-based
negative stiffness matrix into our Rayleigh-Ritz Model. This is done by adding a negative
strain energy term to represent the strain energy dissipated by the crack. We use the strain
energy release rate, a variable meant to represent the loss of strain energy per unit area
given by [20]:
πΊ =πΎ2
πΈ, ππππππ = ββ« 2π‘ πΊ(π , πππππ) ππ
2πππππ
0
β β« 2π‘ πΊ(π , πππππ§π) ππ2ππππ
0
= ββ«2π‘ππ2π
πΈ ππ
2π
0
= ββ« 2π‘πππΈ(π§π )2 ππ2π
0
13
Where a is the crack length in either the beam or the piezoelectric strip, ππππππ is the
potential energy stored in the crack, b is the width of the piezoelectric strip and of the
beam, t is the thickness of the ply or flange, kappa is the curvature at the base of the beam,
and z is the distance from the center to the extremity of the cross-section. Note the
multiplication by two since bending fatigue will induce a crack on bottom and top. Since
π§ =π
2 and π (0) = π1
β²β²(0)πΎ1 + π2β²β²(0)πΎ2 = πΎ1 + πΎ2:
ππππππ = ββ«1
2πΈπ2π‘ππππππ(πΎ1 + πΎ2)
2ππππππ
2πππππ
0
β β«1
2πΈπ2π‘πππππ(πΎ1 + πΎ2)
2πππππ
2ππππ
0
ππππππ = βπΈπ2π‘(πππππ2 + ππππ
2 )(πΎ1 + πΎ2)2
πππππππ
ππΎπ= β2πΈπ2π‘(πππππ
2 + ππππ2 )(πΎ1 + πΎ2) = 0
We can finally put this in matrix form:
[π]{πΎ}Μ + [πΆ]{πΎ}Μ + [πΎ]{πΎ} β [πΎπ]{πΎ} = {πΉ}
Where:
πΎπ,ππ = 2πΈπ2π‘(πππππ2 + ππππ
2 )
As we can see, crack growth quadratically reduces the stiffness of the beam. Plugging in
some cursory values, we can see that the loss in stiffness rapidly goes from trivial to
substantial, long before the crack becomes critical, meaning this will have significant
effects on the outcome of the simulation.
3.0 Numerical Simulation Setup
3.1 Assumptions and Default Model Features
To model the myriad of features affecting GPEH performance and durability, we will
simulate a general GPEH device with custom device parameters in Python. The device will
respond to a synthetic wind signal based on real anemometer data, from which tip
displacement, stress, voltage, and crack length with be calculated until the device fails.
While the code is set up to be endlessly customizable to test a variety of factors, some basic
design factors were arbitrarily chosen ahead of time to keep the code simple.
The first design factor to be chosen was the tip-body cross-section. As it is meant to
respond to wind signals in all directions, a symmetric body which satisfied the Den Hartog
criterion for the most angles of attack is desired. After analyzing section aerodynamic
coefficients for a number of regular polygons, it was determined that equilateral triangles
14
have the most galloping βcoverageβ. Given sufficient windspeed, an equilateral triangle
will gallop if the wind blows normal to one of its sides, plus or minus 40 degrees, resulting
in a coverage of 67% [1].
The next decision would be the GPEH structural geometry. As PZT is the main
piezoelectric material being studied for this project, a cantilevered beam with short strips of
piezoelectric material adhered to all sides is considered standard and optimal for brittle
piezoelectric materials. The cantilevered beam cross-section must be symmetric, unlike the
simple rectangular cross-section used by Sirohi & Mahadik [17] and curved surfaces would
complicate the application of piezoelectric material. Also, a non-hollow cross-section
would make the beam very stiff unless very compliant material is used. Therefore, the
cross-section is decided to be a hollow, square tube. While less important for aerodynamic
analysis, this detail is extremely important to determine the parameters for crack
propagation. For initial simulation run-throughs, the beam was chosen to be made of 7075
Aluminum Alloy and the piezoelectric material was chosen as PZT-5H.
For the probabilistic wind model, a month of 1 Hz anemometer data on the coast of Froya,
Norway was sampled [6].
While the GPEH setup used in [17] was designed to only gallop in response to a single
wind-direction, the multi-directional GPEH used in the simulation would look something
like the below figure.
Cantilever Support Piezoelectric Strips
Triangular Prism
Tip Bodies
Beam
Rigid
Connections
Figure 3.1.1:
GPEH Diagram
Incoming Wind Galloping
Motion
15
3.2 Probabilistic Wind Model
An important aspect of generating this synthetic wind signal was capturing both speed and
direction, and additionally simulating persistence, or the tendency for wind gusts to stay
constant for an arbitrary duration. If the wind was modeled with a basic Weibull or
Rayleigh probability distribution [21] which was randomly sampled every second, the
resulting signal would be erratic and noisy. Galloping occurs best under steady wind
conditions as the device gradually ramps up to the steady-state amplitude. Thus, modeling
persistence is key to ensuring the maintenance of galloping motion for extended periods.
For this, the key choice was to use a Markov Chain Model to independently measure wind
speed and direction [14]. Rather than viewing the entire wind history at once, a Markov
Chain generates a state-dependent probability distribution at each time step. For example, if
the windspeed is currently 5 m/s, the Markov Chain Model estimates what the wind speed
will be one second later, which will very likely be in the range of 4.9 to 5.1 m/s rather than
0 or 10 m/s. Assuming the probability for a wind vector to change is dependent only on the
wind vector itself, the distribution is assumed to be stable and sufficient data sampling will
converge to a single Markov Chain Model.
To create the model, a data time series is reviewed two steps at a time to give us the data
pair (π₯π , π₯π+1). By breaking the data range into bins from 0 to π₯πππ₯, each with size ππ₯, each
of these data points can be stored into a 2D histogram of size (π₯πππ₯
ππ₯ ,
π₯πππ₯
ππ₯) known as the
transition matrix. After all the data is stored into the histogram, each row of the matrix
represents a probability distribution.
Figure 3.2.1: Transition Matrix Example [14]
To create a synthetic wind signal, then, an initial data point is assumed, that windspeed is
associated with a row in the matrix, a random sample from that distribution is made, and
16
that result becomes the state after one time step. This process is repeated with each new
state to generate a random walk through the Markov space. While each probability
distribution will have a very low variance, the probability that a wind speed will remain
stable vanishes over time due to the Law of Large Numbers. Below is a comparison of the
sample data with the synthetic wind signal.
