dehaeze21_desig_compl_filte/mohit/paper.org

Sensor fusion , Optimal filters , $\mathcal{H}_\infty$ synthesis ,
Vibration isolation , Precision

* Introduction
The sensors used for measuring physical quantity often works well within
a limited frequency range called as the bandwidth of the sensor. The
signals recorded by the sensor beyond its bandwidth are often corrupt
with noise and are not reliable. Many dynamical systems require
measurements over a wide frequency range. Very often a variety of
sensors are utilized to sense the same quantity. These sensors have
different operational bandwidth and are reliable only in a particular
frequency range. The signals from the different sensors are fused
together in order to get the reliable measurement of the physical
quantity over wider frequency band. The combining of signals from
various sensor is called sensor
fusion \cite{hua04_polyp_fir_compl_filter_contr_system}. The resulting
sensor is referred as "super sensor" since it can have better noise
characteristics and can operate over a wider frequency band as compared
to the individual sensor used for
merging \cite{shaw90_bandw_enhan_posit_measur_using_measur_accel}.

Sensor fusion is most commonly employed in the navigation systems to
accurately measure the position of a vehicle. The GPS sensors, which are
accurate in low frequency band, are merged with the high-frequency
accelerometers. Zimmermann and
Sulzer \cite{zimmermann92_high_bandw_orien_measur_contr} used sensor
fusion to measure the orientation of a robot. They merged inclinometer
and accelerometers for accurate angular measurements over large
frequency band.
Corke \cite{corke04_inert_visual_sensin_system_small_auton_helic} merged
inertial measurement unit with the stereo vision system for measurement
of attitude, height and velocity of an unmanned helicopter. Min and
Jeung \cite{min15_compl_filter_desig_angle_estim} used accelerometer and
gyroscope for angle estimations. Baerveldt and Klang \cite{baerveldt97}
used an inclinometer and a gyroscope to measure the orientation of the
autonomous helicopter. The measurement of the 3D orientation using a
gyroscope and an accelerometer was demonstrated by Roberts et
al. \cite{roberts03_low}. Cao et
al. \cite{cao20_adapt_compl_filter_based_post} used sensor fusion to
obtain the lateral and longitudinal velocities of the autonomous
vehicle.

Sensor fusion is also used for enhancing the working range of the active
isolation system. For example, the active vibration isolation system at
the Laser Interferometer Gravitational-Wave Observatory
(LIGO) \cite{matichard15_seism_isolat_advan_ligo} utilizes sensor
fusion. The position sensors, seismometer and geophones are used for
measuring the motion of the LIGO platform in different frequency
bands \cite{hua05_low_ligo}. Tjepkema et
al. \cite{tjepkema12_sensor_fusion_activ_vibrat_isolat_precis_equip}
used sensor fusion to isolate precision equipment from the ground
motion. The feedback from the accelerometer was used for active
isolation at low frequency while force sensor was used at high
frequency. Various configurations of sensor fusion for active vibration
isolation systems are discussed by Collette and
Matichard \cite{collette15_sensor_fusion_method_high_perfor}. Ma and
Ghasemi-Nejhad \cite{ma04_frequen_weigh_adapt_contr_simul} used laser
sensor and piezoelectric patches for simultaneous tracking and vibration
control in smart structures. Recently, Verma et
al. \cite{verma21_virtual_sensor_fusion_high_precis_contr} presented
virtual sensor fusion for high precision control where the signals from
a physical sensor are fused with a sensor simulated virtually.

Fusing signals from different sensors can typically be done using Kalman
filtering \cite{odry18_kalman_filter_mobil_robot_attit_estim, ren19_integ_gnss_hub_motion_estim, faria19_sensor_fusion_rotat_motion_recon, liu18_innov_infor_fusion_method_with, abdel15_const_low_cost_gps_filter, biondi17_attit_recov_from_featur_track}
or complementary
filters \cite{brown72_integ_navig_system_kalman_filter}. A set of
filters is said to be complementary if the sum of their transfer
functions is equal to one at all frequencies. When two filters are
complementary, usually one is a low pass filter while the other is an
high pass filter. The complementary filters are designed in such a way
that their magnitude is close to one in the bandwidth of the sensor they
are combined with. This enables to measure the physical quantity over
larger bandwidth. There are two different categories of complementary
filters --- frequency domain complementary filters and state space
complementary filters. Earliest application of the the frequency domain
complementary filters was seen in Anderson and
Fritze \cite{anderson53_instr_approac_system_steer_comput}. A simple RC
circuit was used to physically realize the complementary filters.
Frequency domain complementary filters were also used
in \cite{shaw90_bandw_enhan_posit_measur_using_measur_accel, zimmermann92_high_bandw_orien_measur_contr, baerveldt97, roberts03_low}.
State space complementary filter finds application in tracking
orientation of the flexible links in a
robot \cite{bachmann03_desig_marg_dof, salcudean91_global_conver_angul_veloc_obser, mahony08_nonlin_compl_filter_special_orthog_group}
and are particularly useful for multi-input multi-output systems.
Pascoal et al. \cite{pascoal00_navig_system_desig_using_time} presented
complementary filters which can adapt with time for navigation system
capable of estimating position and velocity using GPS and SONAR sensors.

