Flat Panel Speakers Final Report - Final Draft
Flat Panel Speakers Final Report - Final Draft
Flat Panel Speakers Final Report - Final Draft
Luke Humphry
Kings College
Fourth-year undergraduate project in Group C
2006/2007
I hereby declare that, except where specifically indicated, the work submitted herein
is my own original work.
Signed:____________________________________ Date:____________________
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page i
Technical Abstract
The flat panel speaker is a relatively new technology that disregards the normal rules of
loudspeaker design by encouraging resonance in the speaker cone. These speakers have
been commercially available for some 10 years, yet little is known about their method of
operation. A single company (NXT) licenses flat panel speaker technology to
manufacturers, but has given scant explanation as to how these devices work. Though
there is beginning to be academic interest in these speakers, current computational
techniques (Finite Element Analysis, Boundary Element Method) are extremely slow when
analysing vibrating panels with such high degrees of freedom. Due to the slow nature of
these methods it has not been possible to perform investigative research to understand
the method of operation of such panels. By using a new computational method that
describes the surface displacement of the panel using Jinc function wavelets, the
radiated power spectrum can be computed for a panel in a fraction of time previously
needed. Using this method it has been possible to fully analyse flat panel speakers and
for the first time provide an explanation for how they work to the academic community.
To analyse the acoustical properties of a flat panel speaker, a computer
simulation of such a device was designed and built within MatLab. This model contains
two distinct submodels 1) a simulation of the structural dynamics of a flat panel
speaker, and 2) a prediction of the power radiated from the surface of this panel when
coupled with surrounding air. Both submodels contain simplifying assumptions about the
nature of panels being simulated, but due the separate nature of the submodels it is a
simple procedure to add in more complexity to one area of the model in the future. For a
typical flat panel speaker the complete model can compute a 4000 point power spectrum
over the working range (20Hz 20kHz) in approximately 3 hours. Until recently this would
have taken many days.
This model was used to compute the power spectrum of a real flat panel speaker.
The speaker was dismantled and its panel carefully measured to find out all the
parameter values need to run the model. The panel was found to have high damping and
high modal density, and critical frequency of 23kHz, above the working range of the
speaker. The resulting power spectrum was found to be extremely flat above low
frequencies. By changing the damping of the simulated panel and computing the radiated
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page ii
power spectrum for different damping values, it was found that the real panel was close
to radiating like an infinite panel, with no wave/boundary interactions. For this behaviour
the driving point radiates all acoustic power, acting like a monopole. The real panel
radiated power approximately 3x in magnitude that of an infinite plate. The difference
was found to be due to a reverberant bending-wave field in the panel area around the
drive point. An approximate value for the radiation from this reverberant field was
calculated using Statistical Energy Analysis (SEA), and when added together with the
power radiated from the drive point gave extremely close agreement with the model
predicted power spectrum. These two methods of panel radiation are normal not seen
together, but the combination of high damping (producing infinite plate behaviour), high
modal density (creating a reverberant field), and being below the critical frequency
(meaning no dominant modes) results in both radiation sources being present in this
case. Both sources of radiation have a highly dispersed pattern, a known feature of flat
panel speakers. The dual radiation theory was validated by calculating the SEA
approximate power spectrum of other flat panel speakers and then checking with results
predicted by the full Jinc function model. Excellent agreement was seen for a range of
panels.
The method of operation of a single pair of flat panel speakers is now understood,
and this dual radiation method is thought to hold for a wide range of flat panel speakers.
The SEA approximation provides an extremely quick way of computing the approximate
radiated power spectrum for a flat panel speaker (<15minutes). This approximation is
typically within 3dB of the true value due to the flat nature of the true spectrum and so is
accurate enough for most purposes. The SEA approximation is a new tool in flat panel
speaker design and due to its speed allows for new design methods, namely
optimisation.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page iii
Contents Page
Technical Abstract .................................................................................................................... i
Section 1: Introduction..................................................................................................... 1
1.1 Background................................................................................................................1
1.1.1 Conventional Loud Speakers............................................................................1
1.1.2 Distributed Mode Loudspeakers...................................................................... 1
1.1.3 Current Understanding of Flat Panel Speakers...............................................2
1.2 Motivation for Work...................................................................................................3
1.2.1 A New Analysis Technique ................................................................................3
1.2.2 Explaining the Physics of Flat Panel speakers ................................................ 3
1.3 Research Strategy .....................................................................................................4
1.3.1 Designing a Model Speaker..............................................................................4
1.3.2 Investigate real speakers.................................................................................. 4
1.4 Structure of Report....................................................................................................4
Section 2: Building an Accurate Model ........................................................................... 6
2.1 Structural Dynamics Model ...................................................................................... 6
2.1.1 Theory.................................................................................................................6
2.1.2 Modelling Technique.........................................................................................6
2.1.3 Resulting Investigative Tools ............................................................................8
2.2 Coupling Panel to Surrounding Air ...........................................................................9
2.2.1 Theory.................................................................................................................9
2.2.2 Modelling Technique...................................................................................... 11
2.2.3 Tools Developed............................................................................................. 12
2.3 Optimisations.......................................................................................................... 13
2.3.1 General Optimisation Methods ..................................................................... 13
2.3.2 Outstanding Problems ................................................................................... 14
Section 3: Investigating the Loudspeaker ..................................................................... 15
3.1 Finding Loudspeaker Panel Parameters .............................................................. 15
3.1.1 Dismantling the Speakers ............................................................................. 15
3.1.2 Stiffness Test .................................................................................................. 16
3.1.3 Laser Vibrometer Testing............................................................................... 18
3.1.4 Modal Parameters.......................................................................................... 20
3.2 Model Results......................................................................................................... 21
3.2.1 Radiation Efficiency Plot ................................................................................ 21
3.2.2 Complete Power Spectrum............................................................................ 22
3.3 Explaining the Radiation Method.......................................................................... 23
3.3.1 Infinite Plate Theory ....................................................................................... 23
3.3.2 Effect of Changing Damping.......................................................................... 24
3.3.3 Accounting for Reverberant Field.................................................................. 25
3.3.4 Applying Dual Radiation Theory to Alternate Panels.................................... 28
3.3.5 Computational Use of Dual Radiation Method............................................. 28
3.4 Analysis of Theory................................................................................................... 29
3.4.1 Desirable Properties of DML.......................................................................... 29
3.4.2 Comparison with Conventional Loud Speakers ........................................... 30
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page iv
Section 4: Conclusions................................................................................................... 32
4.1 Novel Modelling Method........................................................................................ 32
4.2 Findings................................................................................................................... 32
4.2.1 Panel Radiation Method ................................................................................ 32
4.2.2 Design Parameters......................................................................................... 32
4.2.3 Dual Radiation Method of Computation....................................................... 33
4.2.4 Comparison with Conventional Loud Speakers ........................................... 33
Section 5: Future Work................................................................................................... 34
5.1 Improving Model Accuracy..................................................................................... 34
5.2 Additions to the Model ........................................................................................... 34
5.3 Alternate Model ...................................................................................................... 35
5.4 Panel Optimisation................................................................................................. 35
Section 6: Acknowledgements....................................................................................... 36
Section 7: References.................................................................................................... 37
Section 8: Appendices.................................................................................................... 38
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 1
Section 1: Introduction
1.1 Background
1.1.1 Conventional Loud Speakers
Electrically powered loudspeakers have existed for over 100 years, and are omnipresent
in our lives we find them in laptops, radios, mobile phones, train address systems to
name only a few. Conventional loudspeakers rely on piston-like cones to provide the
mechanical motion that creates sound waves. This is not new technology it was in
1925 that C. Rice and E. Kellogg first devised the piston-cone arrangement that
dominated loudspeaker design for the rest of the 20
th
century. Much academic and
commercial effort has gone into making these cones as rigid as possible through
improved materials and adhesives, eliminating any resonances. These resonances are
the main cause of unwanted audio colouration and smear.
Despite the age of this technology the design is by no means perfect and piston-
cone loudspeakers suffer from one key problem: at high frequencies the sound radiated
tends to become very directional they beam. This means that it is extremely difficult
to evenly fill a space with sound radiated from conventional speakers, especially given
the variations in directivity at different frequencies.