Original Synthetic
Figure 3.2.2: Original vs Synthetic Model, Windrose
3.3 Workflow Algorithm and Code Features
3.3.1 General Layout
The general flow of the code is given by the flowchart below. First, the provided wind data
is analyzed and stored into a Markov model using the method described in Section 3.2.
17
Then, the material properties of all materials involved are loaded in through several
βMaterialβ class constructors. Likewise, the dimensions of the cantilever beam and
piezoelectric strips are provided through another constructor class. A for-loop is used to
simulate and re-simulate the device under the same wind model, material selections, and
cross-section dimensions, but the length of the cantilever beam is varied each time. This is
done as length is the simplest way to vary the overall beam stiffness without majorly
altering the underlying geometry and mechanics. The simulation is fed these inputs and
spits out a results object which is then visualized in post-processing.
The Rayleigh-Ritz ODE for structural dynamics is solved using the Newmark-Beta method
with an adaptive time step algorithm. There is additionally a function which reads the wind
vector as well as the beam tip velocity to calculate aerodynamic loads, which are evaluated
on each time step. Originally, this stored the results in one large time series for plotting
results and checking for normal functioning. Once this was verified, however, the code was
changed to only store three data points at a time. This minimizes data storage requirements
as it is the minimum amount of data required to identify peaks and valleys.
When a peak or valley (a reversal) is detected, it is appended to a deque (βdouble-ended
queueβ) object. This is data structure is a necessary improvement over a standard array,
which requires many redundant steps to resize as new items are added. When a new item is
appended to a deque, it includes a pointer to the previous element in deque, and thus we are
able to traverse the list by following the list of pointers and not wasting time endlessly
copying entries. The deque structure is used extensively to store data variables of unknown
size which continuously pile up until device failure.
When three or more reversals are detected, the rainflow-counting algorithm, a standard
algorithm for reducing strain data into discrete cycles [25], is called upon to pair peaks and
valleys into cycles or half-cycles. When a cycle is detected, the peak and valley data points
are appended to a βcyclesβ object and the resulting stress and voltage amplitude values are
calculated from that. The additional stress cycle is passed to the crack growth equation and
the values for current crack lengths are increased. This then updates the stiffness matrix
used in the Newmark-Beta solver for the next time step.
This loop continues until a crack length is recorded to exceed the width of the beam.
18
Figure 3.3.1: Full Program Flowchart
Load Material Properties
For πΏ = πΏ1: πΏ2 Run Wind Model Calculate [K], [M], etc.
π¦ = π¦0
π¦π+1 = π(π¦π , πΉπ)
πΉ = π΄πππ πππππ
Re-size Ξπ‘ πΈππππ β₯ π?
Store value.
Reversal Detected?
Store reversal.
Cycle Detected?
Calculate performance
variables. Store.
ππ
ππ= πΆ(ΞπΎ)π
Update Stiffness Matrix
πΎπππ₯ β₯ πΎπΌπ? Rapid Crack Growth Failure?
End Simulation.
No
Yes
Yes
Yes
No
No
No
Yes
Yes No
Write to File
19
3.3.2 Newmark-Beta Method
The Newmark-Beta Method [26] is a numerical integration technique designed to
accurately solve stiff structural dynamics systems. Rather than relying on the first-order
οΏ½ΜοΏ½ = π΄π₯ form, the equation of motion is given in the second-order form:
[π]οΏ½ΜοΏ½ + [πΆ]οΏ½ΜοΏ½ + [πΎ]π₯ = πΉ
Our Rayleigh-Ritz method derived previously is incomplete, however, and must include the
systemβs electromechanical coupling. This consists of two circuits, each consisting of two
piezoelectric strips connected to a single load resistor:
[π]οΏ½ΜοΏ½ + [πΆ]οΏ½ΜοΏ½ + [πΎ]π₯ β ππ = πΉ
πΆποΏ½ΜοΏ½ +π
π + ποΏ½ΜοΏ½ = 0
Where R is the resistance of the load resistor, V is the output voltage, πΆπ is the capacitance
of the piezoelectric material, and π is the electromechanical coupling factor (see [Source]
for calculation).
These can be appended to our pre-existing matrices:
πβ² = [[π] 4π₯4 04π₯2
02π₯4 02π₯2] , πΆβ² = [
[πΆ]4π₯4 04π₯2
ππ1(πΏ) ππ2(πΏ) 0 0
0 0 ππ1(πΏ) ππ2(πΏ)πΆππΌ2π₯2
]
πΎβ² =
[ [πΎ0 β πΎπ]4π₯4
βπ/2βπ/2
00
00
βπ/2βπ/2
02π₯4
1
π πΌ2π₯2 ]
, πΉβ² = (πΉ4π₯1
02π₯1)
The Newmark-Beta Method relies on the Extended Mean Value Theorem to calculate the
next step. After deriving the applied aerodynamic forces from the aerodynamic model [26]:
πΎπ+1 = π΄β1π΅
οΏ½ΜοΏ½π+1 =1
π½(Ξπ‘)2(πΎπ+1 β πΎπ) β
1
π½Ξπ‘οΏ½ΜοΏ½π β (
1
2π½β 1) οΏ½ΜοΏ½π
οΏ½ΜοΏ½π+1 = οΏ½ΜοΏ½π + (Ξπ‘)(1 β πΌ)οΏ½ΜοΏ½π + (Ξπ‘)πΌοΏ½ΜοΏ½π
Where:
20
π΄ =1
π½(Ξπ‘)2[π] +
1
π½(Ξπ‘)[πΆ] + [πΎ]
π΅ = (πΉ) + [π] β (1
π½(Ξπ‘)2πΎπ +
1
π½Ξπ‘οΏ½ΜοΏ½π + (
1
2π½β 1) οΏ½ΜοΏ½π) + [πΆ]
β (πΌ
π½(Ξπ‘)πΎπ β (1 β
πΌ
π½) οΏ½ΜοΏ½π β Ξπ‘ (1 β
πΌ
2π½) οΏ½ΜοΏ½π)
And (πΌ, π½) are method parameters. The most common parameters to use are (0.5, 0.25), as
this method is unconditionally stable [26].