The noise characteristics of the super sensor are governed by the norms
of the complementary filters. Therefore, the proper design of the
complementary filters for sensor fusion is of immense importance. The
design of complementary filters is a complex task as they need to tuned
as per the specification of the sensor. In many applications, analytical
formulas of first and second order complementary filters are
used \cite{corke04_inert_visual_sensin_system_small_auton_helic,jensen13_basic_uas}.
These filters are easy to tune and simple to implement using an analog
circuit \cite{moore19_capac_instr_sensor_fusion_high_bandw_nanop,cite:yong16_high_speed_vertic_posit_stage}.
However, these low order complementary filters are not optimal, and high
order complementary filters can lead to better
fusion \cite{jensen13_basic_uas,shaw90_bandw_enhan_posit_measur_using_measur_accel}.

Several design techniques have been proposed to design higher order
complementary filters.
Pascoal \cite{pascoal00_navig_system_desig_using_time} used linear
matrix inequalities (LMIs) \cite{boyd94_linear} for the design of time
varying complementary filters. LMIs were also used by Hua et
al. \cite{hua04_polyp_fir_compl_filter_contr_system} to design finite
impulse response (FIR) filters for the active vibration isolation system
at LIGO.
Plummer \cite{plummer06_optim_compl_filter_their_applic_motion_measur}
proposed an optimal design method using the $\mathcal{H}_{\infty}$
synthesis and weighting functions representing the measurement noise of
the sensors.

Although various methods have been presented in the literature for the
design of complementary filters, there is a lack of general and simple
framework that allows to shape the norm of complementary filters. Such a
method would prove to be very useful as the noise of the "supper sensor"
and its dynamical characteristics depend on the norm of the filters.
This paper presents such a framework based on the $\mathcal{H}_\infty$
norm minimization. The proposed method is quite general and can be
easily extended to a case where more than two complementary filters
needs to be designed. The organization of this paper is as follows.
Section [[*Complementary filters requirements][2]] presents the design
requirements of ideal complementary filters. It also demonstrates how
the noise and robustness characteristics of the "super sensor" can be
transformed into upper bounds on the norm of the complementary filters.
The framework for the design of complementary filters is detailed in
Section [[*Design formulation using $\mathcal{H}_\infty$ synthesis][3]].
This is followed by the application of the design method to
complementary filter design for the active vibration isolation at LIGO
in Section [[*Application: Complementary Filter Design for Active
Vibration Isolation of LIGO][4]]. Finally, concluding remarks are
presented in Section [[*Concluding remarks][5]].

* Complementary filters requirements
** Introduction                                                      :ignore:
Complementary filters provides a framework for fusing signals from different
sensors. As the effectiveness of the fusion depends on the proper design of the
complementary filters, they are expected to fulfill certain requirements. These
requirements are discussed in this section.

** Complementary characteristics
Consider a case where two different sensors are used for measuring the
same quantity, $x$ in different frequency range. The inherent dynamics
of the sensors is represented by transfer functions $G_1(s)$ and
$G_2(s)$. The two sensor also have uncorrelated noise characteristics
given by $n_1$ and $n_2$. The signals from these two sensors are fused
using complementary filters $H_1(s)$ and $H_2(s)$. The architecture of
sensor fusion using complementary filters is shown in
Figure [[#fig:fusion_super_sensor][1]]. The resulting sensor, termed as
"super sensor", can have larger bandwidth and better noise
characteristics in comparison to the individual sensor. This means that
the super sensor provides an estimate $\hat{x}$ of $x$ which can be more
accurate over a larger frequency band than the outputs of the individual
sensors. Based on Figure [[#fig:fusion_super_sensor][1]], the estimate
of the physical quantity as measured by the super sensor can be written
as $$\label{eq:comp_filter_estimate}
  \hat{x} = \left(G_1 H_1 + G_2 H_2\right) x + H_1 n_1 + H_2 n_2$$

#+caption: [fig:fusion_super_sensor]Block diagram of sensor fusion with
complementary filters
[[file:figs/fusion_super_sensor.pdf]]