1.1.2 Distributed Mode Loudspeakers
10 years ago the first Distributed Mode loudspeakers (DMLs) were developed by New
Transducers Ltd (known as NXT). These speakers turn convention on its head by actively
encouraging resonances in the loudspeaker cone. A thin flat panel is used and is
generally excited using a magnetic coil to set up bending waves throughout the panel.
These DMLs (or flat panel speakers as they are often called) are the first new idea in
speaker technology in over 60 years, and have some exciting properties: a very
dispersed radiation pattern at all frequencies; and their slender dimension. Figure 1
below shows the dispersed nature of this new technology.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 2
Figure 1 Typical Radiation Patterns for Conventional and Distibuted Mode Loudspeakers
The sound quality of flat panel speakers is not yet as good as the best piston-cone
speakers, but their slender dimensions and ability to evenly fill a room with sound are
resulting in them being used widely for home cinema, computer audio, and public
address systems.
1.1.3 Current Understanding of Flat Panel Speakers
Though NXT design and licence DML technology, they give little explanation as to the
physics of DML sound radiation. The most in-depth description of the theory comes in
NXTs original introductory article for the audio industry [Azima, 1998]. Here they state
that by exciting such panels so as to set up bending waves, a series of complex and
highly random waves are produced. These waves are so complex as to be almost
random, and so each point on the panel acts like a small independent cone radiating
uncorrelated sound waves. Due to the complex and uncorrelated nature of resonance,
adjacent sound waves do not cancel one another out, and so we observe a very diffuse
radiation pattern, approaching that of a point source.
There is beginning to be academic interest in this kind of loudspeaker, but there
is not thought to be any further explanation as to how these flat panel speakers work
other than that given by NXT. Most academic work has been concerned with novel uses
for this technology and not in understanding the underlying physics.
Conventional loudspeakers show a high
degree of directivity, beaming at high
frequencies.
Distributed mode loudspeakers show a far
more dispersed radiation pattern.
[Images from Elac Electroacoustics Website]
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 3
1.2 Motivation for Work
1.2.1 A New Analysis Technique
To analyse a thin walled structure such as a flat panelled speaker it is natural to analyse
the air-structure coupling to calculate the sound power radiated into the surrounding air.
In the past two decades it has been convention to analyse complex structures using
either the boundary element method (BEM) or finite element method (FEM). These
methods are well understood and can be applied to a large number of general
structures, but even with advances in computing power they remain rather slow. This is
especially problematic for high frequency analysis, where short wavelengths necessitate
very many degrees of freedom.
A new method has been developed that uses Jinc function wavelets to analyse
the power radiated from planar structures [Langley, 2007]. This method is not as
general as BEM or FEM but for simple planar structures allows for much faster analysis
of the acoustical power radiated from their surface. We shall not discuss the Jinc
function method in any depth, but in summary the method is to describe the surface
displacement of the structure by using a grid of Jinc function wavelets [see Appendix 1
for description of Jinc function]. These wavelets are radially symmetric, allowing for a
simple and analytic calculation of the dynamic stiffness of the structure, from which the
radiated acoustic power can also be calculated analytically. It is this analytic nature that
allows for much faster computation than all previous methods and suits the analysis of
flat panel speakers well.
1.2.2 Explaining the Physics of Flat Panel speakers
Though flat panel speakers have now been on the market for over a decade, there has
been no clear explanation for the governing physics. NXT likely understand the physics,
but their brief explanation is without any physical or mathematical rigour to back up their
claims. These loudspeakers do reproduce sound well, as countless commercial products
and reviews testify, but prior to this research no-one in the academic community
understood why. NXT are an intellectual property based company, and seem to be
working hard to ensure their I.P. is kept secure to the point where DML license holding
manufacturers cannot advertise the full specifications of their DML products. Through
use of the fast Jinc function method it was felt that the time was now right to investigate
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 4
how these loudspeakers operate and try to provide an explanation to the academic
community at large.
1.3 Research Strategy
1.3.1 Designing a Model Speaker
It is rather difficult to analyse DMLs experimentally, as has been documented previously
[Hill & Mapp, 2003]. For this research an accurate computer model of a flat panel
speaker was designed. Firstly, a model that simulated the structural dynamics of a flat
plate was needed how the plate responds to a given force input. A program was
written that simulated the displacement response to a given harmonic force input at any
frequency. This model was made to be general enough to allow for many changes in
plate properties. To predict the acoustic power radiated from such a plate, the model
plate was coupled to the surrounding air space. This involved using the Jinc function
method to predict the dynamic stiffness and so the acoustic power.
1.3.2 Investigate real speakers
Having designed an accurate model, a real set of flat panel speakers was dismantled
and measured to find out the physical properties of a typical set of flat panel speakers.
These properties were fed into the model to allow the acoustic power to be predicted.
This led to some interesting discoveries and theories. These theories were investigated
and validated by changing various properties of the model speakers and re-analysing
the predicted acoustic power output.
1.4 Structure of Report
This report documents the stages of development of the research introduced previously,
and the resulting conclusions. The research consisted of 2 main phases, and so shall
this report:
1) Designing and building the dynamic response model of a flat panel speaker, and
coupling this to the surrounding air
2) Using this model to analyse flat panel speakers
The first section shall give the underlying theory behind the models used, how the theory
was implemented, and what investigative tools result from each model. The second
section will deal with four issues: how the parameter values of a real flat panel speaker
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 5
were found, what results the model predicted for these values, what the implications of
these results are, and how any theories about flat panel speakers can be validated. In
academic papers it has become common to refer to flat panel speakers as distributed
mode loudspeakers (DMLs), and this term shall be used from here on in this document.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 6
Section 2: Building an Accurate Model
2.1 Structural Dynamics Model
2.1.1 Theory
A DML is little more that a flat plate with some boundary conditions given by the frame it
rests in. Simply supported plates have the simplest mode shapes of all plates, so it is
sensible to begin by modelling this type. The model was designed to be general enough
to allow different edge conditions to be imposed in future, with all subsequent parts of
the model still performing correctly. A simply supported flat plate has mode shapes and
frequencies given by [Crighton et al, 1992, p589]:
2
2
, sin sin ) , (
=
y x
nm
y x
nm
L
m
L
n
h
D
L
y m
L
x n
y x
(1a,b)
where L
x
and L
y
describe the planar dimensions of the plate, h the thickness, the
density of the plate material, and D the bending stiffness. The response of this plate to a
forced sinusoidal displacement, F, at a point (x
o
, y
o
), and at some frequency, , can be
found using a summation of mode shapes:
( )
+
=
n m nm
o o nm nm
i
y x y x F
y x u
2 2
1
) , ( ) , (
) , (
(2)
where defines the coefficient of damping of the plate. It is convention to normalise the
mode shapes by the system mass before using the infinite summation. All modes were
multiplied by a normalising factor before summation:
y x
L hL
NF
2
= (3)
2.1.2 Modelling Technique
Using MatLab as our programming environment, it is simple to calculate the mode
shapes given by Eqn. (1a) directly. But to compute the panel displacement at any given
frequency how do we decide the limits to use for our summation in Eqn. (2), which is in
theory infinite?
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 7
Using a wavenumber space approach to represent different modes, it is known
that for modes close to frequency the mode with the maximum n value occurs when m
is 1, and vice versa [Norton & Karczub, 2003, p210], see Figure 2 below.
Figure 2 Modes in Wavenumber Space
It was found that to calculate the response of the plate one should sum over modes up
to and including
1.5n(max),1.5m(max)
, subsequent modes making no practical difference to
the response due to the factor
nm
2
becoming extremely large and further away from
2
.
The values n(max) and m(max) can be calculated by rearranging Eqn. (1b) and setting m
or n to 1 respectively.
A program has been written which lays a mesh on the plate and computes all the
mode shapes needed to find the response of a plate up to a given maximum frequency.
The frequency range of interest to us is 20 20,000 Hz, the range of human hearing.
The mode shapes are computed and then stored in a 4D array (n x m) x (x x y) to
allow them to be called quickly for later parts of the program, namely to find the
response at any given frequency using the summation over modes given by Eqn. (2). For
a maximum frequency of 20kHz, we need to perform a summation of the order of 40x40
(nxm), dependent on plate parameters. When calculating the response at high
frequencies this summation is rather time consuming, but the program has been written
to sum over fewer modes at low frequencies dramatically speeding up calculations.