3.3.3 Adaptive Time-Stepping
As wind conditions are subject to change and thus the precision of the ODE solver vary, it
is beneficial to have a method of adapting the time step. Borrowing from [22], I
implemented a method for estimating the error for each time step and either increasing or
decreasing the time step accordingly to stay within an upper and lower bound for relative
error. While most ODE methods only decrease time step according to error estimates, that
is inefficient for simulations like this meant to run for long periods of time. When wind is
high, time steps should be short as negative damping causing all kinds of chaos in the
beam. But once the wind dies down, this short time step is unnecessary, and much runtime
can be saved by checking and speeding up our simulation when possible. The adaptation
algorithm is featured in Figure 3.2.1.
Using the formula for error estimation [22]:
|π| = (π½ β1
6) (Ξπ‘)2|οΏ½ΜοΏ½π+1 β οΏ½ΜοΏ½π|
The next time-step is calculated according to the Newmark-Beta method and the error is
estimated. If that error lies within the acceptable bounds, that proposed data point is
accepted and the simulation moves forward. Otherwise, the code enters a while loop where
the time-step is continuously decreased or increased and the next data point is continuously
re-evaluated until the resulting error is acceptable.
This method has the possibility of being inefficient when the bounds for error are narrow,
as it would result in repeated re-evaluations that are otherwise unnecessary.
3.3.4 Fracture Toughness Modeling
Critical to determining the point of failure is the modelling of stable or unstable crack
growth. While crack growth is regularly determined by the linear Paris-Erdogan formula,
the case when the maximum stress intensity factor exceeds fracture toughness must be
studied as well, as the system risks immediate unstable crack growth. When the stress in a
21
particular cycle is expected to exceed fracture toughness, the simulation is temporarily
paused and the following flowchart is initiated:
Figure 3.3.2: Rapid Crack Growth Flowchart
First, the crack length is increased by an arbitrarily small amount, the stiffness matrix is
reduced accordingly, and the Newmark-Beta solver re-calculates the current time step with
the new stiffness matrix. This is repeated until either the stress intensity factor falls below
the fracture toughness or the crack length exceeds the critical amount. In the first case of
stable crack growth, the new stress intensity factor is plugged into the Paris-Erdogan
formula and the simulation continues. In the latter case of unstable crack growth, the
simulation ceases.
4.0 Numerical Simulation Results
The following material properties were used across all runs of the simulation [2,10-13,16-
18]:
Beam (7075 Aluminum) Properties
Side Length 1.5 in
Wall Thickness 0.0625 in
Youngβs Modulus 68 GPa
Density 2,710 kg/m^3
Air Density 1.23 kg/m^3
Structural Damping Ratio, ΞΆT 0.01
Fatigue (Parisβ Law) C 88.7 Pa-m1/2
Fatigue (Parisβ Law) m 3.14
Fracture Toughness 14 MPa-m1/2
[K(a)]
π¦π+1 = π(π¦π, πΎ, πΉ)
π π+1 = π(π¦π+1) a
ππΌπΉ = π(π, π )
πΎπππ₯ β₯ πΎπΌπ? π = π + da ππ
ππ= πΆ(ΞπΎ)π
End
Yes No
22
Base Strip (PZT-5H) Properties
Thickness 0.015 in
Youngβs Modulus 63 GPa
d33 180 pC/N
Self-Resistance 1E6 Ξ©
Self-Capacitance 15 pF
Load Resistance 1E3 Ξ©
Electromechanical Coupling, ΞΈ 662.7 ΞΌV / (m/s)
Fatigue (Parisβ Law) C 9.27 Pa-m1/2
Fatigue (Parisβ Law) m 11.8
Fracture Toughness 0.9 MPa-m1/2
4.1 Qualitative Results
Running the simulation for varying cantilever beam lengths with intermittent print-outs of
the data as it was being processed was immediately illuminating. Firstly, the amount of
time in-simulation varied dramatically, with a number of runs lasting less than an hour
before the hypothetical device would fail. The element of randomness involved in the
probabilistic wind model evidently introduces a lot of variation in fatigue life, implying
multiple runs would be required for gathering a more robust dataset. Overall, however,
shorter (and therefore stiffer) beams had significantly longer fatigue lives. Some runs of the
simulation stopped abruptly as the amount of data being stored exceeded the system
capacity. This can be fixed in the future by writing all data to an output file as soon as it
gathered, but for now a sufficient amount of data was gathered from the code and enough to
note important trends.
Notably, the critical crack growth behavior (when exceeding fracture toughness) always
was unstable. Rapidly increasing stresses at the base of the beam were always identified in
the minutes prior to failure, implying that aeroelasticity serves to amplify crack compliance
and rapidly fracture the beam. While crack growth was monitored on both the metallic
beam and the piezoceramic strip, the piezoceramic strip was always the site of critical crack
growth, as expected.
Unfortunately, the code tended to run quite slowly. As a data read-out was printed once
every simulated minute, these statements tended to be displayed every one to three seconds.
While a factor of 20 to 60 times is still good, the runs which integrated over 48 (simulated)
hours ended up taking a few hours. Taking this with the fact that repeated runs are required,
this code is evidently not the most efficient way to develop and test optimization strategies.
This long runtime could be explained by low system resources and thus could have been
improved on a more powerful machine.
23
As the code currently stands, it is a single time-integration thread consisting of an
alternating series of computationally expensive processes which must occur linearly. Thus,
little can be done through parallel programming to speed this up. However, running
multiple simulation instances in parallel on a machine which can handle the storage and
processing requirements would absolutely perform faster.
4.2 Quantitative Results
First, two plots were compiled demonstrating the applied windspeed distribution and the
general trends between windspeed, maximum stress, and beam length.
Figure 4.2.1: Windspeed Overall Probability Distribution
24
Figure 4.2.2: Stress, Windspeed, and Beam Length
(Note: Data points in Figure 4.2.2 are only a portion of all data points recorded) At first,
Figure 4.2.2 should raise some alarm as we expect a hyperbolic relation between the two,
with zero galloping below a critical wind-speed, a sharp increase in windspeed, and a
roughly linear relation as windspeed increases. Here, we instead have something either
linear or quadratic. The reason for this βsmearingβ of data points near the cut-in windspeed
is because, though galloping cannot be sustained below the cut-in speed due to finite
damping, there will still inevitably be wind-induced oscillations. Thus, our code simulation
is accounting for load histories which a simple hyperbolic curve would not. The tendency
for near-zero stresses to be recorded at high wind speeds also demonstrates the moments
when the beam is transitioning from one steady-state amplitude to another.