The complementary property of filters $H_1(s)$ and $H_2(s)$ simply
implies that the summation of their transfer functions is equal to
unity. That is, unity magnitude and zero phase at all frequencies
(filters which satisfies only the magnitude condition are referred as
"magnitude complementary filter pair"). Therefore, a pair of strict
complementary filter needs to satisfy the following condition.

$$\label{eq:comp_filter}
  H_1(s) + H_2(s) = 1$$

** Noise characterization
In order to compute the noise characteristics associated with the
estimate $\hat{x}$, it is first assumed that the dynamics of the
individual sensors are perfect: $$\label{eq:perfect_dynamics}
  G_1(s) = G_2(s) = 1$$

The output of the super sensor, $\hat{x}$, based on the block diagram
shown in Figure [[#fig:fusion_super_sensor][1]] can be written as
$$\label{eq:estimate_perfect_dyn}
  \hat{x} = x + H_1 n_1 + H_2 n_2$$

The complementary are operating only on the noise component of the
individual sensor. Thus, this sensor fusion architecture permits to
filter the noise of both sensors without introducing any distortion in
the physical quantity to be measured. The estimation error, $\delta x$,
of the super sensor can be written as $$\label{eq:estimate_error}
  \delta x \triangleq \hat{x} - x = H_1 n_1 + H_2 n_2$$

The power spectral density (PSD) of the super sensor's estimation error
is given by $$\label{eq:noise_filtering_psd}
  \Phi_{\delta x} = \left|H_1\right|^2 \Phi_{n_1} + \left|H_2\right|^2 \Phi_{n_2}$$
where, $\Phi_{\delta x}$ is the PSD of estimation error, $\Phi_{n_1}$
and $\Phi_{n_2}$ are the PSDs of the noise associated with the
individual sensor. It can be seen that the estimation error's PSD
depends on the PSD of the noise in individual sensor as well as the norm
of the complementary filters. Therefore, by properly shaping the norm of
the complementary filters, it is possible to minimize the noise of the
super sensor noise.

** Robustness requirements
In the previous subsection, the inherent sensor dynamics were ignored.
However in the real system, the sensor dynamics is not equal to unity.
In such cases, the output of the sensor is normalized using a filter
whose transfer function is equal to the inverse of the sensor dynamics.
There are two major concerns in using inversion. First being the sensors
may not have been calibrated properly and the actual sensor dynamics is
not exactly compensated by the inverse filter. The second problem is
that the inversion of sensor dynamics can result in an improper transfer
function and hence may not be physically realizable. We here suppose
that the sensor dynamics can be inverted using a proper and stable
transfer function $\hat{G}_i(s)$. However, we suppose there exists a
normalization error since $\hat{G}_i^{-1}(s) G_i(s) \neq 1$. This
normalization error can be represented using frequency dependent
multiplicative uncertainty
(Figure [[#fig:sensor_fusion_dynamic_uncertainty][2]]). In
Figure [[#fig:sensor_fusion_dynamic_uncertainty][2]], $\Delta_i(s)$
satisfies $\|\Delta_i(s)\|_\infty \le 1$ and $|w_i(s)|$ is a frequency
dependent weighting function that represents the uncertainty
corresponding to the normalization error.

#+caption: [fig:sensor_fusion_dynamic_uncertainty] Representation of
normalization error in sensor fusion using multiplicative uncertainty
[[file:figs/sensor_fusion_dynamic_uncertainty.pdf]]

Based on Figure [[#fig:sensor_fusion_dynamic_uncertainty][2]], the super
sensor dynamics can be written as
$$\label{eq:super_sensor_dyn_uncertainty}
  \frac{\hat{x}}{x} = 1 + w_1(s) H_1(s) \Delta_1(s) + w_2(s) H_2(s) \Delta_2(s)$$

The dynamics of the super sensor now depends on the weighting functions
($w_1(s),w_2(s)$) and the complementary filters ($H_1(s),H_2(s)$).