When calculating the response at a given frequency, a unit input force is assumed, ie
F = 1 in all cases.
For a frequency , modes with a
resonant frequency near mark out
an arc in the wavenumber space.
Modes with maximum values for n
and m are given along kx and ky axes
respectively.
n(max)
m(max)
[Original Image reproduced from Norton and Karczub, 2003]
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 8
2.1.3 Resulting Investigative Tools
Using this model for the response of a plate, three key tools were developed to aid
investigation of DMLs. The first is an animation (slowed down) of the harmonic response
at any frequency. This allows us to view the deformation and wave patterns of the plate.
This animation is simply performed by multiplying the complex displacement given by
the model with a rotating exponential function, and plotting only the real part of the
resulting displacement. The second tool is also an animation but this time animates the
displacement shape as we scan up in frequency. This is extremely instructive in showing
us how the nature of deformation changes with frequency. To perform this animation we
must re-compute the response at each frequency step, and plot the resulting real
displacement. At all frequencies the level of displacement is rather small (<0.1 mm for
200 x 200 mm panel) so the animation displacements are scaled to give the viewer a
clear idea of the mode shapes and wave patterns.
Figure 3 Snapshot of Harmonic Animation Tool
The third tool is one that computes the mobility transfer function from one point on the
plate (force input) to another (velocity output). This lets us quickly visualise what the
plate is doing at various frequencies, and gives an insight into some of the key
parameters such as modal density and modal overlap factor. To validate the model we
used the transfer function. We first computed the parameters modal density, modal
overlap factor, and critical frequency. These are dependent on the plate dimensions and
This snapshot is taken from the MatLab animation tool developed to view the response of a
panel. The above response is found for a 20x12x0.1cm steel panel being excited at 2000Hz.
The x and y axis describe the mesh grid position, and the z axis describes the normalised
plate amplitude.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 9
materials and can be easily calculated [see Appendix 2 for equations]. These
parameters are the basis for statistical element analysis (SEA) and have allowed the
model to be verified as correct against SEA predictions [see Appendix 3 for one such
verification]. The model and SEA predictions are in agreement and so we can trust the
model output.
2.2 Coupling Panel to Surrounding Air
2.2.1 Theory
When analysing the acoustical properties of a vibrating structure, it is the power radiated
as sound waves into the surrounding air that is of key interest. The Rayleigh integral
describes the pressure in space as a result of a vibrating planar structure, from which
we can calculate the total power radiated:
x x x x x =
=
r d w
r
e
p
A
r ik
a
a
, ) (
2
) (
2
(4)
The pressure, p, at a location x in space is defined in terms of the complex displacement
w(x), the air density
a
, acoustic wave number k
a
, and the frequency of vibration .
Calculating the power radiated by integrating over the volume is cumbersome and time
consuming, but we can greatly simplify the problem by using the Jinc function method.
We can calculate the time averaged power radiated by the plate by using the following
equation:
) (
2
, ) Im(
2
*
n n
T
u P x a a D a = =
(5a,b)
where a
n
is a generalised coordinate, calculated trivially from the displacement u(x
n
)
given in Eqn. (2) by employing a Jinc function representation. For a baffled plate the
imaginary part of the dynamic stiffness matrix, D, can be represented as:
c
k r r k
k
k c
D
s j i ij ij a
s
a a a
ij
= = = 2 , ), ( sinc
8
) Im(
4
2
x x
(6a,b,c)
where
a
, c
a
, k
a
, c define respectively: the density of air; speed of sound in air; acoustic
wave-number; and size of mesh spacing. The variable r
ij
is
the distance between two
points i and j on the plate. It is as a result of the Jinc function being radially symmetric
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 10
that we can use this simple formula for calculating the dynamic stiffness matrix, and so
simply calculate the radiated power. This method computes the total power radiated in
the near field, by using the air pressure at every point on the plate due to surface
displacements. The total power in the near field is necessarily the same as the total
power in the far field, and so this method provides a quick way of computing the far field
power, which is of interest when analysing the acoustical properties of DMLs.
When analysing the power radiated from a flat plate it is instructive to have some
idea of the region in which it is operating. The key parameter here is the critical
frequency, the frequency at which the acoustic wave number equals the plate bending
wave number. The combination of critical frequency and excitation frequency tell us how
efficiently different modes will radiate. Maximum efficiency occurs when modes are
excited at their coincidence frequency or above (see Figure 4 below), coincidence
frequency being the frequency at which bending waves fit acoustic waves, ie their wave
numbers match.
Figure 4 Typical Radiation Efficiency for a Single Mode
If we excite a plate below the critical frequency we find that modes with high radiation
efficiency resonate well below the excitation frequency and so do not have high
amplitude. Resonant modes have high amplitude but low efficiency, and so no particular
coincidence
frequency
Low efficiency
below coincidence
High efficiency
above coincidence
n for
n > critical
n for
n < critical
A typical mode has low radiation efficiency below its coincidence frequency. If
resonance occurs below coincidence, low efficiency will occur at resonance. The
mode will resonate below its coincidence if n > critical. Similarly, modes that
resonate above critical will resonate with high efficiency. The plot shown is for the
first mode (1,1) of a 19cmx19cm steel plate.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 11
modes dominate the response. If we excite the plate above the critical frequency
however, we find the resonant modes have high amplitude and high efficiency and so
will be dominant. It is this change in behaviour that makes the critical frequency of such
importance to the investigation of flat plate sound radiation.
2.2.2 Modelling Technique
A program has been written in MatLab to perform the calculations defined in Eqns. (5)
and (6) to find the power radiated from the plate at a given frequency. It is a testament
to the simplicity of the Jinc function method that once the displacement u(x
n
) has been
found using the previous model, only a few lines of codes allow us to find the power
radiated. To populate the matrix D it was decided to find a matrix R such that
r
ij
= |x
i
x
j
|. Originally this was performed as a double for-loop but this proved extremely
slow. R has dimensions of (a x b) x (a x b), where a and b are the number of points in the
mesh in the x and y directions respectively. For a mesh of 20 x 20 this leads to dim(R) =
400 x 400. The final size of R is proportional to the mesh resolution c to the power 4,
and so not only is R very large, but it can increase in size very quickly for a small
increase in resolution. A fast way of computing R was needed.
MatLab is fast at computing matrix calculations, and copes well with complex
numbers, but is rather slow at for-loops. To avoid using for-loops a program to populate
R was developed that described the x and y position of each mesh point by using real
and imaginary components, and used MatLabs in-built (and highly optimised) abs
function to find the distance between pairs of complex points [see Appendix 4 for
program code]. This neat method resulted in a speed up of over 100x when computing
R, and so allowed the power at any given frequency to be calculated in a fraction of a
second. The matrix R was stored, meaning that after the first power calculation the
power could then be calculated at other frequencies by recomputing u(x
n
) and using
Eqns (5) and (6) with R as before.
To validate the power calculations, plots of non-dimensional frequency (k
non-d
)
versus non-dimensional radiation efficiency () were found for the first 6 modes of
vibration. K
non-d
and are found directly using:
2
2
2
2
*
3 2
4 2
, ) ( sinc
8
y x
a
d non a
T
y x a
L
m
L
n
c
k k
L L c
c
+
= =
a R a
(7a,b)
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 12
where L
x
and L
y
define the plate dimensions in both planar directions, and all other
symbols are as described previously. We can see that is directly proportional to the
power calculated using Eqns. (5) and (6) and so if we calculate using the same steps
as we do the power then is a suitable variable to validate our coupled model. Plots of
vs k
non-d
were performed for a square steel plate and the results validated against
benchmark results found using the FFT method [Williams & Maynard, 1982]. The model
and benchmark results show excellent agreement and so we can take the model as
predicting the radiated power accurately [see Appendix 5 for graph].
2.2.3 Tools Developed
Using the air-coupled model of a DML, two key tools were developed to aid investigation
of these speakers. The first is a program that computes and plots the power transfer
function, found by calculating the radiated acoustic power at a series of frequencies.
This allows us to see what kind of acoustical response different DMLs have, ie smooth
and flat (this is desirable for loudspeakers) or modal and jagged. Initially a strange
phenomenon was observed at a high frequency the transfer function would suddenly
rise in amplitude. The reason for this is not due to the physics of the plate, but is in fact
due to a breakdown in the Jinc function method. This occurs when the mesh grid being
used over the plate is of too large a spacing to describe high wave numbers.