This figure implies a good correlation between beam length and cut-in speed, with longer
beams featuring an earlier cut-in speed. Additionally, longer beams feature higher stresses
at the same windspeed due to this earlier galloping onset. Therefore, though longer beams
are less stiff, they are able to harvest more energy because of their susceptibility to
aeroelastic excitation. Data at higher windspeeds is much more sparse, and therefore
difficult to comment on. The lack of data is both due to the low probability of occurring and
the high chance of failure. The high data variance could be a result of the increased
stiffness of the equations of motion in high-wind environments creating more simulation
error.
25
The following graphs will compile the history of all relevant performance variables. Rather
than overlaying histograms, which vary significantly in number of data points, or through
normalized probability densities, which tend to clump together, the data is displayed as a
cumulative density function to better delineate individual lines.
Figure 4.2.3: Tip Displacement Distribution
26
Figure 4.2.4: Stress Distribution
27
Figure 4.2.5: Voltage Distribution
The first characteristic which jumps out of these plots is the differences in skew, with the
πΏ = 5 π being a clear outlier. This larger than normal than skew is the result of a very short
simulation time (1 hour) and thus more an anomaly of the wind model than the structural or
aerodynamic models.
Looking at the other curves, we see a general trend regarding median and skewness, which
both initially increase with beam length from the cases of πΏ = 3 π to πΏ = 4.25 π, but then
decrease (excluding the πΏ = 5 π case). Evidently, this points to some βsweet spotβ around
πΏ = 4 π where the beam is compliant enough to sufficiently exploit galloping, but not so
compliant that it becomes unable to generate sufficient stresses. As the πΏ = 4 π run lasted
longer than that of any longer beams, it was still able to deliver a higher power output.
Therefore, there are certain sets of system parameters in which both the durability and
performance of materials increases, indicating an optimal design!
This ideal length will be dependent on the material properties and the wind conditions, so
once materials, geometry, and location are chosen for a GPEH device, this code can
determine the ideal length for the cantilevered section, making the design customizable to
regional conditions. However, there is still significant variation in the figures above and
continued runs of the code will be required for more definitive solutions.
5.0 GPEH Analysis and Optimization Strategy
5.1 Analytical Model for GPEH Performance
To better understand the tradeoff between performance and durability in the context of the
GPEH device, it is necessary to develop an analytic model of GPEH performance. For
example, we need to develop a formula for determining the steady-state galloping
amplitude of such a device based on the wind speed and device parameters.
For that, we turn to Bibo & Daqaq [4], who use a Taylor-series based aerodynamic
coefficient model to identify a steady-state galloping response. This method has the user
represent the net transverse force coefficient in terms of a Taylor series [4].
πΆπ¦ = πΆβπππ πΌ + πΆππ πππΌ, πΆοΏ½ΜοΏ½ = β π΄π (οΏ½ΜοΏ½
π)
ππ
π:πππ
+ β π΄π (οΏ½ΜοΏ½
π)π οΏ½ΜοΏ½
|οΏ½ΜοΏ½|
π
π:ππ£ππ
, π β₯ 1
This method uses a single lumped-mass (i.e. single natural frequency) estimation of the
system and assumes steady flow. All these parameters can be normalized through the
formulae:
πβ =ππ
ππ, πβ =
οΏ½Μ οΏ½π
ππ, οΏ½Μ οΏ½ =
π
πππ·π‘, π =
ππΏπ‘π·π‘2
4π
28
Where a is the steady-state amplitude, ππ is the beam natural frequency, π is the
equivalent lumped mass at the tip, ππ is total damping ratio, and (πΏπ‘, π·π‘) are the length and
aerodynamic chord of the tip bodies. The user-provided Taylor series, unique to the cross-
section geometry of the galloping body, can then be manipulated to provide us with the
condition for steady-state equilibrium [4]:
πβ2πΆπβ
πββ 1 = 0
Where:
πΆπβ = β [4π΄π
π(
πβ
2πβ)π
β(β1)
π2βπ
π + 1 β 2π
π/2
π=0
]
π:ππ£ππ
+ β π΄π (πβ
2πβ)π
(π + 1
1
2(π + 1)
)
π:πππ
Thus, for a given (normalized) windspeed, the steady-state amplitude of the galloping body
is given as the real root of a polynomial. Performing this for an equilateral triangular cross-
section, we get the following galloping response curve:
Figure 5.1.1: Normalized Windspeed vs Normalized Tip Displacement for an Equilateral
Triangular Tip Body
Which can be approximated by a hyperbola of the form:
πβ2
π1β2 β
πβ2
π2β2 = 1
De-normalizing this formula finally gives us a hyperbolic formula for galloping response
where:
29
π2
π12 β
π¦π‘ππ2
π22 = 1
The conclusions of this method will be tested with the results of the simulation, which
factors in a variety of confounding variables that are not so easily modeled.
5.2 Optimization Strategy
5.2.1 Effect of Material Choice on Efficiency and Durability
5.2.1.1 Piezo-ceramics versus Piezo-polymers
Before analyzing all design parameters together, weβll focus on the choice of piezoelectric
material and see how different classes of material can greatly improve device durability.
Generally, common piezoelectric materials fall into two categories. The first are ceramics
with high dielectric constants, high stiffnesses, low ductility, and poor fatigue resistance
[10, 13]. These offer great efficiency in converting strain energy to electrical energy, but
their poor compliance requires they are only applied in short, thin plies to prevent brittle
fracture. Additionally, their lack of intrinsic crack growth resistance mechanisms ensures a
finite, and often short, fatigue life. For example, Sirohi & Mahadik [17] noted visible
fatigue damage to the PZT plies in the GPEH device after conducting wind tunnel tests.
Secondly, there are piezoelectric semi-crystalline polymers (often thermoplastics) with low
dielectric constants, low stiffness, high ductility, and exceptional fatigue resistance [18].
These are very poor at converting strain energy to electrical energy, but their substantial
compliance means the devices themselves can be much more compact while still featuring
a lower cut-in windspeed than their ceramic-based counterparts. In particular, the
thermoplastic polyvinylidene di-fluoride (PVDF), has a π33 value one-sixth that of the most
common piezoceramic, lead zirconate titanate (PZT), but has been shown to feature an
endurance limit at about 50% of its ultimate tensile strength, whereas PZT will eventually
fail at any stress level [10,18].
For this reason, PVDF is a clear champion when designing durable energy harvesters,
especially if a stopping mechanism can be employed to prevent bending stresses exceeding
the endurance limit. Such a setup would result in an infinite service life while still being
able to deliver half of its maximum achievable power.