The robust stability of the fusion can be studied graphically (refer
Figure [[#fig:uncertainty_set_super_sensor][3]]). The frequency response
of the fusion output is plotted in a complex plane. The unity transfer
function leads to a point $(1,0)$ located on the real axis. The
uncertainty associated with first sensor at a particular frequency is
represented by a circle with the center at (1,0) and radius $|w_1H_1|$.
The uncertainty associated with the second is also represented using a
circle centered at any point on the circle representing uncertainty
associated with the first sensor and radius equal to $|w_2H_2|$.
Therefore, the overall uncertainty of the fusion is represented with a
circle centered at (1,0) and radius equal to $|w_1H_1|+|w_2H_2|$. The
maximum phase difference that can result from the fusion is found by
drawing a tangent from the origin to the uncertainty circle of super
sensor. Mathematically, the maximum phase difference at frequency
$\omega$ that can result from fusion is given by
$$\label{eq:max_phase_uncertainty}
  \Delta\phi(\omega) = \arcsin\left( |w_1(j\omega) H_1(j\omega)| + |w_2(j\omega) H_2(j\omega)| \right)$$

#+caption: [fig:uncertainty_set_super_sensor]Robustness analysis of
sensor fusion in the complex plane. The uncertainty associated with the
super sensor dynamics are represented with a solid circle while those
associated with individual sensors are represented with dashed circles.
[[file:figs/uncertainty_set_super_sensor.pdf]]

A constraint on the maximum phase difference at a given frequency can be
incorporated in the design problem using the following equation:
$$\label{eq:max_uncertainty_super_sensor}
  \max_\omega \big( \left|w_1 H_1\right| + \left|w_2 H_2\right|\big) < \sin\left( \Delta \phi_\text{max} \right)$$
where $\Delta \phi_\text{max}$ is the maximum allowable phase
difference. It can also be inferred from the above equation that the
magnitude of the complementary filter ($|H_i|$) should be tuned to a
smaller value at the frequencies where the magnitude of weighing
transfer functions ($|w_i|$) representing sensor uncertainty is large.

* Design formulation using $\mathcal{H}_\infty$ synthesis
** Introduction                                                      :ignore:
In this section, the shaping of complementary filters is expressed as an
optimal $\mathcal{H}_{\infty}$ synthesis problem. The synthesis goal is
to shape the frequency response of the filters such that they satisfy
the design requirements presented in Section [[*Complementary filters
requirements][2]].

** Synthesis problem formulation
The first step is to formulate the filter design problem as a
generalized plant-controller structure \cite{boyd91_linear}. The
generalized plant and controller structure for complementary filters
design is shown in Figure [[#fig:h_infinity_robust_fusion][4]]. In the
figure, $P(s)$ is the generalized plant, $u$ is the "control input", $v$
is the "measured output" and $H_2(s)$ is the controller (filter) to be
designed. The regulated outputs of the generalized plant, $z_1$ and
$z_2$, are given by $$\begin{split}
    z_1 &= W_1(s)(1-H_2(s)) w = W_1(s) H_1(s) w \text{ by defining } H_{1}(s) \triangleq 1 - H_{2}(s)\\
    z_2 &= W_2(s) H_2(s) w
  \end{split}$$ where $w$ is the "exogenous input" to the plant,
$W_1(s), W_2(s)$ are the weighting functions for shaping the
complementary filters.

The dynamics of the generalized plant can be written as
$$\label{eq:generalized_plant}
  \begin{bmatrix} z_1 \\ z_2 \\ v \end{bmatrix} = P(s) \begin{bmatrix} w\\u \end{bmatrix}; \quad P(s) = \begin{bmatrix}W_1(s) & -W_1(s) \\ 0 & W_2(s) \\ 1 & 0 \end{bmatrix}$$

The weighting functions are chosen based on the specifications and
requirements set for the complementary filters (discussed in
Section [[*Design of weighting functions][3.2]]). The objective of the
optimization is to design a filter $H_2(s)$ such that the following
conditions are satisfied $$\label{eq:comp_filter_problem_form}
  \begin{split}
    \left| \frac{z_{1}}{w} \right| &= |1-H_2(s)| \le \frac{1}{|W_1(s)|} \\
    \left| \frac{z_{1}}{w} \right| &= |H_2(s)|   \le \frac{1}{|W_2(s)|}
  \end{split}, \quad \forall \omega \in \mathbb{R}^{+}$$

#+caption: [fig:h_infinity_robust_fusion]Generalized plant controller
structure for the design of complementary filters
[[file:figs/h_infinity_robust_fusion.pdf]]

Based on Figure [[#fig:h_infinity_robust_fusion][4]], the
$\mathcal{H}_{\infty}$ synthesis problem for the complementary filters
can be stated as

#+begin_quote
  Find a stable transfer function, $H_2(s)$, which takes measured
  output, $v$, as input and generates a control input, $u$, such that
  the $\mathcal{H}_\infty$ norm of the generalized plant from exogenous
  input, $w$, to the regulated output, ${[z_1,z_2]}^T$ is less than
  unity.
#+end_quote

Mathematically, the synthesis objective can be written as
$$\begin{split}
                        &\left\|\begin{matrix} \left[1 - H_2(s)\right] W_1(s) \\ H_2(s) W_2(s) \end{matrix}\right\|_\infty \le 1 \\
    \Longleftrightarrow & \left\|\begin{matrix} H_1(s) W_1(s) \\ H_2(s) W_2(s) \end{matrix}\right\|_\infty \le 1; \quad H_1(s) \triangleq 1 - H_2(s)
  \end{split}
  \label{eq:hinf_syn_obj}$$ The above optimization problem can be
efficiently solved in Matlab \cite{MATLAB2009} using Riccati formulae,
linear matrix inequality based method or maximum entropy method.