Figure 5 Breakdown of Jinc Function Method at High Frequencies
The Jinc function method breakdown is seen as a large rise in predicted
radiated power at 14kHz. This breakdown occurs when the wavelength
of resonant modes is shorter than half the grid spacing of the wavelet
mesh used to describe the plate.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 13
The Jinc wavelets can only be used to describe bending waves that have a half
wavelength larger than grid spacing of the mesh. A conservative value for the maximum
frequency of bending waves that can be depicted using the mesh is given by:
2
2 / 1
max
=
c h
D
(8)
For the plate used to compute the power spectrum shown in Figure 5 above, this
maximum frequency turns out to be
max
= 14kHz, above which the spectrum goes awry.
A simple check was added to the program to stop the spectrum being computed above
this breakdown frequency, ensuring accurate (but sometimes truncated) results. This
model takes anywhere from a few minutes to a few hours to produce a 4000 point
power spectrum. The variation in time is as a result of the number of mesh points. Even
when taking a few hours this is far faster than FEA or BEM computations. The speed of
computation allows for many models to be run in order to fully analyse the effect of
different parameters on the power spectrum.
The second tool is one that computes the non-dimensional radiation efficiency for
each mode, and then plots this as a surface in the modal domain (ie plots
n,m
at (n,m)
for all modes used within our model). This allows us to see how close each mode is to
resonating at maximum efficiency, ie when
res
=
coincidence
. By visualising this we see
which region of operation the plate will be in at low and high frequencies, ie is
res
greater or less than
critical
, with the repercussions resulting from being in either region
discussed in section 2.2.1. We can use this tool to understand the mode of operation of
DMLs. This tool typically takes around 15minutes to run, dependent on the number of
modes used within the model and so allows for quick analysis of many different panels.
2.3 Optimisations
2.3.1 General Optimisation Methods
To ensure that the models developed were of practical use when investigating DMLs, the
models needed to be fast. Conventional computational methods such as FEM or BEM
can take days or even weeks of continuous calculation, we needed to do this in minutes
or hours. Various common MatLab optimisation techniques were used to speed up the
code. These included:
1. Vectorising equations where possible to avoid the use of for-loops.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 14
2. When for-loops were necessary, rewriting functions in-line to avoid the overhead
of calling these functions
3. Splitting in-built functions into sections that must be included within for-loops,
and sections that can be performed out with the loop, to avoid unnecessary
re-computation
4. Using an optimised package of exponential based functions (exp, sine, sinc etc).
These improve upon MatLabs inbuilt functions by not being so general purpose
they are optimised for double type data only.
These optimisations saw a marked improvement in the speed of the programs, the total
time needed to compute a power transfer function was decreased by 10x.
2.3.2 Outstanding Problems
There are two outstanding problems with the code in terms of how well optimised it is.
The first is that it can require a large amount of RAM. This is as a result of the size of R,
which is regularly over 250MB in size. Thus, for Eqns. (5) and (6) we need approximately
4x250MB = 1GB of free RAM to allow the power prediction program to run completely
within memory, which is critical to ensuring fast computation. Few desktop computers
have enough free RAM to run these programs in this manner, and so a server was used
that contained 6GB, thankfully more than enough for the plates investigated.
The other outstanding problem is that MatLabs sinc function is not very fast and
is the bottleneck in computing Eqns. (5) and (6). This has been improved by optimising
in the manner discussed in the previous section but even after re-writing sinc inline and
using an optimised sine function, computing sine(k
a
R) takes 75% of the time of the
whole program. This is clearly the first point of call if the program needs to be optimised
further, though the author has reached the limits of his optimising tricks.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 15
Section 3: Investigating the Loudspeaker
3.1 Finding Loudspeaker Panel Parameters
3.1.1 Dismantling the Speakers
To investigate DMLs it was necessary to determine the parameters of some real
speakers to feed into the model. A pair of second hand Packard Bell FSP-100 DMLs
were sourced and carefully taken apart [see Figure 6 below]. Each speaker is comprised
of:
1. A stiff plastic panel, with fake dust cover on front side
2. A rigid steel frame, with concave perforated back plate
3. Electrically driven magnetic voice coil, positioned centrally
The plastic panel is a section of two-ply corrugated polypropylene board, corrugations
running lengthways along the panel. This suggests that the panel will not have uniform
stiffness, as assumed in the model so far. The panel is simply glued around its outer
edge into the frame.
Figure 6 The Specimen DML
The panel was prised out of the frame using a flat blunt knife, and care was taken to
avoid deforming either the panel or frame in the process. The mass of the steel frame is
sufficiently large compared with that of the panel that we can assume it is rigid in space
during operation. Also, the glue that holds the panel in place is not very strong, and is
A Packard Bell FSP-100
flat panel PC speaker.
Dimensions approximately
20x12cm. Pen shown for
reference.
Perforated concave metal
back plate. Voice coil
shown to top right. Voice
coil sits in circular recess
in centre of back plate
The thickness of the
speaker reaches a
maximum of 15mm in
centre of concave back
plate.
Corrugated plastic panel.
Circular mark seen in
centre due to voice coil.
Soft synthetic cloth dust
cover covers front side of
panel and is glued tight on
back as shown.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 16
only applied on the outer few mm of the panel, so our assumption of a simply supported
boundary condition seems plausible. One should note however that the weak glue will
likely be dissipative and result in a large degree of boundary damping. The panel was
measured and the resulting parameters are shown in Table 1 below:
Table 1 Specimen Panel Parameters
Parameter Symbol Value
Length
Lx
187 mm
Width
Ly
~128 mm *
Excitation Point (xo, yo) (93,64) mm
Thickness
h
4 mm
Mass
M
18.25 g
(Density) () (194 kg/m3)
* The panel has long edges that are slightly concave width varies between
121mm in centre and 131mm at top and bottom
The density was assumed to be uniform and calculated using the other measured
parameters. The point of excitation was found to be in the very centre of the panel.
3.1.2 Stiffness Test
A key parameter needed in the model is that of D, the bending stiffness. This was found
by performing a standard bending stiffness test applying a load and measuring the
resulting deflection [see Figure 7 below]. This was performed in both directions, x and y,
to find the variation in stiffness between the two directions due to the plastic corrugated
ribs. Care was taken to apply an even load across the free end by using a rigid brass
plate to allow the panel to be treated as a beam. Deflection measurements were taken
quickly to avoid errors due to a small amount of plastic deformation that was observed.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 17
Figure 7 Images of Panel Stiffness Test
If the load is applied evenly and the panel treated like a beam, we can use the
equation for a clamped beam:
3
3
x
y x
Wl
EI L D = = (9)
where W is the applied load, the resulting deflection, and l
x
the length of the free
section of the panel. Similarly this equation can be applied in the y direction. By plotting
against W and taking the gradient we can then gain an estimate for the bending
stiffness D. Plots in both directions are shown below in Figure 8; as expected the
relationship is remarkably linear.
Brass plate placed at edge of panel to ensure
even spread of load. Load applied at midpoint
of panel edge. An evenly spread load avoids
any torque and allows bending beam
equation to be used.
Load applied at edge of panel and resulting
displacement measured in vertical direction.
Measurement taken quickly to avoid plastic
deformation.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 18
Figure 8 Results From Stiffness Test Deflection vs Load
The deflections measured are far greater than those experienced during normal
operation of the DMLs, but the plots are linear enough that we can assume the stiffness
calculated here will be the same for small deflections. The results in both directions are
shown below:
Table 2 Panel Stiffness Values
Direction Stiffness Symbol Value (Nm)
Lengthways
D
x
0.80
Widthways
D
y
0.22
Average
D
0.51
As expected the stiffness is different in each direction, though in the same order of
magnitude. To simplify future analysis we assume that the stiffness D is uniform, and
has a value that is the mean of the values found experimentally.