That being said, there are instance where PVDF is still not a practical material choice, as
the achievable amounts of power can often be orders of magnitude lower. For samples of
PZT and PVDF strained to the same extent, the power output is given by the ratio:
Powerπππ
Powerπππ·πΉ=
π33,πππππππ
π33,πππ·πΉππππ·πΉ=
π33,ππππΈπππ
π33,πππ·πΉπΈπππ·πΉ
30
As PZT offers a much higher value for both π33 and πΈ, PZT could be the clear champion in
applications where durability is not necessary.
5.2.1.2 Performance Optimization of a Piezo-Ceramic-Polymer Composite
While both materials have their own trade-offs when used on their own, we will see that
when used together in a composite piezoelectric structure, they are able to offer higher
efficiency than when used separately.
Consider an inverted sandwich lay-up, with the piezo-ceramic in the interior and the piezo-
polymer on the exterior. This design is chosen as it is able to maximize compliance and
ensure the material can easily gallop. A true sandwich layup with the stiffer ceramic plies
on the exterior would be exceptionally stiff and would rarely gallop in normal wind
conditions.
Borrowing from Classical Laminate Theory, when can use the following graphs to
represent strain, stress, and voltage in response to an applied curvature. Here, a filled in
space to the right of the vertical line indicates a positive value while a space to the left
represents a negative value.
Figure 5.2.1.2.1: Bending Diagram
As we can see, the voltage between plies would equalize, and the rate of voltage drop
across each ply is dependent on both the stress experienced by the ply and the
electromechanical coupling coefficient. There is thus a drastic voltage drop across the
piezo-ceramic which slowly increases through the piezo-polymer plies.
Ξπ = [ππππ·πΉπΈπππ·πΉ (1
2βπππ + βπππ·πΉ) +
1
2πππππΈπππβπππ] π = οΏ½Μ οΏ½π
Where π is the electromechanical coupling constant, πΈ is the Youngβs Modulus, π is the
curvature, and β is the ply thickness. Note that βPVDFβ and βPZTβ will be used as
subscripts to denote any general piezo-polymer or piezo-ceramic, respectively. Here, οΏ½Μ οΏ½ is
Strain Stress Voltage + + +
- - -
31
the effective coupling constant between applied curvature and resultant voltage. Thus, οΏ½Μ οΏ½ =
πππππΈπππβπππ.
From basic geometry, we see that βπππ = 2βπππ·πΉ + βπππ and from Classical Laminate
Theory, we can determine:
πΈπππ =πΈππππΌπππ + 2πΈπππ·πΉπΌπππ·πΉ
πΌπππ + 2πΌπππ·πΉ
πΈπππ βπΈπππ (1 +
18πΌ3) + πΈπππ·πΉ(1 + 8πΌ3)
2 + 8πΌ3 π€βπππ πΌ =
βπππ·πΉ
βπππ
The approximately-equal sign is used as we simplified the calculation of area moment of
inertia for the PVDF plies assumes the PVDF plies are much thicker than the PZT plies, or
πΌ > 1. We can therefore solve for the equivalent piezoelectric constant of this layup:
ππππ =οΏ½Μ οΏ½
πππππΈπππ=
(8πΌ2 + 2)ππππ·πΉ
π½ (1 +1
8πΌ3) + (1 + 8πΌ3)+
(8πΌ3 + 2)/(2πΌ + 1)
(1 +1
8πΌ3) +1π½
(1 + 8πΌ3)ππππ
π€βπππ π½ =πΈπππ
πΈπππ·πΉ
We can lastly normalize this with another variable πΎ =ππππ
ππππ·πΉ. Assuming large values for
(πΌ, π½, πΎ):
ππππ β8πΌ3 + 2
8πΌ3 + π½ + 1(π½πΎ
2πΌ+ 1) ππππ·πΉ
As (π½, πΎ) are fixed constants, we optimize this design by varying πΌ. We can find the
optimal layup by setting πππππ
ππΌ to zero and solving for πΌ in the numerator.
πΌ4 + (1
4π½πΎ + 1 β
π½πΎ
6(π½ β 1)) πΌ3 +
1
4(π½πΎ + 1)πΌ2 β
π½πΎ
24(π½ β 1)= 0
Assuming π½
π½β1β 1, π½πΎ β« 4, π½ β«
2
3 we can fully simplify to:
πΌ4 +1
4π½πΎπΌ3 +
1
4π½πΎπΌ2 β
1
24πΎ = 0
Quartic formulas are not easily solved analytically, but using a numerical solver with the
imagined ratios below, we arrive at:
π½ = πΎ = 100, πΌπππ‘ = 3.1, ππππ₯ = 1,148.4ππππ β 11.5ππππ
32
Note that, while the actual voltage drop in the optimal scenario is still lower than that of a
pure PZT layup, but the efficiency is an order of magnitude higher. If we plot ππππ for the
example above, we can note an interested behavior:
Figure 5.2.1.2.2: Piezo Coupling vs Ply Thickness Ratio
The effective piezoelectric constant begins at that of PZT for πΌ = 0, falls slightly as it is
replaced with polymer, then drastically rises to a maximum value, then finally decaying to
that of PVDF as πΌ approaches infinity. We have evidently stumbled upon an ideal regiment
where the load-sharing induced by materials of differing Youngβs moduli is perfectly
calibrated with their differing piezoelectric coupling coefficients to most efficiently convert
strain energy into electrical energy.
5.2.2 Methods of Analytically Correlating Design Parameters and Performance
Variables
5.2.2.1 Probability-Based Model
The windspeed applied to the GPEH mechanism can be modeled as following a Weibull
distribution, with model coefficients unique to the specific region the GPEH is placed. If
we use our hyperbolic formula from Section 5.1, we can then model the GPEH tip
displacement with roughly the shape plotted below.
33
Figure 5.2.2.1.1: Example Probability Distribution for Galloping Tip Displacement
Note that the hyperbolic relation [4] further exaggerates the positive skew of the
distribution while the existence of the cut-in windspeed provides a discrete probability mass
that the steady-state tip displacement equals zero. If we assume a constant mode shape
equal to the first eigenmode, then the curvature and stresses at the base of our cantilevered
beam will be linearly correlated with the tip displacement, meaning we can relate stress in
our piezoelectric material in one of the following ways:
π2
π02 β
π2
π12 = π2
2, π = π1β(π
π0)2
β π22, π = π0β
π2
π12 + π2
2
The variable we are seeking to optimize is the total lifetime energy, that is, the average
power rate times the device fatigue life, or: πΈπππππ¦ = π33πππππ(ππππ). At first glance,
there is a trivial maximum to this variable. As power increases linearly with applied stress
level while the number of fatigue-cycles decreases faster-than-linearly as stress increases.