** Design of weighting functions
The choice of weighting function governs the shape of the designed
complementary filters. Therefore, it is very important that the design
specifications are appropriately transformed into the weighting
functions. The choice of weighting functions is also constrained by the
following factors

1. Only proper and stable transfer functions can be used as weighting
   functions

2. As the order of the designed filter is equal to the sum of the orders
   of the weighting functions, the order of the weighting function needs
   to be reasonably small to ensure the physical implementation of the
   designed complementary filters. This also reduces the computational
   cost of the optimization problem.

3. The complementary property of the filter imposes a fundamental
   limitations on the weighting functions. The imposes a restriction
   that the magnitude of the filters $H_1(s)$ and $H_2(s)$ cannot be
   made small simultaneously at the same frequency.

The specifications of the complementary filters are typically expressed
using the following parameters --- low frequency gain, high frequency
gain, slope (order of the filter) and the crossover frequency. We
propose a weighting function that allows to translate the above
requirements by setting simple parameters: $$\label{eq:weight_formula}
  W(s) = \displaystyle\left( \frac{
      \hfill{} \displaystyle\frac{1}{\omega_0} \sqrt{\frac{1 - \left(\displaystyle\frac{G_0}{G_c}\right)^{\displaystyle\frac{2}{n}}}{1 - {\left(\displaystyle\frac{G_c}{G_\infty}\right)}^{\displaystyle\frac{2}{n}}}} s + \left(\frac{G_0}{G_c}\right)^{\displaystyle\frac{1}{n}}
    }{
      \left(\displaystyle\frac{1}{G_\infty}\right)^{\displaystyle\frac{1}{n}}\displaystyle \frac{1}{\omega_0} \sqrt{\displaystyle\frac{1 - \left(\displaystyle\frac{G_0}{G_c}\right)^{\displaystyle\frac{2}{n}}}{1 - \left(\displaystyle\frac{G_c}{G_\infty}\right)^{\displaystyle\frac{2}{n}}}} s + \left(\displaystyle\frac{1}{G_c}\right)^{\displaystyle\frac{1}{n}}
    }\right)^n$$ where, $G_0 = \lim_{\omega \to 0} |W(j\omega)|$ is the
low frequency gain, $G_\infty = \lim_{\omega \to \infty} |W(j\omega)|$
is the high frequency gain, $\omega_c$ is the crossover frequency,
$G_c = |W(j\omega_c)|$ is the absolute gain at the crossover frequency
and $n$ is the order of the filter. As an illustration, the magnitude of
the frequency response of the weighting function with the parameters
$G_0 = 0.001$, $G_\infty = 10$, $\omega_c = \SI{10}{Hz}$, $G_c = 2$,
$n = 3$ and having high pass characteristics is shown in
Figure [[#fig:weight_formula][5]].

#+caption: [fig:weight_formula]Magnitude of the frequency response of
the weighting function obtained using
equation [[#eq:weight_formula][[eq:weight_formula]]] with the parameters
$G_0 = 0.001$, $G_\infty = 10$, $\omega_c = \SI{10}{Hz}$, $G_c = 2$,
$n = 3$
[[file:figs/weight_formula.pdf]]

** Verification
The proposed methodology for the design of complementary filters is now
applied on the following example.

*Example.* Design complementary filters with the merging frequency
around 10 Hz. The low pass filter should have DC gain of 0.001 and slope
-2 above the merging frequency. The high pass filter should have a slope
of 3 below the merging frequency and 0.001 high frequency gain.