3.1.3 Laser Vibrometer Testing
The final parameter that is needed to run the model is the damping coeffiecient, , of
the panel. To find the damping of a vibrating structure it is common to excite the
structure and perform modal analysis on the resulting transfer function. The panel is
extremely light and so attaching even a small accelerometer will result in too large a
A linear relationship is seen in both planar directions. Lines of best fit and equations
are shown above. The predicted y axis intercept approaches zero for both cases, and
so we can assume that stiffness is also linear at smaller deflections than the range
measured experimentally.
y =0.0076x - 2E-05
y =0.013x +8E-05
0
0.005
0.01
0.015
0.02
0.025
0 0.5 1 1.5 2 2.5
Load (N)
D
e
f
l
e
c
t
i
o
n
(
m
)
widthways
lengthways
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 19
change in mass for accurate results. With this in mind, a laser vibrometer was used to
measure the surface response due to excitation, in this case performed by a short tap
from an impulse hammer. The hammer tap was applied at the same point in the panels
centre that the electromagnetic voice coil had been attached, to try and excite the same
modes in testing as in normal use.
Figure 9 Laser Vibrometer Testing
The laser vibrometer measured the response of the central point ie a driving point
response. The model predicted and experimental mobility transfer functions agree well
for low frequencies [see Appendix 6]. All predicted modes within the model are shown
experimentally, with small differences in resonant frequency being due to the different
edge conditions imposed upon the panel for each method. Some extra modes are found
experimentally that are not shown in the model spectrum. Some of these are anti-
symmetric modes that occur because the impact did not occur exactly on the centre of
the plate, and so were excited experimentally but not within the model. The few
remaining unaccounted modes are due to resonances of the whole system, ie the
system of the panel and the elastic suspension ropes. At frequencies above 1khz the
experimental spectrum becomes less clear and unsuitable for modal analysis.
Panel suspended by elastic chord
through top corners. Laser aimed at
reflective sticker placed in centre of
panel.
Impact hammer used for test, released
to swing under gravity to give impulse.
Tip of medium hardness used.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 20
We can use the low frequency region to find the damping of the panel. This was
performed by looking at the sonogram of the first few seconds after impact and
calculating the damping at a range of frequencies by using the exponential decay of
each. The damping decreases with increasing frequency, to an asymptote of
approximately 0.014 [see Appendix 7 for detail]. It is thought that the high damping at
low frequencies is predominantly due to boundary damping (longer wavelengths are
affected by boundaries more strongly), and that at high frequencies the internal
damping of the panel dominates. In this experimental case the boundary damping is due
to friction and small impacts of the elastic cord in each corner hole. When the panel is in
the speaker frame the boundary damping will be caused by the glue holding it in place
along its edge. This is very tacky and will be highly dissipative, greatly increasing the
damping at low frequencies. The internal damping that dominates at high frequencies is
high for a plastic panel (0.014). This was thought to be due to the fake dust cover on the
surface of the plastic panel. The soft fabric is pulled tightly over the front of the panel
a crude damping treatment as well as an aesthetic addition.
The model uses a constant damping value, and this was assumed to be the
mean of the damping values recorded experimentally = 0.05, but in reality the panel
will have higher damping at low frequencies (due to glue on edges) and lower damping
at high frequencies (internal damping).
3.1.4 Modal Parameters
Some parameters that give us an insight into how the panel is operating are those used
within SEA, specifically modal density, n
d
, critical frequency,
c
and modal overlap factor,
m
d
. For a flat plate these can all be calculated directly [see Appendix 2]. Using the
dimensions and material values found for the real panel, these modal parameters were
found, see Table 3 below.
Table 3 Modal Parameters for Experimental Panel
Parameter Name Symbol Value
Modal Density
n
d
0.0152 modes/Hz
Critical Frequency
c
23.1kHz
Modal Overlap Factor
m
d
7.61 @ 10kHz
The panel has a very high critical frequency, slightly above that of the operating region of
loud speakers. These means that all modes will be operating below coincidence, though
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 21
at high frequencies some modes will be approaching coincidence. The modal density is
high, with approximately one mode every 65Hz. This results in a high value for the modal
overlap factor: we expect to see a smooth transfer function especially at high
frequencies. It should be noted however that due to the driving point being positioned in
the centre of the panel anti-symmetric modes will not be excited and so the modal
density in practise will be half that calculated here.
3.2 Model Results
3.2.1 Radiation Efficiency Plot
As found in section 3.1.4 the panel operates with all modes below coincidence. To see
exactly how close to coincidence each mode operates, the tool that computes a plot of
non-dimension radiation efficiency for each mode was used. The resulting image is
shown below:
Figure 10 Radiation Efficiency of Modes of Experimental Panel
This plot shows the radiation efficiency
n,m
for all modes used within the model (ie
1.5n(max) x 1.5m(max) ), and thus shows many modes that are above coincidence.
These modes, however, resonate well above 20kHz and so will not contribute much to
the overall response. The important point to note from the plot is that the transition to
coincidence at 23kHz is very steep, and so all modes below 20kHz have a comparable
The above surface plots shows a ridge at 23kHz, where modes resonate at their coincidence
frequency. All modes below 20kHz are shown in the dark blue region, resonating well below
coincidence. Modes at 20kHz are found on the n axis at (24,1), and m axis at (1,16).
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 22
low radiation efficiency value. Thus, there should be a fairly even value of radiated
power across a large frequency range, with all resonant modes radiating at a similar
efficiency below coincidence. This would not have been the case had the critical
frequency been below 20kHz. We also see from the plot the high modal density, with
approx 300 modes below coincidence at 23kHz. Again this will result in a smooth power
spectrum, as resonances merge into one another.
3.2.2 Complete Power Spectrum
The key result for analysing the method of operation of vibrating panels is the power
spectrum for radiated power. Using the tools developed previously this was computed
for the real panel parameter values found in section 3.1. The resulting spectrum is
shown below:
Figure 11 Modelled Power Spectrum of Experimental Panel
The key features of this spectrum are:
1. Distinct modes below 1kHz
2. A very flat response above 3kHz, almost horizontal at 10kHz
3. A gentle rise in radiated power at high frequencies
The first feature is to be expected, as at low frequencies the modal overlap factor is very
low giving rise to distinct modes. The last of these features can be explained by looking
Model predicted power spectrum for experimental panel. After an initial low frequency
modal region the spectrum is remarkably flat.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 23
at radiation efficiencies in Figure 10 above. At high frequencies our modal summation
will include some modes that are approaching coincidence, and so the overall radiated
power will increase. What is interesting here is quite how flat the spectrum becomes at
10kHz.
3.3 Explaining the Radiation Method
3.3.1 Infinite Plate Theory
One possible mode of operation of the panel is that it has such high damping that it
radiates much like an infinite plate. For an infinite plate (or finite plate with extremely
high damping) the waves that are created by the excitation point decay quickly and do
not interact with a boundary. The resulting wave pattern is of all waves travelling
outwards like ripples in a pond. In this scenario all sound power will be radiated from the
driving point, which will act as a point source. If the DML operates as a point source it
explains why we do not see any beaming for this type of speakers.
For an infinite plate radiating below its critical frequency, the power radiated from
the driving point can be calculated directly [Fahy, 1987 p94]:
2
2
inf
) ( 4 h c
F
P
a
a
= (10)
where F is the force of excitation (assumed to be 1 for all calculations). This value was
calculated using the experimental panel material properties, and was compared with the
model predicted power:
Figure 12 Power Spectrum and Infinite Plate Radiation
The power radiated from an
infinite plate is constant at
all frequencies, radiating
from the driving point like a
point source. The power
from an infinite plate with
the same properties as the
experimental panel is lower
than that of the panel but in
the same order of
magnitude.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 24
Though the power from an infinite plate is of a similar magnitude to that of the model,
the model is approximately 3x greater. Some other method of radiation is occurring, but
of comparable magnitude to that of an infinite plate.
3.3.2 Effect of Changing Damping
Previous results showed that the panel radiates power of a level comparable to that of
an infinite plate. It was questioned as to exactly how close the behaviour of the panel
was to that of an infinite plate. An infinite plate can be realised in practise by using a
finite plate with extremely high damping. With this in mind, the damping of the model
panel was varied from 0.25% to 50% and the radiated power calculated using the power
spectrum tool. The mean power in the region 7.5 17.5kHz (the flat region) was
calculated for each panel and plotted against damping.