Therefore, the optimal design is one which consistently undergoes practically zero stress,
with a trivial power output but a nearly infinite fatigue life.
Obviously, such a design would be impractical in most applications as whatever payload is
meant to be powered would have some minimal power requirement. Thus, we have to
approach this problem already assuming some average power output. In other words,
πΈ[π] = π = ππππ π‘. This assumption can be made as the formula for average power output:
πΈ[π] = π33πΈ[π ] = β« π33π π(π )ππ
β
0
Since the formula for power is linearly dependent on stress, this simplification can be made.
If we wish to calculate fatigue life via Minerβs Rule, however, the relation is non-linear and
thus must be evaluated in the form:
πΈ[π] = β« π(π)π(π)ππβ
0
P
34
Where π(π) is the expected number of life cycles associated with an arbitrary applied
stress and π(π) is the probability distribution of stress, loosely based off the Weibull model
of windspeed. What this tells us is that highly skewed stress distributions, where the
applied stress in the piezoelectric material is low for the majority of the time, will have
significantly longer fatigue life, as the average is weighed more heavily by high-cycle
fatigue than low-cycle fatigue. In summary, our goal for the design parameters is to
maximize skewness of the stress distribution.
The existence of the cut-in speed is significant in this. Standard methods of stiffening the
cantilever beam would greatly enhance the skew of the distributionβs right tail, but will also
result in a higher cut-in speed, increasing the discrete probability that the stress level is zero
for a given time. The extent to which changes to cut-in windspeed affect stress mean and
skewness are highly dependent on the Weibull model parameters. Therefore, the only way
to fully track these opposing factors is to develop a formula for the mean and skewness of
the stress distribution based on the wind model and the relation between windspeed and
stress. To transform a random variable π into another π through the relation π = π(π)
where π has the probability function π(π), we use the Law of the Unconscious Statistician:
πΈ[π] = β« ππ(πβ1(π₯))ππβ
ββ
Thus we must write the windspeed in terms of the stress:
π = βπ1π2 + π2
Note: the (π1, π2) values here are not the same as those used previously, but both
represent the hyperbolic relation between windspeed and stress. The key distribution
parameters are therefore [21]:
ππππ = πΈ[π] = β« πβ
0
π(π)ππ
= β« πβ
π0
π
ππ(π1π
2 + π2)12(πβ1) exp [β(
βπ1π2 + π2
π)
π
] ππ
ππ‘π. π·ππ£.= βπΈ[π2] β πΈ[π]2
ππππ€πππ π =πΈ[π3] β 3(ππππ) β (ππ‘π. π·ππ£. )2 β (ππππ)3
(ππ‘π. π·ππ£)3
Where:
35
πΈ[π2] = β« π2β
0
π(π)ππ = β« π2β
π0
π
ππ(π1π
2 + π2)12(πβ1) exp [β(
βπ1π2 + π2
π)
π
] ππ
πΈ[π3] = β« π3β
0
π(π)ππ = β« π3β
π0
π
ππ(π1π
2 + π2)12(πβ1) exp [β(
βπ1π2 + π2
π)
π
] ππ
Where (π, π) are Weibull model parameters [21], π0 is cut-in windspeed, and (π1, π2) are
simplified design parameters, though in a different from those derived previously. To try
and solve these integrals, we will begin by writing a formula for the general nth raw
moment:
πΈ[ππ] = β« ππβ
π0
π
ππ(π1π
2 + π2)12(πβ1) exp [β(
βπ1π2 + π2
π)
π
] ππ
Making the following substitution:
π(π) =(π1π
2 + π2)k/2
ππ
πΈ[ππ] = β«π
π1ππβ1π1/k exp[βπ] ππ
β
π(π0)
Note that we can re-write π2 in terms of this new variable π easily:
π2 =π2
π1π2/k β
π2
π1
πΈ[ππ] = β«π
π1(
π2
π1π2/k β
π2
π1)
12(πβ1)
π1/k exp[βπ] ππβ
π(π0)
To solve an integral of this form, we can make use of the upper incomplete gamma function
[8]:
Ξ(π, π) = β« π₯πβ1 exp[βπ₯] ππ₯β
π
Since 1
2(π β 1) is not always an integer, we need to perform a general binomial expansion
which works for non-integer powers:
36
(π₯ + π¦)π = β(ππ) π₯πβππ¦π
β
π=0
Therefore:
πΈ[ππ] = β«π
π1π1/k exp[βπ]β(
1
2(π β 1)
π)
π(πβ1)βπ
π1
12(πβ1)βπ
π1π(πβ1)βπ (β
π2
π1)
πβ
π=0
ππβ
π(π0)
πΈ[ππ] = β« β(β1)π (1
2(π β 1)
π)
ππβππ2π
π1
12(π+1)
π(π/πβπ)
β
π=0
exp[βπ] ππβ
π(0)
πΈ[ππ] = β(β1)π (1
2(π β 1)
π)
ππβππ2π
π1
12(π+1)
β
π=0
Ξ(1 +π
πβ π, π(π0))
Note that this reduces to a finite sum for odd values of n, and for πΈ[π2], we can still
approximate the solution with a finite number of terms.
The next step is to evaluate this formula at π = 1,2,3, solve for mean, standard deviation,
and skewness. Finally, we would be left with two equations:
πΈ[π] = π1(π·ππ πππ ππππππ ) = ππππ π‘.
ππππ€πππ π = π2(π·ππ πππ ππππππ )
Where optimizing skewness while keeping mean stress constant leaves us with a relation
between design parameters which optimizes durability while still maintaining an acceptable
power output. The goal of the user is to first assume all material properties are constants,
and that only the dimensions of the beam, piezoelectric plies, etc, are customizable. This
should provide the user with a single optimal design. The process can be repeated with new
material properties to find a new optimal design. The best of these optimal designs is finally
used and constructed. Unfortunately, the formulae here are highly involved and I was
unable to make further progress on this method. However, the method is now fully laid out
and ready to be implemented.
While the formula for skewness is extremely convoluted, keeping the mean stress constant
is relatively simple.
πΈ[π] =π
π1Ξ (1 +
1
π, π(π0)) = ππππ π‘.