The first step is to design the weighting functions that translate the
above requirements. To do so,
equation [[#eq:weight_formula][[eq:weight_formula]]] is used. The
parameters corresponding to $W_1(s)$ and $W_2(s)$ are listed in
Table [[#tab:weights_params][1]]. The obtained transfer functions of the
weighting functions are $$\begin{split}
    W_1(s) &= \dfrac{1000 (s+34.55)^2}{(s+3455)^2}\\
    W_2(s) &= \dfrac{0.1 (s+87.43)^3}{(s+4.058)^3}
  \end{split}$$

Using these weighting functions, the generalized plant is evaluated
using equation [[#eq:generalized_plant][[eq:generalized_plant]]]. The
optimal complementary filters are obtained by solving the optimization
problem given by equation [[#eq:hinf_syn_obj][[eq:hinf_syn_obj]]]. The
complementary filters obtained after optimization are $$\begin{split}
    H_1(s) &= \frac{10^{-8} (s+6.6\times 10^9) (s+3450)^2 (s^2 + 49s + 895)}{(s+6.6e^4) (s^2 + 106 s + 3\times 10^3) (s^2 + 72s + 3580)}\\
    H_2(s) &= \frac{(s+6.6\times 10^4) (s+160) (s+4)^3}{(s+6.6\times 10^4) (s^2 + 106 s + 3\times 10^3) (s^2 + 72s + 3580)}
  \end{split}$$

The obtained complementary filters are of order 5 which corresponds to
the sum of the orders of the weighting functions used. The frequency
responses of the designed complementary filters are shown in
Figure [[#fig:hinf_filters_results][6]]. It can be seen that the
designed filters fulfills all the design specifications and hence
demonstrates the effectiveness of the designed methodology (more complex
real life example is taken up in Section [[*Application: Complementary
Filter Design for Active Vibration Isolation of LIGO][4]]).

<<tab:weights_params>>
| Parameter              | $W_1(s)$ | $W_2(s)$ |
|------------------------+----------+----------|
| $n$                    | $2$      | $3$      |
| $G_c$                  | $0.5$    | $0.5$    |
| $G_\infty$             | $1000$   | $0.1$    |
| $G_0$                  | $0.1$    | $1000$   |
| $\omega_c$ [$\si{Hz}$] | $11$     | $10$     |
#+caption: [tab:weights_params]Parameters used for $W_1(s)$ and $W_2(s)$

#+caption: [fig:hinf_filters_results]Bode plot of the weighting
functions and designed complementary filters
[[file:figs/hinf_filters_results.pdf]]

** Extension to set of three complementary filters
In certain applications, more than two sensors are used to measure the
same quantity and can be merged together to form a "super sensor". In
such case, a set of three (or more) complementary filters is required.
It is here shown that the proposed method can be generalized for the
design of a set of arbitrary number of complementary filters. The
control objective is now to design of a set of $n$ complementary filters
($H_i(s), i=1,\cdots,n$) which satisfy the following conditions
$$\label{eq:hinf_problem_gen}
  \begin{split}
    &\sum_{i=0}^n H_i(s) = 1 \\
    &\left| H_i(s) \right| < \frac{1}{\left| W_i(s) \right|}
  \end{split}$$ Here, we extend the method to a case of three
complementary filters. The generalized plant controller setup for this
case is shown in Figure [[#fig:comp_filter_three_hinf][7]]. The
synthesis objective is to design filters $H_2(s)$ and $H_3(s)$ such that
the $\mathcal{H}_\infty$ norm from exogenous input $w$ to regulated
output vector $[z_1,z_2,z_3]^T$ is less than unity. That is,
$$\label{eq:hinf_syn_obj_three}
  \begin{split}
    &\left\| \begin{matrix} \left[1 - H_2(s) - H_3(s)\right] W_1(s) \\ H_2(s) W_2(s) \\ H_3(s) W_3(s) \end{matrix} \right\|_\infty \le 1\\
    \equiv &\left\| \begin{matrix} H_1(s) W_1(s) \\ H_2(s) W_2(s) \\ H_3(s) W_3(s) \end{matrix} \right\|_\infty \le 1; \quad H_1(s) \triangleq 1 - H_2(s) - H_3(s)
  \end{split}$$

#+caption: [fig:comp_filter_three_hinf]Generalized plant controller
setup for designing a set of three complementary filters using
$\mathcal{H}_\infty$ synthesis
[[file:figs/comp_filter_three_hinf.pdf]]

To validate this synthesis method, let's take an example where 3 sensors
are merged together. The three sensors are working in the following
three frequency range --- below 1 Hz, between 1Hz to 10Hz and above
10Hz. The weighting functions used for the synthesis are $$\begin{split}
    W_1(s) &= \dfrac{1000 (s+3.141)^2}{(s+314.1)^2}\\
    W_2(s) &= \dfrac{2200 (s+62.83)^2 (s+6.283)^2}{(s+6283)^2 (s+0.06283)^2}\\
    W_3(s) &= \dfrac{0.1 (s+87.43)^3}{(s+4.058)^3}
  \end{split}$$ The complementary filters are obtained by solving the
optimization problem given by
equation [[#eq:hinf_syn_obj_three][[eq:hinf_syn_obj_three]]]. The
frequency response of the designed filters and the weighting functions
are shown in Figure [[#fig:three_complementary_filters_results][8]].