Figure 13 Power vs Damping for Experimental Panel
At high damping values the radiated power tends to a limit. The relationship of radiated
power to damping was modelled as an inverse power law plus a constant offset, and
using a curve fitting tool this offset was found to be 4.8x10
-4
W. This agrees favourably
with the infinite plate value found using Eqn. 10, Pinf = 4.6x10
-4
W. The graph shows us
that the damping value of 0.05 used within the model of the real plate puts us in a
region that is quite close to exhibiting infinite plate behaviour in terms of the power
Inverse power law observed between radiated power and damping. An
asymptote of P = 4.8x10
-4
W is found at high damping values.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 25
radiated. It was thought that the radiation from the driving point must account for at
least some of panel radiation.
By using the animation tool we find that the real panel is vibrating in a rather
special manner in the flat region:
Figure 14 Experimental Panel Response at 10kHz
Thin plates usually vibrate in one of two distinct manners: if damping is high then the
response is like an infinite plate, or if damping is low then the waves reflect off edges to
create a semi-random reverberant field. The reverberant field consists of multiple
uncorrelated modes superimposed on one another. The direct field experienced in an
infinite plate can be thought of as consisting of multiple modes with high correlation ie
all tracking the central driving point. The model panel exhibits both types of behaviour:
the driving point being much higher in amplitude than the surrounding displacements,
but the surrounding displacements being semi-random and all of similar amplitude.
Does the reverberant field radiation account for the difference between our model
predicted radiation and that of an infinite plate?
3.3.3 Accounting for Reverberant Field
To analyse the power radiated by the reverberant field a method for calculating it was
needed. The reverberant field radiation at any frequency can in fact be calculated
directly using an SEA approximation [Lyon R, 1975, p234]:
The panel exhibits infinite plate behaviour (large response at driving point) and also reverberant
behaviour (semi-random wave pattern in surrounding region).
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 26
p rad rev
E P = (11)
where P
rev
defines the acoustic power radiated due to the reverberant field,
rad
the
radiation loss factor resulting from coupling the panel with the surrounding air, and E
p
the energy within the vibrating plate. The energy can be eliminated by noticing that a
similar SEA equation holds for the power input to the plate at the drive point:
in
rad
rev p in
P P E P
= = (12a,b)
where refers to the plate loss factor, or damping, as before. We can calculate the input
power directly [Lyon R, 1975, p49]:
h L L
n
P
y x
d
in
4
= (13)
To find the radiation loss factor
rad
we find there is a direct formula for baffled plates
[Fahy 1987 p237]:
M
L L c
y x a a
rad
= (14)
where defines the radiation efficiency as before. Having computed at resonance for
all the modes of interest in our model (section 3.2.1) this data was used to estimate at
a given frequency of interest, allowing
rad
to be calculated and so P
rev
found.
The radiation efficiency was estimated by taking an average efficiency of those
modes with resonant frequencies close to the given frequency of excitation, . Only
modes close to the frequency of excitation will respond strongly, and they will all have a
similar radiation efficiency (as shown by Figure 10 above). Thus, it was thought to be a
satisfactory assumption to take the mean efficiency of these modes when calculating .
Modes close to the frequency of excitation were defined as having resonant
frequencies within 10%. This was felt to be a reasonable approximation for the
region of modes that would contribute towards the displacement of the panel. Using this
approximate value for at each frequency a power transfer function for P
rev
was found,
as shown in Figure 15 below:
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 27
Figure 15 Model Predicted and Theoretical Power for Experimental Panel
The power radiated due to the reverberant field is slightly too low to account for all the
power from the panel, but when the driving point radiation is added the combined power
is strikingly similar to that predicted by the model. The theoretical combined power is
smooth (save for some noise due to using a discrete sample of modes) and does not
show the detail of the model prediction, but shows the same general trend. The lack of
detail is due to using an SEA approximation when calculating the radiation from the
reverberant field. The agreement between this theoretical spectrum and that predicted
by the model validates the theory that the panel operates with both driving point and
reverberant field radiation being significant to the total power output, as noted when
looking at the displacement pattern (see Figure 14). The combined theoretical power is
slightly lower than the model predicted power at high frequencies. The model is thought
to be correct and so the theoretical value is under-predicting at high frequencies. This is
due to the region of modes used to find at high frequencies. Modes slightly higher than
this region will not have a large amplitude but will be at coincidence or above and so will
be radiating extremely efficiently. If the region used is increased from 10% to 15%
we include these efficient modes in our calculation of and see a much better
agreement between the model predicted and theoretical radiated power. However, if the
region is increased much further then the theoretical value over predicts at high
frequencies. The correct region to use when calculating depends on critical frequency
By adding together theoretical values for infinite plate and reverberant field behaviour
excellent agreement with model predict results is seen.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 28
and modal density, and will vary from plate to plate, hence choosing this region is
somewhat of an art.
3.3.4 Applying Dual Radiation Theory to Alternate Panels
It has been shown that for the DML analysed the acoustic radiation can be split up into
two distinct radiation methods that can be easily computed using an SEA approximation.
To validate the theory that these two methods occur somewhat independently and sum
to give the total power, the theoretical result was plotted against the model predictions
for some other panels. The two very different panels were modelled one small stiff
panel (where the radiation from the reverberant field dominates total radiation), and one
large less stiff panel (where the driving point radiation dominates). For both cases we
see excellent agreement between theoretical radiated power and model predicted
power [see Appendix 8 for plots]. It is important to note that we have modelled one
panel with 1/4 of the area of the original, and one panel with 1/10 of the stiffness of the
original, yet both panels still have a significant contribution from both types of radiation,
as was found for the original panel. It is thought that this dual radiation method is found
for panels of a wide range of parameter values, and so one can surmise that this is the
radiation method operating in all DMLs.
3.3.5 Computational Use of Dual Radiation Method
The dual radiation method for calculating the radiated power from DMLs uses a
computationally trivial direct calculation if is known. This SEA approximation of the
power spectrum shows none of the detail of the power spectrum predicted by the full
model. However, as the full model shows, DMLs quickly have a high modal overlap
factor as the frequency increases, and so settle down to give a very flat response. The
response is so flat that above the initial modal low frequency region the direct method
for calculating the power is typically within 3dB of the model predicted value. It should
be noted that the extremely high edge damping experienced by the panel when glued
into the speaker frame has not been included in the model. The low frequency region is
expected to have higher modal overlap than predicted by the model and so the initial
modal region will be contained to lower frequencies. Many DMLs are rated as operating
from 300Hz 20kHz, and so it is expected that the edge damping lowers the modal
region to be confined to below 300Hz. Thus, the dual radiation method will show a good
approximation to the model for all frequencies above 300Hz, ie the working range of the
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 29
speaker. This means that the approximate power response for the full working range of
a DML can be calculated in 15mins; 15mins to calculate for each mode, and negligible
time to compute P
tot
from . This is an extremely useful tool when designing and
analysing such speakers.
To speed up the computation still further a fast method of computing the
radiation efficiency is needed. There is one such direct method for finding the average
radiation efficiency that is extremely fast to compute [Leppington et al, 1982]. This has
the added benefit of computing the average radiation efficiency, ie we do not need to
guess a suitable region of modes over which to average . This method will allow the
power spectrum for a flat panel to be computed in a fraction of a second.
3.4 Analysis of Theory
3.4.1 Desirable Properties of DML
For accurate sound reproduction we require loud speakers to have a flat radiated power
spectrum. Through analysing a real DML using the computational model developed in
this research it has been found that the acoustic response of this DML is indeed very flat
over a large range of frequencies. As discussed above, the modal low frequency region
will likely be confined to an even lower frequency when the panel is sitting in the frame,
and a relatively flat response will be seen from 300Hz 20kHz. The power that radiates
from the drive point contributes to the flat response by being constant at all frequencies,
but also has the desirable property of radiating power evenly in all directions, as if it
were a monopole. The power radiated due to the reverberant field is also thought to be
highly dispersed. A single mode will radiate sound with some directivity, and will have a
specific radiation pattern. The high modal density of the DML ensures that many modes
are active, each with a different radiation pattern. When these patterns are added
together the resulting radiation pattern is thought to be averaged over space, and so will
not have a high directivity. The modal density of the plate is constant with frequency and
so it is not expected that the degree of directivity will vary much once into the working
region of the speaker. This is highly desirable when trying to fill a space with even sound,
and is the main improvement over beaming conventional speakers.