By parametrizing this and solving for the parametric relation:
37
π0 = π(π0) =π2
k/2
ππ, π1 =
π
π1, πΈ[π] = π1Ξ (1 +
1
π, π0)
ππΈ[π]
ππ0=
ππ1
ππ0Ξ (1 +
1
π, π0) + π1π0
1/kπβπ0 = 0,
ππ1
ππ0= β
π1π01/k
exp[βπ0]
Ξ (1 +1π
, π0)
Separating variables and integrating yields:
π1 = Ξ(1 +1
π, π0) , πΈ[π] = Ξ2 (1 +
1
π, π0) = ππππ π‘.
Therefore:
π0 = ππππ π‘, π0 = ππππ π‘, π2 = ππππ π‘.
In short, to keep mean power output constant, we need to keep the cut-in windspeed
constant.
5.2.2.2 βLifetime Efficiencyβ-Based Model
As the probabilistic model results in a difficult solution that is not analytically solvable, itβs
possible that another optimization method could tackle the same issue more simply. This
time, instead of maximizing absolute power output, which is heavily dependent on the scale
of the particular GPEH device and local wind conditions, we could instead try to maximize
a new variable weβll call βLifetime Efficiencyβ, given by:
πΏ. πΈ. = πππππ§πππππππ
Where:
πππππ§π =πππ‘ππ πΈππππ‘πππππ πΈπππππ¦ πΈπ₯π‘ππππ‘ππ
πππ‘ππ ππ‘ππππ πΈπππππ¦
πππππ =πππ‘ππ ππ‘ππππ πΈπππππ¦ πππ πΆπ¦πππ
πππ‘ππ ππππ πΈπππππ¦=
β«12πΈπΌπ (π₯)2 β
ππ
2π12ππ΄π‘π3
And π΄π‘ is the surface area of the tip body, π is the beam curvature, and N is the number of
cycles the device can endure before fatigue failure. In short, this variable represents the
total energy harvested in terms of how many seconds of pure wind power would generate
the equivalent amount of power. When a specific power requirement is not so critical,
designing in terms of energy efficiency can ensure the GPEH device operates cost
effectively even when its output is less substantial.
38
Piezoelectric efficiency is mostly a function of the piezoelectric coupling constants and the
portion of the beam which is made of piezoelectric material, which is not as customizable
as the aeroelastic efficiency, thus most of our efforts will focus on that variable. Using
statics, we can find the displacement across a cantilever beam in response to a tip load in
terms of the tip displacement.
π€(π₯) = π€π‘ππ (3
2(π₯
πΏ)2
β1
2(π₯
πΏ)3
) , π (π₯) = 3π€π‘ππ (1 βπ₯
πΏ)
Using our hyperbolic model of galloping displacement:
π€π‘ππ = π2β(π
π1)2
β 1
Where (π1, π2) are design parameter variables to be examined later on. Plugging this in:
π (ππ‘ππππ πΈπππππ¦) = β«9
2πΈπΌπ2
2 [(π
π12) β 1] (1 β
π₯
πΏ) ππ₯
πΏ
0
=3
2πΈπΌπΏπ2
2 [(π
π22) β 1]
Next, we need to divide this by the period of oscillation:
π =ππ
2π=
π32
2ππΏ2β
πΈπΌ
ππ΄
Where π3 is the non-dimensional coefficient unique to the natural frequency of a cantilever
beam.
πππππ =3
4π
π22π3
2
πππππ΄π‘
(πΈπΌ)3/2
πΏ(ππ΄)1/2
1
π3[(
π
π1)2
β 1]
As we can see, the aeroelastic efficiency is the difference between two inverse polynomial
terms, which results in a well-known graph as shown below:
Figure 5.2.2.2.1: Aerodynamic Efficiency vs Windspeed (Sketch)
π
π π0 β3π0
39
Note the efficiency intercepts zero at the cut-in windspeed, rises sharply to a maximum
efficiency at π = β3 π1, then decays back to zero. Therefore, our device is most efficient at
windspeeds just above the cut-in speed, but as power increases about linearly at higher
windspeeds while available wind power increases cubically, our device becomes noticeably
less efficient.
The next step is to expand the condensed (π1, π2) terms given by:
π12 =
4ππ΄ππππ
πππππ·π‘π1
2, π22 =
4ππππ΄ππ
πππππ·π‘π2
2
Where ππ is the total damping ratio between all mechanical and electrical components, π·π‘
is the aerodynamic chord of the system such that π΄π‘ = πΏπ·π‘, and (π1, π2) are the original
normalized hyperbolic fit constants unique to the equilateral triangular cross-section as
discussed in Section 5.1. The maximum efficiency is then given by:
πππππ,πππ₯ =π3
4πβ3
π22
π12
(πΈπΌ)5/4
(ππ΄)3/4πΏβπππππ·π‘ππ
From this we can see that short, stiff beams will have the highest maximum efficiency.
This makes sense as stiff structures are able to store more strain energy.
Next, we will analyze the N variable. The ASTM-E379-10 standard [24] fits all S-N curves
to the log-log linear fit given by:
log π = βπ logπ + π, π = 10b/aπ-1/a
Where maximum stress is here given by:
π = πΈπ§π (0) = 3πΈπ§π2β(π
π1)2
β 1
Where z is the distance from the center of the cross-section to the extremity. Plugging
everything in, we are finally left with:
πππππ β π = 10b/a(3πΈπ§π2)-1/a
3π22π3
4
π
(πΈπΌ)2ππ
πΏ4(πππππ·π‘)2[
πππππ·π‘
4ππ(ππ΄)πππ12 πβ1 β πβ3]
1β12π
This tells us that the behavior of our lifetime efficiency variable varies depending on the S-
N curve fit. When π >1
2, which is to say that the S-N curve decreases sharply and fatigue
life is less sensitive to applied stress, then the efficiency variable dominates and the shape
of the curve takes the form:
40
Figure 5.2.2.2.1: Efficiency versus Windspeed for Fatigue-Insensitive Materials
When π <1
2, which is to say that the S-N curve decreases gradually and fatigue life is more
sensitive to the applied stress, then the fatigue life dominates and the shape of the curve
takes the form:
Figure 5.2.2.2.2: Efficiency vs Windspeed for Fatigue-Sensitive Materials
41
Note that, as the windspeed approaches cut-in speed, the efficiency approaches zero, but the
number of cycles approaches infinity faster and thus prevails. The location of the extrema
in both forms is the same, but the a-value determines whether it is the maximum or
minimum efficiency. The materials where π <1
2 are clearly the most desired as they
approach infinite lifetime efficiency at high speeds and at low (near cut-in) speeds.