#+caption: [fig:three_complementary_filters_results]Bode plot of the
weighting functions and designed set of three complementary filters
[[file:figs/three_complementary_filters_results.pdf]]

* Application: Complementary Filter Design for Active Vibration
** Introduction                                                      :ignore:
Isolation of LIGO
Gravitational waves can help in detection various astrophysical events
occurring in our universe. This can also pave a path to validate
theories built around the existence of gravitational waves. However, the
detection of these waves is an arduous task owing to the extraordinary
small strain experienced by the earth due to gravitational waves.
Various methods have been proposed for their detection, out of which
laser interferometers are the most popular ones. Laser interferometers
offers large projection range and high displacement sensitivity. Among
the existing detector, Laser interferometer gravitation-wave observatory
(LIGO) is the most sensitive operational detector. LIGO consists of two
longs arms, referred as beam tubes, that are placed orthogonal to each
other. The arms of the LIGO accommodates a Michleson interferometer with
a cavity (Fabry-Perot). The mirrors at the extremity of the cavity serve
as inertial test masses which responds to the strain induced due to the
gravitational waves. The optics of the LIGO are suspended like a
pendulum. The schematics of the LIGO are shown in
Figure [[#ligo_schematics][9]].

#+caption: [ligo_schematics]Schematics of LIGO
[[file:figs/ligo.png]]

The isolation of the terrestrial interferometers is necessary in order
to isolate the motion of the suspended inertial masses from the seismic
ground motion. The vibration isolation system attenuates the effect of
disturbances on the motion of the suspended masses in the gravitational
wave band. The other sources that can affect the sensitivity of the LIGO
are thermal noise, human activities, tidal motion etc. The suspensions
of the LIGO needs to serve two purpose --- isolation and alignment. The
alignment is also important in order to ensure that the interferometer
beam is targeted at the center of the suspended mass. The current
vibration isolation system for LIGO has seven different stages. In the
first stage, hydraulically actuated external pre-isolators are used for
attenuating large ground motions. Two stages of active electromagnetic
isolation system are placed next to pre-isolators. This is followed by
four stages of pendulum based passive isolation system.

In the active isolation stage of LIGO, different sensors are used to
sense the same physical signal in different frequency range. For
example, seismometers are used to sense the position of the platform in
the frequency band 0.5--10 Hz while geophones are employed above 10 Hz.
The signals recorded from different sensor are fused using complementary
filters \cite{hua05_low_ligo,hua04_polyp_fir_compl_filter_contr_system}.
The stringent requirements on these filters complicate their design.
Hua \cite{hua05_low_ligo} proposed complementary FIR filters which were
synthesized using convex optimization. The designed FIR filters were
found to be compliant with the design specifications. However, the order
of the designed filter was very high, which limits its application to a
practical system. In this section, we demonstrate the design of
complementary filters with the same specification using the proposed
method based on $\mathcal{H}_\infty$- synthesis.

** Design specifications
The design specification of the complementary filters (as listed out
in \cite{hua05_low_ligo}) are as follows:

1. In the frequency range $0$-$\SI{0.008}{Hz}$: the high pass filter's
   magnitude should be less than $8 \times 10^{-4}$.

2. For frequency range $\SI{0.008}{Hz}$-$\SI{0.04}{Hz}$: slope of the
   high pass filter is equal to three.

3. Between $\SI{0.04}{Hz}$-$\SI{0.1}{Hz}$ frequency range: the high pass
   filter's magnitude should be less than $3$.

4. For frequencies above $\SI{0.1}{Hz}$: the low pass filter's magnitude
   should be less than $0.045$.

The specification of the complementary filters are shown graphically by
dashed black lines in Figure [[#fig:ligo_weights][10]].

** Weighting Functions Design
As the synthesis objective of the complementary filters is described by
Eq. [[#eq:hinf_problem_gen][[eq:hinf_problem_gen]]], it is clear that
the weighting functions should be chosen such that their inverse
magnitude represent the maximum allowed norm of the complementary
filters. This can be done manually using by combining poles and zeros or
using useful formulas such as
Eq. [[#eq:weight_formula][[eq:weight_formula]]]. It is important to note
that the order of the filters should be kept reasonable small in order
to keep the computational cost of the optimization reasonable. This will
also ensure that the designed filters are realizable in the physical
world. The transfer function representing weights should also be stable
and minimum phase.

The weighting function corresponding to the low pass filter, $w_L(s)$,
is here taken as Type I Chebyshev filter. The order of the weighting
function for low pass filter is set as 20.