The one undesirable property of DMLs is the relatively high value of 300Hz for
the start of the working region. This is dependent on the fundamental wavelength and
damping of the panel. For the Packard Bell speakers analysed here the fundamental
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 30
mode is at 110Hz. To reach as low as 20Hz we would need an extremely large plate, and
so most manufacturers choose to use DMLs in conjunction with a separate (piston-cone)
subwoofer.
3.4.2 Comparison with Conventional Loud Speakers
It is interesting to compare the efficiency of DMLs and normal piston-cone loud
speakers. For conventional loudspeakers the efficiency is rather difficult to estimate.
Often efficiency is defined by a sensitivity rating, the sound level (ie acoustic intensity) at
1m from the speaker when a 1W input is given. This is typically of the order of 90dB for
conventional loudspeakers. Using the model predicted power for the Packard Bell DMLs
analysed, we see that the locally averaged acoustic power is never less than 1.2x10
-3
W.
This power is as a result of P
in
= 0.1W, found by Eqn. 13. It was assumed that the
magnetic voice coil driving the panel has an efficiency of 20%, and so the total electrical
power in to the speaker was calculated as being 0.5W. The efficiency of the panel and
voice coil were assumed to be constant over the working range, and so the total
radiated power from a 1W electrical input was found to be 1.2x10
-3
/0.5 = 2.4x10
-3
W.
For DMLs the radiated sound is extremely dispersed, and so we assume that it radiates
equally in all directions, giving an intensity of 83dB at 1m from the panel. This compares
well with rated specifications of commercial DMLs, which are typically 85dB
[Wharfedale, n.d.].
The fact that conventional speakers typically have a slightly higher sensitivity
rating (90dB vs 85dB) should not be mistaken for thinking that they are more efficient.
DMLs project evenly over an entire sphere, but convention loud speakers project in a
wide cone shape. The sensitivity rating of a conventional speaker is taken in the middle
of this cone, where the intensity is highest. If we assume that a conventional speaker
projects acoustic power evenly over a cone of 90, we find that the projection area at
1m is 0.15x that of the whole sphere at 1m. To convert the intensity of the DML to be
comparable with the sensitivity of a conventional speaker we must normalise by dividing
by this factor, giving a normalised DML sensitivity of 83+8 = 91dB. Rather large
assumptions have been made in calculating the normalised sensitivity of the DML, but
we can at least see that the efficiency of flat panel and conventional speakers is of the
same order of magnitude.
The key difference between DMLs and conventional speakers is their directivity
or lack of it. This difference is seen most obviously in the current uses of DMLs in
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 31
environments where an even sound level is required: hospital p.a. systems, home
cinema, car cabins. Conventional loud speakers are still typically used for high end
music stereos, dj monitor speakers, and any environment where having a directed
sound field and a resulting sweet spot is desirable. Using the previous rough
calculation for sensitivity it is interesting to note that there has not been a rapid uptake
of DMLs for portable audio devices which we might expect due to their slender
dimensions. Portable devices have a finite supply of power (the batteries) and so must
be efficient. Though DMLs are of comparable efficiency to conventional speakers, the
fact that they are not directed means that the sound intensity for a given power input is
lower than that in the centre of projection of a conventional speaker. Portable devices
are typically used by only a few people, and are of low volume so reflections do not
dominate the sound levels, it is the direct field that is key here. If the projection angle of
a conventional speaker is sufficiently large to include these few listeners it will be more
efficient to use a conventional speaker than a DML, with the same resultant intensity for
the listeners. There are some fantastic benefits to using flat panel speakers, but not in
every situation.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 32
Section 4: Conclusions
4.1 Novel Modelling Method
A novel method based on Jinc function wavelets has been employed to calculate the
power radiated from a DML. This method improves upon FEA and BEM by taking far less
time to compute, especially at high frequencies. It is thought that this is one of the first
uses of this new method, and the first time DMLs have been modelled in this way. The
method has been shown to be a powerful tool in practical vibro-acoustic problems.
4.2 Findings
4.2.1 Panel Radiation Method
The key finding of this research is that the sound emitted from a small DML is due to two
wave-field phenomena point source radiation emanating from the central driving point,
and scattered radiation from a reverberant bending wave field. The two methods of
radiation are of similar magnitude, and both have a dispersed radiation pattern. This
research has resulted in the full analyse of one specific DML, but by simulating other
panels it has been shown that the dual radiation theory of sound production holds for
many panels of varying dimensions. It is thought that the dual radiation theory
developed for this particular DML can be applied to a wide range of DMLs, possibly all.
NXT have previously provided a brief explanation for how DMLs can radiate power
in such a dispersed manner. The explanation involves the idea of a chaotic bending
wave field resulting in uncorrelated sound being projected in all directions. There is a
semblance of truth in their description, though it is overplayed. The chaotic wave field
mentioned is simply a reverberant field, a result of using panels with high modal density.
By using a panel with high damping the driving point also contributes significantly to the
radiated power, though this is not mentioned by NXT. The author suspects that high
damping elicits less of a buzz in trade magazines than chaos theory.
4.2.2 Design Parameters
The key parameters that define the response of the panel are the SEA parameters,
namely: modal density and critical frequency. The experimental panel analysed was
shown to have high modal density (1 mode every 65Hz), and a critical frequency above
the working region (23kHz). This value of the critical frequency results in resonant
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 33
modes being below coincidence, and so no single mode dominates the acoustic
response. This multimodal radiation results in a dispersed radiation pattern. The high
modal density is important for two reasons 1) it ensures the reverberant field includes
many modes and so produces a dispersed radiation pattern, and 2) it results in a high
modal overlap factor, giving the flat response that is so critical to the performance of any
loud speaker. Now that the method of radiation is understood these parameters can be
used as a design tool, optimising parameters with reference to some desired value of
critical frequency and modal density.
4.2.3 Dual Radiation Method of Computation
A fast method for calculating the power spectrum for a DML has been developed. This
method uses an SEA approximation for the power emitted from a reverberant field and a
direct calculation for the power from an infinite plate. By accepting the dual radiation
theory, we can compute an approximate result for a DMLs power spectrum in minutes.
This approximate result shows the general trend of the power spectrum but not the
detail. The real power spectrum is so smooth at all medium and high frequencies,
however, that this approximate result is typically within 3dB of the real value, even
tracking the rise in power seen at high frequencies.
In this research a method was used that allowed a 4000 point power spectrum to
be computed in 15 minutes. Another method is known that uses a direct calculation for
radiation efficiency at resonance [Leppington et al, 1982], and will allow an
approximation for the radiated power of a flat panel to be computed in seconds. For
many design situations the value of being able to compute the approximate spectrum in
seconds will far outweigh the value of having a precise spectrum .
4.2.4 Comparison with Conventional Loud Speakers
DMLs improve upon conventional speakers by their size and lack of directivity, and are
ideal for filling a space with uniform sound. They do not improve upon conventional
speakers in terms of efficiency, and so are better suited to home cinema and p.a.
systems than portable audio despite their small size. DMLs can only reproduce sound
as low as their fundamental frequency, for small panels this is approximately 100Hz.
Unless extremely large panels are used DMLs are generally paired with a conventional
piston cone subwoofer.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 34
Section 5: Future Work
5.1 Improving Model Accuracy
The model of a DML that has been developed through the course of this research is very
general and leaves us with much scope for future work. The model used made
simplifying assumptions about plate damping, stiffness and boundary conditions. Future
work could involve developing a more accurate model by using a non-uniform bending
stiffness value, non-constant damping value, and different boundary conditions to
compute mode shapes. Finding an accurate model for damping is non-trivial, especially
when the panel has peculiar boundary conditions (tacky glue). The resulting mode
shapes can be used to compute the radiated power without any changes to the power
spectrum tool.
Further accuracy can be gained by altering the power spectrum tool to compute
the power radiated for an un-baffled panel. The method for calculating the power
spectrum for an un-baffled or even partially baffled plate is known [Langley, 2007], but
will require a new power spectrum computer program. This program takes the same
inputs as the power program used in this research (namely the mode shape at a given
frequency) and so can be used as a direct replacement of the power spectrum tool used
here for a baffled plate.
5.2 Additions to the Model
Theories about the possible radiation pattern emitted from a DML have been
guessed, with reference to the two modes of radiation found to be dominant. It would be
useful however to compute this radiation pattern directly and visualise it in 3D space.