Research on S-N curve fits has shown that aluminum alloys belong to the π < 0.5 class
[27]. As we can see, stiffening the structure will increase all values in either graph:
Figure 5.2.2.2.3: Lifetime Efficiency Graph vs Stiffness
Thus, the optimal strategy after choosing a durable material is to design an infinitely stiff
structure with a high cut-in speed but which performs perfectly for windspeeds above cut-
in. Obviously, this is not always a practical option, and structures seeking to optimize
lifetime efficiency should design their structure to have a cut-in windspeed aligned with an
arbitrary cumulative probability. For example, say we were to design a GPEH device such
that it spent 50% of the time galloping, which we would do by setting the cut-in windspeed
to the mean windspeed.
6.0 Conclusion
The various mechanisms of galloping-based piezoelectric wind energy harvesters were
studied, including aerodynamics, structural dynamics, piezoelectricity, and critically,
fatigue mechanics. These phenomena were modeled numerically and analytically, and
strides toward a formula for durability optimization were tried, focusing on both physical
and probabilistic variables. A code simulation was performed for a basic GPEH device with
varying system parameters and provided useful insights to how these correlate with
performance metrics. This code currently is inefficient and requires significant runtime
improvements, but is able to account for a myriad of factors and in that way offers
something over a standard analytical model.
Stiffness
Stiffness
a > 0.5 a < 0.5
42
While this project intended to conclude by testing the derived optimization formulae with
the program results, all optimization methods proved partially inconclusive, requiring
extensive dimensional analysis to better simplify the problem. However, with the next steps
of the derivation process clearly laid out, this research could be continued in a future
project using all the formulae provided in this paper.
7.0 Bibliography
1 Alonso, G & Meseguer, J. (2006). A parametric study of the galloping stability of
two-dimensional triangular cross-section bodies. Journal of Wind Engineering and
Industrial Aerodynamics, 94(4), 241-253. ISSN 0167-6105.
https://doi.org/10.1016/j.jweia.2006.01.009
2 ASM Material Data Sheet. [Webpage]. www.asm.matweb.com
3 Bathias, Claude & Pineau, Andre. (2013). Fatigue of Materials and Structures. Pg.
238. Wiley. ISBN 1118623371.
4 Bibo, Amin & Daqaq, Mohammed. (2015). An analytical framework for the design
and comparative analysis of galloping energy harvesters under quasi-steady
aerodynamics. Smart Materials and Structures. Vol 24.
5 Daniel, Isaac & Ishai, Ori. (2006). Engineering Mechanics of Composite Materials,
2nd Edition. Oxford University Press. ISBN 0-19-515097-X.
6 Domagalski, Piotr & Sætran, Lars. (2016). Frøya wind data (1Hz). [Data set].
Zenodo. https://zenodo.org/record/3403362#.YJMmEaFOmM8
7 Forman et al. (2005). Fatigue Crack Growth Database for Damage Tolerance
Analysis. FAA. Retrieved 6 July 2019. www.tc.faa.gov/its/worldpac/techrpt/ar05-
15.pdf
8 Incomplete Gamma Functions. Digital Library of Mathematical Functions.
[Webpage]. https://dlmf.nist.gov/8#PT2
9 Liu, A. F. (2005). Mechanics and Mechanisms of Fracture: An Introduction. Figure
A7.10, pg. 406. ASM International.
10 Material Properties. [Webpage]. https://support.piezo.com/article/62-material-
properties
11 Moreto et al. (2015). Environmentally-assisted fatigue crack growth in AA7050-
T73511 Al Alloy and AA2050-T84 Al-Cu-Li Alloy. Materials Research. Vol. 18.
12 Medved et al. (2004). Corrosion Pit Size Distributions and Fatigue Lives.
International Journal of Fatigue, 26(1), 71-80.
13 Okayasu et al. (2010). Fatigue failure characteristics of lead zirconate titanate
piezoelectric ceramics. Journal of the European Ceramic Society, 30(3), 713-725.
14 Papaefthymiou, George & Klockl, Bernd. (2008). MCMC for Wind Power
Simulation. IEEE Transactions on Energy Conversion, 23(1), 234-240.
43
15 Paris, P. C.; Erdogan, F. (1963). A critical analysis of crack propagation laws.
Journal of Basic Engineering. 18 (4): 528β534.
https://doi.org/10.1115%2F1.3656900
16 Salz, Christopher & Hoffman, Mark. (2005). Cyclic Crack Growth in PZT Under
Mechanical Loading. Journal of the American Ceramic Society. 88(5). 1331-1333.
17 Sirohi, Jayant & Mahadik, Rohan. (2011). Piezoelectric wind energy harvester for
low-power sensors. Journal of Intelligent Material Systems and Structures, 22(18),
2215β2228.
18 Solef PVDF Design & Processing Guide. V2.7. Figure 44.
19 Stress Intensity Factor. (May 5, 2021). Engineering Library [Website].
https://engineeringlibrary.org/reference/fracture-mechanics-stress-intensity-factor
20 Sun & Jin. (2012). Elastic-Plastic Fracture Criteria. Fracture Mechanics. 171-187.
21 Weibull Distribution. (2001). Encyclopedia of Mathematics. [Website].
https://www.encyclopediaofmath.org/index.php?title=Weibull_distribution
22 Zeng et al. (1992). A posteriori local error estimation and adaptive time-stepping for
Newmark integration in dynamic analysis. Earthquake Engineering & Structural
Dynamics, 21(7), 555-571.
23 Zuo et al. (2015). Fatigue Life Prediction under Variable Amplitude Loading using
a Nonlinear Damage Accumulation Model. International Journal of Damage
Mechanics, 24(5), 767-784.
24 ASTM E739-10 (2015). www.astm.org
25 Rainflow. V 3.1.0. PyPI. [Github Repo]. https://pypi.org/project/rainflow/
26 Lindfield, George & Penny, John. (2019). 5.11.2: Newmarkβs Method, Solutions of
Differential Equations. Numerical Methods 4th Ed.
27 Yang et al. (2019). Effects of Prebending Radii on Microstructure and Fatigue
Performance of Al-Zn-Mg-Cu Aluminum Allot after Creep Age Forming. Metals.
Vol 9.