The weighting function for the high pass filter, $w_H(s)$, is designed
in such a way that its magnitude response is as close as possible to the
design specifications. This was achieved using a combination of high-,
low- and band-pass filters in the particular frequency band. The overall
order of the weighting function for high pass filter is 7.

The magnitude responses of the inverse of the designed weighting
functions and their comparison with the specifications are shown in
Figure [[#fig:ligo_weights][10]]. It can be seen that the inverse of the
designed weights, shown in solid blue line for high pass filter and
solid red line for low pass filter, are close to the specifications
shown in black dotted line.

#+caption: [fig:ligo_weights]Specifications and weighting functions
magnitudes
[[file:figs/ligo_weights.pdf]]

** $\mathcal{H}_\infty$ synthesis of complementary filters
The complementary filters are designed using $\mathcal{H}_\infty$
synthesis based on the architecture shown in
Figure [[#fig:h_infinity_robust_fusion][4]]. The generalized plant is
obtained by substituting the transfer functions of the designed
weighting functions in
equation [[#eq:generalized_plant][[eq:generalized_plant]]], i.e.,
$W_1(s)=w_H(s)$ and $W_2(s)=w_L(s)$. The objective of the
$\mathcal{H}_\infty$ synthesis is given by
equation [[#eq:hinf_syn_obj][[eq:hinf_syn_obj]]]. The optimization
problem in solved in Matlab \cite{MATLAB2009} using Ricatti method. The
frequency responses of the designed optimal complementary filters are
shown in Figure [[#fig:comp_fir_ligo_hinf][11]]. The order of the
filters obtained using $\mathcal{H}_\infty$ optimization is $27$.

Now, we compare the designed complementary filters with the FIR filters
designed by Hua \cite{hua05_low_ligo}. The transfer function of the FIR
filter, $G(\omega)$, is the Fourier transform of its coefficients,
$g(n)$. That is, $$G(\omega) = \sum_n g(n)e^{-j2n\pi\omega}$$ The FIR
filter synthesis was formulated as a convex optimization problem. The
objective of the optimization problem was to find the filter's
coefficients such that their norm are below the specified upper bounds.
The optimization problem was solved using
SeDuMi \cite{sturm99_using_sedum} and the obtained order for the FIR
filters is $512$. The bode plot of the FIR filters are shown with dotted
lines in Figure [[#fig:comp_fir_ligo_hinf][11]]. It can be seen that
frequency responses of the designed complementary filters matches quite
well with those of the FIR filters. The designed complementary filters
are of much lower order and can be implemented with less computational
cost can the FIR filters. The proposed methodology for the design of
complementary filters can be effectively employed to obtain physically
realizable filters.

#+caption: [fig:comp_fir_ligo_hinf]Bode plot of the filters designed
using $\mathcal{H}_\infty$ synthesis and FIR
filters \cite{hua05_low_ligo}
[[file:figs/comp_fir_ligo_hinf.pdf]]

* Concluding remarks
The measurements from the sensors are reliable only within its
bandwidth. The signals from different sensors are usually fused in order
to measure a physical quantity over larger bandwidth. The sensor
obtained after fusion is called as super sensor as it has superior noise
characteristics and wider bandwidth. Complementary filters are used for
the combining the signals from different sensors. A new framework based
on $\mathcal{H}_\infty$ synthesis has been presented in this paper to
aid the design of complementary filters. The method presented allows to
shape the complementary filters based on the design specifications. The
task of filter design is posed as an $\mathcal{H}_{\infty}$ synthesis
problem. The design specifications of the systems are transformed in the
form of weighting functions. These weighting functions are used in the
optimization problem to constraint the filter response in a frequency
band. The method has also been demonstrated for designing a set of three
complementary filters. The design frame is general, simple to implement
and can easily be extended to difference scenarios of sensor fusion. The
effectiveness of the method is demonstrated for a real life application
where complementary filters are designed for active vibration isolation
of Laser Interferometer Gravitational-Wave Observatory (LIGO). The
filters designed with the proposed method have been with compared with
the finite impulse response (FIR) filters. It was found that the filters
designed using $\mathcal{H}_\infty$ have lower order compared to FIR
filters. The designed filters are physically realizable and have lesser
computational cost compared to FIR filters. The proposed method can be
effectively used to shape complementary filters based on design
specifications. The method can be further be extended for the design of
robust complementary filters with desired noise characteristics
considering uncertainties in the sensor dynamics. This is the focus of
our future research.

* Acknowledgment
  :PROPERTIES:
  :CLASS: unnumbered
  :END:
The authors would like to acknowledge the help received from the French
Community of Belgium for funding the FRIA Grant of Thomas Dehaeze (Grant
No. FC 31597).