This can be done using either the Rayleigh Integral (slow but accurate), or by an
extension of the Jinc function method (fast but possible inaccurate).
Further to this, being able to fully auralise the radiated signal would allow us to
listen to the response of a particular DML design, at some point in space. To listen to
the radiated sound we need both phase and magnitude information, though the model
gives only magnitude. If a minimum phase assumption is applied, we can regain the
phase information by using the Hilbert transform.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 35
5.3 Alternate Model
The response of DMLs has been found to be very smooth due to a high modal density,
and so using the fast method of computing radiated power (by addition of driving point
and reverberant field radiation) is a valid approximation. Currently this takes in the
region of 15minutes to compute, due to using the (already fast) Jinc function method to
calculate the radiation efficiency. If the direct method [Leppington et al, 1982] is used a
tool which give the power spectrum in less than a second will be possible.
5.4 Panel Optimisation
One outcome of this project is a very fast method for computing the power spectrum of
such a DML. With such a fast tool it is now possible to use optimisation in the design of
DMLs. Possible criteria for optimisation include total power output, critical frequency,
modal density, and bandwidth of low frequency modal region. The latter of these criteria
will need to use the full Jinc function based model, for the other three the fast SEA
model will suffice. Parameters that can be included in optimisation procedures include
damping, stiffness, panel dimensions, and boundary conditions. The possibility of using
non-uniform plates and varying the drive point position is another area of investigation.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 36
Section 6: Acknowledgements
Prof Robin Langley for steering me in the right direction throughout the project
Tim Hughes who saved me a great deal of time when using the laser vibrometer
Prof Jim Woodhouse for use of his data logging and modal analysis MatLab program
Prof Malcolm Atkinson for help conquering the subtleties of Word
Prof Frank Fahy for a chance conversation that resulted in interesting analysis results
Marcel Leutenaggar for an optimised sine function that speeded up MatLab by 2x
Gail and Mum who harass me when I slip into procrastination mode
The Flat for organised fun
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 37
Section 7: References
1. Azima, H (1998): NXT Technology when a little chaos is good for you, originally published in
Audio Magazine.
Available at: http://www.emgeton.com/manual_speakers/NXT%20Technology.pdf
[cited 16
th
May 2007].
2. Cremer L, Heckl M & Petersson B.A.T (2005): Structure Borne Sound, 3
rd
Edition, Springer-Verlag
Berlin Heidelberg, Chapter 7 p. 454
3. Crighton D.G, Dowling A.P, Ffowcs-Williams J.E, Heckl M & Leppington F.G, (1992): Modern
Methods in Analytical Acoustics Lecture Notes, Springer-Verlag London Ltd, Chapter 20 p. 589
4. Elac Electroacoustics (n.d.): DML Distributed Mode LoudSpeaker
Available at: http://www.elac.com/en/products/imago/dml.html
[cited 22
nd
May 2007]
5. Fahy F (1987), Sound and Structural Vibration, Academic Press London, pages 94 & 237
6. Hill N & Mapp P (2003): An overview of suitable measurement and assessment techniques for
distributed mode loudspeakers, The Journal of the Acoustical Society of America, Vol. 114, Issue 4,
p. 2410
7. Langley, R (2007): Numerical evaluation of the acoustic radiation from planar structures with
general baffle conditions using wavelets, The Journal of the Acoustical Society of America - Vol
121, Issue 2, p. 766-777
8. Leppington F.G, Broadbent E.G & Heron K.H. (1982): The Acoustic Radiation Efficiency of
Rectangular Panel, Proceedings of the Royal Society of London. Series A, Mathematical and
Physical Sciences, Vol. 382, No. 1783, p. 245-271.
9. Lyon, R (1975): Statistical Energy Analysis of Dynamical Systems, MIT Press, pages 49 & 234
10. Norton M & Karczub D (2003), Fundamentals of Noise and Vibration Analysis for Engineers, 2nd
Edition, Cambridge University Press, Chapter 3 p. 210
11. Wharfedale n.d.: PPS1 DML product specification
Available at: http://www.wharfedale.co.uk/model.php?model_id=55#
[cited 25
th
May 2007]
12. Williams & Maynard (1982): Numerical evaluation of the Rayleigh integral for planar radiators
using the FFT, The Journal of the Acoustical Society of America Volume 72, p. 2020-2030
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 38
Section 8: Appendices
Appendix 1 Description of Jinc Function
The Jinc function is defined as:
x
x J
x jinc
) (
) (
1
= ,
where J
1
(x) is a Bessel function of the 1
st
kind and satisfies
2
1
) ( lim
0
=
x jinc
x
.
The graph below displays a 1D jinc function for values of x from -15 to 15.
Figure A1 1D Jinc Function
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 39
Appendix 2 Equations Used to Calculate Modal Parameters
Modal Density:
D
h
L L
n
y x
d
4
=
Critical Fequency:
D
h
c
a c
2
=
Modal Overlap Factor:
d d
n m =
[Cremer & Heckl, 2005, p302 & 496]
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 40
Appendix 3 Mobility Transfer Function SEA Verification
Figure A2 Mobility Transfer Function, Model predicted and SEA Approximation
The SEA approximation for velocity is found using the following equation:
2
) ( 2
) (
y x
d
SEA
L hL
n
v =
where nd defines the modal density of the plate.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 41
Appendix 4 make_R code printout
f unct i on R = make_R( Lx, Ly, c)
%makes an ar r ay R, whi ch cont ai ns t he di st ance bet ween t he i ' t h and j ' t h
%poi nt s i n t he vect or i sed f or mof t he pl at e posi t i on.
%Lx, Ly, and c ar e used t o def i ne t he gr i d, bef or e vect or i si ng.
[ B, A] = meshgr i d( ( 0: c: Ly) , ( 0: c: Lx) ) ; %make meshes of x and y posi t i on
AB = A + B*i ; %combi ne x and y posi t i on i nt o si ngl emat r i x, def i ni ng t he x
and y coor di nat es of ever y mesh poi nt as x + i y
AB = AB( : ) ; %vect or i se mat r i x i nt o si ngl e r ow vect or
ABmat = r epmat ( AB, 1, si ze( AB) ) ; %make ver t i cal mesh of x and y posi t i ons
R = abs( ABmat - ABmat . ' ) ; %f i nd di st ance bet ween poi nt s
r et ur n
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 42
Appendix 5 Power Validation using Non-dimensional Radiation Efficiency
Figure A3 Radiation Efficiency vs Non-dimensional Frequency
The above results are found using the Jinc function model, and are plotted atop
benchmark results found using the FFT method of Maynard and Williams (marked with
x). FFT results kindly supplied by Prof. Langley. Excellent agreement is seen, validating
the accuracy of the model.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 43
Appendix 6 Experimental vs Predicted Transfer Function
Figure A4 Mobility Transfer Function found using Model and Laser Vibrometer
The first three model modes are predicted at frequencies coinciding with modes found
experimentally. The model panel is excited exactly in its centre, but the experimental
panel shows small resonances of anti-symmetric modes due to not being excited exactly
in its centre. The experimental result comes from a laser vibrometer that was un-
calibrated, hence we see an offset between the two spectra.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 44
Appendix 7 Damping with frequency
Figure A5 Fitting Curve to Experimental Damping Data
MatLab Curve Fitting Tool Output:
Gener al model :
f ( x) = a*exp( b*x) +c
Coef f i ci ent s ( wi t h 95%conf i dence bounds) :
a = 0. 1487 ( 0. 131, 0. 1664)
b = - 0. 002897 ( - 0. 003768, - 0. 002027)
c = 0. 01422 ( 0. 003852, 0. 02459)
Goodness of f i t :
SSE: 0. 0002492
R- squar e: 0. 9811
Adj ust ed R- squar e: 0. 977
RMSE: 0. 005262
For a line of best fit that gives a high R
2
value we find that the asymptote, c, is at c =
0.014. This is an estimate for the internal damping of the panel.
Luke Humphry, Kings College Cambridge May 2007
Analysis of Flat Panel Speakers Page 45
Appendix 8 Theoretical Power Plots
Figure A6 Power Spectrum for 95x65mm
2
Panel
Figure A7 Power Spectrum for Pane with 0.1x Stiffness